prompt
stringclasses 1
value | text
stringlengths 39
387k
|
|---|---|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
The rapid loss of biodiversity in the last century has opened a debate on the consequences for ecosystem functioning. Therefore, understanding whether there is a relationship between species diversity and ecosystem processes is a key priority in the face of major global changes [1] , [2] , [3] . One of the most explored relationships has been between plant species richness and productivity, a process determining ecosystem carbon (C) pools and fluxes, and closely linked to ecosystem C sequestration [4] , [5] . Most studies conducting manipulative experiments have found a positive effect of species richness on productivity [2] , [6] . However, as these experiments are conducted in simplistic settings (e.g. even-aged species with short life cycles), there is controversy whether this effect holds in structurally more complex natural systems.
Forest ecosystems are major terrestrial C sinks, with a larger capacity to remove atmospheric C than previously thought [7] . Wood production is one of the main components of atmospheric C sequestration in the biosphere, with a high spatial variation depending on biotic, environmental and management factors [8] . Given the global interest in mitigating the consequences of greenhouse gases in the atmosphere, and the need for biodiversity conservation, it is necessary to determine to what extent wood production is reduced by the loss of tree species diversity, and to pinpoint differences among forest types [4] , [9] , [10] .
The tree species richness-productivity relationship has been investigated in forests by analyzing forest inventory data [11] , [12] , [13] , [14] , experimentally by manipulating tree species diversity in plantations [15] , [16] , [17] , [18] , [19] and by simulation modeling [20] , [21] . Studies based on forest inventory data have the potential for testing whether there is a positive relationship between tree species richness and wood production in the “real world”. However, such studies must control for the spatial heterogeneity of forest structure and confounding environmental factors such as climate [14] . To date, most studies have been conducted within certain climatic regions and for particular monospecific-mixed assemblages (e.g. [22] , [23] ), while only few have encompassed large environmental gradients including a variety of forest types (cf. [12] , [13] , [14] , [24] ).
By using unpublished data from more than 55000 forest inventory plots across Europe, we constructed Structural Equation Models (SEM) to test for the direct and indirect dependence of wood production on tree species richness while accounting for stand structure and climatic factors. The hypotheses tested were:
Wood production is positively and directly related to tree species richness.
Wood production is positively and directly related to the richness of functional tree types. The rationale for this is that tree functional types represent main differences in tree life-history and resource use. Therefore, ecosystem functioning might be as related to tree type richness as to species richness per se [4] .
Wood production indirectly increases with tree species richness through a positive effect of tree species richness on tree stand basal area. Our rationale for this hypothesis is that because most European forests have been largely managed in the past, they are predominantly early successional secondary forests (i.e. young forests) that have not reached maximum size and still accumulate carbon [25] . Under these circumstances, stand basal area is expected to be positively associated with local tree species richness [22] , [26] .
The positive association between wood production and tree species richness still remains when controlling for differences in climatic conditions. Our prediction is that mean annual precipitation and mean annual temperature have a parallel influence on both wood production and tree species richness [27] .
Materials and Methods
Database and selected variables
We collated forest inventory datasets from five European countries (France, the Netherlands, Spain, Sweden and Switzerland) on the basis of their quality and accurate evaluation of aboveground wood production. With the exception of France, inventories have been conducted in permanent plots surveyed from 1983 to 2009. We selected pairs of contiguous surveys ranging from 5 to 13.5 year periods. The French forest inventory is based on temporary plots where the volume growth of each tree over the last five years is estimated retrospectively based on radial and height growth measurements. In France, only data for the Alps and the Jura Mountains (southeast France) were available for this analysis. The basic criteria of plot selection were the lack of human intervention during contiguous surveys, and that all trees in the plot had been measured above a diameter at breast height (DBH) threshold (Table S1). Detailed information on inventory data for each European country is summarized in Table S1 [28] .
For each selected plot, we assigned the forest type according to the European Environmental Agency classification (EEA 2006). In total, our dataset included 55265 plots of 11 European forest types ( Table 1 ). Tree species were also classified into four coarse tree functional types: evergreen conifers, deciduous conifers, evergreen broadleaved -sclerophyllous- and deciduous broadleaved trees). For each plot, tree species richness and tree type richness were calculated. The number of tree species per plot (tree species richness) ranged from one to ten. On average, 49.39% of the plots were mixed with two and three tree species mixtures being the most common (28.21% and 13.56%, respectively). Less than 1% of the plots had more than six tree species. Boreal, hemiboreal and broadleaved evergreen forest plots had a maximum of five tree species. The highest tree species richness was found in mesophytic deciduous forests (ten species per plot), and in floodplain forests and exotic plantations (nine species per plot). The number of tree types (hereafter tree type richness) ranged from one to three. Most commonly, plots had only one tree type (68.76%). Plots with three tree types were rare (2.32%). Table 1 provides information on the number of monospecific and mixed plots for each forest type.
10.1371/journal.pone.0053530.t001
Table 1
Main characteristics of forest plots.
Acidophilous oak
Alpine coniferous
Beech
Boreal and hemiboreal
Broadleaved evergreen
Coniferous Mediterranean
Exotic plantations
Floodplain
Mesophytic deciduous
Non-riverine pioneer
Thremophilous deciduous
Number of plots (mono/ mixed)
14/105
5655/7064
563/1826
515/2504
7114/3285
9627/4294
1254/2358
43/162
2381/4167
110/644
692/888
Countries
*
NL
SPA/FRA/SWI
FRA/NL/ SPA/SWI
NL/SWE/SWI
FRA/SPA
FRA/SPA
FRA/NL/SPA/ SWE/SWI
FRA/NL/SPA/SW
FRA/NL/SPA/SWI
FRA/NL/SPA/ SWE/SWI
SPA/SWI
Annual precipitation (mm)
811.8±28.7
910.3±302
1263±335.3
646.5±110.2
669.3±157.7
586.2±217.1
1262±409.5
1040.3±440.3
1052.2±291.4
831±336
726.2±214.9
Annual temperature (°C)
9.52±0.19
9.3±2.1
8.7±1.4
3.6±3
14±2.2
13.5±1.9
10.4±3.4
11.2±2.3
10.6±1.6
6±4.3
12±1.8
Dominant species
Betula pendula, Quercus robur
Abies alba, Pinus nigra, P. sylvestris
Fagus sylvatica
Picea abies, P. sylvestris
Quercus ilex, Q. suber
P. halepensis, P. pinaster, P. pinea
Eucalyptus globules, Picea abies, P. radiata
Alnus glutinosa, Populus nigra, Salix spp.
Q. petraea, Q. pubescens, Q. pyrenaica
Betula spp., Populus alba, Populus nigra, Populus tremula
Fraxinus angustifolia, Q. faginea
Stand basal área (m 2 /ha)
21.4±7.4
21.7±14
27.3±13.1
22±11
7.4±5.7
12.7±10.2
26.2±17.1
16.2±10.9
17.3±11.4
17.5±11.4
9.4±7.8
Plot size (ha)
0.05±0
0.08±0.06
0.09±0.07
0.03±0
0.12±0.07
0.1±0.06
0.07±0.06
0.12±0.07
0.07±0.06
0.05±0.04
0.1±0.07
Tree species richness/plot
3.3±1.3
1.8±1
2.6±1.3
2.2±0.7
1.4±0.6
1.4±0.7
2.2±1.3
3±1.8
2.3±1.4
2.6±1.2
1.8±0.9
Tree type richness/plot
1.5±0.5
1.4±0.5
1.4±0.5
1.7±0.5
1.3±0.5
1.2±0.5
1.5±0.5
1.3±0.5
1.3±0.5
1.6±0.5
1.5±0.6
Wood production (t/ha/yr)
2±2.3
2±1.7
3.1±2
2.5±2.1
0.49±0.47
1.2±1.3
4.4±4
2.6±2.1
2±1.67
2.6±2.1
0.7±0.7
*
Country nomenclature: France (FRA), The Netherlands (NL), Spain (SPA), Sweden (SWE) and Switzerland (SWI). Values indicate means (±SD).
plot as follows:
In inventories based on permanent plots, for each living tree with a minimum DBH of 4–12 cm depending on the country (Table S1), the species identity was noted and tree volume ( V ) was calculated with species-specific functions of DBH and H fitted on field data from the respective countries as: where f is the form factor of each species. Wood biomass ( B ) was estimated as:
where Dw is tree wood density of the species.
The annual increase in aboveground biomass of surviving trees s (BG s ) was measured as:
where Bs 1 is the biomass of a surviving tree measured in the first survey (1) and still alive in the second survey (2) and t is the time elapsed between the two surveys.
Aboveground wood production per plot ( WP ) was estimated as: where N alive is the number of surviving trees in the plot and BG s their respective annual increase in aboveground biomass. N recruit is the number of recruited trees during the two contiguous surveys (i.e. trees reaching the minimum DBH of 4–12 cm to be included in the survey), Bi 2 is their aboveground biomass and t is the time elapsed between the two surveys.
In France, BG s were computed with an estimation of volume growth over the last five years for each tree alive on the plot at the time of measurement ( VG s ) and Dw :
VG s was estimated by functions based on five years radial growth (determined from a tree core sample), H and height growth over five years [29] .
Mean annual temperature and mean annual precipitations were assigned as climatic variables to each plot (temperature and precipitation, hereafter) based on available interpolated climatic maps for each country.
Statistical analysis
First, for each forest type, we developed Generalized Linear Models to test for differences in wood production among tree species richness using the PROC-GENMOD procedure in SAS (version 9.2, SAS Institute Inc., Cary, NC, USA) with a normal error distribution and identity link function [30] , and plot area as a covariate. When differences among tree species richness were significant, pair-wise differences of Least Square means (LS means) were tested. Likewise, we tested for differences in wood production among tree type richness.
To select the appropriate variables to be included in the Structural Equation Modeling (SEM) [31] , we performed a stepwise regression analysis to test for the correlation of wood production with tree species richness, tree type richness, stand basal area, temperature and precipitation for each forest type, respectively.
The species richness-productivity relationship might vary with the spatial grain (i.e. plot size), the spatial extent (i.e. local, landscape, regional, continental or global), and also the ecological association scale (e.g. within or across community types) of the study [32] . Our forest surveys were conducted at local spatial scales, across a whole continent, and within 11 different forest types. Plot size ranged from 5 to 25 m radius, and was not always the same across forest inventories. Plot sampling areas were, however, within the size range considered appropriate for vegetation studies of European forests [33] and in forest inventories [28] . Therefore, our analysis captured tree alpha diversity across plots of similar size ( Table 1 ). Following recommendations to investigate how the richness-productivity relationship changes across climatic gradients [34] , we did not extrapolate the number of tree species to the regional scale but maintained the plot as the sample unit while the geographical extent was enlarged by incorporating plots from several countries. Values for tree species richness, stand basal area and wood production were standardized per unit sampling area prior to the analysis.
Finally, SEM was used to test the above hypotheses. A SEM was constructed for each forest type. The model contains causal relationships among variables ( Fig. 1 ), represented by single-headed arrows, and a correlational relationship between the two climatic variables that is represented by a double-headed arrow connecting temperature and precipitation. Direct effects of one variable on another are indicated by an arrow linking the two variables (e.g. tree species richness on wood production in Fig. 1 ), while indirect effects are those linked by an intermediate variable (e.g. tree species richness on wood production through tree type richness in Fig. 1 ) (see [35] for a detailed description of SEM procedures).
10.1371/journal.pone.0053530.g001
Figure 1
Structural Equation Model (SEM) for tree wood production.
Single arrows represent causal paths (i.e. simple regressions between variables), whereas the double-headed arrow denotes correlation between mean annual precipitation and temperature. U n values represent unexplained variance in each endogenous variance. The letters on each arrow indicate the standardized regression weights (path coefficients) between variables. Path coefficient values for each European forest type are given in Table 3 .
Due to large sample sizes in each forest type and the assumption of multivariate normality, standardized path coefficients were estimated using maximum likelihood techniques [35] , [36] . We tested for both univariate and multivariate normality, applied transformations when necessary and examined for influential outliers (squared Mahalanobis distance, [37] ). When normality assumptions were not met as a consequence of large sample sizes (i.e. alpine, broadleaved evergreen and coniferous Mediterranean forests), bootstrapping was used to evaluate statistical significance of each path coefficient [38] , [39] . Subsequently, the goodness-of-fit was determined to test the degree to which the aprioristic SEM fits the sample data [40] . Since the commonly used chi-square test for the absolute model fit is sensitive to sample sizes and multivariate normality assumption of the input variables [40] , the Comparative Fit Index (CFI) was used which does not depend on sample size as much as the chi-square test [41] . Values of CFI can range between 0 and 1, with values ≥0.90 confirming a good model fit.
For each forest type, we calculated the standardized regression coefficients associated with each path. These values represent the amount of change in one variable given a standard deviation unit change in the other one. We also calculated the coefficient of determination (R 2 ) for each variable as an indication of the contribution of the model to the variation of that variable. The unexplained variance (u) of the model to each variable was also indicated (Table S2).
For models with CFI values ≥0.90, differences of path coefficients among forest types were determined through Multigroup analyses [36] , [42] . SEM and Multigroup analyses were performed using the AMOS.18.0 software [38] .
Results
Wood production was higher in mixed compared to monospecific forests of the same type as indicated by values falling above the line of unity in all forest types, except in acidophilous oak forests for which values were lower ( Fig. 2 ). On average, wood production was 24.38% higher in mixed than in monospecific forests.
10.1371/journal.pone.0053530.g002
Figure 2
Tree wood production in pairs of monospecific and mixed forests.
Values indicate means (±SE). Each point represents a different European forest type. The dashed line represents the line of unity.
Taken alone, wood production increased with tree species richness, at least from monospecific to mixed plots with 3–4 species, and then the relationship reached an asymptote ( Fig. 3 ). In alpine forests, wood production increased up to six species, while in non-riverine pioneer forests maximum wood production was already reached in two species forests. In acidophilous forests, wood production decreased from monospecific to mixed plots with 3–5 species, while productivity in plots with 6–8 species was not significantly different from the monospecific ones. Similarly, wood production increased with tree type richness with the exception of floodplain, mesophytic deciduous and non-riverine pioneer forests, where the relationship was not significant, and acidophilous oak forests for which plots with only one tree type were more productive than with two tree types ( Fig. 4 ). Alpine forests had a hump-shaped relationship, with two tree type forests being more productive than one and three tree type forests.
10.1371/journal.pone.0053530.g003
Figure 3
Tree wood production with increasing tree species richness.
Values indicate LS means (±SE). Different letters above columns indicate significant differences between stands with different species richness according to GENMOD-procedure in SAS. n.s. = not significant.
10.1371/journal.pone.0053530.g004
Figure 4
Tree wood production with increasing tree type richness.
Values indicate LS means (±SE). Different letters above columns indicate significant differences between stands with different species richness according to GENMOD-procedure in SAS. n.s. = not significant.
For each forest type, wood production was related to all variables tested in the stepwise analysis ( Table 2 ). Stand basal area was the most important variable, explaining 54–84% of the variance in wood production. Overall, climatic variables were stronger determinants of wood production compared to tree species or tree type richness, and tree type richness explained more variance in wood production than tree species richness.
10.1371/journal.pone.0053530.t002
Table 2
Stepwise procedure on the relationship of abiotic and biotic variables with tree wood production.
Abiotic variables
Biotic variables
Forest types
Temperature
Precipitation
Stand basal area
Tree species richness
Tree type richness
Acidophilous oak
0.41
0.42
0.54
0.28
0.29
Alpine coniferous
0.51
0.61
0.81
0.43
0.54
Beech
0.69
0.69
0.81
0.48
0.65
Boreal and hemiboreal
0.54
0.59
0.62
0.53
0.54
Broadleaved evergreen
0.47
0.55
0.78
0.29
0.58
Coniferous Mediterranean
0.43
0.53
0.81
0.27
0.46
Exotic plantations
0.47
0.56
0.70
0.32
0.52
Floodplain
0.59
0.51
0.80
0.37
0.52
Mesophytic deciduous
0.57
0.64
0.84
0.58
0.53
Non-riverine pioneer
0.48
0.55
0.70
0.56
0.57
Thermophilous deciduous
0.48
0.54
0.78
0.20
0.50
For each forest type we indicate the adjusted R 2 for each variable taken alone. All variables tested were also related to wood production across all 11 European forest types.
All the above variables were included in the SEM and were retained in the model. The CFI of the SEMs were ≥0.90 in all forest types except for broadleaved evergreen (0.89), non-riverine pioneer (0.78) and thermophilous deciduous (0.85) forests (Table S2). On average, 47% of the variance in wood production was explained by the model, with highest values in coniferous Mediterranean forests (68%), and alpine coniferous and mesophytic deciduous (>55%); and lowest values (19%) in boreal and hemiboreal forests (Table S2).
Tree species richness had a low direct effect on wood production (path 3b, Table 3 ). However, in almost all forest types, stand basal area increased with tree species richness (path 3a, Table 3 ), and stand basal area was the variable with the largest positive effect on wood production (path 4, Table 3 ). Therefore, the effect of tree species richness on wood production is mainly indirect by increasing stand basal area. Tree type richness increased wood production in some forest types, namely alpine coniferous, coniferous Mediterranean, broadleaved evergreen and exotic plantations. However, path coefficients were small (path 5, Table 3 ) and of a similar magnitude to tree species richness.
10.1371/journal.pone.0053530.t003
Table 3
Structural equation modelling (SEM) path coefficients.
Path coefficients
1a
1b
1c
2a
2b
2c
3a
3b
3c
4
5
c1
Acidophilous oak
0.77
***
−0.09
ns
0.00
ns
−0.37
***
−0.03
ns
0.22
*
0.00
ns
−0.15
ns
0.23
*
0.48
***
−0.11
ns
0.53
***
Alpine coniferous 1
0.06
***
0.09
***
−0.17
***
0.47
***
0.12
***
0.19
***
0.11
***
0.00
ns
0.47
***
0.72
***
0.03
***
−0.43
***
Beech
−0.04
*
0.21
***
−0.06
**
0.33
***
0.14
***
0.12
***
0.10
***
−0.02
ns
0.41
***
0.64
***
0.03
ns
−0.4
***
Boreal and hemiboreal
−0.08
**
0.27
***
0.35
***
−0.05
*
−0.03
ns
−0.02
ns
0.18
***
0.04
ns
0.65
***
0.27
***
0.00
ns
0.61
***
Broadleaved evergreen 1
−0.39
***
−0.06
***
−0.27
***
0.03
***
0.05
***
0.34
***
−0.05
***
0.06
***
0.35
***
0.7
***
0.10
***
0.11
***
Coniferous Mediterranean 1
−0.06
***
0.03
***
−0.1
***
0.14
***
0.19
***
0.25
***
0.48
***
0.01
ns
−0.07
***
0.76
***
0.03
***
−0.15
***
Exotic plantations
−0.58
***
0.24
***
−0.19
***
−0.19
***
0.15
***
0.3
***
0.06
***
−0.13
***
0.39
***
0.71
***
0.06
***
−0.65
***
Floodplain
−0.39
***
0.13
*
0.04
ns
0.28
***
0.017
ns
−0.08
ns
0.14
*
0.05
ns
0.25
**
0.71
***
−0.11
*
−0.32
***
Mesophytic deciduous
−0.13
***
0.043
***
−0.06
***
0.31
***
0.1
***
0.2
***
0.25
***
0.08
***
0.24
***
0.7
***
−0.07
***
−0.36
***
Non-riverine pioneer
−0.22
***
0.17
***
0.22
***
0.1
*
−0.07
ns
0.00
ns
0.33
***
0.03
ns
0.27
***
0.51
***
0.05
ns
0.69
***
Thermophilous deciduous
−0.18
***
−0.03
ns
−0.1
***
−0.24
***
0.13
***
0.3
***
−0.07
*
0.10
*
0.28
***
0.69
**
−0.03
ns
0.16
***
For each forest type we indicate the standardized regression weights of the paths according to the nomenclature indicated in Figure 1 .
1
Forest data was analyzed through bootstrapping . Significance of the path coefficients: *P<0.05, ** P<0.005, ***P<0.0001, ns = not significant.
Temperature increased wood production in most forest types (path 1b, Table 3 ). On the contrary, temperature had almost always a negative effect on tree species richness except in acidophilous and alpine coniferous forests where it was positive (path 1a, Table 3 ). Precipitation increased wood production in most forests, except in acidophilous oak, boreal and hemiboreal, floodplain and non-riverine pioneer forests where the relationship was not significant (path 2b, Table 3 ). Precipitation also increased species richness, except in acidophilous oak, boreal and hemiboreal, exotic plantations and thermophilous deciduous forests where it was negative (path 2a, Table 3 ).
Not only was the direction of the relationship between climatic variables and wood production different compared to that of species richness, it also differed in magnitude. That is, even within a forest type the effect of climate on tree species richness and wood production could be in opposite directions, be significant for one variable and not significant for the other, or of different magnitude. For example, in acidophilous oak forests, temperature and precipitation had a non-significant effect on wood production, but temperature increased tree species richness (77% of the variation explained) while precipitation affected tree species richness negatively (37% of the variation explained). Multigroup analyses revealed that path coefficients among forest types were significantly different (Table S3). However, differences between forest types were dependent on the path under consideration (Table S4).
Discussion
We found a positive relationship between tree richness and wood production in most European forest types. Our analysis is the first to describe this relationship at the local scale for the largest dataset across a continent, encompassing a wide range of climatic conditions. This result is in line with other regional studies showing higher productivity in mixed compared to monospecific forests [43] , [44] . We found European mixed forests to be on average 24% more productive than monospecific forests. Although we do not have precise information on the management history of these forests, most of our study plots were not plantations but natural forests. Moreover, even if some might be plantations they had not been managed during the inventory measurement periods. This indicates that the positive relationship between species richness and productivity is found in structurally complex woody systems, encompassing a wide range of environmental conditions [45] .
As also found in other ecosystems, in many forest types maximum wood production was reached at medium levels of species richness. There may be several non-exclusive explanations for this pattern. Functional redundancy and niche overlap may occur at high levels of species richness [2] . Therefore, a complete exploitation of available resources for wood production seems to be reached faster in high compared to low species rich forests. Alternatively, the saturation of the tree species richness-productivity relationship may be a consequence of higher levels of evenness in plots of low (i.e. 2–3 species) compared to high tree species richness. Tree species evenness has been found to be a better predictor of wood production than tree species richness [44] . Furthermore, plots of high tree richness are less common than plots of low richness [12] , [13] . Plots of high richness are therefore more variable in wood production due to small sample sizes, but possibly also due to a larger variation in species composition and a lower abundance of rare species.
The positive association between tree species richness and wood production was mediated by an increase in tree stand basal area with species richness. Although stand age was not available, most European forests have an uneven-aged structure, have been highly managed historically, and are at an early seral stage [25] . In these circumstances, stand basal area has not reached its maximum yet [46] and tree species richness is high [47] . Although our study cannot elucidate the ecological mechanisms underlying the positive relationship between tree species richness and wood production, two non-mutually exclusive mechanisms have been hypothesized to drive this observation: the complementarity effect and the sampling effect. The first hypothesis postulates that species rich stands are most efficient in resource use because they contain species with a diverse array of ecological traits such as multilayered canopies or roots at different depths that optimize ecosystem resource use. Complementarity can result from niche partitioning and/or facilitation among species with different traits, decreasing competition in diverse communities [20] , [48] . Alternatively, the positive association might be explained by a sampling effect, whereby species rich stands are more likely to contain and become dominated by at least one species highly efficient in resource use that accounts for most of the production in the community [1] , [49] . Both mechanisms can act simultaneously or there might be transitions between them over large time spans [6] . Moreover, their importance might depend on the forest type. For example, in climatically stressful Mediterranean conditions, mixed forests containing species of low productivity might achieve higher wood production because of species niche partitioning in water use [12] . On the other hand, in many European forests, traditional management has favored economically important species and highly productive varieties (e.g. exotic trees). When abandoned and colonized with other tree species, these stands might still remain highly productive because of the sampling effect of highly productive trees. Long term experimental tree plantations are needed to test the mechanisms underlying the positive signal between tree species richness and wood production and how it might change over time [50] .
The positive relationship between tree species richness and wood production mediated by an increase in basal area remained significant when climatic factors were included in the models. This indicates that climatic differences are not the sole explanation for differences in wood production along a gradient of species richness. Moreover, our analysis shows that the influence of temperature and precipitation has on wood production are highly dependent on forest type. Our analyses also reveal that climate does not influence wood production and tree species richness in parallel [27] .
In more than half of the forest types, wood production was positively related to tree type richness. However, often there were no significant differences between two- and three-tree type mixtures. In some forest types, the relationship was not significant, negative, or hump-shaped. This idiosyncrasy was unexpected as we had predicted tree type richness to be functionally as relevant as species richness. The low number of tree types in European forests (i.e. evergreen conifers, deciduous conifers, evergreen broadleaved and deciduous broadleaved) is possibly the cause of these inconsistencies among forest types. Moreover, tree types are possibly too coarse to underpin differences in functional traits responsible for wood production. Tree species richness might better reflect functional trait diversity than the tree type richness used in our study. Biodiversity categories based on growth forms are “soft traits” that may mask within-group variability of traits [51] . Recent studies have shown that functional diversity indices based on traits relating to reproduction, growth, successional status and resource use perform better than indices of species diversity [13] , [52] . However, due to the large variation in species composition in European forest inventories, there is still not enough information on functional species traits for many species, especially Mediterranean and alpine tree species.
Overall, our study shows for the first time across a continent that local tree wood production is positively associated with local tree species richness in many forest types, even when controlling for climatic variation. Although wood production is just one process of the global C cycle, tree growth is the principal forest C flux contributing to atmospheric CO 2 sequestration by the biosphere [53] . Our results suggest that preserving forests with a high alpha diversity could substantially increase C sequestration at the local scale by increasing wood production. Thus, forest related biodiversity issues, although neglected until now, should be incorporated in management and policy plans for C sequestration.
Supporting Information
Table S1
Main characteristics of the forest inventories.
(DOC)
Table S2
Results of the structural equation model (SEM).
(DOC)
Table S3
Goodness-of-fit statistics for multigroup analyses.
(DOC)
Table S4
Summary of multigroup comparison among forest types for single path coefficients.
(DOC)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
The aim of genetic mapping studies is to identify loci that underlie phenotypic variation. Genetic mapping studies are critical for improving crops through marker-assisted breeding and for our understanding of the relationship between genotype and phenotype [1] . Genome wide association (GWA) mapping [2] and genomic selection (GS) [3] are increasingly being adopted for crop improvement and they often require large numbers of genetic markers. One of the main challenges in agricultural genetics is to access and use the tremendous genetic variation present in germplasm collections and in the wild, as crop species are far more diverse than the vertebrate systems used in biomedical research. To do this, approaches for applying next generation sequencing technology to non-model systems need to be developed [4] .
The first step towards GWA and GS is to discover large numbers of genetic markers, generally single nucleotide polymorphisms (SNPs), across the genome. This initial step of large-scale SNP discovery is already underway in several organisms. For example, in humans the International HapMap Project currently boasts over 3 million SNPs ( http://www.hapmap.org/ ), and similar projects are in progress for Arabidopsis thaliana ( http://walnut.usc.edu/2010 ), rice ( http://irfgc.irri.org ) and maize ( http://www.panzea.org/ ). While previous SNP discovery initiatives relied on laborious and relatively expensive sequencing and genotyping platforms, SNP discovery has become less time consuming and much more cost-effective since the introduction of next-generation sequencing (ABI's SOLiD, Illumina's Genome Analyzer and Roche's 454). SNP discovery using next-generation sequence data is still in its infancy, but several studies have already demonstrated that large numbers of high quality SNPs can be identified in a cost effective manner using next-generation sequence data [5] – [9] . Deep sequence coverage across many samples is generally desired in order to identify high quality SNPs. To achieve an increase in coverage, the portion of the genome that is sequenced can be reduced by constructing reduced representation libraries (RRLs). RRLs are generated by digesting each sample with a common restriction enzyme before sequencing and they have been useful for large-scale SNP discovery in several organisms [8] – [11] .
After large-scale SNP discovery, it is crucial to gain an understanding of the pattern of linkage disequilibrium (LD) and population structure in the species of interest. The strategy underlying GWA and GS is to genotype enough markers across the genome so that functional alleles will likely be in LD with at least one of the genotyped markers [12] . Thus, an assessment of the rate of LD decay is essential in estimating the number of SNPs required for GWA and GS studies. For example, it has been shown that 500,000 SNPs provide reasonable power for GWA in humans [13] and that 140,000 SNPs provide reasonable coverage of the 125 Mb Arabidopsis thaliana genome [14] . An evaluation of population structure in the species of interest is also crucial: it allows the selection of germplasm for a mapping population that will maximize genetic diversity, and thus the number of QTL that can be detected. Numerous studies have recently used genome-wide SNP data to characterize patterns of population structure in domesticated species as a starting point for GWA and GS [15] – [17] .
Here we describe the initial steps we have taken towards genome-wide genetic mapping studies in the world's most economically important fruit crop, the grapevine (genus Vitis ). The grapevine is a long-lived woody perennial consisting of dozens of species whose natural habitat spans the northern hemisphere [18] . The cultivated grapevine, V. vinifera , represents one of the earliest domesticated fruits [19] and there are currently ∼19 million acres under vine ( http://faostat.fao.org/ ). Previous characterizations of the genetic structure of the grapevine have been restricted to small numbers of microsatellites [20] or a few hundred informative SNPs [21] – [23] . The grapevine is diploid, has a relatively small genome size (475 Mb) and was recently sequenced by two independent groups [24] , [25] . Genetic mapping in the grapevine has relied almost exclusively on linkage mapping, which is time-consuming because of the grapevine's long generation time (generally 3 years). These considerations make the grapevine an ideal candidate for assessing the utility of next-generation sequencing and genotyping arrays in characterizing genome-wide patterns of genetic diversity of a high-diversity, domesticated plant species in order to move rapidly towards GWA studies.
Here we describe a simple and rapid procedure for identifying hundreds of thousands of SNPs from 11 V. vinifera cultivars and 6 wild Vitis species. From these data, we assess patterns of segregation within and between V. vinifera and wild Vitis species and provide the most comprehensive analysis of LD decay in V. vinifera to date. We also describe the design of a SNP genotyping array for the grapevine that assays 8898 SNPs (the Vitis9KSNP array). We show that the Vitis9KSNP array provides sufficient high-quality genotypes to successfully capture the genetic structure within and between the V. vinifera cultivars and wild Vitis species. Our analyses suggest that the use of SNP arrays for WGA studies will be inadequate for high-diversity plant species in which LD decays rapidly, as in the grapevine. We suggest a stronger focus on experimental design in the anticipation that future mapping populations will be cost-effectively whole-genome sequenced in the near future.
Results
We generated reduced representation libraries (RRLs) from 17 grapevine DNA samples (10 cultivated V. vinifera varieties, 6 wild Vitis species and the reference genome (inbred Pinot Noir) – see Table S1 for details on samples) by digesting each sample with the restriction enzyme HpaII , which has proved useful in the generation of RRLs by others [26] , [27] . The generation of RRLs permits high-coverage sequencing of a small, similar fraction of the genome across samples. Each RRL was sequenced on a single lane of Illumina's Genome Analyzer to produce 57.3 million 36-bp reads (2.6 Gb of DNA sequence). We trimmed off the last 4 bases of each read and aligned the 32 bp reads to the reference genome using ELAND (Illumina Inc). In total, 68% of the reads successfully mapped to the reference genome: 57% mapped uniquely, 11% mapped to multiple locations (repetitive) and 32% provided no alignment (no match). Figure 1 provides a summary of the alignment results and the proportion of reads carrying the HpaII sequence tag across the 17 samples.
10.1371/journal.pone.0008219.g001 Figure 1
Alignment results of Illumina GA reads to the grapevine reference genome.
The total number of reads generated for each sample is found in the box to the right. The upper bars in the barplot indicate the proportion of reads belonging to each of the categories in the legend. Reads aligning with 0 to 2 mismatches were included for SNP discovery. Reads mapped repetitively and reads with no match were discarded. The lower bars (dark grey) show the proportion of reads beginning with the HpaII tag. The wild Vitis samples are shown in italics.
The sequencing was clearly enriched for successfully digested fragments as 81% of the sequence reads began with the HpaII sequence tag (CGG). Figure 2 summarizes the extent to which the sequencing of the RRLs resulted in higher than expected coverage of a small fraction of the genome. We observed a strong enrichment of reads mapping to HpaII digested fragments between 40 bp and 250 bp ( Figure 2A ), which is likely the result of PCR and cloning biases in the Illumina system. In addition, we compared the observed coverage to the coverage expected if no enrichment procedure had been performed ( Figure 2B ). Our enrichment procedure resulted in more bases covered at 0x and ≥8x than expected if no enrichment procedure had been performed ( Figure 2B ; see Methods for details). Thus, the use of RRLs concentrated the sequencing on a smaller portion of the genome which provided high enough coverage for reliable SNP calling.
10.1371/journal.pone.0008219.g002 Figure 2
HpaII digestion results in an enrichment of genomic regions with high read coverage.
Panel A presents two overlapping fragment size distributions. The size distribution of fragments from an in silico HpaII digestion are shown in blue and the size distribution of the in silico digested fragments to which reads were successfully mapped is shown in orange. Panel B compares the observed amount of the genome sequenced at each level of coverage to the expectation at random. The random expectation was generated assuming that coverage follows a Poisson distribution (see Methods for details). The inset in gray demonstrates that the observed coverage begins to exceed the random expectation at 8x coverage. SNPs were called from positions with ≥10x coverage.
After aligning all of the reads to the reference genome and applying some preliminary filters (see Methods ), we identified 469,470 SNPs, which we refer to as our 470K SNP set. Figure 3 demonstrates that SNPs were infrequent within the first 3 bp of reads and enriched towards the ends of reads in our 470K SNP set. The former observation is explained by our library preparation procedure: 81% of reads begin with the CGG-tag and we are therefore unlikely to observe polymorphism within the first 3 bp of reads. The latter observation, however, is consistent with the effects of sequencing error: errors are concentrated towards the ends of reads [5] , [8] , [28] . This suggests that our 470K SNP set contains false positives which are disproportionately represented at the ends of reads. We found that implementing a strict filter that disregards evidence of polymorphism from the ends of reads resulted in unacceptably high false negative SNP call rates. We therefore investigated several methods that help eliminate the observed read position effect. We found that the two most effective methods were the application of a quality score (Q) score threshold and a threshold on the p-value from a genotypic contingency test. The genotypic contingency test is applied to the read counts at a particular SNP (reference vs. alternative allele across samples) which are represented as a contingency table (see Supplementary Methods S1 for details). Figure 3 demonstrates that these methods are effective in eliminating the bias of SNP discovery towards the ends of reads. Selecting SNPs with average Q scores ≥20 and contingency test p-values ≤0.01 results in a set of 71,397 SNPs which we refer to as the 71K SNP set. The 470K and 71K SNP sets are publicly available at [ ftp://brie4.cshl.edu/pub/vitis_plosone_2009_snps/ ]. SNPs were most often called with coverage from fewer than all 17 accessions. In the 71K SNP set, for example, 95% of SNPs were assayed from ≥7 accessions (see Figure S1 ). Figure 4 presents the degree of shared polymorphism between the European cultivated V. vinifera cultivars and the wild Vitis species for the 71K SNP set.
10.1371/journal.pone.0008219.g003 Figure 3
Quality score (Q) and genotypic contingency test thresholds eliminate read position bias during SNP calling.
The 470K SNP set is enriched with SNPs identified from the ends of reads. Panel A demonstrates that this read position bias can be eliminated by applying a Q score threshold. Panel B demonstrates that the genotypic contingency test also improves SNP calling.
10.1371/journal.pone.0008219.g004 Figure 4
Segregation of SNPs in the 71K SNP set within and between V. vinifera and wild Vitis species.
The proportion of SNPs polymorphic only within V. vinifera is 68.5%. The proportion segregating only within wild Vitis species is 53.1%. A substantial proportion (24.3%) of SNPs shows evidence of segregation within both V. vinifera and the wild Vitis species. Only 2.7% of SNPs appeared fixed between V. vinifera and wild Vitis .
To assess patterns of LD decay in Vitis , we used a set of simple rules to call genotypes from the Illumina GA sequence data (see Methods ). We restricted our analysis to the 10 cultivated V. vinifera samples, as each of the wild Vitis species was represented by a single sample and there may be significant differences in LD decay between species. Levels of LD are generally low in V. vinifera (r 2 <0.2) even at short physical distances ( Figure 5A ). To determine at what distance LD decays to background levels, we calculated background LD as the degree of LD between SNPs on different chromosomes. We then compared background levels of LD to the observed pattern of LD decay up to 40 kb. Figure 5A demonstrates that while LD is generally low across all distances it remains above background levels to ∼10 kb. To formally test at what distance LD is no longer distinguishable from background LD levels, we compared the observed distribution of r 2 values in each bin to the 20,000 r 2 values generated from comparisons of SNPs on different chromosomes using a Mann-Whitney U test (see Methods for details). Figure 5B shows that p-values for these comparisons are consistently highly significant out to ∼10 kb and then begin to decay towards non-significant values.
10.1371/journal.pone.0008219.g005 Figure 5
LD decay in the grape.
Panel A shows the observed LD decay compared to background LD across 40 kb. LD was calculated as the median r 2 in bins of 1000 comparisons. The background LD is the median r 2 from 20,000 comparisons between SNPs on different chromosomes. Panel B shows the –log10 p-values from comparing the distribution of observed r 2 values within each bin to the distribution of background r 2 values generated from comparisons between SNPs on different chromosomes using a Mann-Whitney U test (see Methods for details).
We designed a custom Infinium SNP genotyping array (Illumina) that assays 8898 SNPs selected from the 470K set by relying on several criteria described in Table S2 and Supplementary Methods S1 . We refer to this SNP array as the Vitis9KSNP array. To date, we have genotyped 156 samples with the array and the 94 pairwise comparisons between replicate samples give an average concordance of 99.75%. We compared genotype calls from the Illumina sequence data to genotype calls from the Vitis9KSNP array for the 17 samples (see Methods for details on genotype calling). For 36,904 genotypes called from both datasets, we observe 97.7% concordance. Table 1 summarizes these concordance results by genotype class.
10.1371/journal.pone.0008219.t001 Table 1
Concordance of SNP genotype calls.
Vitis9KSNP array
homozygous reference
heterozygous
homozygous alternative
homozygous reference
24083 (65.26)
285 (0.77)
10 (0.03)
heterozygous
80 (0.22)
4158 (11.27)
41 (0.11)
homozygous alternative
29 (0.08)
408 (1.11)
7810 (21.16)
Concordance was assessed for 36,904 SNPs called from both Illumina GA sequence data and the Vitis9KSNP array. Concordance is found along the diagonal and the remaining cells represent different categories of non-concordance. The values inside each cell refer the number of SNPs in that category, followed by the percent value in parentheses. The most common type of non-concordance is found in cases where a SNP is called homozygous from the Illumina data but is called heterozygous from the array data.
To investigate patterns of population structure, we performed principal components analysis (PCA) on 14,325 SNPs from the Illumina GA sequence data, which were chosen without regard to the pattern of segregation among the 17 wild and cultivated grapevines ( Figure 6A ; see Methods for details). In Figure 6A , the first PC, which accounts for 20.7% of the variance, separates wild from V. vinifera accessions, while the second PC differentiates among wild species. The exception is V. sylvestris , the wild ancestor of the domesticated V. vinifera , which clusters with the V. vinifera varieties. We also performed PCA on genotype data generated from the Vitis9KSNP array for the same set of 17 samples ( Figure 6B ). In Figure 6B , the first PC separates wild from V. vinifera as in Figure 6A . The second PC, however, differentiates among V. vinifera varieties.
10.1371/journal.pone.0008219.g006 Figure 6
Principal components analysis (PCA) plots from grapevine SNP data.
The first two PCs are shown with the proportion of the variance explained by each PC in parentheses. Panel A shows a PCA plot generated from 14,325 SNPs called from the Illumina GA without regard to segregation pattern. Panel B shows a PCA plot from the Vitis9KSNP array data, whose SNPs were chosen purposely to distinguish among V. vinifera cultivars.
Discussion
WGA and GS studies have generally concentrated on a small number of organisms with established genotyping arrays. With the decreasing costs of DNA sequencing and genotyping, we anticipate that there will be interest in moving rapidly towards GWA and GS studies in organisms for which relatively little genetic data currently exists. Particularly in plants, the Germplasm Repositories of the United States Department of Agriculture currently house over 500,000 different accessions, presenting an enormous amount of genetic diversity to be catalogued and an incredibly large inventory of genetic variation waiting to be discovered and used. In the present study, we provide a framework for rapidly and cost-effectively moving from very few genetic resources, to genome-wide characterization of a species of great economic and cultural interest, the grapevine.
We generated ∼2.6 Gb of DNA sequence using the Illumina Genome Analyser, a substantial proportion (32%) of which did not align successfully to the available genome sequence ( Figure 1 ). Some of these unmatched reads likely come from portions of the genome that are not represented in the current genome, as the genome sequence is not complete. In addition, genetic variation among samples (e.g. highly divergent haplotypes, structural and copy number variation) may result in unaligned reads to the reference genome. For example, the inbred Pinot Noir, which is the identical DNA sample used to generate the reference genome, provided the highest number of successfully aligned reads as expected ( Figure 1 ). Reads from the distantly related wild Vitis species matched less often than the cultivated varieties, however several cultivars (e.g. Plavac Mali) showed a low proportion of matches. The variation across samples in the proportion of matches could be due to numerous factors, including variation in exogenous DNA contamination or quality of the sequence data.
Three lines of evidence strongly suggest that our genomic reduction procedure was successful. First, 81% of the sequenced reads begin with the HpaII tag ( Figure 1 ). Thus, most of the sequence we obtained came from fragments successfully digested by HpaII . Second, there is an excess of reads that map to HpaII fragments 40–250 bp in length and a deficit of reads mapping to HpaII fragments 0–30 bp in length ( Figure 2A ). It is known that fragments between 50–250 bp are preferentially amplified on the flow cell of Illumina's GA and this explains our enrichment of reads mapping to fragments in that size range. Finally, Figure 2B demonstrates that the sequencing of RRLs successfully produced an excess of bases with high coverage (≥8x) compared to what is expected without any genomic reduction procedure. Overall, we sequenced 26.4% of the 290 Mb assembled genome to ≥1x coverage and obtained no sequence from ∼73.6% of the assembled genome (i.e. 0x coverage). SNPs were identified only from positions with ≥10x and ≤1000x coverage, which represented only 2.3% of the assembled genome. (A very small portion of the genome was sequenced at >1000x coverage (0.01%)). Although we call SNPs from only 2.3% of the assembled genome, the generation of an equivalent amount of sequence data without an enrichment step would have made large-scale SNP discovery impossible as the required coverage would not have been obtained.
Our genomic reduction procedure and subsequent sequencing enabled the identification of 470K putative SNPs. The excess of evidence for polymorphism at the ends of reads in our 470K SNP set closely resembles the previously described distribution of errors across read positions: the sequencing error rate increases towards the ends of reads [28] . This suggests that the simple SNP calling procedure we implemented to generate the 470K set often does not accurately distinguish between true SNPs and error ( Figure 3 ). Our use of SNP calling criteria based on quality score and the genotypic contingency test (see Methods for details) eliminated this read position bias and resulted in our 71K SNP set. It is also worth noting that indels at the ends of reads may not inhibit alignment and can in some instances be mistaken for SNPs in downstream analyses. SNP calling from short-read sequence data is currently in its infancy, and more sophisticated algorithms exist [6] , [29] and will continue to be developed. The fact that the grapevine is highly heterozygous and significantly more genetically diverse than many of the organisms in which SNPs have been called from short-read sequence data [5] , [7] , [8] , makes SNP calling more challenging. In addition, our genome reduction procedure makes it impossible to eliminate the effects of PCR bias as we expect reads to begin and end at the same positions. However, we have demonstrated that a set of simple heuristics can generate a useful data set rapidly and without excessive computational demands. The generation of 71k high-quality SNPs represents a significant enhancement of current genetic resources available to the grape genetics community.
We find relatively few fixed differences (2.7% of SNPs) and a considerable degree of shared polymorphism (24.3% of SNPs) between V. vinifera and wild Vitis species ( Figure 4 ). The wild Vitis species are primarily from North America, but results remain largely the same when Vitis amurensis , the only Eurasian wild species in the present study, is excluded from analysis (data not shown). Moreover, this high degree of shared polymorphism is likely an underestimate since polymorphism was often missed due to low read counts. Despite being geographically isolated for more than 20 million years, there is strong evidence of significant degrees of shared polymorphism between North American wild grapevine species and European cultivated grapevines. This observation supports the view that grapevine species have maintained large effective population sizes for millions of years and that, despite having undergone domestication and breeding, V. vinifera cultivars still harbor variation that dates back tens of millions of years.
We found that LD decays to background levels at inter-SNP distances of ∼10 kb ( Figure 5 ). Consistent with previous reports [21] , [30] , levels of LD in V. vinifera are low, even at short inter-SNP distances. The median r 2 for SNPs within 50 bp of each other is only 0.18, for example. This striking observation suggests that the effective population size of the domesticated grapevine is extremely large and historical recombination has fragmented the V. vinifera genome into very short haplotype blocks. The rapid breakdown of LD in V. vinifera , together with the presence of shared polymorphism between V. vinifera and wild Vitis species, suggests that grapevine domestication did not involve a severe population bottleneck. Future work assessing levels of diversity and LD decay in V. sylvestris , the ancestor of V. vinifera , will allow us to quantify more accurately the severity of the domestication bottleneck in the grapevine.
The consequence of the observed rapid LD decay is that genetic mapping in the cultivated grapevine will not follow other organisms' paths towards genome-wide mapping studies. To date, the path towards GWA and GS has begun with genotyping microarrays that carry tag SNPs, SNPs that effectively capture neighboring variants through LD [31] . The grapevine, however, has such low LD that most functional alleles would not be tagged by a genotyped marker from an array-based assay. Thus, we anticipate that whole-genome sequencing will be required for well-powered genome-wide approaches in the grapevine. There are two other reasons why this is a reasonable way to move forward. First, we found that the quality scores from the Vitis9KSNP array are influenced by the number of SNPs present in the probe sequence ( Figure S2 ). This observation suggests that it may be difficult to obtain high-quality genotype data using genotyping microarrays on high-diversity plant species. Second, because the grape is a long-lived perennial that generally produces fruit 3 years after planting, the focus should now be on establishing a mapping population that effectively captures the diversity within the grapevine, paying careful attention to experimental design (e.g. number of replicates, number of environments, etc.). It is likely that by the time sufficient phenotype data is collected from such a mapping population, the sequencing costs will be minimal compared to the costs of establishing and phenotyping the population. Thus, we argue that it is most effective to now concentrate on establishing grapevine mapping populations that will allow for well-powered genetic mapping studies in the future and to exploit the anticipated low future costs of whole-genome sequencing.
To assess the genetic structure of the grapevine, we have designed the Vitis9KSNP array which we are currently using to genotype ∼1200 V. vinifera and ∼250 wild Vitis species from the USDA's grape germplasm collection. We selected SNPs discovered by Illumina GA sequencing to include on the array based on a number of criteria ( Table S2 and Supplementary Methods S1 ) and observed 97.7% concordance between genotype calls from the Illumina GA data and the genotype calls from the Vitis9KSNP array ( Table 1 ). Table 1 demonstrates that the most common type of error (82% of errors) involves cases in which a SNP is called homozygous from the Illumina GA data but is called heterozygous from the array data. The likely reason for the excess of non-concordant genotypes in these two classes is the presence of polymorphism in HpaII sites: an allele at a SNP will not be sequenced if it is linked to an allele that disrupts the HpaII site at the start of the sequence. Thus, calling heterozygotes from RRLs is necessarily complicated by the presence of polymorphism within the restriction site, especially in highly heterozygous species like the grapevine. Overall, however, the high concordance rates suggest that the array is providing genotypes that are consistent with the Illumina GA sequence data.
Designing a SNP array to assess the genetic structure of an entire genus is challenging; only a few SNPs that show fixed differences between two species may be necessary to distinguish between them. We intentionally introduced an ascertainment bias during SNP selection for the Vitis9KSNP array and favored SNPs that segregate within the cultivated V. vinifera , but also chose a smaller set of SNPs that show fixed differences between each wild species and the V. vinifera samples ( Table S2 ). Selecting SNPs for the array strictly based on quality without regard to segregation patterns results in large numbers of SNPs differentiating the wild Vitis species. This is apparent in the PCA plot generated from 14,325 SNPs chosen without regard to the pattern of segregation among wild and cultivated grapevines ( Figure 5A ). For this unbiased SNP set, there is essentially no differentiation among V. vinifera until PC4, which accounts for only 7.4% of the variance ( Figure S3 ). When PCA is performed on the same set of samples using the biased set of SNPs from the Vitis9KSNP array, PC1 distinguishes between wild Vitis species and V. vinifera , and PC2 accounts for 11.8% of the variance and provides clear separation of the V. vinifera cultivars ( Figure 5B ). The exception is the wild species V. sylvestris , the known progenitor of V. vinifera [18] , which is found close to the V. vinifera as expected. Inclusion of additional samples that we have genotyped with the array demonstrates that the Vitis9KSNP provides power to distinguish between V. vinifera , hybrids and wild species ( Figure S4 ) and even resolves relationships among diverse wild species ( Figure S5 ).
The relative positions of the V. vinifera samples along PC2 in Figure 5B suggest that geography may have an influence on the genetic structure of the domesticated grapevine as PC2 reflects the longitude from which these cultivars are believed to have originated. For example, cultivars from Western Europe (Pinot Noir, Gewurztraminer, Riesling, Ehrenfelser and French Colombard) are concentrated at the top of PC2 while cultivars of eastern origin are found at the bottom of PC2 (Plavac Mali from Croatia; Kadarka from Hungary; Muscat of Alexandria from Egypt; Malvasia from Greece and Thompson Seedless from Iran). The V. sylvestris sample in Figure 5B is from Tunisia, so its position along PC2 is also consistent with the longitudinal gradient. Only a small number of accessions have been analyzed here and the results from our analyses of Vitis9KSNP array data from the entire USDA grape germplasm collection promises to provide a more in-depth view of the genetic structure of the cultivated grape.
Having assessed the diversity of the grapevine using a whole-genome sequencing approach as well as a genotyping array, it is evident that the choice between using either of these two technologies depends very much on the purpose of the study at hand. The design of a high-quality genotyping array with millions of SNPs for GWA in the grapevine is, arguably, an impossible task because of the difficulties associated with assaying diversity across such a diverse genus. It is our view that next-generation sequencing should and will be primarily utilized for GWA studies in high diversity crop species. On the other hand, customized SNP arrays, such as the Vitis9KSNP in this study, will be valuable for preliminary assessments of germplasm collections and for breeders to verify their material.
Methods
SNP Discovery by Illumina GA Sequencing
Genomic DNA was extracted with DNeasy Plant Mini Kits (Qiagen) from young, lyophilized leaves, cambium tissue or leaf bud tissue. Details about the 17 DNA samples are provided in Table S1 . DNA samples were amplified with bacteriophage Phi29 DNA polymerase provided in the Genomiphi whole-genome amplification kit (GE Healthcare). We performed a genome complexity reduction step by fully digesting each sample with the restriction enzyme HpaII (recognition sequence = CCGG) to generate reduced representation libraries (RRLs). HpaII is a methyl-sensitive enzyme, but the genome amplification step prior to restriction digestion eliminates methylation and HpaII therefore behaves as a non-methyl-sensitive enzyme in this case. The standard library preparation for Illumina's 1G Genome Analyzer was then performed for each RRL with one alteration: size selection by gel excision was not performed as our experience suggests that it makes no difference in sequence quantity or quality (Ed Buckler, unpublished data). Each RRL was sequenced on a single lane of the Genome Analyzer with 36 cycles to produce 57.3 million reads. The sequences generated in this study have been submitted into the NCBI short read archive (SRA accession: SRA009211.21). Each 36 bp read was first shortened to 32 bp (a requirement for the alignment tool) and aligned to the grape reference genome [24] using Illumina's ELAND alignment tool. In this manner, we detected 2,271,594 positions in the genome where 2 or more alleles were observed (i.e. putative SNPs).
To obtain a robust set of SNPs from this set of 2,271,594 putative SNPs, we implemented a series of preliminary filters. First, we rejected a putative SNP if the read count for the minor allele(s) was ≤5% of the total read count. This filter aims to distinguish between sequencing error, which should be found at low frequency, and true polymorphism. While this filter likely rejected true low-frequency SNPs in some cases, this is of little concern since we were primarily concerned with identifying intermediate-frequency SNPs. Some putative SNPs were covered by >50,000 reads. Putative SNPs covered by extremely high read counts are more likely to be non-allelic, i.e. the result of paralogy: although a set of reads may align to a single genomic location according to the genome sequence, they in fact are derived from multiple genomic locations that are misrepresented as a single sequence in the currently available genome sequence. To mitigate the paralogy problem, we implemented a second filter whereby putative SNPs were rejected if the total read count was >1000. This second filter also aids computational speed. Third, we implemented an arbitrary read count requirement and rejected SNPs with total read counts <10. Finally, when 3 or 4 alleles were observed, we rejected putative SNPs if the sum of the 3 rd and 4 th most common alleles was ≥2% of the total read count. We then considered only the two most common alleles as we are only interested in identifying bi-allelic SNPs. The implementation of these preliminary filters resulted in 469,470 SNPs, which we refer to as our 470K SNP set. From the 470K SNP set, we identified a 71,397 high-quality SNPs which we refer to as the 71K SNP set. The 71K SNP set was established by choosing SNPs from the 470K set with average Q scores ≥20 and genotypic contingency test p - values ≤0.01. See Supplementary Methods S1 for a detailed explanation of the genotypic contingency test.
Coverage Analysis
A significant proportion (31.1%) of the grape genome sequence has not been assigned to a chromosome. Another 7.9% of the genome is assigned to chromosomes, but not anchored to a chromosomal location. For our coverage analysis, we considered only the 60.9% of the genome sequence that is assigned and anchored to locations on chromosomes 1 to 19. We refer to this portion of the genome as the “assembled genome”.
A total of 17,326,203 reads (554,438,492 bp) were successfully mapped to the assembled genome. We generated the observed coverage distribution by calculating the coverage for every base in the assembled genome (see Figure 3B ). The observed number of bases with no coverage was 234,673,000 bp. Bases can have no coverage because no reads mapped to their location, or because reads cannot be mapped to their location. The latter scenario applies to bases that are unknown (i.e. bases assigned ‘N’ in the genome sequence) and for bases that lie within repetitive regions. We subtracted the number of unknown bases (12,848,811 bp) and the number of bases within repetitive regions (31,282,949 bp) to obtain a more accurate observed number of bases with no coverage (190,541,240 bp). We obtained an estimate of the amount of repetitive sequence in the assembled genome from http://www.genoscope.cns.fr/externe/Download/Projets/Projet_ML/data/annotation/repeats/ .
To generate the expectation from sequencing at random without the use of RRLs, we followed the Lander-Waterman model whereby coverage follows a Poisson distribution if sequence is obtained at random from the genome [32] . Similar to the manner in which we obtained the observed number of bases with no coverage above, we calculated the “mappable portion” of the assembled genome by subtracting the number of unknown bases and the number of bases within repetitive regions from the total number of bases in the assembled genome. Thus, we consider 14.6% of the assembled genome essentially unmappable and exclude it from our calculation of the Lander-Waterman coverage distribution. The random coverage distribution was generated from a poisson distribution with λ = 2.14, where λ is the mean coverage. The mean coverage was obtained by dividing the 286,454,112 bp of sequence that maps to the assembled genome by the 258,954,041 mappable bases of the assembled genome.
Segregation Patterns
We assessed the pattern of segregation within and between V. vinifera and wild Vitis species using read count data from the 71K SNP set. For this analysis, V. sylvestris , the wild progenitor of V. vinifera , was included in the V. vinifera group. SNPs with ≥1 read carrying the reference allele and ≥1 read carrying the alternative allele within V. vinifera were identified as “segregating” or “polymorphic” within V. vinifera . The same criteria were applied to the wild Vitis species. Fixed differences were identified as SNPs with one allele present exclusively in V. vinifera and the other allele present exclusively in the wild Vitis species.
LD Decay
We called genotypes from the raw Illumina GA read data as follows. A genotype was called only if the read count for an individual at that locus was ≥4 reads. Individuals were called homozygous if they carried ≥4 reads for one allele and 0 reads for the other allele. Individuals with ≥4 reads carrying both alleles were called heterozygous. For the analysis of LD decay, only the 10 V. vinifera samples were included. D' is an unreliable measure of LD with small sample sizes and we therefore only present r 2 values. SNPs with ≥2 missing genotypes were excluded. Singleton SNPs were excluded. Using these criteria enabled us to include 16,486 SNPs and provided sufficient resolution to assess LD decay. The genotype calls are likely sufficiently reliable since comparisons between r 2 values generated from this SNP calling method and from the stricter SNP calling method described below under “ Vitis9KSNP array ” were highly correlated (r 2 = 0.95, p<1×10 −15 ). A table of r 2 values and their respective inter-SNP distances was sorted by inter-SNP distance. We calculated the median r 2 in sequential bins of 1000 observations along this table and plotted this value against the mean inter-SNP distance for each bin. Background LD was assessed by calculating 20,000 r 2 values between pairs of SNPs on different chromosomes. Pairwise LD was calculated using the R package “genetics” which incorporates maximum likelihood phase estimates into the estimation of LD [33] .
Vitis9KSNP Array
We called genotypes from the Illumina sequence data and compared them to genotype calls from the Vitis9KSNP array. We attempted to find a set of rules for calling genotypes from the Illumina sequence data that would provide a sufficient number of SNPs for comparison while minimizing the false positive rate. An individual was called a homozygote at a locus if there were> = 5 reads from that individual mapping to that locus and all these reads carried the same allele at that locus. An individual was considered heterozygous at a SNP if it had ≥8 reads mapping to the position and if it passed the heterozygosity test (see Supplementary Methods S1 for details of the heterozygosity test). Genotypes were considered missing data if they failed these conditions. This genotyping scheme results in 820,612 genotype calls from the Illumina sequence data. Genotypes from the Vitis9KSNP array were called using Illumina's BeadStudio software. Our observations suggest that larger sample sizes improve genotype calling. We therefore included 139 samples in addition to the 17 samples sequenced by the Illumina GA when calling genotypes with BeadStudio. Only high-quality genotype calls are useful in assessing concordance between data sets. We therefore visually inspected genotype cluster plots in Beadstudio and decided on a set of strict quality thresholds (GenCall score≥0.5; GenTrain score≥0.7) for SNP calling. The use of these thresholds resulted in 69,078 genotype calls from the Vitis9KSNP array. The total number of genotypes called from both the Illumina sequence data and the Vitis9KSNP array was 36,904.
Principal Components Analysis
Principal components analysis (PCA) was performed using the prcomp command in R [34] . Genotypes were called with the BeadStudio software. Genotype calling included 139 samples in addition to the 17 samples sequenced by the Illumina GA. From visually inspecting genotype clusters, we decided on the following genotype quality thresholds for PCA analysis: GenCall score≥0.15 and GenTrain score≥0.5. We excluded SNPs with call rates <0.8 and SNPs that were monomorphic. The application of these criteria resulted in a set of 5840 SNPs used for PCA analysis.
We called genotypes for SNPs in the 71K SNP set from the Illumina sequence data. To do so, we employed the SNP calling criteria described under the heading “LD decay” of the Methods section above. SNPs called in <14 of the 17 samples were excluded. This resulted in a set of 14,325 SNPs for PCA analysis.
Supporting Information
Methods S1
Supplementary Methods
(0.10 MB PDF)
Figure S1
The distribution of assayed accessions for the 470K and 71K SNP set. In many cases, reads covering a SNP are only obtained from a fraction of the total number of samples sequenced. The histograms partition SNPs by the number of accessions from which reads were obtained.
(0.01 MB PDF)
Figure S2
The effect of neighboring polymorphisms on array-based SNP call quality. Each SNP on the Vitis9KSNP array is queried by a probe sequence that is complementary to the 50 bp of sequence adjacent to each SNP. SNPs within this adjacent probe sequence may reduce probe-sequence hybridization and thus result in poor quality SNP calling. The GenTrain score, along the Y-axis, is a metric of SNP quality assigned to every SNP on the Vitis9KSNP array by Illumina's BeadStudio software. The number of SNPs from the 71K set within each SNPs' probe sequence is shown along the X-axis. The boxplot demonstrates that the GenTrain Score decreases as the number of SNPs present in the probe sequence increases. Thus, obtaining reliable genotype calls using SNP arrays in highly diverse species will be challenging.
(0.16 MB PDF)
Figure S3
Plots of the first 10 PCs generated from 14,325 SNPs chosen without regard to the pattern of segregation among wild and cultivated grapevines. The proportion of the variance explained by each PC is in parentheses above each plot.
(0.08 MB PDF)
Figure S4
A PCA plot of 100 grapevine accessions. The SNP data were generated from the Vitis9KSNP array and only the first 2 PCs are shown. The proportion of the variance explained by each PC is shown in parentheses. The V. vinifera , hybrid Vitis cultivars and wild Vitis species are easily distinguishable along PC1. PC2 distinguishes among V. vinifera cultivars. V. sylvestris , the ancestor of V. vinifera , is found among the V. vinifera cultivars as expected.
(0.03 MB PDF)
Figure S5
A PCA plot of 50 wild Vitis accessions. The SNP data were generated from the Vitis9KSNP array and only the first 2 PCs are shown. The proportion of the variance explained by each PC is shown in parentheses.
(0.03 MB PDF)
Table S1
Additional information on grape DNA samples used in the present study
(0.20 MB PDF)
Table S2
Criteria used to choose the 8988 SNPs assayed by the Vitis9KSNP custom genotyping array
(0.09 MB PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Nasopharyngeal carcinoma (NPC) is a common epithelial malignancy in southern China. The highest incidence has been reported in Guangdong province, where the rate is approximately 20 per 100,000 people per year [1] , [2] . Radiotherapy alone has become the standard treatment for early stage disease, and chemoradiotherapy for the advanced NPC [3] . Biologically different from other squamous cell cancers of the head and neck, approximately 95% of these cases were undifferentiated carcinomas with the highest incidence of distant metastases [4] , [5] . Once metastasis is diagnosed, the overall survival of patients is very poor after palliative chemotherapy. Furthermore, patients with distant metastasis at initial diagnosis had been demonstrated with a significantly shorter survival when compared with those with subsequent metastases [6] – [10] .
However, patients with distant metastasis at initial diagnosis do not behave in a uniform manner. It is hence not surprising to see significantly variable results between studies of similar therapeutic approaches in patients with metastatic NPC [11] , [12] . Although palliative chemotherapy has been demonstrated as the most effective way with high objective response rates, recurrence frequently occurs after chemotherapy ceases. However, the application of radiotherapy of the primary tumor remains controversial because of their short life expectancy and radiation-induced complications [11] – [13] .
Therefore, determining the prognostic factors of survival outcomes in NPC patients with distant metastasis at initial diagnosis could help to select those patients who would most benefit from comprehensive treatment including radiotherapy of the primary tumor by retrospectively analyzing patients’ clinical characteristics, treatment modalities and survival. These results might contribute to management of treatment and exploration of avenues of further research.
Materials and Methods
Patients and selection criteria
Between January 2001 and December 2010, 271 NPC patients presenting with distant metastases at initial diagnosis were referred to our cancer center. The selection criteria were as follows: (1) pathologically confirmed NPC in the nasopharynx, (2) diagnosis of distant metastasis based on physical examination and imaging, (3) receiving at least one anti-cancer treatment including the chemotherapy and the radiotherapy, (4) complete follow-up and clinical data, including laboratory and imaging data. Patients with other malignancies or unstable cardiac disease requiring treatment were excluded. Of the 271 NPC patients, 37 patients were excluded from the survival analysis, including 14 cases because of missing clinical data and 23 cases because of refusing any treatment, leaving 234 patients for evaluation. The clinicopathological data of the 234 patients are presented in Tables 1 , 2 , 3 .
10.1371/journal.pone.0108070.t001 Table 1
Clinical characteristics.
Characteristics
N(%)
Gender
Female
32(14)
Male
202(86)
Age (years)
<48
116(50)
≥48
118(50)
Karnofsky performance score (KPS)
<80
30(13)
≥80
204(87)
Histology
WHO Type 2
8(3)
WHO Type 3
226(97)
Bony metastasis
Present
157(67)
Absent
77(33)
Liver metastasis
Present
75(32)
Absent
159(68)
Lung metastasis
Present
36(15)
Absent
198(85)
Distant nodal metastasis
Present
27(12)
Absent
207(88)
No. of metastatic sites
Single
52(22)
Multiple
182(78)
10.1371/journal.pone.0108070.t002 Table 2
Laboratory characteristics.
Characteristics
N(%)
Haemoglobin (g/L)
<120
29(12)
≥120
205(88)
Lactate dehydrogenase (LDH) (IU/L)
≤245
154(66)
>245
80(34)
Alkaline phosphatase (ALP) (IU/L)
≤110
182(78)
>110
52(22)
VCA-IgA
Negative
13(5)
Positive
221(95)
10.1371/journal.pone.0108070.t003 Table 3
Treatment characteristics.
Characteristics
N(%)
Treatment modality
Chemotherapy alone
94(40)
Chemoradiotherapy
140(60)
Chemotherapy regimen
Cisplatin+fluorouracil
124(53)
Paclitaxel+cisplatin
110(47)
Chemotherapy response
Progression of disease
24(10)
Stable disease *
79(34)
Complete remission+Partial remission †
131(56)
Chemotherapy cycles
1–3 cycles
74(32)
≥4 cycles
160(68)
*52 patients received RT to primary lesions and 27 patients did not received RT; † 88 patients received RT to primary lesions and 43 patients did not received RT.
Ethical Review Committee of Sun Yat-Sen University Cancer Center has approved the project. Written consent was given by the patients to be stored in the hospital database.
Pre-treatment evaluation
All patients had a pre-treatment evaluation including complete history, physical examination, hematology and biochemistry profiles, Epstein-Barr virus (EBV) serology, chest radiographs, sonography of abdomen, whole-body bone scan and magnetic resonance imaging (MRI) of head and neck regions. A titre of more than 1∶20 was considered to be positive for the VCA-IgA antibodies as adopted in previous study on the marker [9] . Patients were evaluated according to the 2002 American Joint Committee on Cancer (AJCC) TNM stages.
Treatment
The treatment modalities were determined according to the experience of our center and the acceptance of the patients. Radiotherapy of the primary tumor was generally administrated to those patients who achieved disease control of the metastatic lesions after chemotherapy. It was also administered to reduce serious symptoms caused by the primary tumor that affected the quality of life. Among the 234 patients, 94 patients received chemotherapy alone (CT), and 140 patients received chemoradiotherapy (CRT).
All the patients were treated with cisplatinum-based chemotherapy. The median number of cycles of chemotherapy was 5 (range, 1–14).
Among of the patients who received RT, 116 (82.9%) were treated with conventional techniques, 20 (14.3%) underwent intensity-modulated radiotherapy (IMRT) and 4 underwent three-dimensional conformal radiotherapy (3D-CRT). Details regarding the RT techniques have been previously reported [14] – [15] . One hundred and seventeen patients received a radiation dose ≥66 Gy and 23 patients underwent a dose <66 Gy. The median dose was 70 Gy (range, 40–78 Gy).
Fifty-five patients received local therapy to metastases, including 39 patients received radiotherapy to bone lesion (30–66Gy/10-33f), 10 received radiofrequency ablation (RFA) and 3 received interventional embolization of liver lesions, and 3 received surgery of lung lesions.
Treatment evaluations and follow up
Imaging of the metastasis was performed after every two courses of chemotherapy, and then every 3 months during follow-up. Objective response was measured according to the Response Evaluation Criteria in Solid Tumors (RECIST). The evaluation of bone metastasis was based on the imaging findings of re-calcification shown in CT and the decreased concentration in the whole bone scanning and the clinical evidence of the pain relief.
Patients were followed up by direct telecommunication mean or by checking the clinic attendance records. The overall survival (OS) was defined as the duration from the date of diagnosis to the date of death from any cause or the censoring of the patient at the date of the last follow-up. The median follow-up for the whole was 22 months (range, 2-125).
Statistical analysis
Statistical analysis was performed using SPSS 13.0 package. Overall survival (OS) was analyzed using the Kaplan-Meier method and was compared using the log-rank test. Univariate and multivariate analysis were performed using the Cox proportion hazards model. The multivariate analyses were undertaken with both forward and backward stepwise procedures for identifying variables correlated with overall survival. Covariates included patients’ characteristics (Karnofsky performance score, gender and age), laboratory parameters (hemoglobin, lactate dehydrogenase, alkaline phosphatase and the EBV serology), metastatic features (extension and response to chemotherapy) and treatment approaches (number of chemotherapy cycles, radiotherapy of the primary tumor and local therapy of metastases). Furthermore, the relationship of response to chemotherapy and various factors was tested by logistic regression model. A two-tailed P-value <0.05 was considered statistically significant.
Results
Treatment response and overall survival
One hundred and fifty-four patients had been dead by the final evaluation date. The main cause of death was progression died of metastatic lesions, which occurred in 137/154 (89.0%) patients; 15/154 (9.7%) patients died of local failure and 2/154 (1.3%) die of cardiac disease. The median OS time was 22 months (range, 2-125 months), and the 1-year, 2-year, 3-year overall survival rates were 82.2%, 51.3% and 34.1%, respectively.
Of the 234 patients, 10/234 (4.3%) achieved complete response (CR) of metastatic lesions, 121/234 (51.7%) achieved partial response (PR), 79/234 (33.8%) had stable disease (SD) and 24/234 (10.2%) had progressive disease (PD). The overall response and disease control rates were 56.0% and 89.8%, respectively. Logistic regression analysis showed that the following factors were significantly associated with poor response to chemotherapy (PD+SD): KPS <80 (P = 0.016); liver metastasis (P = 0.001); LDH>245 IU/L (P = 0.023); and number of chemotherapy cycles <4 (P<0.001).
Toxicities
Two of the patients died of treatment-related toxicity including one with severe infection caused by the grade IV leucopenia and one with the hepatic failure during chemotherapy exhibited. In total, 45.3% developed grade III–IV leucopenia or neutropenia and 16.7% exhibited grade II–III toxicity with vomiting and nausea. Among the patients receiving RT, the most significant toxicity was the grade 3/4 mucositis with a rate of 40.5%, and the skin reaction with a rate of 25.0%. All patients completed the full course of RT.
Univariate analysis and Multivariate analysis
The result of univariate analysis and multivariate analysis are summarized in Table 4 and Table 5 . The negative prognostic factors in the univariate analysis for OS were as follows include KPS<80 (P<0.001), LDH>245 (P<0.001), ALP>110 (P<0.001), Liver metastasis (HR = 2.204, P<0.001), and Multiple metastases (P<0.001). CT alone (P<0.001), Chemotherapy cycles<4 (P = 0.001), Poor response to chemotherapy (P<0.001), and Without local therapy to metastatic lesions (P<0.001) were also associated with poor OS in the univariate analysis.
10.1371/journal.pone.0108070.t004 Table 4
Univariate analysis of variables correlated with overall survival.
Characteristic
Univariate Analysis
P
HR (95% CI)
Gender, men vs women
0.096
1.536(0.927–2.545)
Age, <48 vs ≥48
0.787
0.957(0.698–1.314)
KPS, <80 vs ≥80
<0.001 a
4.712(3.018–7.358)
Liver metastasis, yes vs no
<0.001 a
2.204(1.598–3.039)
Lung metastasis, yes vs no
0.377
0.819(0.525–1.276)
Bone metastasis, yes vs no
0.754
0.948(0.681–1.321)
Distant nodal metastasis, yes vs no
0.800
1.069(0.636–1.798)
Number of involved site,>1 vs 1
<0.001 a
2.648(1.678–4.178)
Haemoglobin, <120 vs ≥120
0.933
1.021(0.624–1.672)
Serum LDH,>245 vs ≤245
<0.001 a
2.554(1.843–3.538)
Serum ALP,>110 vs ≤110
<0.001 a
2.124(1.497–3.014)
VCA-IgA, Positive vs Negative
0.370
0.734(0.374–1.443)
Local therapy to metastases, no vs yes
<0.001 a
2.565(1.657–3.970)
Treatment modality, CT vs CRT
<0.001 a
3.058(2.202–4.247)
Response to chemotherapy, PR+CR
Baseline
SD
<0.001 a
2.251(1.583–3.202)
PD
<0.001 a
6.455(3.876–10.735)
Chemotherapy cycles, <4 vs ≥4
0.001 a
1.783(1.280–2.484)
HR: hazard ration; CI: confidence interval; CT: Chemotherapy CRT: Chemoradiotherapy; PD: Progression of disease; SD: Stable disease; PR: Partial remission; CR: Complete remission; a Statistically significant. 10.1371/journal.pone.0108070.t005 Table 5
Multivariate analysis of variables correlated with overall survival.
Variables
HR(95%CI)
P
Clinical and Laboratory Characteristic
KPS, <80 vs ≥80
4.077(2.481–6.700)
<0.001 a
Liver metastasis, yes vs. no
1.652(1.140–2.393)
0.008 a
Number of involved site, >1 vs 1
2.106(1.288–3.444)
0.003 a
Serum LDH, >245 vs ≤245
1.686(1.187–2.395)
0.004 a
Treatment Characteristic
Treatment modality, CT vs CRT
2.066(1.440–2.964)
<0.001 a
Chemotherapy cycles, <4 vs ≥4
1.748(1.223–2.499)
<0.001 a
Response to chemotherapy, PR+CR
Baseline
SD
2.338(1.591–3.437)
<0.001 a
PD
3.370(1.947–5.833)
<0.001 a
HR: hazard ration; CI: confidence interval; CT: Chemotherapy CRT: Chemoradiotherapy; PD: Progression of disease; SD: Stable disease; PR: Partial remission; CR: Complete remission; a Statistically significant.
The multivariate analysis show that the significant prognostic factors for poor survival were KPS<80 (HR = 4.077, P<0.001), LDH>245 (HR = 1.748, P = 0.004), Liver metastasis (HR = 1.652, P = 0.008), and Multiple metastases (HR = 2.106, P = 0.003). CT alone (HR = 2.066, P<0.001), Chemotherapy cycles<4 (HR = 1.748, P<0.001), and Poor response to chemotherapy(PD group, HR = 6.455, P<0.001; SD group, HR = 2.251, P<0.001) were also associated with poor OS in the multivariate analysis. Patients with good performance status (KPS≥80) survived longer than those with poor performance status (3-year OS: 37.9% vs. 4.4%). Patients with normal LDH level had a better survival than those with high LDH level (3-year OS: 44.7% vs. 13.7%). The 3-year OS rate for patients with liver metastasis was poorer than those without liver metastasis (14% vs. 45.7%). Patients with single metastasis had a better survival than those with multiple metastases (the 3-year survival rates: 65.8% vs. 25.9%). Furthermore, the therapy related factors were also associated with OS. The 3-year OS rate for patients receiving chemotherapy cycles <4 was poorer than those receiving chemotherapy cycles ≥4 (23.2% vs. 39.1%). The 3-year survival of patients receiving CRT was 48.2%, better than those receiving CT alone with only 12.4%. Patient with response to chemotherapy of metastatic lesions also show better survival with 3-year OS rate of 38.0% for patients with PR or CR, and 14.2% for patients with SD, and none for patients with PD. These results are shown in Figure 1 .
10.1371/journal.pone.0108070.g001 Figure 1
Overall survival rates according to KPS (a), liver metastasis (b), number of metastatic site (c), radiotherapy of primary tumor (d), response to chemotherapy (e), number of cycles of chemotherapy (f) and LDH (g).
For patients who achieved CR or PR after chemotherapy of metastatic lesions, multivariate analysis showed that radiotherapy of the primary tumor was an independently significant favorable prognostic factor (HR = 0.435, P = 0.001). Significantly improved survival was achieved by radiotherapy of the primary tumor in these patients (3-year OS rate 59.6% vs. 20.3%, P<0.001, Figure 2a ). For patients who achieved SD after chemotherapy of metastatic lesions, multivariate analysis also showed that radiotherapy of the primary tumor was an independently significant favorable prognostic factor (HR = 0.363, P = 0.001). Significantly improved survival was achieved by radiotherapy of the primary tumor in these patients (3-year OS 24.7% vs. 0%, P = 0.003, Figure 2b ).
10.1371/journal.pone.0108070.g002 Figure 2
Overall survival rates for patients who achieved CR or PR after chemotherapy of metastatic lesions (a), for patients who achieved SD after chemotherapy of metastatic lesions (b).
Discussion
For patients presenting with distant metastases at initial diagnosis, the optimal treatment strategy remains a subject of debate [11] – [13] . The benefits of systemic chemotherapy have been demonstrated in some studies and considered as the only possibly curative option. Platinum-based combination regimen achieves high response rates and is the most widely used regimen [11] , [16] , [17] . For the number of cycles of chemotherapy was an independent factor associated with survival, it was important for patients receive a sufficient number of cycles. However, it was still uncertain regarding the optimal cycles of chemotherapy. In a retrospective study involving 20 long-term disease-free survivors with metastatic NPC reported by Fandi et al. [11] , the results showed that approximately six cycles of chemotherapy were required. In the current study, the cut-off point of number of cycles was evaluated by the Receiver operating characteristic (ROC) and the patients with at least four cycles of chemotherapy had a significantly better survival than those with less than four cycles. The results indicated the importance of sufficient chemotherapy for patients with metastatic NPC. However, owing to the retrospective nature of this study, it was still hard to determine the optimal cycles of chemotherapy. Furthermore, the response of metastatic lesions to chemotherapy was demonstrated as a significant predictor of OS. The overall response rate (CR and PR) after chemotherapy was 56.0%, and poor response was associated with KPS <80, liver metastasis, LDH>245 IU/L and number of chemotherapy cycles <4, suggesting that these factors could be potential predictors of treatment response. The response of metastatic lesions to chemotherapy also plays a key part in the consideration of the treatment choice. The results indicated that patients with CR or PR were recommended for a more progressive treatment as this could significantly improve survival.
In the clinical practice, the most controversial issue for NPC patients initially with metastases was the application of radiotherapy to the primary tumor for the uncertain indications in the guideline of NCCN (National Comprehensive Cancer Network), which posed great challenge for the oncologists [12] , [13] . It was often considered as inappropriate to give a prolonged course of radiotherapy to patients with stage IVC NPC because of their short life expectancy and serious late complications in the past era. However, due to the improvements in radiation techniques and increasing efficacy of platinum-based combination regimen, some studies show that the local control of primary tumor following the radiotherapy would improve the quality of life and contribute to prolonged survival. In a retrospective analysis of 125 NPC patients initially with metastases reported by Yeh et al. [13] , the 2-year OS rate was 24.0% when they received radiotherapy alone when compared to 10% in those who received chemotherapy alone, and it also showed that the local control of the primary tumor improved the quality of life because of the reduced necrosis, bleeding and severe headaches. In the current study, the application of radiotherapy after chemotherapy was a positive factor associated with survival. The 3-year OS of patients receiving radiotherapy after chemotherapy was up to 48.2%, significantly higher than those receiving chemotherapy alone with only 12.4%. However, the survival benefit may be also related to the selection for radiotherapy. Therefore, it was very important to select the patients who would most benefit from the radiotherapy. In the subgroup analysis, we found the radiotherapy could significantly improve the survival of patients who achieved the CR or PR of metastatic lesions after chemotherapy with a 3-year OS rate of 59.6%. Even though for patients who achieved SD after chemotherapy of metastatic lesions, significantly improved survival was achieved by radiotherapy of the primary tumor in (3-year OS 24.7% vs. 0%, P = 0.003).
These findings indicated that excellent local control may help reduce the tumor burden and the risks of death caused by progression of primary tumor, especially for the patients with CR/PR or SD of metastatic lesions after chemotherapy. Furthermore, the improvements of radiation technique such as the application of Intensity-modulated radiotherapy (IMRT) may further improve the treatment benefit.
Part of our results were consistent with those reported by Toe et al. [6] , liver metastasis was associated with poor survival. In the current study, the 3-year OS rate of patients with liver metastasis was only about 14.0%, significantly poorer than those with other metastasis included the lung, bone or distant nodal metastasis with a 3-year OS rate of 43.7%. In the retrospective analysis of 379 NPC patients with subsequent metastases reported by Hui et al. [7] , the lung metastasis alone was demonstrated as a positive factor of survival and long-term survival was possible for those patients. The reason for poor survival of liver metastasis may relate to the rich blood supply of liver and the low rate of the response to chemotherapy. Furthermore, the patients with single metastasis exhibited the excellent survival with 3-year OS rate of 65.8%, while only 25.9% for those patients with multiple metastases. It may be the sub-group of long-term survival after aggressive approach to treatment.
Elevated levels of LDH also demonstrated as a negative prognostic factor, which may be associated with large tumor burden, tumor extension and high risk of metastasis [18] – [20] . Serum LDH levels twice normal levels are rarely seen in loco-regional disease but are commonly observed in NPC patients with liver metastasis or multiple organ metastases. Studies have found that NPC patients with elevated baseline LDH levels were more likely to develop liver metastasis after treatment. In the study of Jin et al. [20] , elevated LDH levels were reported in over 55.0% of patients with metastatic NPC, the relative risk to die increased with LDH>245 IU/L by the factor 1.8. In our study, the 3-year OS rate of patients with normal level of LDH was about 47.7%, significantly higher those with elevated LDH levels with a 3-year OS rate of 13.7%. More than 60% of patients with liver metastasis had elevated levels of LDH. Furthermore, elevated LDH was also associated with poor response of metastatic lesions to chemotherapy. Pretreatment serum level of LDH may be a potential predictor.
This retrospective analysis has several weaknesses. First, the circulating EBV DNA load has been demonstrated as an independent prognostic factor in disseminated NPC [21] . However, only small part of patients’ EBV DNA data was collected in our study, therefore we had excluded the factor to avoid the potential bias. Second, treatment modality has an impact on survival outcome in patients with disseminated NPC at initial diagnosis. Since the treatment modalities were selected according to the physician’s policy of practice in our study, it is inevitable to cause selection bias when we identify prognostic factors for patients with distant metastases at initial diagnosis.
Conclusion
In this study we identified some negative prognostic factors for patients with distant metastases at initial diagnosis which included poor performance status, elevated levels of LDH, liver metastasis and multiple metastases. We also found that chemotherapy alone, chemotherapy cycles<4 and poor response to chemotherapy were associated with poor OS. It can help to select the appropriate patient for more progressive treatment of a combination of chemotherapy and radiotherapy. Long-term survival is possible for patients with less negative prognostic factors. Prospective randomized studies are needed to optimize treatment strategy.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
In Japan, the number of child abuse consultations handled by child guidance centers nationwide has continually increased in the past 30 years, exceeding 200,000 cases for the first time in the financial year (FY) 2020. Prevention of child abuse requires cooperation among multiple organizations [ 1 ], including medical, healthcare, and welfare institutions [ 2 ], and social welfare organizations [ 3 ]. In Europe and the U.S., Nierop et al. [ 4 ] discovered a relationship between stress during pregnancy and postpartum depression. They pointed out the importance of putting in specific efforts from the pregnancy period onward to ensure abuse prevention. In a randomized controlled trial, Olds et al. [ 5 ] found that prenatal home visits by nurses reduced child maltreatment and revealed the importance of maternal involvement during pregnancy. Furthermore, Ashraf et al. [ 6 ] state that health care providers can identify risk factors and signs of abuse in the medical setting. Additionally, referrals to community resources, parenting education, and other preventive measures must be incorporated into clinical practice. In Japan, there are several mother-child support programs for the prevention of child abuse, such as “Healthy Start Oita” [ 7 ] and “Suzaka Trial” [ 8 ], which provide seamless support during pregnancy, childbirth, and the postpartum period, per the characteristics of each region (for example, whether the government and obstetric hospitals can easily collaborate on a daily basis).
In 2011, the Okayama Prefecture started its operation of the “Contact system for support for mothers and children of concern during pregnancy” (hereinafter referred to as the “Okayama model”) [ 9 ], strengthening seamless support for pregnant and postpartum women of concern (PPWC) in collaboration with obstetric facilities and the community. The Okayama model is characterized by early recognition of risk factors at obstetric facilities by focusing on medical and social backgrounds and efforts to provide support in cooperation with multiple professions throughout pregnancy. According to the Ministry of Health, Labour, and Welfare’s FY 2018 Welfare Administration Report [ 10 ], the number of municipal consultation responses in the Okayama Prefecture was 850 in FY 2018 (3.54 consultations per thousand population per year), compared with 1,641 consultations in FY 2012 (6.29 consultations per thousand population per year), which represents a nationwide increase. However, the number of child abuse consultation responses has reduced. Kobayashi [ 11 ] analyzed child abuse and death cases and concluded that the key institutions for abuse prevention involved health and medical centers, hence stating the importance of cooperation between the two.
A multidisciplinary approach through partnership represents a standard for preventing child abuse [ 12 , 13 ]. Early interventions with pregnant women at risk of child abuse are considered effective in preventing child abuse [ 14 ]. It is at administrative agencies, such as obstetric facilities where women are diagnosed with pregnancy and health centers where maternal and child health handbooks are issued, that public health nurses and midwives have contact with pregnant women at risk of child abuse or pregnant women about whom they feel a vague sense of alertness that “causes them to be concerned about something.” Public health nurses (PHNs) and midwives must share information to provide continuous support for preventing abuse if either institution identifies PPWC. Both PHNs and midwives provide ongoing support to expectant mothers: the PHNs through long-term life support for pregnant and postpartum women and their families, and the midwife in her role as a close supporter of women’s health. Hence, both are important partners in primary care for pregnant and postpartum women, and together they may be able to identify health problems and signs of potential child abuse.
Stolper et al. [ 15 ] stated that a feeling that “there is something wrong here,” a vague and intuitive sense of alertness, helps child health nurses become alert to situations that may lead to child abuse or maltreatment. Furthermore, a U.S. study by child abuse pediatricians (CAPs) reported that the diagnosis of child abuse is procured by a combination of intuitive responses elicited by family encounters and social information obtained outside those encounters [ 16 ]. Due to differences in their respective specialist disciplines, PHNs and midwives may have discrepancies regarding the identification of PPWC. Clarifying these differences is helpful for PHNs and midwives to understand each other’s perspectives and cooperate.
There are individual differences in the ability of health visitors to find and contact medically and socially high-risk pregnant women, depending on the person in charge [ 17 ]. There are cases wherein the affected individual may not receive support. Adachi et al. [ 18 ] clarified that whether a pregnant woman needs support is determined by the competence of the PHN. Similarly, it has been reported that there are individual differences in the ability of obstetric nurses to assess the risk of abuse depending on whether or not they have experience in caring for mothers about whom there are concerns related to child maltreatment [ 19 ]. Differences in the quality of their support (such as being able to recognize the risk of abuse) are influenced by their years of experience [ 20 ]. Experienced PHNs and midwives may have some standard assessment points when determining whether they are “concerned” about pregnant women and need to provide support. Clarifying these points will reduce the likelihood of such cases being overlooked due to individual abilities.
The latest research on child abuse prevention has focused on the characteristics of at-risk pregnant and postpartum women and the social and medical factors related to child abuse. So far, no research has elucidated the intuitive concern to pay attention to situations that may lead to child abuse from the perspective of PHNs and midwives who provide direct support to pregnant and postpartum women.
Therefore, this study aimed to clarify the characteristics of PPWC, as observed by experienced PHNs and midwives, in support of child abuse prevention. Our findings will promote abuse prevention and support from early pregnancy, resulting in the PHNs and midwives working together to ensure that these women in need of support are not overlooked.
Operational definitions
Pregnant and postpartum women of concern
The implication of PPWCs is that the PHNs and midwives feel concerned about them while providing support during pregnancy and worry about the possibility of issues leading to child abuse.
The perspective of child abuse prevention
This refers to the perspectives of PHNs and midwives regarding awareness of the risk of child abuse due to the background factors, attitudes, and moods of pregnant and postpartum women, and on starting preventive support early on.
Materials and methods
Research design
This was a qualitative descriptive study with an inductive approach.
Research participants
The participants included PHNs and midwives working at municipal health centers and obstetric medical institutions actively utilizing the contact system for the Okayama model in operation in the Okayama Prefecture. We requested cooperation in the study by written and verbal means from the general PHNs and the director of the nursing department of each institution. We also received recommendations of those who have had five years or more of experience in supporting mothers and children, including child abuse prevention.
Data collection
The participants’ data were collected from August to November 2019. Semi-structured interviews were conducted based on an interview guide with the PHNs and midwives who agreed to participate at locations that were designated by the participants. Face-to-face interviews were conducted by the first author. The interview guide was developed after discussion with the faculty and graduate students in the field of adult and child health nursing, specializing in maternal support. During the interviews, the interviewees were asked to recall cases and situations where they felt “concerned” about pregnant and nursing mothers under the premise of abuse prevention and to describe their experiences and reasons for feeling concerned. The duration of the interviews with the PHNs was 54 minutes 46 seconds ± 10 minutes 59 seconds (mean ± SD) and with the midwives was 55 minutes and 16 seconds ± 10 minutes and 15 seconds (mean ± SD). All interviews were recorded on an integrated chip (IC) recorder with the consent of the participants and were transcribed verbatim, maintaining anonymity.
Data analysis
Data were analyzed using qualitative inductive analysis methods. During the analysis, personally identifiable information was anonymized. Data analysis was conducted simultaneously with data collection, with interviews transcribed immediately after each interview. The interview transcripts were cross-checked among the researchers. The contexts of the concerns that the PHNs and midwives felt about expectant mothers were extracted from the transcripts. Coding was performed while considering these contexts so that the meaning of the narratives could be understood, and subsequent categorization was carried out based on the similarities and differences. After aggregating similar categories and examining their relationships, the categories were grouped. The codes were extracted then compared by two co-researchers with experience in qualitative research. The authors specialize in community nursing, and their expertise in supporting children and their families to live safely and healthily in the community was helpful in capturing concerns about mothers during the coding process. When opinions differed during the categorization process, the researchers repeatedly reviewed the results until a consensus was reached. They checked for any gaps between the intentions of the participants and the interpretation of the data, and presented the results to the participants to confirm the accuracy of the content.
Ethical considerations
This research was approved by the Institutional Review Board of the Okayama University Graduate School of Health Sciences (D19-1). We briefed the participants verbally and in writing regarding the study’s purpose, the provision of voluntary and free withdrawal from participation, protection of personal information ensuring anonymity of the data provided, data storage method, and information regarding the publication of research results. After this, written consent was obtained.
Results
Overview of research participants
The research participants were ten PHNs working at municipal health centers in the Okayama Prefecture and ten midwives working at obstetric medical institutions, comprising 20 people in total. The number of years of experience as PHNs was 21.0±5.4 (mean±S.D.) years and 22.0±11.3 (mean±S.D.) years as midwives ( Table 1 ).
10.1371/journal.pone.0281362.t001
Table 1 Background of the participants.
Public Health Nurses (n = 10)
Midwives (n = 10)
Age
Under 35 years
1
2
35–39 years
0
2
40–49 years
7
1
50 years and older
2
5
Number of years of experience
21.0±5.4[7–28] a
22.0±11.3[6–40] a
Annual number of births at affiliated worksite
Less than 100
1
100 to 199
3
200 to 299
2
300 and over
4
Number of deliveries at affiliated worksite
Less than 100
1
100 to 499
5
500 to 999
1
1,000 and over
3
This table demonstrates the age of the PHNs and midwives, their years of work experience, the annual births reported by PHNs, and the deliveries reported by the midwives at their respective affiliated worksites.
a Mean±S.D. [range]
PPWC as seen by PHNs
We extracted four main categories, 12 subcategories, 32 subordinate categories, and 168 codes during the data analysis ( Table 2 ).
10.1371/journal.pone.0281362.t002
Table 2 Characteristics of pregnant and postpartum women of concern as observed by public health nurses.
Main Category
Subcategory
Subordinate Category
ID
Difficulties in daily life circumstances
Daily life foundations are unstable
Having economic instability
A,B,C,E,H
Difficulties in getting accustomed to the area due to transfer
D,E,G
Lack of ability to support parents’ families
Parents’ households are in financial distress
E,H
Family members at parents’ households have health issues
H,G
Difficult to receive support from surrounding people
Poor relationship with parents
B,C,E,H
Refusing support or involvement
B,C,D,I
Having nobody to rely upon except parents
E,G
Cannot envision life plan after giving birth
No planning for pregnancy
A,B,C,E,F,G,I,J
Cannot envision daily life after giving birth
B,I,H
No progress in preparations for life after giving birth
H,I
In a dirty living environment
Garbage scattered in room and not cleaned up
G,H
Unsanitary child-rearing space
A,H
PHNs’ discomfort with the mannerisms of the pregnant women
Difficult to communicate with
No progress in verbal exchange, and feeling that no conversation is established
A,B,C
Poor facial expression
B,E,J
Few remarks initiated by the individual
A,B,C,D,H,J
Insufficient comprehension
C,E,J
Having a unique way of thinking and mood
Having particular views on pregnancy and childbirth style
D,H
Dressing in an unkempt manner
G
Having particular views on unique health methods and ways of thinking
F,H
Having unique mood
A,C,F,G,H
Having mental instability
Having risk of mental instability
A,B,E,F,H,J
Having emotional instability
A,D,F,H
Have difficulty in child-rearing behavior
Feelings are not directed toward the fetus/child
Having behaviors and attitudes that do not celebrate pregnancy
A,B,E,F,G,H,I
Being unable to feel affection for fetus/child
B,E,F,G,I
Not showing care and attention toward child
B,I
Not changing drinking or smoking habits for fetus
B,G,I
Having inappropriate child-rearing attitude toward older child
Having cold attitudes toward the older child
F,H
Being unable to take care of older child
G,H,I,J
Feeling unsure about child-rearing techniques
Feeling anxious about child-rearing and not feeling confident
B,F,G
Having insufficient child-rearing skills and knowledge
B,C,G,H
Have multiple risk factors recognized by an assessment tool
Multiple risk factors checked by an assessment tool
Multiple risk factors checked by organization’s own risk indicators
A,B,C,J
Multiple risk factors checked through a common form: “Contact form for support for mothers and children of concern during pregnancy”
A,B
This table demonstrates the four categories identified from the interviews with PHNs: have “difficulties in daily life;” “a sense of discomfort of not feeling like a normal pregnant woman;” “difficulty in child-rearing behavior;” and “multiple risk factors checked by objective indicators using an assessment tool.”
Difficulties in daily life circumstances
This category covers the health workers’ awareness that the pregnant or nursing mother had difficulties in her family background, living environment, socioeconoomic background, history, etc., and that stable living conditions were not in place, which caused them to be concerned about the mother’s situation. This category comprised five subcategories: “daily life foundations are unstable,” “lack of ability to support parents’ families,” “difficult to receive support from surrounding people,” “cannot envision life plan after giving birth,” and “in a dirty living environment.”
The findings showed that the pregnant and postpartum women’s unstable financial conditions and “difficulties in getting accustomed to the area due to transfer” resulted in unstable daily life foundations. Furthermore, even if they wanted their parents’ support, their situations showed a lack of ability to support the family home through factors such as the “lack of ability to support parents’ families” due to “financial distress” and “health issues” among the family members at the parents’ households. Furthermore, there were cases where the pregnant or parturient women refused “support or involvement” from PHNs or surrounding people. These women refused counseling when they were informed that they were pregnant and tried to return home after receiving the Mother and Child Health Handbook. They also had a “poor relationship with their parents” due to histories of childhood abuse and “having nobody to rely upon except parents”, which resulted in it being “difficult to receive support from surrounding people.” There were also cases where women were pregnant or parturient without plans for the daily life arrangements necessary for pregnancy or childbirth. Thus, a concern was raised regarding those who could not “envision a life plan after giving birth.” Furthermore, there were concerns raised by PHNs regarding “difficulties in daily life” among these women or their socioeconomic backgrounds as they lived in “dirty living environments”, with “garbage scattered in their rooms” without being cleaned or “unsanitary child-rearing spaces.”
PHNs’ discomfort with the mannerisms of the pregnant women
This category represented the discomfort that PHNs felt toward pregnant and postpartum women during their interactions at the time of pregnancy notification due to the unique mood or behavior of the women presented. This involved three subcategories: “difficult to communicate with,” “having a unique way of thinking and mood,” and “having mental instability.”
Concerns about “having a unique mood” were mentioned, with PHNs stating, “I am not exactly sure, but intuitively feel concerned” (PHN:C) about the unusual appearance of the woman’s hair or their complicated relationships with their companions at the time of pregnancy notification. There were also concerns about “having particular views on pregnancy and childbirth style,” such as a strong desire for a painless home delivery and “having particular views on unique health methods and ways of thinking.” These involved the beliefs of women wanting to control chronic illnesses through alternative healing powers. PHNs also mentioned concerns regarding the women “having mental instability,” such as “emotional instability,” as they had sad facial expressions, cried easily, or were easily swayed by their symptoms of mental illness. Meanwhile, PHNs witnessed cases wherein the pregnant women did not speak a word during the interview at the time of pregnancy notification, and their parents answered all questions. Sometimes, the conversations lacked factual statements regarding financial aspects, childcare supporters, and chronic illnesses, leading to “few remarks initiated by the individual.” These factors indicated “discomfort of not feeling like a normal pregnant woman,” with the PHNs sensing that it was “difficult to communicate with” the women due to their “poor facial expressions” or lack of “progress in verbal exchange” stemming from their inconsistent words and actions.
Have difficulty in child-rearing behavior
This category indicated concerns by PHNs that pregnant and postpartum women may not be able to perform appropriate child-rearing behaviors due to their lack of interest and involvement in their fetuses or babies, and their lack of confidence and skills in raising them. This was composed of three subcategories, namely, “feelings are not directed toward the fetus/child,” “having inappropriate child-rearing attitude toward the older child,” and “feeling unsure about child-rearing techniques.”
Concerns raised by PHNs included the women “having behaviors and attitudes that do not celebrate pregnancy,” with the nurses mentioning that the women did not view the pregnancy positively because it was unwanted. One PHN mentioned that the women “were unable to do the other things they wanted to do because of the pregnancy” (PHN:G). Other aspects included not finding their child cute and “being unable to feel affection for fetus/child,” resulting in the women strongly prioritizing themselves and “not changing drinking or smoking habits for the fetus.” Moreover, pregnant women expressed feelings of not being interested in or concerned about the fetus/child, such as not hugging the child even when it cried and “not showing care and attention toward the child.” PHNs also raised concerns about the women “having cold attitudes toward the older child” by ignoring their other biological children or step-children and treating them in an aggressive manner. This was described in the following narratives: treating the older child as if they are dirty and not letting him/her touch the baby, intentionally ignoring him/her even when he/she cried, and brushing away the step-children when they came near the baby. In addition, the PHNs expressed concern about the women “being unable to take care of the older child,” for example, through cleanliness and health management, due to insufficient child support. These women were viewed as “having inappropriate child-rearing attitude toward the older child.” A typical example was:
The woman would say that they are pregnant and they are having a difficult time, so they cannot do household chores, and they would ask the older sister in the upper grade of elementary school to even skip school to do household tasks or take care of the baby (PHN:J).
Meanwhile, PHNs mentioned concerns about the women “feeling anxious about child-rearing and not feeling confident.” One participant mentioned that although the pregnant or parturient woman felt affectionate toward the fetus or child, they are “ultimately not giving affection, or because they do not receive affection themselves, they do not know how to do it and feel anxiety” (PHN:B). There was also concerns about the women “feeling unsure about child-rearing techniques” because they did not know how to raise their children as they had “insufficient child-rearing skills and knowledge.”
Have multiple risk factors recognized by an assessment tool
This category indicated the PHNs’ concern toward those pregnant or parturient women with multiple risk factors for child abuse from objective indicators, such as medical records used for interviews or checklists within the organization. This category was composed of “multiple risk factors checked by an assessment tool.”
There are concerns of the woman being at high risk when there are multiple factors on the contact form for support for mothers and children [of concern during pregnancy] from the obstetrics facility, such as a history of mental illness, being in a step-family, or being of advanced maternal age. (PHN:A)
PPWC as seen by midwives
We extracted four main categories, nine subcategories, 33 subordinate categories, and 178 codes during the data analysis ( Table 3 ).
10.1371/journal.pone.0281362.t003
Table 3 Characteristics of pregnant and postpartum women of concern as seen by midwives.
Main Category
Subcategory
Subordinate Category
ID
Mental and physical safety of mother is in jeopardy
Having risk of childbirth that jeopardizes maternal safety
Being careless about managing own physical condition
a,b,c,f,i,j
Suspected domestic violence leading to miscarriage
a,d,f,g
Being fixated on desired childbirth style
b,d,i
First visit or hospital transfer after 30 weeks of gestation
h,j
Previous experience of childbirth with no prenatal care
a
Having mental instability
Feeling depressed and having sad facial expressions
b,c,j,h
Having emotional instability
d,h,f,g
Having history of mental illness
a,c,d,g
Have difficulty in child-rearing behavior
Feelings are not directed toward the fetus/child
Not showing care and attention toward child
a,b,c,g
Trying to prioritize self over raising child
a,b,d
Being unable to feel affection for child
a,c,j
Being unable to accept pregnancy
b,h
No progress in preparations for childbirth and life after giving birth
Using or considering livelihood protection or hospitalized midwifery system due to financial instability
a,c,d,e,f,j
Not feeling like she will become a mother
e,h,j
Not having a fixed residence
a,f
No progress in preparation of items for childbirth
e,h
Having inappropriate child-rearing attitude toward older child
Having cold attitudes toward the older child
b,i,g,j
Older child is unkempt in appearance
c,e
Difficulties due to child-rearing not progressing as expected
Feeling confused due to being unable to raise child as expected
a,b,d,i,j,h
Lack of knowledge regarding child-rearing is evident
b,d,f,i,j
Having many minor questions
b,d,h
Fixating on child-rearing by the book
b,d,h,j
Feeling unsure about child-rearing techniques
j,h,g,b
Have difficulties in maintaining relationships with surrounding people
Difficult to receive cooperation from surrounding people
Having minimal cooperation from husband or partner
c,g,j,I,d
Being unkempt and having a detached mood from surrounding people
e,f,j,d
Refusing to have other people step in to individual affairs
a,e
Having a small number of visitors and visits during hospitalization
a,j,e
Having unreliable parents due to disagreement or history of abuse
h,c,j
Difficult to communicate
Few remarks initiated by the individual
d,j,e,b
Having disjointed conversations
b,d,f,e
Poor comprehension
a,e
Have multiple risk factors recognized by an assessment tool
Multiple risk factors checked by an assessment tool
Multiple risk factors checked through a common form: "Contact system for support for mothers and children of concern during pregnancy"
a,b,c,d
Multiple risk factors checked by organization’s own risk indicators
b,e,h,j
This table demonstrates the four categories identified from the interviews with midwives: “mental and physical safety of mother is in jeopardy;” “have difficulty in child-rearing behavior;” “have difficulties in maintaining relationships with surrounding people;” and “have multiple risk factors recognized by an assessment tool.”
Mental and physical safety of mother is in jeopardy
This category referred to midwives feeling that pregnant and postpartum women were at risk of being unable to have safe births. It comprised two subcategories, namely “having risk of childbirth that jeopardizes maternal safety” and “having mental instability.” Midwives mentioned concerns regarding cases where they were unable to keep track of the pregnancy until just before giving birth, such as “previous experience of childbirth with no prenatal care,” “first visit or hospital transfer after 30 weeks of gestation,” or when the pregnant or parturient woman was “being fixated on a desired childbirth style,” such as a painless home delivery, without considering her body’s safety. Furthermore, midwives felt concerned about the mother’s mental and physical safety being in jeopardy if they had negative emotional expressions, like “feeling depressed and having sad facial expressions,” or psychological issues, like “having mental instability.”
Have difficulty in child-rearing behavior
This category referred to the midwives’ concerns about pregnant or parturient women experiencing difficulties in raising children because of their lack of affection or involvement with their children and siblings. It comprised four subcategories: “feelings are not directed toward the fetus/child,” “no progress in preparations for childbirth and life after giving birth,” “having an inappropriate child-rearing attitude toward older child,” and “difficulties due to child-rearing not progressing as expected.”
Midwives felt concerned that the women did not have affection toward their children as they were “unable to accept the pregnancy” or that they were not thinking about their daily life after childbirth at all and making “no progress in preparations for childbirth and life after giving birth” because they did not want to acknowledge that they would become mothers.
Furthermore, midwives mentioned concerns about difficulties occurring when child-rearing was not progressing as expected, such as “feeling confused due to being unable to raise the child as expected.” This was because the challenges of raising a child exceeded the women’s expectations and were different from what they had anticipated, as seen in the statement: “She became pregnant with infertility treatment, and it was all well and good at childbirth, but she cried about not thinking it would be so tough, and she does not want to take care of the baby” (MW:d). The women were also “fixating on child-rearing by the book” and “having many minor questions,” as seen in the statement: “They would thoroughly read through the child-rearing book and immediately contact nurses when something does not go exactly as mentioned and what they should do” (MW:d). Midwives also expressed concern about “difficulty in child-rearing behavior.” For instance, they witnessed pregnant or postpartum women using harsh words or actions toward their older child in the waiting rooms or wards, with the women “having cold attitudes toward the older child,” or “having inappropriate child-rearing attitudes toward older child” when the “older child is unkempt in appearance.”
Have difficulties in maintaining relationships with surrounding people
This category referred to midwives’ concerns that the women lacked an immediate supporter who could cooperatively help them, rejected support, or did not have any relationship with surrounding people. This comprised two subcategories: “difficult to receive cooperation from surrounding people” and “difficult to communicate.” Midwives felt concerned about the women engaging in child-rearing and household work by themselves after being discharged from the hospital due to “having minimal cooperation from husband or partner,” “unreliable parents due to disagreements or history of abuse,” or “having a small number of visitors and visits during hospitalization.” One participant mentioned, “The husband has night shifts, so the woman is always thinking about how to stop [the baby from] crying” (MW:g). Furthermore, the midwives expressed concern about the pregnant or parturient woman finding it “difficult to receive cooperation from surrounding people” because they “refusing to have other people step in to individual affairsrefuse to let others into their personal matters:” “They have an atmosphere of not wanting others to get involved, such as, “It is fine, I will do everything by myself’” (MW:a).
The midwives also expressed concern about “difficulties in maintaining relationships with surrounding people” because of difficulties in communication, such as in cases where there were “few remarks initiated by the individual,” as seen in the statement: “The woman usually does not have conversations with the midwife, and even if she comes [to the medical examination] with her mother, it is just the mother talking, and there are few reactions from the woman herself” (MW:j). There were also instances of “having disjointed conversations,” as seen in the statement, “The answer I get is slightly different from what I asked. She likes giving lots of answers to things she is interested in. She does not respond for the important parts” (MW:d).
Have multiple risk factors recognized by an assessment tool
This category referred to concerns regarding pregnant and postpartum women with multiple risk factors related to child abuse, using objective indicators such as maternity interviews during initial visits and checklists used within facilities during maternity examinations. This category comprised “multiple risk factors checked by an assessment tool.” A typical example is as follows:
I try to make a template and pick up people who are likely to require follow-ups during pregnancy so that I can continuously do so […] I try to keep an eye on people throughout their pregnancies when they have several risk factors (MW:h).
Discussion
PPWC as seen by PHNs and midwives
Characteristics of how PHNs and midwives viewed PPWC included determining the target women by maximizing their specialist strengths in their respective professions and viewing PPWC as targets for support.
Common aspects of PPWC, as seen by both PHNs and midwives, were those considered so-called specified pregnant women [ 21 ], who did not display affection toward the fetus/child because of undesired or unexpected pregnancies; those lacking “support from surrounding people” as they were unmarried, single mothers, or not obtaining cooperation from the husband; and those with child-rearing problems, such as “having inappropriate child-rearing attitudes toward the older child.” Obstetrics and medical institutions are expected to provide information on specified pregnant women to administrative institutions as support targets for abuse prevention. This study showed that both PHNs and midwives commonly perceived specified pregnant women as targets for support and were conscious of their relationship with pregnant and postpartum women.
Furthermore, PHNs were characterized by their focus on the daily life background, child-rearing ability, and environment of the pregnant and postpartum women with high social risk, such as “having a family background where parents’ home cannot be relied upon” or having difficulties in daily life due to unstable daily life foundations. The nurses viewed childcare in the context of a stable lifestyle. Previous research [ 22 , 23 ] reported that child-rearing supporters for pregnant and postpartum women from pregnancy to the first month after childbirth were mainly immediate family members, such as the women’s husband, mother, and mother-in-law. The present results indicate that the absence of people to rely on can pose a threat to the mental and physical stability of pregnant women.
Meanwhile, midwives focused on maternal health management and stable fetal and child-rearing skills for safe delivery. Their concerns were regarding pregnant and postpartum women with medical risks, such as their mental and physical safety, and those who “have difficulty in child-rearing behavior” that may continue after being discharged from the hospital. This included those not thinking at all about life after childbirth and making “no progress in preparations for childbirth and life after giving birth,” those with anxiety due to mental and physical changes from lifestyle changes around the child experienced in the early postpartum period, and those with a lack of knowledge regarding breastfeeding skills and childcare. Thus, the midwives’ perspectives regarding child-rearing behaviors focused on life after being discharged from the hospital. In this manner, the strengths of the specialties [ 24 ] of PHNs, who are close to the community, and midwives, who specialize in pregnancy and childbirth, were maximized when identifying the pregnant and postpartum women.
Specified pregnant and postpartum women with high social risks have many overlapping elements, and they may also include pregnant women with high medical risks [ 25 ]. Thus, child abuse may be prevented by supporting pregnant women who may or may not require child-rearing support due to various factors, whether or not these factors may lead to child abuse. Sharing feelings of “concerns” toward pregnant and postpartum women based on the perspectives of PHNs and midwives may allow pregnant women to be continuously given support without being overlooked. The characteristics of a PHN, who captures the daily life background and child-rearing environment of PPWC, as well as those of a midwife, who focuses on maternal safety and child-rearing skills, need to be mutually understood by the other profession. These two professionals also need to notice the risks that lead to abuse, sharing this information among supporters as soon as they become aware.
Matsubara [ 26 ] indicated that the mothers and children deemed “of concern” by the PHNs in 18-month child health examinations might be concerned about unique aspects that do not fit into the general image held by PHNs, or aspects concerning a minority of people. Even among the PPWC, as seen by the PHNs in this study, the PHNs felt that certain women with unique moods or thoughts “have the discomfort of not feeling like a normal pregnant woman.” Ozawa et al. [ 27 ] indicated that the job of a PHN is to guide the health conditions of an individual toward a better direction and to foster their ability to maintain an active life. When a PHN feels “concerned for some reason,” then this indicates that the possibility of a problem is present and that the individual needs assistance. Furthermore, valuing the “concerning aspect” and process of verification will lead to an improvement in the quality of on-site practice. Therefore, future studies should verify whether a pregnant or parturient woman who is considered “of concern” in this study truly requires support, as well as determine the effect of providing support from the point where a concern for the woman arises.
Suggestions for collaborative support to PPWC provided by PHNs and midwives
Methods for determining PPWC for both the PHNs and midwives included identifying “concerning” aspects not only from subjective information, such as behavior and attitude during interviews with the women, but also from objective information, such as the questionnaire used at the time of the interview and contents of the risk assessment index. Since 2011, Okayama Prefecture has been using the “Contact form for support for mothers and children of concern during pregnancy” as a communication tool between obstetric medical institutions and administrative institutions using the Okayama model [ 9 ]. This system has been in operation throughout the Okayama Prefecture, with the current study also including using a contact form that was unique to the Prefecture. In recent years, risk assessments have been conducted using indicators that were independently created by child abuse prevention committees within obstetric and medical institutions. The importance of continuous collaborative support at related institutions, such as medical and administrative institutions, has been recognized [ 28 ]. Wada [ 29 ] stated that in the field of obstetrics medical care for pregnant women, there are individual differences in the “concerns” of midwives depending on the staff in charge. However, preparing a standard framework and dealing with issues as a team instead of relying on individual sensibilities can change the response from “being concerned” to “noticing” issues. Furthermore, activity reports regarding initiatives for preventing child abuse in the perinatal medical fields indicated the effectiveness of systematic risk determination using checklists and the promotion of collaboration with health institutions [ 30 ]. Both PHNs and midwives using common risk assessment indicators allow for the possibility of noticing that PPWC may require support. Consequently, if both the PHNs and midwives share their observations, they will not overlook the PPWC in need of support, establishing a support system from an early stage and serving as the first step for providing continuous monitoring support [ 13 ].
Yamaguchi et al. [ 31 ] reported the following as examples of midwives’ concerns when a mother and child in the early postpartum period are discharged from the hospital: “single concern,” such as support from surrounding people, child-rearing techniques, and mental illness complications; and “multiple concerns,” which are combinations of such single concerns. “Single concerns” represent issues that mothers and children generally encounter, and care tends to focus on these concerns such that it is easy to provide care that is tailored to each mother and child. Meanwhile, “multiple concerns” are highly subjective and involve various factors, including the environment. Thus, there is a wide variety of care recipients and content, and time is required until the effects begin to improve. The present results reveal that both PHNs and midwives viewed those who “have multiple risk factors recognized by an assessment tool” as PPWC and demonstrated their respective expertise in responding to such women with multiple factors, suggesting the importance of collaborative support. A wide range of factors, such as inadequate prenatal check-ups [ 32 ], poverty, and housing instability [ 33 ], were consistent with previous studies. These findings suggest the importance of both parties demonstrating their expertise and working together to support pregnant and nursing mothers who are concerned about a combination of these factors. Yamazaki [ 34 ] indicated a need for related professions to gain a common understanding regarding contact methods between medical and health institutions and that information sharing will progress effectively through a mutual understanding of the differences in perceptions between the PHNs and midwives.
Meanwhile, regarding collaboration issues between the PHNs and midwives, Hattori et al. [ 35 ] reported that the midwives were worried about whether their perspectives on “mothers and children of concern” were appropriately communicated to the PHNs. Many reports mention that turf issues and the division of roles hinder good inter-agency collaboration [ 36 , 37 ]. Furthermore, Karata et al. [ 38 ] stated that feedback of information from other institutions is essential for developing collaborations after the nurses at obstetrics and medical facilities provide information on “parents and children of concern.” In the future, it will be necessary to determine how the PHNs and midwives view each other’s characteristics regarding PPWC and how they provide support while investigating the ideal way for effective collaboration between the two professions.
Research limitations and future issues
This study has some limitations. Only ten PHNs and ten midwives participated, and their workplaces were limited to a single prefecture. Furthermore, we targeted PHNs and midwives with five or more years of experience in this study. However, both groups had an average of over 20 years of experience. Hence, their identification of PPWCs likely differed according to their years of experience. Future tasks involve the development of indicators to support the prevention of child abuse from the early stages of pregnancy without overlooking pregnant and postpartum women needing support, regardless of the number of years of experience of the nurses and midwives. Moreover, the construction of an effective collaborative support model for PHNs and midwives is needed to prevent child abuse.
Conclusion
The study’s results revealed that each professional had their perspectives of determining target women by using their respective specialties and that had a common perspective of determining specified pregnant women as PPWC. PHNs focused on childcare in a stable lifestyle, wherein the foundations involved the daily life backgrounds of the pregnant or parturient woman. Contrarily, midwives focused on the health management of mothers and stable fetal and child-rearing skills, alongside perspectives on child-rearing behavior after discharge from the hospital. Future research must determine how each professional views the other’s characteristic perspectives for providing support, alongside investigating the ideal way for effective collaboration between these two professions.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
The biological system involves tens of thousands of genes/proteins that are tightly regulated in a complex network [ 1 – 3 ]. Interactions and regulations in the network are highly dynamic. They change substantially in different cell types, developmental stages, or in response to environmental conditions [ 4 ]. Gene expression and similar types of data, such as proteomics and metabolomics data, represent outcomes of the dynamic regulatory network. Changes in the underlying regulation patterns can often result in changes in correlation between genes.
In many gene expression profiling datasets, the cellular states or sub-classes are not observed directly. Thus dynamic correlation needs to be inferred from the data. Once successfully extracted from the data, the dynamic correlation patterns can in-turn help deduce hidden cellular states and sub-classes. The most common dynamic correlation takes following form: the correlation between a pair of genes g i and g j is reliant on the value of an unobserved variable Z , i.e. cor( g i , g j ) = f(Z) , where f() is an unspecified monotone function. Z can be the activity of a specific regulator in the system, or it can be the reflection of cellular states resulting from the collective activities of multiple regulators. Because gene expression is tightly controlled in the cell, the same Z variable can govern the dynamic correlation of many gene pairs.
Given the complexity of cellular regulations, systematically studying dynamic correlation is challenging. First, as the biological system is organized in a modular manner [ 5 ], there could be multiple Z variables that govern the dynamic correlation of different groups of genes. Secondly, the underlying cellular states may not manifest into biological/clinical observations, making most of the Z variables unobservable. Hence the major interest is to find the unobserved Z variables. To this end, Li has developed the Liquid Association (LA) approach, which uses genes as proxy measurements of the unobserved Z variables [ 6 , 7 ]. The method scans through all possible gene triplets to find potential dynamic correlations. Similar approaches that utilize genes as mediators [ 8 , 9 ], integrative analysis utilizing Liquid Association [ 10 , 11 ], as well as statistical theory of Liquid Association [ 12 ] were later developed.
Although using genes as surrogate measurements of the Z variables can reveal some important local regulatory mechanisms, a more global approach to dynamic correlation could discover critical regulation mechanisms that penetrate multiple biological processes, or help identify hidden sub-groups in the samples. To this end, using the original LA or similar approaches is not effective due to the following reasons. First, scanning through all possible gene triplets is computationally intensive. Second, a genome-scale scan yields large numbers of significant gene triplets, causing difficulties in the interpretation. Given the LA score is calculated in a symmetric manner among the three genes involved, discerning which gene reflects cellular states could be tricky. Third and the most important, measurements in the genes may not be good indicators of the true underlying cellular states.
In this study, our purpose is to find latent signals that govern the dynamic correlation of a large number of gene pairs. The key differences between our approach and screening by Liquid Association are: (1) We do not assume the signals that control the dynamic correlation of gene pairs are contained in any gene; (2) We are only interested in finding the dominating dynamic correlation signals that impact large numbers of gene pairs, but not local signals that govern only a small number of gene pairs. (3) Compared to screening all gene triplets by Liquid Association, the method is magnitudes faster.
To develop such a method, the biggest difficulty is we do not know a priori which gene pairs are dynamically correlated. To solve this problem, we design a new metric, named Liquid Association Coefficient (LAC), to effectively and efficiently screen all gene pairs for potential dynamic correlations. From gene pairs that are most likely to be dynamically correlated, we provide a simple and straight-forward solution for quickly finding the latent dynamic correlation signals. The procedure is named DCA: Dynamic Correlation Analysis. We refer to the latent signals found by DCA as Dynamic Components (DCs).
We demonstrate the performance of the method using extensive simulations. In real biological datasets, the method can identify latent signals that are biologically meaningful and not found by existing methods. In a single cell RNA-seq dataset, DCA was able to separate more cell types and shed light on the biological functions that drove the separation. In the Cancer Genome Atlas (TCGA) breast cancer (BRCA) dataset, DCA found new interesting subgroups in the subjects that are related to patient survival outcome. In a merged cell cycle dataset, the method recovered signals pertaining to the original experimental grouping, as well as biological processes that differentiate the experiments, shedding lights on the side-effects of the data-generation process.
Results
Behavior of the Liquid Association Coefficient (LAC)
In this study, a new metric was defined to rank all pairs of variables in the data matrix. The purpose of the LAC was to help identify gene pairs that were most likely to have the relationship of dynamic correlation, without knowing the underlying physiological states that govern the dynamic correlation. Gene pairs with such relations should receive high LAC score, while other gene pairs, either independent or correlated, should receive low scores.
The LAC requires all variables to have mean zero and standard deviation 1. As illustrated in Fig 1A , if both variables X and Y followed the standard normal distribution marginally, and one-third of the (X,Y) pairs were positively correlated, one-third of the (X,Y) pairs were negatively correlated, and another one-third uncorrelated, then the absolute values would be positively correlated, and the LAC tended to be large ( Fig 1A , left column). On the other hand, when X and Y were truly independent or simply correlated, the LAC tended to be small.
10.1371/journal.pcbi.1006391.g001
Fig 1
The liquid association coefficient (LAC).
(a) Illustration of LAC using examples. Left column: dynamic correlation with an unknown conditioning factor. When the factor is low, x and y are negatively correlated; when the factor is high, x and y are positively correlated. Second left column: independent case. Right two columns: correlated case. In all the cases, the marginal distribution of X and Y are standard normal. (b) Empirical distributions of LAC score under conditions of dynamic correlation, simple correlation, or independence. The densities are based on 1000 simulations. In the dynamic correlation cases, one-third of the data points follow a bivariate normal distribution with mean ( 0 0 ) and variance-covariance matrix ( 1 ρ ρ 1 ) , one-third follow a bivariate normal distribution with mean ( 0 0 ) and variance-covariance matrix ( 1 − ρ − ρ 1 ) , and another one-third follow independent standard normal distributions. In the correlated case, all data points follow a bivariate normal distribution with mean ( 0 0 ) and variance-covariance matrix ( 1 ρ ρ 1 ) .
We further conducted a simulation study to examine the empirical distribution of LAC under different circumstances. As illustrated in Fig 1B , when the two variables were dynamically correlated, the distribution of the LAC score was centered at a positive value ( Fig 1B , blue curves). The higher the correlation level, the higher the mean ( Fig 1B , left to right panels). The higher the sample size, the less the spread ( Fig 1B , different line types). At the same time, in the independent and correlated cases, the LAC scores were centered around zero if the first definition of LAC is used. Using the second definition, the LAC was still centered around zero in the independent case, and the center was negative in the correlated case ( Fig 1B , lower panels). Intuitively for the correlated case, when taking the absolute value, the range of | x | became smaller than x itself, while the spread of data points around the trendline stayed the same. This meant the correlation between the absolute values tended to become smaller than the original, resulting in a negative LAC score.
Simulation study
We conducted an extensive simulation study to evaluate the method’s capability to recover latent dynamic correlation signals. Each simulated dataset was made of multiple modules, each of which was regulated by a single underlying dynamic correlation factor. To simulate a module of genes that have dynamic correlation conditioned on the same factor, we first simulated the latent factor z by sampling the standard normal distribution. For the conditional correlation pattern, we simulated three different fashions separately: (1) E ( XY | z ) = (Φ −1 ( z )−0.5) × 2; (2) Truncate z at -3.2 and 3.2, and then E ( X Y | z ) = s i g n ( z ) × | z / 3.2 | ; (3) Truncate z at -3.2 and 3.2, and then E ( X Y | z ) = s i g n ( z ) × | z / 3.2 | 4 .
We then generated 10 pairs of seed vectors ( x , y ) such that each random variable followed the standard normal distribution marginally, and between a pair of X and Y, their correlation was dependent on z. The details of generating an (x , y) pair were as follows:
For each z value, we found the conditional correlation value ρ z = E ( XY | z ) between X and Y based on the three scenarios above, for example, ρ z = E ( XY | z ) = (Φ −1 ( z )−0.5) × 2;
We sampled one data point from the two-dimensional Gaussian distribution with mean vector ( 0 0 ) and variance-covariance matrix ( 1 ρ z ρ z 1 ) ;
We iterated steps (1) and (2) through all N values of the z vector, to obtain N pairs of (x , y) . Together they made the two vectors that were dynamically correlated conditioned on Z .
For each z vector, after repeating the above process and generating 10 pairs of such seed vectors, we used the following procedure to generate the observed expression vectors:
We randomly selected one pair from the seed vector pairs;
We added Gaussian noise to the selected seed vector to generate one pair of simulated genes;
We repeated steps (1) and (2) until the desired number of simulated genes were generated.
In each simulation dataset, multiple z vectors were generated. From each of the z vectors, a group of genes that were dynamically correlated conditioned on the z vector were generated. In addition, noise genes were generated by sampling from the standard normal distribution. The number of noise genes was equal to the total number of genes involved in dynamic correlation modules.
In order to mimic the situation where the data are highly skewed and zero-inflated as in RNA-seq data, we also conducted another set of simulations. First, we simulated data with normal marginal distribution using the procedure above. Then for each simulated gene, we randomly drew one gene from the TCGA BRCA dataset that had less than 75% zero values, and matched the quantiles of the simulated gene to those of the real gene using the interpolating quantile normalization procedure described in [ 13 ]. This approach forced each simulated gene to have the same marginal distribution as a real gene.
After generating 50 datasets in each simulation setting, we compared DCA with six other methods. The first was screening by Liquid Association. Conceptually, this would involve computing the LA score for all possible gene triplets, and then selecting the top LA scouting genes that were involved in the highest numbers of triplets with high-LA scores. However, the heavy computational cost of screening through all ( p 3 ) triplets made it impractical to actually conduct the computation on all simulated datasets. Rather, the “LA screening” results were obtained by directly selecting the genes that had the highest absolute correlation with the true hidden factors, one gene for each factor. Given LA screening can only find signals that are in the genes, the simulation obtained the best possible results of LA screening, which showed the upper limit of how well LA screening could recover the dynamic correlation signal. But such results may not be attainable in actual computation. The other methods compared were dimension reduction methods: Principal Component Analysis (PCA), Independent Component Analysis (ICA), t-Distributed Stochastic Neighbor Embedding (t-SNE), kernel PCA with degree-two polynomial kernel, and kernel PCA with radial basis function kernel.
For comparison, assuming there were K true latent factors, each method was allowed to find the top K+2 factors, except for LA screening, which found K factors as discussed above. Among the found factors, K of them were paired with the latent factors, by sequentially seeking the highest absolute correlation coefficient between any found factor and latent factor. We then calculated the absolute correlation coefficient between the hidden factors and their paired true hidden factors, and found the average absolute correlation for each simulation setting as the indication of how well the latent factors were recovered.
In setup 1, when the marginal distribution of gene expression was normal, DCA recovered the latent signals when signal to noise ration (S/N) and the sample size were moderate to high ( Fig 2A ). When the number of modules increased ( Fig 2A , left to right), the capability to recover the latent factors decreased for lower sample sizes. At the same time, other methods failed to recover the latent factors. As the likelihood to generate spurious correlation was higher at smaller sample sizes, the dotted curves (low sample size) of LA screening and other methods were higher than the corresponding dashed and solid curves (higher sample sizes). However, this only reflected spurious correlations, rather than actual recovery of true signals.
10.1371/journal.pcbi.1006391.g002
Fig 2
Simulation result–the average absolute Spearman correlation between latent factors and their corresponding found factors.
(a) The marginal distributions of gene expression levels were normal. (b) The marginal distributions of gene expression levels mimicked real RNA-seq data. Row sub-plots: number of genes in each module; Columns subplots: the number of modules; Line type: sample size; line color: method used for latent factor recovery. Given the heavy computational cost, the “LA screening (upper limit)” results were obtained by directly selecting the genes that have the highest absolute correlation with the hidden factors, meaning the values plotted are the best possible, but may not be attainable in actual computation.
The same trend held true for the datasets in which the marginal distribution of gene expression values mimicked the real data ( Fig 2B ). With the highly skewed and zero-inflated data, the faithful recovery of the hidden signals required more sample size and higher signal to noise ratio, compared to normally distributed data. Nevertheless, DCA was the only method that was capable of recovering the hidden signals.
Setup 2 was a weaker LA relationship than setup 1. As expected, the average absolute correlation was lower compared to setup 1 ( S1 Fig ). However the overall trend was the same–DCA recovered part of the latent variables, while other methods failed to recover the latent variables.
In setup 3, there were less extreme correlations between X and Y compared to setup 1. But at the same time, there were less low-correlation X-Y pairs. Overall the performance was similar to setup 1 ( S2 Fig ). In this setup, DCA performed better than setup 1 when sample size was small. It still recovered the latent variables at small sample sizes when the total number of modules were small ( S2 Fig , left columns). Again the other methods failed to recover the latent variables. Overall, in all three setups, our method could faithfully recover the hidden dynamic correlation signal when the sample size and signal strength was sufficient.
Real data analysis—single cell RNA-seq dataset of small intestinal epithelium cells
We used the single cell RNA seq data from the GSE92332 dataset [ 14 ]. The dataset contains measurements in mouse small intestinal epithelium cells under both normal condition and enteric pathogen treatments. For pattern detection we used the normal cells only. The data contains the measurement of 20108 genes measured in 1522 cells falling into nine types that were defined by known cell type-specific marker genes. For pattern detection using DCA, we removed genes with >25% zero counts. Given the sequencing depth, the remaining matrix contained 3041 genes.
We first examined the scores of the top latent factors ( Fig 3 ). The score of each latent factor was a vector of 1522 values, corresponding to the 1522 cells. Every point in a subplot of Fig 3 represents a cell. As shown by the color of the points using the cell type information, the first 5 DCs clearly separated 4 types of the cells from the rest, and the separation was quite clear ( Fig 3 , lower-left sub-plots). As a comparison, the first 5 PCs only separated 2 cell types from the rest ( Fig 3 , upper-right sub-plots). Although there are some separations between the cell types when the points are colored by cell type, without the coloring, we would not be able to delineate the cell types clearly from the point patterns.
10.1371/journal.pcbi.1006391.g003
Fig 3
Scatter plots of the first five DCs (lower left) and PCs (upper right) from the GSE92332 single cell RNA-seq data. The cells were colored using cell types based on known markers .
We next examined the biological processes whose differential correlation separated the cell types ( Fig 4 ). DC1 mostly separated paneth cells from the rest. The function of paneth cells is mostly the secretion of anti-microbial proteins and peptides [ 15 ]. As shown in Fig 4A , the biological processes associated with DC1 were clearly concentrated in protein synthesis and energy production, which indicated protein/peptide biosynthesis was a critical functional aspect that separates the paneth cells from the rest.
10.1371/journal.pcbi.1006391.g004
Fig 4
Major biological processes associated with the DCs.
(a) DC1, (b) DC2, (c) DC3, and (d) DC5. Gene pairs were selected using fdr threshold of 0.01. Biological process pairs were selected using a p-value threshold of 0.001 and fold-change of 4. All were limited to biological processes with 50 or more connections, except for DC2, for which the limit was 100 due to the existence of excessive connections.
DC2 mostly separated tuft cells from other cells. Tuft cells had long been considered a sensory cell. Only recently was tuft cell determined to be an important cell in innate immune response. Tuft cell secretes IL25 to stimulate the proliferation of innate lymphoid cells (ILC2s), and forms a feed-forward loop with ILC2s to generate type 2 immunity [ 16 ]. Our results showed many immune regulation processes and signaling processes were among the top biological processes associated with DC2 ( Fig 4B ). The results strongly agreed with the immunological function of tuft cells.
DC3 separated enterocytes from the other types of cells. Enterocytes are intestinal absorptive cells. The top biological processes included “digestion” and “potassium ion transmembrane transport” ( Fig 4C ), which includes the sodium-potassium pumps that are essential for the co-transport mechanism to absorb glucose and amino acids into the blood stream [ 17 ]. Some processes related to macromolecule biosynthesis were also among the highly connected.
DC5 separated goblet cells from the rest. Goblet cells secrete mucins, which are large glycoproteins, in order to protect the mucous membrane. Unsurprisingly, the major biological process that was associated with DC5 was protein glycosylation ( Fig 4D ). Interestingly, most other highly connected biological processes were immune-related functions. Some studies have started to confirm that the goblet cells actually have major immune functions [ 18 ], such as working as antigen retrievers [ 19 ]. The results here indicated a number of immune processes were activated at the transcription level.
From the pattern detection perspective, if the cell types were hidden, DCA clearly extracted more meaningful information to help differentiate the cell types, as well as points to important pathways that cause the distinction. In most real applications of dimension reduction, information such as sample grouping are not available. We next examined the TCGA breast cancer (BRCA) dataset to see if the method can extract any new insights from the data.
Real data analysis—TCGA breast cancer data
The TCGA BRCA data contains the measurement of 20532 genes by deep sequencing in 762 subjects with breast cancer. After removing genes with >20% zero readings, 17728 genes remained in the study. Similar to the single cell RNA-seq data, DCA captured signals that were distinct from traditional methods. Here we focus our discussion on three of the DCs, as they are clearly linked to estrogen receptor (ER) status. Fig 5A shows the plot of the factor scores of these three DCs, each point corresponding to a subject. DC1 largely separated ER-positive and ER-negative samples, which agreed with the second principal component very well ( Fig 5B ). On the other hand, in the space spanned by DC3 and DC7, ER-positive samples were tightly clustered in the middle, while part of the ER-negative samples were spread widely ( Fig 5A , S3 Fig ). No PCs captured a similar structure in the data ( S4 Fig ).
10.1371/journal.pcbi.1006391.g005
Fig 5
Results from the TCGA BRCA dataset.
(a) Scatter plots of DC1, DC3, and DC7 scores. The points are colored based on the ER status of the subjects. DC1 separates ER+ and ER-, while DC3 and DC7 have a wide spread only for the ER- subjects. (b) DC1 captures similar information as the second principal component. (c) Kaplan–Meier curves of the ER-negative subjects, red: absolute factor score > 0.05.
Further analyses showed that among the ER-negative subjects, those with more extreme scores in either DC3 or DC7 showed a different survival characteristic than those in the center ( Fig 5C ). The subjects with more extreme scores tended to have a much higher chance of dying earlier, while in long follow-ups the remaining subjects tended to survive longer, albeit supported by relatively few data points.
Functionally, the biological processes that showed excessive dynamic correlations conditioned on DC3 were centered around two main themes ( Fig 6A ). The first was protein sumoylation and stress response. Sumoylation is a post-translational modification that often occurs in response to cellular stress [ 20 ]. Many oncogenes and tumor suppressors are functionally related to sumoylation [ 21 ]. The second main theme was cell differentiation and tissue development that were related to several types of tissues, indicating a dysregulation in the cancer cells.
10.1371/journal.pcbi.1006391.g006
Fig 6
Biological process pairs with excessive dynamic correlations related to DCs 3 and 7.
Gene pairs were selected using fdr threshold of 0.01. Biological process pairs were selected using a p-value threshold of 0.001 and fold-change of 3. For simplicity, only nodes with connections above a certain threshold are shown. Node sizes reflect the total number of connections of each node. (a) Biological process pairs associated with the 3 rd DC. (b) Biological process pairs associated with the 7 th DC. Inset: scatterplot of LUMP (leukocytes unmethylation for purity) vs DC7 score. The correlation coefficient is -0.35.
The biological processes associated with DC7 were mostly immune response processes ( Fig 6B ). Patterns of immune cell infiltration has been linked to the prognosis and treatment response of breast cancer [ 22 ]. The changed expression patterns of mostly immune-related genes in these samples were likely reflective of a certain immune cell infiltration pattern that had implications in prognosis [ 23 ]. We took the cell purity estimates based on LUMP (leukocytes unmethylation for purity) criterion, which was based on the average of 44 non-methylated immune-specific CpG sites and largely reflected immune cell infiltration [ 24 ]. As shown in the inset of Fig 6B , samples with high DC7 scores were those with low purities estimated by LUMP, while samples with low DC7 scores were a subset of those with higher purity scores. How these samples differ from the other high purity samples is an interesting point for future studies. Similarly, beside the three DCs that we discuss here, most of the other DCs showed clear functional implications, but require extra study beyond this manuscript to elucidate their biological meaning.
Real data analysis—the yeast cell cycle microarray dataset
Thirdly, we analyzed the well-studied Spellman cell cycle gene expression data [ 25 ]. The dataset has been analyzed by many authors. The purpose of the analysis here was to demonstrate that DCA can extract information that was not reported before, yet clearly meaningful, and provided novel biological insights.
The cell cycle dataset includes four time-series experiments of the yeast cell cycle, each using a different method of synchronization. The total dimension is 6178 genes by 73 samples. Missing values were imputed by the K-nearest neighbor (KNN) method [ 26 ]. When all four time series datasets were combined into a single dataset, traditional methods such as PCA and SPCA [ 27 ] extracted signals that were consistent across the four time series ( S5 Fig ), but not signals that separated the four time series, except the first PC that captured an oscillating signal which was an artifact in the CDC15 time series data [ 28 ].
Applying DCA to the combined cell cycle data yielded factors that were distinctly different. Most of the DCs clearly differentiated one of the four time series from the rest ( S6 Fig ). Here we focus our discussion on three of the factors. Fig 7 shows the plots of the scores of the three DCs, each point representing a sample. The first DC had high scores for samples from the CDC15 experiment only. It has been documented that an oscillating signal is present in the CDC15 data across many genes, causing an elevated level of correlation overall [ 28 ].
10.1371/journal.pcbi.1006391.g007
Fig 7
Some example Dynamic Components from the cell cycle data.
Colors: the four cell cycle experiments. Red: alpha factor; green: CDC15; blue: CDC28; purple: elutriation.
The second DC only had extreme scores for some of the samples of the elutriation experiment. A closer examination revealed the DC showed a sine-wave pattern in the elutriation samples ( Fig 7 ). An examination of the data revealed a strong dynamic correlation pattern between genes associated with this DC. Selecting biological processes pairs that had excessive dynamic correlation links between them, we found that the processes were focused on rRNA biogenesis and ribosome assembly ( Fig 8A ). Much more positive/negative correlations were shown between genes in these biological processes when the DC2 score is low, which corresponded to half of the samples in the elutriation experiment. While all the other three experiments were based on block-and-release cell cycle synchronization, the elutriation process separates synchronized cells based on their size, shape and mass [ 29 ]. The results here indicated that protein biosynthesis tended to be better synchronized in the elutriation samples compared to the other three experiments.
10.1371/journal.pcbi.1006391.g008
Fig 8
Biological process pairs with excessive dynamic correlations related to DCs 2 and 5.
Gene pairs were selected using fdr threshold of 0.01. Biological process pairs were selected using a p-value threshold of 0.001 and fold-change of 2. For simplicity, only nodes with connections above a certain threshold are shown. Node sizes reflect the total number of connections of each node. (a) Biological process pairs associated with the DC2. (b) Biological process pairs associated with the DC5. (c) Example plots of gene pairs with LA relation with DC5. Red points: samples in the lower 33% of DC5 score; blue points: samples in the upper 33% of DC5 score.
For the fifth DC, samples in the CDC28 experiment had lower scores, while the alpha factor samples had higher scores, with a smaller magnitude ( Fig 7 ). This indicated that some gene pairs had a reverse correlation pattern between the two experiments, which was intriguing given both experiments used block-and-release to synchronize cells. Some recent studies have shed light on the metabolic behavior of the yeast cells under the alpha factor or CDC28 cell cycle arrest. Under the alpha factor treatment, the central metabolic fluxes are at a high level, and the cellular metabolism tend to be respiratory even when glucose is abundant [ 30 ]. The cell cycle CDK Cdc28 regulates both the cell division processes and metabolic processes. Under the CDC28 inhibition, the cells accumulate glycogen and trehalose to extremely high levels [ 31 ]. Given the different characteristics of the two cell cycle arrest mechanisms, it is understandable that after the release of cell cycle arrest, the cells proceed from very different metabolic situations, and metabolism will adapt to those situations. Functionally, we observed the highly connected biological processes mostly involve small molecule metabolism and transport ( Fig 8B ). Two typical pairs of genes are shown in Fig 8C , where clear dynamic correlation is observed.
Unlike traditional methods such as PCA and SPCA that identified commonalities, the DCA approach tended to find signals that differentiate the four underlying experiments, and revealed some important biological processes that behaved differently between the experiments. Given the existing knowledge on the dataset, these results validated that DCA extract new and meaningful information.
Discussion
In this study, we developed a new method to detect major dynamic correlation signals from large gene expression matrices. A new measure of dynamic correlation between a pair of variables, the Liquid Association Coefficient (LAC), was developed to facilitate the discovery of the dynamic correlation signals. We used eigen value decomposition to find the DCs after the top gene pairs that were likely to be dynamically correlated were found by LAC scores, and a new H matrix was constructed from the gene pairs. Conceptually, other methods used to find latent factors, such as Independent Component Analysis (ICA) [ 32 ], Sparse Principal Component Analysis (SPCA) [ 27 ], Modular Latent Structure Analysis (MLSA) [ 33 ], or various clustering methods could also be applied to the H matrix.
In all three datasets, the latent factors found by DCA showed strong dynamic correlation relations with large numbers of gene pairs. Two of the datasets were RNAseq data, which tend to be skewed in distribution. As we discuss in the first subsection in METHODS, although normality is required of the z vector to make the LA score a valid estimate of the expected derivative of the correlation between the gene pair given Z , and our approach doesn’t have the normality constraint, it still finds z vectors that are correlated with the change of correlation of large numbers of gene pairs, hence recovering the dominant dynamic correlation signals without relying on the distribution assumption of Liquid Association. We examined if any gene could be good surrogates of these latent factors. In the intestinal epithelial data, the highest absolute value of Spearman correlation coefficient between any gene and any of the latent factors was 0.65 ( S7 Fig ). In the BRCA data and Spellman cell cycle data, the correlation coefficients were even lower, with maximum values of 0.43 and 0.55 respectively ( S7 Fig ). These results suggest that using genes as surrogate measurements is not as effective.
On the surface, our method bears some resemblance to kernel PCA with degree two polynomial kernel, in which the kernel is defined as K ( g i , g j ) = (⟨ g i , g j ⟩ + κ ) 2 . However in fact the two methods are very different. In our method, when considering a pair of dynamically correlated genes, what’s involved in the downstream computation is the vector, ( g i 1 g j 1 , g i 2 g j 2 , …, g iN g jN ), instead of the inner product. We further analyzed all three real datasets using the kernel PCA approach. The results clearly showed that kernel PCA with degree two polynomial kernel could not discover the patterns found by our method ( S8 Fig , S9 Fig , S10 Fig ).
Overall, as a new unsupervised learning method for high dimensional data, DCA can extract new and useful information from the data. DCA complements existing dimension reduction methods to reveal more internal structure in the data that could lead to new biological discovery. The method is straight-forward, and the computation is efficient. The R package is available at https://cran.r-project.org/web/packages/DCA/index.html .
Materials and methods
Setup and the overall workflow
The data is in the form of an expression matrix, G p × n , with p genes in the rows and N samples in the columns. We assume that all genes are normalized to have mean 0 and standard deviation 1. Thus the correlation between two genes represented by two row-vectors, g i and g j , r g i , g j is equal to E ( g i g j ).
Here we consider the situation where among the ( p 2 ) gene pairs, a small portion are dynamically correlated. Further, a small group of latent variables Z k , k = 1, …, K , govern the dynamic correlations of the majority of the dynamically correlated gene pairs. Which pair of genes are governed by which latent variable is unknown.
Ideally, we would like to estimate the latent variables, Z k , k = 1, …, K , as well as which gene pair is associated with which latent variable. However, in real datasets, the number of genes p is usually over 10 4 . Subsequently, the number of possible pairs is on the scale of 10 8 , making it nearly impossible to treat the gene pair–latent variable relation as missing value, e.g. using the Expectation-Maximization (EM) algorithm approach. Thus our goal is to develop a heuristic approach that involves dimension reduction to find good approximate solution efficiently.
Following the notations of Liquid Association [ 6 ], given a pair of genes g i and g j , and a latent factor Z , let g ( Z ) = E ( g i g j | Z = z ) denote the conditional correlation of the two genes given Z = z . The LA score is defined as LA ( g i g j | Z ) = Eg ′( Z ), which is the expected change of correlation between g i and g j with respect to Z . As shown in [ 6 ], LA ( g i g j | Z ) = E ( g i g j Z ) and is estimated by 1 N ∑ n = 1 N g i , n g j , n z n , if Z is standard normal. If a new vector h is generated, which is the entry-wise product of g i and g j ,
h n = g i , n g j , n , n = 1 , … , N ,
then under the assumption of normality of Z , LA ( g i g j | Z ) is estimated by 1 N ∑ n = 1 N h n z n , which is proportional to the dot product between the corresponding vectors z and h .
If the pair of genes g i and g j are governed by Z , then LA ( g i g j | Z ) = Eg ′( Z ) has a large absolute value, which means ( z ⋅ h ) 2 is large. On the other hand, if the pair of genes g i and g j are not dynamically correlated with regard to Z , then ( z ⋅ h ) 2 is small. Given the scaling of Z only linearly scales the LA scores, we can add the constraint that the vector z we are seeking is unit-length.
If we can somehow gather all gene pairs that are dynamically correlated, and construct a new matrix H , each row of which being an h vector constructed from a dynamically correlated gene pair, then one good heuristic solution is to seek the z vectors sequentially, by applying eigen value decomposition to the matrix H ′ H , which finds the solution to the following optimization problem:
z 1 = argmax ‖ z ‖ = 1 ∑ m ( z ∙ h m ) 2 ,
z k = argmax ‖ z ‖ = 1 ∑ m ( z ∙ h m ) 2 , s . t . z ′ z l = 0 , l = 1 , … , k − 1 ,
where m indexes all the h vectors. The more gene pairs a latent Z variable regulates, the larger the sum of squared projection length. This way, the top eigen vectors of the H ′ H matrix capture the major signals that regulate the dynamic correlation of the majority of the dynamically correlated gene pairs. We name these vectors Dynamic Components (DCs). They are each of length N , which is the number of samples.
We note that for the quantity 1 N ∑ n = 1 N h n z n to be a valid estimate of LA ( g i g j | Z ), i.e. the expected derivative of the correlation between g i and g j with respect to Z , the normality assumption of Z needs to hold. However, this is not guaranteed in the above estimation procedure. On the other hand, the above procedure seeks z vectors on which large numbers of h vectors have big projections, i.e. projection directions that are correlated with large numbers of r g i , g j . Thus even without the normality assumption, such z vectors are highly correlated with the change of correlation between many dynamically correlated gene pairs, meaning they are good estimates of the latent dynamic correlation signal. At the same time, with many data types, such as RNA-seq or LC/MS metabolomics data, the data itself is highly skewed. There is no reason to believe the underlying latent factors that govern dynamic correlation are normal. Thus loosening the assumption may be beneficial in the discovery of the true latent factors.
To apply this approach, the key is to find the dynamically correlated gene pairs from the ~10 8 possible pairs. We find gene pairs that are dynamically correlated by ranking all pairs of genes using a newly developed metric, Liquid Association Coefficient (LAC), which is described in the next subsection. We should note that we cannot guarantee all dynamically correlated pairs are found, nor there are no noise pairs among the selected pairs. However, with the dimension reduction approach being applied, missing some pairs or including some noise pairs, as long as they do not account for too large a proportion in the H matrix, the main latent factors can still be recovered.
Selecting gene pairs that are likely to be dynamically correlated
For the purpose of selecting informative gene pairs, we define a measure for dynamic correlation between a pair of genes, the Liquid Association Coefficient (LAC), which can take two forms. The first is the correlation coefficient of the squared values of the two genes, minus the correlation coefficient of the original values squared.
ζ i , j = r ( g i 2 , g j 2 ) − r 2 ( g i , g j ) ,
where r () is the Pearson’s correlation coefficient. It has been shown that when both g i and g j follow the bivariate normal distribution with mean ( 0 0 ) , and variance-covariance matrix ( 1 ρ ρ 1 ) , the above quantity converges to zero no matter what value ρ takes.
Alternatively, to reduce the impact of more extreme values, we can use the correlation coefficient of the absolute values of the two genes minus the absolute value of the correlation coefficient:
ζ i , j = r ( | g i | , | g j | ) − | r ( g i , g j ) | .
We compute the matrix of LAC values for all pairs of genes. Notice the computational cost is on the same scale as computing the pairwise correlation matrix. We then select the ( i , j ) pairs whose LAC values are above a certain percentile of all the values in the matrix. In this study, we use top 2.5% or 10 6 pairs, whichever is smaller.
Finding DCs and their associated gene pairs
After selecting the top ( i , j ) pairs, we construct the H matrix, in which each row is constructed from a selected pair of genes. For example, if g i and g j are selected as a pair of informative genes, then the corresponding row of the new matrix is ( g i 1 g j 1 , g i 2 g j 2 , …, g iN g jN ). We then find a sequence of latent factors using eigenvalue decomposition on the matrix H ′ H .
In order to improve the interpretability of the resulting factors, further factor rotations can be conducted to better align the DCs with groups of h vectors (gene pairs). In this study, we used the varimax rotation, which rotates the latent factors in the subspace they span, and seeks to maximize the sum of the variances of the squared loadings of the h vectors on the latent factors [ 34 ].
To find the gene pairs associated with each of the DCs, we first calculate the LAC coefficients for all pairs of genes, and select gene pairs with LAC coefficients belonging to a top percentile (20% in this study). For each selected (g i , g j ) pair, we construct the h vector,
h n = g i , n g j , n , n = 1 , … , N .
For a z vector, we calculate its dot product with all the h vectors that are constructed from the selected pairs,
γ m = ∑ n = 1 N z n h n ( m ) , n = 1 , … , N , m = 1 , … , M
where m indexes the h vectors, and M is the total number of gene pairs used. According to the Central Limit Theorem, the dot products approximately follow a normal distribution when the z vector is independent of an h vector, i.e. a (g i , g j ) pair. As we now consider a large number of gene pairs (20% of all possible pairs, on 10 7 scale), we can safely assume the majority of the gene pairs don’t have a dynamic correlation with regard to a given z vector, while a small portion of the dot products follow another distribution as the corresponding pairs are dynamically correlated with regard to z . Thus together, { γ m } m = 1 M follow a mixture distribution. This is very similar to the considerations in the local false discovery rate (lfdr) literature. We consider the density of { γ m } m = 1 M as a mixture with two components:
f ( γ ) = π 0 f 0 ( γ ) + ( 1 − π 0 ) f 1 ( γ )
where f () is the mixture density for the observed γ statistic, f 0 () and f 1 () are the respective densities of the null (unassociated with z ) and non-null (associated with z ) gene pairs, and π 0 is the proportion of the true null gene pairs. Then the posterior probability that a gene pair belongs to the null distribution is π 0 f 0 ( γ ) f ( γ ) , at any value of the γ statistic.
Given the similarity of the setup, we can simply borrow from the mature local false discovery rate (lfdr) methods. For every z vector, we generate the collection of γ statistics { γ m } m = 1 M , with each element corresponding to a gene pair. We then apply the existing local false discovery rate (lfdr) method to calculate the posterior probability that a gene pair belongs to the null distribution [ 35 ], and threshold the lfdr values to select gene pairs that are dynamically correlated given the latent factor.
Finding biological processes associated with a latent factor
For functional interpretation, we use gene ontology (GO) biological processes. We first select a set of representative GO biological process terms that are of reasonable size and relatively small overlaps, following an existing procedure that considers both the ontology structure and the number of genes assigned to each term [ 36 ]. For the mouse data, we select 428 biological processes with 100~1000 assigned genes each, covering 15161 genes in total. For the human data, we select 423 biological processes with 100~1000 assigned genes each, covering 14414 genes in total. For the yeast data, we select 172 biological processes with 50~1000 assigned genes each, covering 5334 genes in total. From the gene pairs associated with each latent factor, we conduct two types of analyses:
Within-process dynamic correlation
For each biological process, we count the occurrence of gene pairs in which both genes fall into the process. We also calculate the expected number of such gene pairs if all the gene pairs were randomly drawn. We calculate the fold-change by taking the ratio of observed count v . s . the expected count, and p-value using the binomial distribution.
Between-process dynamic correlation
For each pair of selected biological processes, we first remove their overlapping genes. We then count the occurrence of gene pairs in which the two genes fall into the two processes respectively, and calculate the expected number of such gene pairs if all the genes were randomly drawn. We then calculate the fold-change by taking the ratio, and p-value using the binomial distribution. After thresholding the fold change and p-value to select pairs of processes, we visualize the resulting network using Cytoscape [ 37 ].
Supporting information
S1 Fig
Simulation result of setup 2.
(a) The marginal distributions of gene expression levels were normal. (b) The marginal distributions of gene expression levels mimicked real RNA-seq data. Row sub-plots: number of genes in each module; Columns subplots: the number of modules; Line color: sample size; line type: method used for latent factor recovery. Given the heavy computational cost, the “LA screening (upper limit)” results were obtained by directly selecting the genes that have the highest absolute correlation with the hidden factors, meaning the values plotted are the best possible, but may not be attainable in actual computation.
(TIF)
S2 Fig
Simulation result of setup 3.
(a) The marginal distributions of gene expression levels were normal. (b) The marginal distributions of gene expression levels mimicked real RNA-seq data. Row sub-plots: number of genes in each module; Columns subplots: the number of modules; Line color: sample size; line type: method used for latent factor recovery. Given the heavy computational cost, the “LA screening (upper limit)” results were obtained by directly selecting the genes that have the highest absolute correlation with the hidden factors, meaning the values plotted are the best possible, but may not be attainable in actual computation.
(TIF)
S3 Fig
Pairwise scatter plots of DC factors 1, 3 and 7.
Red points: ER-positive; Blue points: ER-negative; Grey points: unknown status.
(TIF)
S4 Fig
Pairwise scatter plots of the first 8 principal components of the BRCA data.
Red points: ER-positive; Blue points: ER-negative; Grey points: unknown status.
(TIF)
S5 Fig
Principal components of the cell cycle data.
(TIF)
S6 Fig
Dynamic Components (DCs) of the cell cycle data.
(TIF)
S7 Fig
The maximum absolute Spearman correlation between each latent factor and any gene in the dataset.
(a) Intestinal epithelial dataset. (b) TCGA BRCA dataset. (c) Spellman cell cycle dataset.
(TIF)
S8 Fig
Kernel PCA results from the mouse intestine single cell RNAseq data.
Degree 2 polynomial kernel was used to generate the results.
(TIF)
S9 Fig
Kernel PCA results from the TCGA BRCA data.
Degree 2 polynomial kernel was used to generate the results.
(TIF)
S10 Fig
Kernel PCA results from the Yeast cell cycle data.
Degree 2 polynomial kernel was used to generate the results.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
The dynamics of single neurons and neuronal circuits are controlled by various types of neural pathways involving excitatory and inhibitory neurons. It is generally suggested that the computation in neuronal circuits requires a balanced excitation (E) and inhibition (I) in synaptic transmission through regulation of neuronal excitability [ 1 – 5 ]. In the cerebellum, fine-tuning of E-I balance also plays an important role in cerebellar development and motor coordination [ 6 – 8 ]. Purkinje cells (PCs), as the only output cells of the cerebellum, leverage E-I balance for controlling their dynamics from different neural pathways, involving excitatory inputs from granule cells (GCs) and inhibition inputs from molecular layer interneurons (MLIs) [ 6 , 8 , 9 ].
A sequence of inhibitory inputs through MLIs can rapidly terminate the membrane depolarization of PCs induced by direct excitatory inputs from GCs [ 10 , 11 ]. In this way, correlated E and I inputs can prevent saturation of the postsynaptic spiking activity and extend its dynamic range for coding of a stimulus. Moreover, developmental changes in the strength of synapses from MLIs to PCs can result in an 11-fold decrease in overall postsynaptic currents [ 12 ]. Such a large variation of MLI-PC synaptic strength implies that the MLI activity must be coordinated to efficiently inhibit the activity of PCs. As a result, even a single MLI can dramatically change the PC activity [ 13 ].
The neuronal circuit in the cerebellum, as one of the densest networks, employs massive synaptic connections installed with their short-term dynamics for computation [ 10 , 14 – 16 ]. One important mechanism for modulating neural excitability is the influence of synaptic transmitter release by synaptic short-term plasticity (STP) at various types of synapses [ 17 – 19 ]. STP can make synaptic efficacy decreased (synaptic depression) or increased (synaptic facilitation) according to inputs from repetitive presynaptic activity [ 20 , 21 ]. Previous work observed that a wide range of synaptic change was correlated strongly with the strength of GC-PC synapses [ 22 ]. As a result, massive mossy fibers from GCs to PCs installed with STP can strongly contribute to the neural dynamics of PCs, which then provide a computational basis for a wide range of behaviors, from motor control to cognition [ 23 , 24 ].
Therefore, the diversity of excitation gated by the STP of GC-PC synapses, modulated by strong inhibition from MLI-PC synapses, can play an important role in the PC dynamics. Recently, it has been demonstrated that coordinated excitation and inhibition from synaptic short-term dynamics converged to PCs lead to a wide diversity of PC firing dynamics [ 18 , 19 ]. However, it is still not clear how the interaction of two neural pathways, the large variation of MLI strength in the feedforward inhibitory pathway, together with the STP of excitation raised in GC-PC synapses in the feedforward excitatory pathway, can influence the E-I balance for changing single PC firing dynamics and network behaviors of the PC circuit.
In this work, we addressed this question using a computational model of a neural network consisting of GCs, MLIs, and PCs. Specifically, we incorporated two neural pathways from GCs to PCs. The first one is the feedforward excitatory pathway from GCs to PCs. The second one is the feedforward inhibition pathway from GCs, via MLIs, to PCs. We aim to clarify the specific contribution of excitatory GC-PC synaptic STP and inhibitory MLI-PC strength to downstream PC dynamics. We show that the nonlinear characteristic of excitatory GC-PC STP dynamics can significantly affect PC dynamics in terms of firing rate, firing phase, and temporal spike pattern, which are modulated by MLI inhibition. In particular, excitatory STP enables nonlinear gain modulation. Notably, we demonstrate that the change of synchronization in the network is governed not only by the E-I balance but also by the synaptic STP depending on input burst patterns, whereas the pause response of the network emerges from the tight interaction of two neural pathways. Together with other recent findings, our results show that the interaction of neural pathways of excitation and inhibition dramatically modulate the neural dynamics of PCs that consequently change their network behaviors.
Methods
Single cell models
PCs and MLIs were modeled as modified integrate-and-fire neurons [ 25 ] that were also used for modelling of cerebellar cells [ 26 ]. The membrane potential V obeys the equation:
C m d V d t = - g L ( V - E L ) - I N a - I n o i s e - g A H P z A H P ( t ) ( V - E K ) - I s y n ( t ) (1)
where C m is membrane capacitance, g L is leak conductance and E L is leak resting potential. The sodium current was given by I Na = − g L Δ T exp[( V − V T )/Δ T ] with Δ T = 3 mV and the firing threshold V T drawn randomly for each neuron using a Gaussian distribution. When the membrane potential reaches the threshold V T at the spike t spk , V is set to 40 mV for a duration of the spike as τ dur = 0.6 ms. After the spike, at t = t spk + τ dur , repolarizing potential is set to V rest , and an afterhyperpolarization (AHP) conductance is activated. The gating variable z AHP follows the dynamics dz AHP / dt = (1 − z AHP )/ x AHP − z AHP / τ AHP . The resource variable x AHP obeys the dynamics d x A H P / d t = - x A H P / τ A H P x + δ ( t - t s p i k e - τ d u r ) , where τ A H P x = 1 ms. The refractory period is set as t ref = 2 ms. To mimic the ongoing activity in our simple point neuron models, a noisy excitatory current I noise = ( V − V E ) g N was injected with a slowly fluctuating conductance g N , described by an Ornstein-Uhlenbeck process, τ N d g N / d t = - g N + σ N τ N b ( t ) , where σ N = 0.12 nS, τ N = 1000 ms, and b ( t ) is white noise with unit variance density.
Similarly, GCs were modeled as previously based on experimental data [ 15 ], whose membrane potential V obeys the equation:
C m d V d t = - g l ( V - E L ) exp ( - ( V - E L ) / 5 ) - I n o i s e - g A H P z A H P ( t ) ( V - E K ) - I s y n ( t ) (2)
where the model components have the similar meanings as the PC and MLI model. All the parameters of neural models of PCs, MLIs, and GCs have the same values, except those listed in Table 1 , where the adjustment of parameters for individual cell types were based on previous studies [ 8 , 15 , 18 , 27 ].
10.1371/journal.pcbi.1008670.t001
Table 1
The parameters of single cell models.
Neuron
C(pF)
g L (nS)
E L (mV)
V T (mV)
V rest (mV)
g AHP (nS)
E K (mV)
τ AHP (ms)
PC
250
12.5
-70
−50 ± 1
-70
4
-100
20
MLI
20
1
-50
−45 ± 2.25
-50
4
-100
20
GC
4.9
1.5
-90
−50 ± 2.5
-65
1
-90
3
Synapse models
For synaptic currents, I syn of MLIs represents the total excitatory input arriving from GCs. I syn of PCs receives excitatory input from GCs and inhibition input from MLIs. I syn of GCs represents excitatory input from mossy fibers (MFs). All the synaptic currents were modeled with a similar form as:
I s y n = g m a x r ( t ) Y ( V - E s y n ) (3)
where E syn = 0 mV is for excitatory AMPA and NMDA synapses, and E syn = −80 mV is for inhibitory GABA synapses. The scaling factor Y is a nonlinear voltage-dependent function for NMDA: Y = 1/(1+ exp (−( V − 84)/38)). Otherwise, Y = 1 for other types of synaptic receptors.
The gating variable r was described by
r ′ = - r / τ d e c a y + α . s ( 1 - r ) s ′ = - s / τ r i s e + R u ∑ k δ ( t - t s p k ) (4)
where, Ru is to represent short-term synaptic plasticity with a simple phenomenological model that describes the kinetics of plasticity, such that it treats short-term depression and facilitation as two independent variables, R and u , respectively [ 20 , 28 ].
R ′ = ( 1 - R ) / τ r e c - R U δ ( t - t n ) u ′ = ( U - u ) / τ f a c + U ( 1 - u ) δ ( t - t n ) (5)
Totally, we mainly used four types of synaptic connections between neurons: the excitatory MF-GC, GC-MLI, and GC-PC synapses, and the inhibitory MLI-PC synapse. All types of synapses show a varying heterogeneity of short-term plasticity with mixed fast and slow time scales. When STP is switched off for GC-PC and MLI-PC synapse, the variables R = 1 and u = U are held fixed without dynamic updates. Synaptic delays in all synapse are included as 1 ms. In addition, MLI-PC synaptic delays are heterogeneous with a Gaussian distribution (mean as 1 ms and SD as 0.2 ms). In case of studying the effect of recurrent inhibition, the MLI-MLI inhibitory synapse was also included. Synaptic parameters for each type of synapse in the model were constrained by experimental measurements as shown in Fig 1B [ 12 , 15 , 22 , 29 ]. Postsynaptic currents recorded with current clamps at PCs from experimental data were fitted by models. Specifically, we used the following data: the postsynaptic GC-MLI current under a 10-spike stimulation clamped at -60 mV for MLIs [ 19 ]; unitary GC-PC current clamped at -70 mV for PCs [ 22 ]; unitary MLI-PC current clamped at 0 mV for PCs [ 12 ]; unitary MLI-MLI current clamped at -50 mV for MLIs [ 30 ]. The parameters values are described in Table 2 . Through the study, two key parameters are focused: the MLI-PC synaptic weight W MLI for varying inhibition strength; the initial efficiency U exc of GC-PC synapses for varying excitation strength. By default, we set W MLI = 3.5 nS and U exc = 0.4 as guided by experimental data in Fig 1 , unless those values mentioned differently in the work below. To further analyze the impact of MLI inhibition on PC network dynamics, we also systematically varied the initial efficiency U inh for MLI-PC synaptic STP.
10.1371/journal.pcbi.1008670.g001
Fig 1
PC dynamics controlled by excitation and inhibition.
(A) Schematic illustration of feedforward excitatory GC-PC short-term plasticity (STP) pathway and inhibitory GC-MLI-PC pathway on a PC. Granular cells (GCs, red), molecular intermediate neurons (MLIs, blue) and Purkinje cells (PCs, black). (B) Postsynaptic currents of four types of synapses from experimental data fitted by models. (C) The PC network with 50 PCs (black), 1000 GCs (red), and 500 MLIs (blue). For illustration, only 3 PCs are shown. (D) PC in response to the GC-PC input. (Left) EPSPs triggered by a single GC spike by varying GC-PC synaptic STP amplitudes U exc (0.05–0.75 with a 0.05 increment). (Middle) EPSPs triggered by a train of 10 spikes at 200 Hz at two different values of U: U exc = 0.06 for facilitation and U exc = 0.42 for depression, with (light blue) and without (purple) STP switched on. (Right) STP described by the ratio EPSP n /EPSP 1 showing facilitation or depression in a train of a varying number of burst spikes under different U (0.02-0.7, fixed burst frequency at 200 Hz). (E) Similar to D but for IPSPs triggered by the MLI-PC input. Single IPSPs induced by different strengths W MLI (0.5–7 with a 0.5 increment).
10.1371/journal.pcbi.1008670.t002
Table 2
Synaptic parameters for each synapse in the model.
Synapse
Strength
Synaptic dynamics
short-term plasticity
Pre-Post
type
g peak (nS)
α (1/ms)
τ rise (ms)
τ decay (ms)
U
τ rec (ms)
τ fac (ms)
GC-MLI
AMPA fast
3.2
3
1
1.5
0.1
100
50
NMDA
9.6
0.35
5
20
0.07
50
100
GC-PC
AMPA fast
0.5
3
1
1.5
U exc
50
400
AMPA slow
0.7
0.3
3
8
U exc
50
400
MLI-PC
GABAA fast
W MLI
3
1
10
0.1
100
800
GABAA slow
1.5 * W MLI
0.35
5
100
0.05
800
100
MF-GC
AMPA fast
1.2
3
0.3
0.8
0.5
12
12
AMPA slow
2.4
0.3
0.5
5
0.5
12
12
NMDA
2.88
0.35
8
30
0.05
-
-
MLI-MLI
GABAA fast
3.1
3
1
2.5
0.5
-
-
PC network model
A network model was set up with 50 PCs, 1000 GCs, 500 MLIs and 500 MFs, where each PC receives synaptic input from 100 GCs and 8 MLIs randomly, each MLI receives input from 4 random GCs. When studying the impact of MLI inhibition on PC network dynamics, we systematically varied two parameters: the number of MLIs targeting a PC for increasing overall inhibition, and the number of MLI-MLI connections for recurrent inhibition before converging to PCs. Therefore, for each PC, there are two input pathways: one direct excitatory input from GC-PC pathway, and another inhibitory input from GC-MLI-PC pathway. The schematic network is illustrated in Fig 1 . The spike firing of each GC was generated by injecting a sequence of MF-like spikes as described previously [ 15 ]. Each individual GC was activated with different types of spike input patterns, such as Poisson firing, regular firing, modulated firing, and burst firing. Irregular Poisson spike trains were generated based on the method described in [ 15 ]. Regular patterns were generated with the same interspike intervals. The input patterns before and after burst are either Poisson or regular spike trains across the population of GCs. The GCs modulation firing inputs were induced by the same probability distribution modulated sinusoidally in time for every trial as in [ 15 ]. It is worth noting that each GC was activated with a different onset time, so that all GCs were activated heterogeneously over the time course, except the burst stimulation, where all GCs were activated at the same time. Simulations were run in C++ VS2015 with a time step of 0.1 ms.
Data analysis
Data collected after simulation was loaded in MATLAB for further analysis. Simulations were run in four major different conditions: the baseline, i.e. , PC model running without MLI inhibition and without STP in GC-PC synapses, and three other conditions with MLI on, GC-PC STP on, or both on.
Gain and offset of the I-O function
PC firing rates (F) as a function of GC inputs were fitted with the following Hill function as previously [ 31 ]:
F ( GC ) = F max 1 + ( GC 50 / GC ) n + F 0 (6)
where n is the exponent factor, F 0 the firing rate offset, and F max the maximum firing rate. GC 50 is the value of GC at which F reaches half maximum. To investigate changes in the input-output relationship of PC firing caused by MLI and/or STP, the change of PC response was quantified by ΔGain calculated as follows [ 31 ]:
Δ Gain = ( F ′ + x - F ′ - x F - x ′ ) (7)
where F′ is the average slope of the fits between 5% and 75% its maximum value. +x and -x denote different conditions of with/without STP and/or MLI, e.g. ±STP or ±MLI. A shift along the input axis corresponds to an additive operation, while a change in slope corresponds to a multiplicative operation, or gain change. Offset shifts (ΔOffset) were defined as the difference between the half-maximum values of the fits in the conditions +x and -x.
Spike train analysis
To characterize the temporal structure of spike trains, we used three characteristics of spike trains as previously [ 32 ]: interspike intervals (ISIs), coefficient of variation (CV) of ISIs), and local regularity CoV 2 = 2|ISI n+1 − ISI n |/(ISI n+1 + ISI n ) for a pair (n, n+1) of ISIs.
Synchronization
To characterize network dynamics, the synchronization index for the whole population of PCs was calculated for burst stimuli. Using spike trains of all PCs with a 1 ms time bin, the coefficient of cross-correlation between any two PCs was calculated, i.e. , CC ij is the correlation coefficient between PC i and j . Then the network synchronization is given by,
K net = 2 ∑ i = 1 N - 1 ∑ j = i + 1 N CC ij N ( N - 1 ) (8)
where N is the number of the PCs in the network. The average K net , based on the firing rate over a time windows, was calculated for a period of post-burst stimulation in different parameters of burst setting. We used 20, 30 and 40 ms for bursts of 2, 5 and 7 spikes at 200 Hz, and 200, 100, 50 and 40 ms for busts of 50, 100, 200 and 300Hz with 10 spikes, respectively.
Pause response
Under the condition of both MLI and GC-PC STP switched on, burst stimulus triggers a prominent pause response after burst in the population spikes of PCs, which is defined by a time interval where there is no spikes in PCs, typically, it is from the offset of burst to the onset of next wave of population spikes in PC network.
Phase shift
To quantify the phase shift in the output PC spike train relative to the frequency of sinusoidal modulation of GC input spike train, the average PC firing rate over the whole population was fitted by a sinusoidal function A sin(2 πft + ϕ ) + C , where C is the offset of firing rate, and the modulation phase shift at frequency f is thus fully specified by the phase shift ϕ of the sinusoidal component, since the initial phase of input stimulus was set as 0 in all simulations.
Results
PC dynamics in response to GC input
PC has a unique feature with high-frequency firing activity as the primary information carrier affecting the activity of downstream neurons. The way a neuron transfers information can be represented by its input-output relationship, which is affected by synaptic inputs. Here we examine how inhibitory MLIs and excitatory GC-PC STP are cooperated to affect the PC dynamic sensitivity in response to the GC input.
For each PC, there are two streams of synaptic inputs coming to its dendrites as illustrated in Fig 1A . There is a feedforward inhibitory pathway, the synaptic GC-MLI-PC connection, which consists of GC-MLI synapses with AMPA and NMDA receptors activated by GCs, and MLI-PC synapses with GABA receptors activated by MLIs. Then each PC also receives another feedforward excitatory pathway, the synaptic GC-PC connection consisting of slow and fast AMPA receptors activated by GCs. In addition, there are recurrent connections between MLIs. These four types of synaptic currents can be constrained by experimental measurement shown in Fig 1B (see Methods ). To study the network dynamics of PCs, we set up a network with 50 PCs, 1000 GCs, and 500 MLIs illustrated in Fig 1C , where all types of neurons were implemented with modified integrate and fire models and synapses were modeled with short term plasticity (see Methods ). The weights of all synapses were drawn from a Gaussian distribution to consider the variability. For each PC, randomly generated network connections were used but with randomized synaptic weights so that the summation of synaptic inputs induces different responses for different PCs ( S1 Fig ).
When a sequence of mossy fiber spike inputs is delivered, GCs are activated to enable two synaptic pathways onto PCs playing different roles in controlling the firing dynamics. For which, we consider two key features, the excitation of the STP in GC-PC synapses in Fig 1D , and the inhibition strength from MLIs in Fig 1E , as previous evidence show that inhibition from a single MLI can efficiently change PC dynamics [ 13 ], and the presynaptic STP can dramatically affect PC activity [ 18 ]. We will show that these two factors contribute to PC activity in different ways.
The single spike response of a PC can be gradually changed by the strength of excitation and inhibition, where synaptic strength is regulated by the synaptic weight together with the initial efficacy parameter U of STP. When varying U exc and W MLI systematically, one can change the amplitude of EPSP (excitatory postsynaptic potential) and IPSP (inhibitory postsynaptic potential) recorded at a PC in Fig 1D and 1E ). The short-term dynamics of GC-PC and MLI-PC synapses depend on the STP parameters, e.g. , the initial efficacy U as in Fig 1D and 1E and time constants, and the input burst frequency ( S2 Fig ). The parameter U controls the amplitude of synapse release from facilitation to depression in terms of short-term dynamics. With a train of burst spikes at 200 Hz, synapses are facilitating when U exc <0.1 and U inh <0.15 in our model. As the MLI inhibition is more prominent with a large variation of the strength [ 13 ], we fixed U inh as 0.1 but changed the weight of MLI-PC synapses to investigate the interaction of inhibition and excitation. However, PC responses vary significantly for different formats of STP dynamics in our study, and the results presented in this work are robust to the change of STP parameters as shown below. We therefore set the default values as
W MLI = 3.5 nS and U exc = 0.4, unless those changed in the following study.
PC input-output relationship modulated by excitatory STP
In the cerebellum, outside inputs are represented by a sequence of mossy fiber spikes conveying efferent information into a high-dimensional, sparse code of a large population GCs in the network [ 16 , 33 , 34 ]. In the default case, there is no MLI inhibition (MLI off) and no GC-PC STP onto PCs (STP off, i.e. , no short term dynamics on GC-PC synapses), while other types of synapses are still dynamic and subjected to the change following the STP rule (See Methods ). We aim to examine how these two factors, MLI inhibition, and GC-PC STP, are interacted to change the balance of excitation and inhibition and modulate the dynamics of downstream PCs.
The first fundamental feature of neural computation is the input-output relationship, i.e. , I-O function, of PC firing. To investigate this, we used a wide range of input frequencies and different types of stimulation, which are represented by GC firing activities. Within the context of the I-O function, there are two types of changes in the shape mostly observed for single neurons. The first one is the additive change, where there is a shift of the I-O curve without changing the shape. Another one is the multiplicative change, or gain change, where the I-O function is reduced or increased by a multiplicative factor.
We first leave out MLI inhibition and study the effect of STP of GC-PC synapses by injecting a sequence of Poisson spike trains at different frequencies. Enabling STP can suppress excitatory input conductance and reduce the number of spikes in PCs ( Fig 2A ), which nonlinearly depends on the GC input, such that total excitation G exc is boosted at lower frequencies but suppressed higher input frequencies ( Fig 2B ). Compared to the baseline, adding MLI inhibition produces an approximate additive change of the I-O function with reduced firing. However, using STP can nonlinearly change the I-O function as in Fig 2C . Our default values (U exc = 0.4 and U inh = 0.1) generate depressing excitation on GC-PC synapses and facilitating inhibition on MLI-PC synapses. It has been suggested from experimental data that these synapses could behave in an opposite way [ 18 , 19 , 35 ]. We implemented our model with an opposite pair of values (U exc = 0.08 and U inh = 0.2) giving facilitating excitation and depressing inhibition, and found the similar results for nonlinear gain control with GC-PC STP ( S3 Fig ). Furthermore, the similar results still hold when the MLI-PC STP was switched off ( S3 Fig ). This confirms the above observation as the effect of STP in GC-PC synapses, rather than the detailed profiles of facilitation or depression in the GC-PC STP. We then characterized the change in slope (ΔGain) and the shift in the half maximal response (ΔOffset) of the I-O relation using fits to Hill functions as previously [ 31 ] (see Methods ). Results in four conditions in Fig 2D reveals that GC-PC STP has a greater effect on gain modulation, which is more prominent when both STP and MLI are switched on. Thus, excitatory STP can greatly enhance gain modulation aided by MLI inhibition.
10.1371/journal.pcbi.1008670.g002
Fig 2
Inhibition-mediated gain modulation of PC firing dynamics enhanced by excitatory short-term plasticity.
(A) (Top) Sum of excitatory conductance G exc onto a PC in the baseline condition (STP off) and a test condition with STP. The input is 100 independent synaptic trains using Poisson stimulation at 50 Hz. (Bottom) PC membrane potential traces with and without STP. Vertical ticks indicate spike times. (B) Average G exc changing over a range of GC rates with and without STP. G exc was averaged over the time course and all GCs connected to a PC. (C) PC input-output relationship in four conditions, with/without STP and/or MLI. Each point is mean±SD (n = 50). Poisson stimulation was used. Lines in (B) and (C) are fits to a Hill function. (D) Heterogeneous gain and offset changes due to STP (±STD) and MLI (±MLI) from fits in (C).
In STP, facilitation needs lower values of U, whereas depression for higher values of U. Thus, fixed U values change scaling of synaptic currents, as well as profiles of STP. Moreover, STP also depends on input frequency, such that at higher input frequencies, dynamical variables of STP, particularly resource variable R saturates with a limited dynamic range to regulate synaptic currents. As a result, differences of gain control is more prominent at lower frequencies. To further characterize the effect of excitation STP and MLI inhibition on the PC I-O function, we systematically varied the strengths of STP U exc and inhibition W MLI , respectively. In the absence of STP, there is a limited effect of inhibition when changing W MLI as in Fig 3A (top). However, when STP is presented, a sequence of increased inhibition results in a dramatic change of PC firing ( Fig 3A , bottom). Under the same amount of inhibition increment, there is a larger step of gain change with STP than the case without STP.
10.1371/journal.pcbi.1008670.g003
Fig 3
The gain change of PC firing significantly depending on GC-PC STP.
(A) PC firing modulated by different levels of MLI inhibition strength without STP (top) and with STP (bottom) at U exc = 0.4. (B) Excitatory conductance modulated by different levels of GC-PC synaptic efficacy U exc without STP (top) and with STP (bottom).
Similarly, we leave out MLI inhibition while changing U exc by turning STP on and off respectively. Without STP, decreasing excitation strength shows a similar profile as increasing inhibition, as shown by the total excitation G exc in Fig 3B (top). The gain change is well displayed, in particular with larger difference for 10-50 Hz GC inputs, whereas less difference at higher frequencies due to saturation of short-term dynamics ( Fig 3B , bottom). This is due to that when STP is present, low synaptic efficacy shows more facilitation whereas high synaptic efficacy shows more depression, which can compensate for the overall change of total excitation (see S4 Fig for lower values of U exc ). These dynamic behaviors are also shown in the profiles of each individual synaptic component of GC-PC synapses ( S5 Fig ). Therefore, these results confirm that GC-PC synaptic STP with the aid of MLI inhibition can significantly enhance gain modulation of PC firing dynamics.
PC firing phase modulated by excitatory and inhibitory pathways
Modeling studies suggest that synaptic short-term plasticity can contribute to the phase shift in neural dynamics in response to time-varying inputs at the single cell level [ 36 , 37 ]. Recent experimental findings in the cerebellum suggest that there is a diversity of phase shifts in different types of cerebellar neurons [ 38 ]. Here we investigate how MLI inhibition and STP of GC-PC excitation affect PC responses under oscillating inputs at the population level.
Sinusoidally-modulated inputs with different frequencies were injected into GCs, then drove the PC population firing activity fitted by a sinusoidal curve shown in Fig 4A . When increasing input driving frequency, there is a corresponding increase in the modulated phase shift. Relative phase shifts are easily seen when these fitted sinusoids are overlapped in Fig 4B in the baseline case. Most of the input frequencies generate a phase delay lag of PC firing, except for very low frequency, here 1 Hz modulation. The input with 30 Hz completely reverses the phase in PC firing. Compared to the baseline, phase shifts induced by MLI and STP are different in Fig 4C . MLI inhibitory inputs directly delay PC spike timing and drive large phase delays in PCs, especially, on higher input frequencies. However, excitatory STP results in a relative decrease in the phase shift in response to the same input. Pairing STP with MLI results in a balanced state where the phase shift shows a compromised change as an average of two factors. By varying excitation U exc and inhibition W MLI , there is a wide range of systematical phase shifts under a sequence of input frequencies from 0 to 30 Hz in Fig 4D . At the low level of synaptic efficacy as 0.05, strong inhibition inputs trigger few spikes. Otherwise, at other synaptic efficacy values, the increased inhibition dramatically delays phase shift. However, the opposite effect is found for STP of excitation.
10.1371/journal.pcbi.1008670.g004
Fig 4
PC phase modulation affected by MLI and STP.
(A) PC population firing rate (n = 50) in response to GC inputs sinusoidally-modulated with 1, 10, 20 and 30 Hz, from 100 independent input spike trains. Solid color lines are fitted with sinusoidal functions. (B) Normalized fitting curves from (A) show phase shifts relative to the input. (C) Phase shift as a function of input frequency in four different conditions (left), and the corresponding changes of phase shifts relative to the baseline (right). (D) Phase shift changed over a range of excitation U exc and inhibition W MLI . (Left) Phase shift over a sequence of input frequencies with different levels of excitation and inhibition in three conditions. (Right) Phase shift changed by combined MLI and STP, where each inner rectangle represents a PC phase shift spectrum over the same sequence of input frequencies (0-30 Hz).
As phase shift is related to the time scale of neural dynamics [ 15 ], we examined the detailed dynamics with varying time constants of STP. Combinations of facilitation and recovery time constants, ranging from a few to hundreds of milliseconds, enable information transmission over a very wide range of interspike intervals [ 39 ]. Fig 5A shows the PC response with different facilitation and recovery time constants at 10 Hz modulation frequency. The change in the phase shift in Fig 5B shows that increasing the recovery time constant results in increased phase shifts at all frequencies. However, increasing the facilitation time constant results in increased phase shifts at a higher modulation frequency (large than 10 Hz), whereas decreased phase shifts at low modulation frequency inputs (less than 10 Hz). Fig 5C shows the systematic change of phase shift for combinations of facilitation and recovery time constants across different input frequencies.
10.1371/journal.pcbi.1008670.g005
Fig 5
PC phase modulated by timescales of short-term dynamics of GC-PC synapses.
(A) Histograms of spikes in GCs (top) and PCs (bottom) in response to an input spike train sinusoidally-modulated at 10 Hz, with three examples of time constants of recovery τ rec (fixed τ fac = 400 ms) and facilitation τ fac (fixed τ rec = 50 ms). (B) Phase shift as a function of input frequency for (left) τ rec and (right) τ fac , with similar settings as in (A). (C) Phase shift in the parameter space of τ rec and τ fac . Each inner rectangle represents a PC phase shift spectrum changing over different modulation frequencies (0-30 Hz).
These results suggest that MLI inhibition is sufficiently strong to increase phase delays, whereas STP of GC-PC excitation results in reduced phase delays. Within the STP, the recovery time constant has a greater impact on phase shift than the facilitation time constant. In this way, the balance between excitation and inhibition cooperatively shapes the modulation of PC responses in the network.
PC temporal spike pattern affected differently by excitation and inhibition
Apart from firing rate, spike temporal pattern has been suggested to play an important role in information processing of neuronal dynamics [ 40 ]. Inhibition can control the temporal window of PC discharge and cause highly variable delays in the membrane potential dynamics of neurons [ 11 ]. Moreover, STP can induce complex contingencies on the temporal patterns of neural activity [ 39 ] and contribute to temporal filtering of synaptic transmission [ 28 ]. Here, by varying inhibition W MLI and excitation U exc , we investigate the impact of these factors on the temporal pattern of PC spikes ( Fig 6 ).
10.1371/journal.pcbi.1008670.g006
Fig 6
PC temporal spike patterns affected by MLI and STP.
(A) Schematic illustration of feedforward MLI inhibition and GC excitation when GC-PC STP is off. (B) PC membrane potential traces gradually delayed by different levels of MLI inhibition with an input of 10 Hz Poisson spikes. (C, D) Similar to (A, B) but for MLI off and STP on with varying excitation at different levels of U exc . (E) Interspike interval (ISI) distribution of the population PC spikes under four different settings with/without excitatory STP and/or MLI (left). (Right) The change of ISI, CV (coefficient of variation), and CoV 2 (local regularity) over a range of GC inputs, averaged over the population of PCs (mean±SD, n = 50).
As expected, stronger inhibition increases a systematical delay of PC action potentials ( Fig 6A and 6B )). However, larger excitation in STP results in contingencies in spike latency such that spike times can be shifted in both ways ( Fig 6C and 6D )). Characteristics of the temporal structure of spike patterns, interspike interval (ISI), coefficient of variation (CV) of ISIs and local regularity (CoV 2 , see Methods ), show a larger variation and higher values with both STP and MLI on ( Fig 6E ). Presumably, these variations could greatly enhance the dynamic range of the PC response. Taken together, these results suggest that the mixture of inhibition and excitation installed with STP can change PC spike patterns in a dynamic coding fashion. Then we will explore this role at the network level.
PC network dynamics in response to burst input
Finally we investigate how MLI and STP reshape high-frequency bursts of GC activity into PC spike activity. Occurring at a much shorter time scale with a few spikes, a burst at a very high frequency of a few hundred Hz can change the PC response remarkably [ 41 ]. Fig 7A shows an illustration of the burst stimulus protocol, where a sequence of 3 or 7 spikes at 100 Hz was delivered to each GC, together with the corresponding traces of EPSPs and EPSCs ((excitatory postsynaptic current)) recorded at a PC. Enabling GC-PC synaptic STP has a strong effect of short-term depression under U exc = 0.4, compared to the baseline case without STP.
10.1371/journal.pcbi.1008670.g007
Fig 7
PC network dynamics in response to burst input.
(A) Schematic illustration of GC-PC and GC-MLI-PC pathways receiving burst inputs. (Left) GCs stimulated by bursts with 3 and 7 spikes at 100 Hz. (Right) EPSPs and EPSCs recorded from PCs for different burst inputs in the baseline condition (red) and with STP on (green). (B) GC spike raster triggered by a burst input of 5 spikes at 200 Hz, only 50 GCs are shown (top). (Bottom) The corresponding spike rasters of 50 PCs (left), and averaged PC population firing profiles in different conditions. Note that the pause response indicated as the time interval between two arrows in the condition of +STP+MLI only.
To study the effect of burst input on the PC network, we used a background 20 Hz Poisson spikes first, then a sequence of 5 spikes at 200 Hz was injected into all the GCs in the network to mimic the synchronization of GCs in the granular layer [ 42 ] as in Fig 7B . As a consequence, the spike pattern and averaged population firing rate of PCs are triggered. Compared to the baseline without MLI and STP, PC responses are differently tuned by MLI and STP, separately or simultaneously. When both MLI and STP are on, PCs show a prominent response feature: following a burst input, there is a silent period of several tens of milliseconds, named as a pause characterized by the length of the time interval of PC activity after the burst [ 43 ]. Pause response remains when switching off the MLI-PI STP in the model ( S6 Fig ), and is not depending on the specific choice of GC-PC STP parameters, even using an opposite pair of the synaptic U values that give facilitating GC-PC synapses and depressing MLI-PC synapses ( S6 Fig ). More detailed analysis shows that during burst inputs, GC-PC synapses become effectively depressed and excitability on PCs is reduced by STP. When MLIs are applied and there is a balanced level of inhibition on PCs, pause can be generated ( S7 Fig ). These results suggest that the interaction of both neural pathways on PCs, resulting in balanced excitation and inhibition, is necessary to generate pause response.
A unique feature of network dynamics is the synchronization of a neural population. Synchronized PC activity, affecting either accurate timings or rate changes in the downstream nuclei [ 44 ], has been suggested to result from the upstream GCs [ 45 ]. We conducted experiments on varying the duration of the burst with a fixed number of spikes, to test the burst effect on the PC response. Consistent with the previous observation that input bursts at frequencies up to 300Hz allow PC firing to persist at the network level [ 41 ], we found that activation of GCs with 10 stimulus spikes at 50, 100, 200 and 300 Hz, can trigger different types of PC firing dynamics, depending on MLI and/or STP included or not. The PC population firing rate tends to be sustained in the baseline, whereas, MLI builds up dynamics over stimulus, STP generates sharp and transient dynamics, and using MLI and STP together makes the dynamics more transient ( Fig 8A and S8 Fig ).
10.1371/journal.pcbi.1008670.g008
Fig 8
PC synchronization controlled by excitation and inhibition.
(A) PC population firing rate in different conditions. (Right) The corresponding MLI population firing rate when MLIs are triggered by GCs. The burst input here is 10 spikes at 50Hz. (B) Time course of PC network synchronization computed from the population PC firing rate, in response to input bursts with 10 spikes at 50, 100, 200 and 300 Hz under different conditions. (Right) The corresponding time course of MLI population synchronization. (C) Similar to (B) but for burst inputs at 200 Hz with 2, 5 and 7 spikes, respectively. Colored shadows indicate burst duration. The background noise are Poisson spikes at 20Hz. The time scale bar in (A) applied in all the plots.
To quantify the synchronization of PC firing in response to burst inputs, the index K net was calculated by cross-correlation of the firing rate profiles between a pair of PCs (see Methods ). Fig 8B and 8C ) shows the temporal change of K net in different conditions with/without MLI and STP, separately or simultaneously, and under different protocols of burst inputs varying either bust duration in terms of the frequency with a fixed number of spikes, or burst spikes with a fixed frequency. Under burst inputs consisting of a sequence of 10 spikes at 50, 100, 200 and 300 Hz, the change of synchronization shows different behaviors in the network of PCs and MLIs. In general, MLIs show a systematically increasing synchronization with higher frequencies, whereas synchronization of PCs varies according to burst frequency. In contrast, PC synchronization is changed systematically with the increasing number of spikes at a fixed frequency ( Fig 8C and S9 Fig ), where the PC network was triggered using the different bursts with 2, 5, 7 spikes at 200 Hz. These results suggest that the interaction of inhibitory MLIs and excitatory STP can modulate synchronization in different ways.
To obtain a detailed map of the change in PC network synchronization triggered by bursts within the parameter space of inhibitory MLI W MLI and excitatory STP U exc , we quantified the gain of synchronization defined as the relative change calculated by the averaged K net of test conditions (with/without MLI and STP) subtracted by that of the baseline. Fig 9A shows that there is a large variety of changes across conditions. Generally speaking, weaker inhibition results in decreased synchronization, except the case for the burst with 2 spikes at 200Hz, where spikes are too few to obtain the significant change. Similarly, we computed the pause response as in Fig 9B . It is clear that the change of pause is more systematic. The network develops a systematic pause with longer duration after burst input with increased MLI inhibition, which is expected due to that inhibition prolongs the pause period. In addition, low excitatory synaptic efficacy tends to enhance pause duration when inhibition strength is fixed, since the excitation is too weak at low values of U exc and inhibition is more stronger.
10.1371/journal.pcbi.1008670.g009
Fig 9
The change of PC synchronization and pause response in the parameter space of excitation and inhibition.
(A) The gain of synchronization changed in different ways with W MLI and U exc under bursts with different spikes and frequencies. Each inner rectangle in each panel represents a single gain value of synchronization under one burst protocol. (Top) Bursts with 2, 5 and 7 spikes at 200 Hz. (Bottom) Bursts with 10 spikes at 50, 100, 200 and 300 Hz. In all plots, note that there is no STP but static excitation with U exc = 0.4 in the last row of the matrix. Similarly, there is no MLI inhibition in the first column of the matrix. Thus the first point at the left-bottom corner of the matrix is the baseline. The gain was defined as the relative change calculated by the average K net compared to that in the baseline. (B) Similar to (A) but for pause response induced by bursts.
So far we varied excitation and inhibition using excitatory STP strength U exc and inhibitory MLI strength W MLI . Total excitation and inhibition are contributed by a set of four parameters, weights of GC-PC and MLI-PC synapses and their STP strengths. We then systematically vary the parameters of STP in both U exc excitatory and U inh inhibitory synapses. In contrast to the nonlinear behaviors shown above, the paired change of U values results in a linear behavior of the gain of synchronization ( Fig 10A ). Such that excitation and inhibition have an opposite effect on the gain: higher excitation reduces gain, whereas higher inhibition increases it. However, the U strength of excitatory plays a dominant role in controlling of pause response as shown in Fig 10B , where there is little effect of U inh . Therefore, a large variation of MLI inhibition strength, rather than short-term dynamics of MLI-PC synapse, is essential to the nonlinear tuning of PC network dynamics.
10.1371/journal.pcbi.1008670.g010
Fig 10
PC synchronization is controlled by STP of both excitation and inhibition, whereas PC pause response is less dependent on STP of inhibition.
(A) The gain of synchronization regulated differently by STP of ML-PC inhibition U inh and excitation GC-PC U exc under bursts with different spikes and frequencies. Each inner rectangle in each panel represents the gain of network synchronization under one burst protocol. (Top) Bursts with 2, 5 and 7 spikes at fixed 200 Hz. (Bottom) Bursts with 10 spikes at 50, 100, 200 and 300 Hz. Note that the gain was defined as the relative change calculated by the average K net with STP subtracted by that without STP. (B) Similar to (A) but for pause duration induced by bursts.
The results above indicate that it is MLI inhibition that plays an important role in PC network dynamics. We then systematically vary inhibition strength by increasing the number of MLI-PC synapses per PC from 1 to 8 MLIs ( Fig 11A ). Consistent with previous results, stronger inhibition generates a large change for PC synchronization. Depending on the burst protocol, there is more or less synchronization changing over MLI inhibition. However, the pause response is systematically large with stronger inhibition. Another feature for MLIs is that there are recurrent connections between MLIs [ 30 ]. We then generate different recurrent connections between MLIs by increasing the number of MLIs-MLIs synapses per MLI, while fixing the inhibition level using 8 MLIs per PC ( Fig 11B ). There is no systematic change across different levels of recurrent MLI connection. Thus, it is the inhibition strength from MLIs that determines PC synchronization and pause response.
10.1371/journal.pcbi.1008670.g011
Fig 11
PC synchronization and pause response controlled by MLI inhibition.
(A) PC network with no recurrent MLI inhibition. (Left) Illustration of the PC network receiving different numbers of MLIs. PC synchronization (top) and pause response (bottom) affected by MLI inhibition with stronger weight or more number of MLI-PC synaptic connections per PC, under burst inputs with different spikes and frequencies. Inner rectangles in each panel represent the gain of synchronization under one burst protocol. The gain of synchronization was defined as the relative change of the average K net . (B) Similar to (A) but for the PC network with recurrent MLIs included, such that each MLI receives 1-8 MLIs recurrently. Here in all plots, each PC receives 8 MLIs.
Taken together, these findings suggest that both synchronization and pause response in the PC network are affected by MLI inhibition and GC-PC synaptic STP with the transition from facilitation to depression. However, the major role in controlling pause response is MLI inhibition, where short-term dynamics of MI-PC synapses and recurrence of MLIs less affect the pause response of PC network dynamics.
Discussion
In this study, using a network consisting of Purkinje cells with their upstream excitatory granular cells and inhibitory molecular layer interneurons under different stimulation protocols of mossy fiber spiking inputs, we explored how the spiking dynamics of PCs are modulated by two neural pathways from GCs to PCs: direct excitatory GC-PC pathway and feedforward inhibitory GC-MLI-PC pathway. We focused on GC-PC synaptic short-term plasticity and strong MLI-PC inhibition to regulate PC spiking activity at the level of single cells and networks. On the single-cell level, the input-output firing dynamics, temporal spiking pattern, and phase shift, are nonlinearly modulated by these two factors. On the network level, the synchronization and pause response are modulated by excitation and inhibition. Notably, nonlinear gain control is achieved by excitatory STP, and pause response is controlled by the interaction of both neural pathways.
Inhibition-mediated PC dynamics modulated by short-term plasticity of excitation
Inhibition is a unique feature for shaping neural dynamics in neuronal circuits [ 46 ]. In the cerebellum, the role of MLIs is significant in shaping PC dynamics [ 13 , 14 ], with strong synaptic strength changing up to 11 folds for postsynaptic PCs [ 12 ]. PCs receive direct MLI inhibitory inputs from stellate cells at pericellular terminals known as spiny dendrites, and also from basket cells that terminate on the smooth shafts of dendrites [ 47 ]. A large variety of inhibitory strength depends on the specific location rather than specific types of cells. Although there are various types of short-term dynamics in MLI-PC synapses [ 18 , 19 ], the determinate role in controlling PC dynamics is more dramatic when turning on or off a single MLI, which can alter downstream PCs significantly shown by in vivo experiments [ 13 ]. MLIs also show spontaneous firing activities in the absence of excitatory inputs [ 48 , 49 ], which make inhibition more prominent to PCs. Thus, it is expected that the effect of the MLI inhibition strength is more significant in the PC network [ 13 ].
On the other hand, the excitation to PCs is controlled by the short-term dynamics of GC-PC synapses [ 18 , 22 ]. There is a large variation of synaptic modifications due to different release probabilities of GCs, such that short-term plasticity of GC-PC synapses can lead to a wide range of changes in release probability [ 22 ]. Therefore, here we stressed MLI inhibition and GC-PC synaptic STP together and studied how short-term plasticity of excitation changes the dynamics of MLI-medicated PCs. MLIs are highly sensitive to excitatory inputs [ 29 ], such that increased weight of inhibition can narrow the time window for the integration of excitatory inputs. As a result, different contributions of MLI and GC-PC STP could affect the time window, especially changing the precise synchronous firing time of PCs.
Input-output function of PC firing
How neurons respond to stimuli is one of the central questions in neural coding. The stimulus-response, or input-output, relationship of single neurons is a fundamental feature for neural dynamics. A change in the gradient input-output function of a neuron, termed gain modulation, is associated with several factors, including shunting inhibition, synaptic noise, and dendritic morphology [ 50 – 53 ]. Here, we investigated the contribution of synaptic input to gain modulation taking into account the effect of MLI inhibition and STP of excitation, which can modulate PC discharges in response to GC inputs. Our findings show that the effect of inhibition is an approximate additive change of gain modulation, which is related to the fast onset of feedforward inhibition that can rapidly limit excitatory postsynaptic potentials to reduce responsiveness to inputs [ 10 , 11 ]. In contrast, excitatory STP enhances gain changes due to highly nonlinear synaptic behavior. However, short-term depression and facilitation modulate gain changes differently, since short-term depression is functionally like low-pass filtering, whereas short-term facilitation acts as high-pass filtering. Thus, the way of gain modulation can be altered by controlling of STP even with the same input.
The functional role of STP in nonlinear gain control is generally demonstrated [ 54 ] and also observed in GC recordings with dynamics-clamp inputs mimicking synaptic spiking inputs [ 31 ]. Consistent with these observations, we found nonlinear gain control exists in a wide range of model settings parameters of short-term dynamics of synapses. Linear PC I-O function could be a specific case of nonlinear gain control under certain conditions [ 55 ]. Our results suggest that the diversity of nonlinear gain control in different degrees could enhance neuronal computation using different coding strategies and make the cerebellar microcircuit as an adaptive filer [ 56 ].
Phase modulation of PC firing
The impact of synaptic dynamics on the phase of neuronal responses is potentially significant in a wide range of neuronal systems, particularly for sensory processing and generating motor outputs in the cerebellum [ 15 , 38 , 57 ]. Here we found that PC phase modulation at the population level can be determined by the input frequency, but can also be adjusted by inhibitory MLI and excitatory STP differently.
Phase modulation by inhibition is sensitive to the input frequency and exhibits in-phase at low frequency and phase lag at high frequency, especially for higher inhibition intensity, which is due to that inhibition can delay spike timing. In contrast, phase modulation by excitatory STP exhibits phase lead at high frequency. As STP can show both depression and facilitation, depressing synapses trigger spikes after a long period of presynaptic inactivity, whereas facilitation synapses are most effective at transmitting at the end of a burst of activity [ 39 ]. Our results indicate that combined inhibition and excitatory STP act on spike times in a nonlinear way, which then extends the PC encoding capability to modulate the phase to downstream neurons.
Synchronization of PC firing
For neural temporal codes involving spike pattern and temporal structure in spike trains, previous studies suggest that PC synchronization has a great effect on the computation carried out by downstream neurons [ 27 , 44 , 45 , 58 , 59 ]. There are several proposed mechanisms contributing to PC synchronization, including GC inputs, ephaptic coupling between PCs, and PC axon collaterals [ 27 , 60 – 64 ]. It is possible that a combination of multiple mechanisms could corroborate to generate the synchronization of PC firing [ 64 ].
One of the main contributions to the synchronization of PC simple spikes is the excitation input from GCs [ 45 , 61 ]. GC synchronization is a general phenomenon that could be induced by Golgi cells [ 65 , 66 ]. Here we delivered a burst protocol mimicking the effect of GC synchronized spikes [ 42 ] and investigated how they are propagated to PCs. We then studied how the precise PC synchronization is affected by the STP of GC-PC synapses and strong inhibition from MLIs. We found that GC bursts can trigger PCs to generate a significant level of synchronization, mainly due to the slow spillover component of GC-PC synapses that provides persisting excitation inputs. As a result, excitation slowly returns to baseline, and there is enough time to accumulate excitation continually approaching the spike threshold. For instance, with a 300 Hz burst, the interspike interval is only about 3 ms that makes the slow spillover component staying at a high level to easily produce a spike for the next incoming stimulus.
In more realistic scenarios, excitation and inhibition are entangled. Our results here suggest that PC synchronization could result from different effects of excitatory STP and MLI inhibition on precise temporal dynamics of simple spikes, which then are modulated by the balance of excitation and inhibition, together with input burst spike patterns, such as the duration and number of spikes in a burst.
Pause response in PC firing
Another unique feature of temporal spiking pattern is the pause response shown here for PCs. A delayed and adaptively timed pause in PCs simple spike firing is thought to play an essential role in the computation performed by PCs [ 67 – 70 ], as well as modulate downstream deep cerebellar nucleus neurons [ 71 ]. This type of response has been ascribed to intrinsic mechanisms and long-term depression of parallel fibers converged to PC synapses [ 43 , 70 , 72 – 75 ].
Here, we investigated how pause response is elicited by GC-PC and GC-MLI-PC neural pathways. The pause is more likely to appear when both neural pathways are incorporated, which is presumably due to that the offset of inhibition is regulated by short-term dynamics of excitability. Burst input can induce a significant pause that is enforced by MLIs. Overall, all the four parameters of MLI-PC inhibition, the number of synapse, number of MLI-MLI recurrent connection, synaptic weight, and synaptic short-term dynamics, could contribute, in different ways, to the change of MLI-PC synaptic transmission. However, consistent with the previous findings [ 74 ], the main factor is the enhancement of MLI-PC synaptic weight that increases the duration of pause response and provides evidence that spike pauses are mainly regulated by MLI. We also observed that MLI-PC and MLI-MLI inputs regulate the pause in an opposite fashion, and the input frequency has no significant effect on pause, while the pause duration increase as the inhibition strength increases, which is consistent with experimental results [ 76 ].
Limitations
Here we explored how PC dynamics can be modulated by synaptic inputs from two neural pathways. It is also known that neuronal dynamics can be modulated by various intrinsic properties. One of which is neuronal morphology. Especially PCs have a complex morphological structure in which parallel fibers and climbing fibers are targeted at different parts of the dendritic tree [ 8 , 77 ]. Although the complex dendritic tree can be reduced to a simple structure while preserving basic characteristics of firing activity [ 78 ], the PC dynamics is suggested to branch-specific due to the distribution of ion channels [ 77 ]. These intrinsic mechanisms can dramatically regulate PC dynamics. For instance, Ca 2+ -activated K + channels behave like a high-pass filter that allows PC to respond GC high frequency inputs [ 79 ]. For MLI inhibitory synaptic inputs on PCs, it is also suggested that the input from stellate cells is often located at spiny dendrites, whereas basket cells are more on the soma and smooth dendrites [ 47 ], which induce different inhibition strengths to PCs. Here we mimicked a wide range of inhibition strengths by varying the strength of MLI-PC synapses. However, these explicit components, dendritic locations of synapses and ion channels, are not explored in this study.
In the cerebellar granular layer, Golgi cells provide inhibition input for GCs and form both feedforward and feedback inhibition loops to regulate GC dynamics [ 80 ]. It has been suggested that GC network dynamics of synchronization and oscillation are related to the feedback loop between GoC and GC [ 65 , 66 , 81 ], and gap junctions between GoCs that can promote synchronous oscillatory patterns [ 26 ]. Therefore, Golgi cells could contribute to the modulation of PC dynamics [ 80 , 82 , 83 ]. Further studies are needed to explore the role of Golgi cells, together with GCs and MLIs, in regulating PC dynamics.
Here we only considered the dynamics of PC simple spikes. Another unique feature is that PCs also receive strong excitation from climbing fibers represented by distinctive responses known as complex spikes, which play an important role in cerebellar learning [ 77 ]. Moreover, climbing fibers also regulate the dynamics of MLIs via synaptic spillover [ 84 ] and neighbor PCs via ephaptic coupling [ 85 ]. Therefore, they could potentially contribute to the PC network dynamics [ 13 ]. Further work is needed to elucidate these unaddressed questions.
Supporting information
S1 Fig
Related to Fig 1 . The variability of PC responses in the network induced by the randomization of parameter setting in the model.
(A) Membrane potential traces of three example PCs triggered by 20 Hz Poisson spikes. (B) The corresponding ISI (interspike interval) distribution. (C) The input-output relationship represented by PC firing a function of input GC spikes. Here the simulation is conducted in the baseline condition, where there is no inhibitory MLI (MLI off) and no excitatory STP on GC-PC synapses (STP off, i.e. , synaptic dynamics is not subjected to the modulation of short term plasticity).
(TIF)
S2 Fig
Related to Fig 1 . The dynamics of short-term plasticity (STP) with different settings of parameters and stimulation protocols.
(A) The STP dynamics with different parameters of time constants: τ fac for facilitation and τ rec for depression, under two settings of initial efficacy U for GC-PC synapses and MLI-PI synapses. (Left) EPSPs triggered by a train of 10 spike at 200 Hz at U exc = 0.06 for facilitation and U exc = 0.42 for depression, with different facilitation time constants ( τ fac ) and recovery time constants ( τ rec ). (Right) Similar to EPSPs but for IPSPs with U inh . (B) STP described by the ratio PSP n /PSP 1 (EPSPs, top; IPSPs, bottom) showing facilitation or depression triggered by a train of different burst spikes at 50, 100 and 300 Hz with a wide range of U values (0.02-1).
(TIF)
S3 Fig
Related to Fig 2 . Nonlinear gain control of the PC I-O curves with different settings.
(A) The profiles of PC firing with an opposite (in contrast to the default values) pair of STP initial efficacy U values: U exc = 0.08 for GC-PC facilitation and U exc = 0.02 for MLI-PC depression, under different conditions. Here STP refers to GC-PC STP, and MLI refers to MLI-PC inhibition. Baseline without STP and MLI (-STP-MLI); STP ON without MLI (+STP-MLI); STP off with MLI but no MLI-PC STP (-STP+MLI(-STP)); STP off with MLI and MLI-PC STP (-STP+MLI(+STP)); STP on with MLI but no MLI-PC STP (+STP+MLI(-STP)); STP on with MLI and MLI-PC STP (+STP+MLI(+STP)). Each point is mean±SD (n = 50). (B) Similar to A but GC-PC synapses are depressing and MLI-PC synapses are facilitating. Here the STP of MLI-PC synapses is switched off, compared to Fig 2 . W MLI-PC = 3.5 nS gives strong inhibition and depresses PC firing (left). The I-O profiles are more visible with W MLI-PC = 1.2 nS. Lines are fits to a Hill function.
(TIF)
S4 Fig
Related to Fig 3 . The PC I-O curves affected by GC-PC STP and MLI inhibition in different ways.
(A) The total excitation input G exc as a function of GC input at various levels of synaptic efficacy U exc (0.01-0.09) without short-term facilitation (STF) (left) and with STF (right). (B) The gain change without and with STP for different levels of inhibition (left), and without and with inhibition for different levels of STP U exc (right), compared to the baseline for each case. The MLI inhibition weights are changed by a scaling factor as shown in the axis index. The default parameter values are used here, such that +STP means U exc = 0.4, and +MLI means W MLI = 3.5 nS.
(TIF)
S5 Fig
Related to Fig 3 . The I-O function of each individual component of GC-PC synapses.
(A) Excitation G exc of AMPA slow component (left) and AMPA fast component (right) as a function of GC input at various levels of synaptic efficacy U exc (0.1-0.75) without short-term depression (STD) (top) and with STD (bottom). (B) Similar to (A) but for lower values of U exc (0.01-0.09) without short-term facilitation (STF) (top) and with STF (bottom).
(TIF)
S6 Fig
Related to Fig 7 . Synchronization and pause in the PC network under different settings.
(A) The PC network shows similar behaviors with (blue) and without (red) MLI-PC STP. (B) Similar to Fig 7 , but with an opposite pair of U values: U exc = 0.08 for GC-PC facilitation and U exc = 0.2 for MLI-PC depression, in contrast to Fig 7 . Without short-term depression in MLI-PC synapses, the inhibition is too strong to suppress PC firing, which indicates the need of a balance of excitation and inhibiting, as also revealed by Figs 9 – 11 .
(TIF)
S7 Fig
Related to Fig 2 and S6 Fig . Time courses of the excitatory conductance and short-term variable.
Time courses of PC population firing rate (red), total excitatory conductance averaged over all PCs (G exc , green) and raster plots of excitatory conductance of each PC, and short-term plasticity R variable averaged over all PCs (blue) and raster plots of R of each PC. The same pairs of U values under different settings as in S6 Fig GC-PC synapses are depressed due to STP, and R variable reduced to be close to 0 during burst inputs.
(TIF)
S8 Fig
Related to Fig 8 . PC firing rate under different burst frequencies.
PC population firing rate under bursts of 10 spikes at 100, 200, and 300 Hz in four different conditions, baseline, with MLI, with STP, and with both MLI and STP. The background Poisson stimulation frequency is 20 Hz. The PC population firing rate tends to be sustained in the baseline. Adding MLI tends to build up dynamics over stimulus, in particular for high frequencies. Adding STP tends to make dynamics transient and decay over stimulus. Using MLI and STP together makes the dynamics more transient, so that the first peak is more prominent.
(TIF)
S9 Fig
Related to Fig 8 . PC firing rate under different burst spikes.
PC population firing rate under bursts of 200 Hz with 2, 5, and 7 spikes in four conditions. The background Poisson stimulation is 20 Hz.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Alternative splicing (AS) is a frequent regulatory process of transcription in animals, plants and fungi [ 1 , 2 ]. Through differential inclusion/exculsion of exons and introns, AS produces multiple variant proteins from a single gene [ 1 , 2 ]. AS and gene duplication (GD) are important sources of genetic innovation and protein diversity [ 3 – 6 ]. Compared to GD, relatively little is known about the evolution of AS and its role in adaptation [ 3 , 7 , 8 ]. Moreover, the relationship and difference between these two evolutionary processes, AS and GD, are largely understudied [ 4 , 9 , 10 ].
Pathogens and parasites recruit diverse virulence genes to overcome their hosts and adapt to coevolution with hosts [ 11 – 14 ]. For example, to ensure successful parasitism, parasitoid wasps often inject venom to manipulate their hosts’ immunity, metabolism, development and even behavior [ 14 , 15 ]. Driven by frequent host shifts and the arms race between parasitoid wasps and their hosts, parasitoid venoms show rapid compositional turnover and high sequence evolutionary rates [ 14 , 16 – 18 ]. Evolutionary processes for parasitoid venom gene recruitment include GD [ 19 – 21 ], AS [ 22 , 23 ], lateral gene transfer [ 24 , 25 ], and single-copy gene co-option for venom functions [ 17 , 26 ].
Serpins ( ser ine p rotease in hibitors) are a widely distributed protein superfamily in all kingdoms of life [ 27 – 29 ]. Serpins contain similar structures with three β-sheets, 7–9 α-helices and an exposed reactive center loop (RCL) on the C-terminus, which determines serpin activity and specificity [ 30 , 31 ]. Most serpins are inhibitors that irreversibly inhibit target enzymes by conformational change [ 32 ], where the hinge region of RCL acts as a “spring” and inserts into β-sheets [ 33 , 34 ]. Through their inhibitory activities, serpins play central roles in proteolytic cascades, e.g., coagulation and inflammation in mammals and the prophenoloxidase (PPO) casecade and Toll pathway in insects [ 28 , 29 ].
In parasitoid wasps, serpins are a common venom component and have been reported in numerous parasitoid venoms [ 17 , 20 – 23 , 35 – 43 ]. Both AS and GD have been reported in the recruitment of serpin into parasitoid venoms [ 21 , 22 ]. Previously, we isolated a venom serpin protein from Pteromalus puparum , a generalist parasitoid wasp that parasitizes pupae of several butterfly species [ 22 ]. This venom serpin is produced by the PpSerpin-1 gene through AS and suppresses the host’s melanization immunity [ 22 ]. In contrast, extensive GD of serpin was reported in the venom apparatus of a parasitoid wasp, Microplitis mediator [ 21 ]. Therefore, parasitoid venom serpins may provide a promising model for comparative studies on the evolution and ecological adaptation of AS and GD.
Here, we analyzed the evolutionary trajectory of PpSerpin-1 and compared the different features of AS and GD in serpin evolution and their associations with the parasitic life strategy. In addition, we demonstrate that PpSerpin-1 is involved in wasp’s immune responses and has been recruited to both wasp venom and larval saliva, to regulate host immunity.
Results
AS contributes to serpin protein diversity with GD and domain duplication in insects
Utilizing the accumulated transcriptomic data, we found that the PpSerpin-1 gene has alternative splicing, with two different N-terminal and 21 C-terminal AS forms (Figs 1A and S1 ). At the N-terminus, one form was predicted to be secreted extracellularly with a signal peptide, while the other was predicted to localize intracellularly without a signal peptide ( Fig 1B ). At the C-terminus, AS occurs at the end of the serpin hinge region with an extra nucleotide G (Figs 1B and S1 ).
10.1371/journal.ppat.1011649.g001
Fig 1
Alternative splicing (AS), gene duplication (GD) and domain duplication contribute to serpin protein diversity in insects.
(a) Gene structure of PpSerpin-1 . Gray indicates constitutive exons. Blue and red indicate N-terminal and C-terminal alternative exons, respectively. (b) Protein structure of PpSerpin-1F predicted by AlphaFold2. PpSerpin-1F is the longest protein isoform of the PpSerpin-1 gene and was chosen as the representative. The gradient colors from blue to red indicate different exons. The bracket indicates the signal peptide. Black arrows indicate alternative translation start sites. The dashed circle indicates the hinge region of PpSerpin-1F. (c) Phylogenetic tree of serpin in insects. The different colors on the labels indicate different insect orders. Black arrows indicate parasitoid wasps. The red and blue dots on branch terminals indicate the presence (extracellular) and absence (intracellular) of the signal peptide, respectively. Black dots on branch terminals indicate N-terminal AS with both extracellular and intracellular isoforms. The red bars indicate the log2 transformed numbers of C-terminal AS events. The gray bars indicate the log2 transformed numbers of serpin domains within a gene. (d) Relationship between the number of C-terminal AS events and gene localization. (e) Relationship between the percentage of C-terminal alternative spliced genes and serpin gene copy number. (f) Relationship between the number of C-terminal AS events and serpin gene copy number. (g) Relationship between the number of C-terminal AS events and sperin domain number. *** p < 0.001; * p < 0.05.
To construct the phylogeny of the PpSerpin-1 gene, a homology search was performed against the insect Refseq_protein database in NCBI using the constitutive region of PpSerpin-1 (for details, see Materials and Methods ), resulting in a total of 1,230 matched genes. Phylogenetic analyses showed that the C-terminal AS of serpin genes was clustered into one clade with 731 genes but was relatively rare in the outgroup ( Fig 1C ). The following analyses were all focused on this clade with enriched C-terminal AS.
Both N-terminal and C-terminal AS are widespread in insects ( Fig 1C ). Out of 731 genes in this clade, 261 genes (35.7%, not including PpSerpin-1 ) have either N-terminal or C-terminal AS. At the N-terminus, 153 out of 731 (21.0%) had N-terminal AS, with 115 out of 153 having both extracellular and intracellular forms. These genes with both extracellular and intracellular forms at the N-terminus had more C-terminal AS than genes with only one N-terminal form, either extracellular or intracellular ( Fig 1D ; Mann–Whitney U test (MWUT); both comparisons: p < 0.001). At the C-terminus, 192 out of 731 (26.3%) genes had C-terminal AS, producing 1406 serpin proteins, with an average of 7.32 serpin proteins per gene. Notably, the number of C-terminal AS events showed rapid fluctuation, suggesting a high turnover rate of exon gain and loss ( Fig 1C ).
Together, AS, GD and domain duplication within genes contribute to serpin protein diversity. Forty-five genes have multiple serpin domains, producing 105 serpin domains with an average of 2.33 serpin domains per gene. The remaining 494 genes, which have neither C-terminal AS nor multiple serpin domains, likely originated by GD. Taken together, AS, GD and domain duplication accounted for 70.1%, 24.6% and 5.2% of the total number of 2005 serpin proteins/domains, respectively. Moreover, AS is negatively correlated with GD and domain duplication. The number of duplicated serpin gene copies in a species is inversely correlated with the percentage of C-terminal AS ( Fig 1E ; Spearman correlation; ρ = -0.58, p = 4.0e -12 ) and with the means of C-terminal variants ( Fig 1F ; Spearman correlation; ρ = -0.59, p = 1.2e -12 ). Both correlations held after phylogenetic correction (both tests: p < 0.001). In addition, AS and domain duplication within genes are mutually exclusive, as no multiple-serpin gene contains C-terminal AS ( Fig 1G ; MWUT; single- vs multiple-domain serpin: W = 19755, p = 4.8e -5 ).
AS genes show divergent sequence features from non-AS genes
Both AS and GD are major sources of protein diversity. To compare their differences in gene sequence characteristics, we divided the single-domain serpins into genes with C-terminal AS (referred to as the AS gene set) and those without (referred to as the non-AS gene set). Here, we focused on C-terminal but not N-terminal AS, as N-terminal AS often changes protein localization but not the mature serpin protein sequence.
For the AS gene set, the splicing positions of C-terminal AS occurred mainly near the end of the hinge region (GSEAAAVT in PpSerpin-1) ( Fig 2A ). Considering the last nucleotide at the end of the hinge region as position 0, 161 out of 192 (83.9%) AS genes spliced at position +1 and 19 (9.9%) at position +4. The AS gene set was more likely to splice at position +1 than the non-AS gene set ( Fig 2A ; χ 2 = 35.81, p = 0.00001). Both AS and non-AS gene sets have G|GTAAGT and TTNCAG|N sequence motifs near C-terminal splicing sites ( Fig 2B ). Position 0 (at codon position 3), +5 (at codon position 2) and +11 (at codon position 2) are more inclined to be G, T and T in the AS gene set than in the non-AS gene set, respectively ( Fig 2B ; p < 0.05).
10.1371/journal.ppat.1011649.g002
Fig 2
AS genes show divergent sequence features from non-AS genes.
(a) Distribution of splicing sites on the AS and non-AS gene sets. The AS gene set represents single-domain serpin genes with C-terminal AS, and the non-AS gene set represents single-domain serpin genes without C-terminal AS. The end of the hinge region is noted as position 0. (b) Sequence logos of AS and non-AS gene sets. The dashed lines indicate splicing positions. The underline indicates the hinge region. (c) Predicted protein structure of PpSerpin-1F with mapped conservation scores of AS and non-AS gene sets. (d) Comparisons of conservation scores between AS and non-AS genes. (e) Correlation of C-terminal AS number and probability of dl and Dif binding motif best hits. (f) Binding motif hit distribution of dl and Dif on AS and non-AS genes. + and—indicate coding and noncoding strands, respectively. *** p < 0.001; ns.: p > 0.05.
For the protein sequence, we divided the serpin proteins into constitutive (present in all isoforms) and C-terminal alternative spliced regions based on the end of the hinge region ( Fig 2B ). Conservation scores were estimated based on alignments and then mapped to PpSerpin-1F as the reference ( Fig 2C ). In the constitutive region, the AS gene set was more conserved than the non-AS gene set ( Fig 2C and 2D ; Wilcoxon signed-rank test (WSRT); V = 985, p < 2.2e -16 ). Conversely, in the alternative region, the AS gene set was more variable than the non-AS gene set ( Fig 2C and 2D ; WSRT: V = 1125, p = 1.5e -05 ). To exclude the effect of reference sequence selection, we also used non-AS genes within the clade (Dmel_SPN55B, NP_524953.1) or in the outgroup (Dmel_SPN27A, NP_001260143.1) as reference sequences, and the conclusions held regardless of reference selection (all comparisons: p <0.001).
Furthermore, we compared the regulatory sequences of the AS and non-AS gene sets. For RNA-binding motifs, no motifs were enriched in the AS gene set compared to the non-AS gene set ( p > 0.05), which may be due to the general short length of RNA-binding motifs and lack of statistical potency, as previously reported [ 44 ]. For binding motifs of transcription factors, two motifs were enriched in the AS gene set compared to non-AS. They were M02712_2.00 dl (dorsal) ( Fig 2E ; p = 0.010, enrichment ratio 8.01) and M03953_2.00 Dif (Dorsal-related immunity factor) ( p = 0.015, enrichment ratio 7.39). dl and Dif are both transcription factors involved in the Toll pathway [ 45 ]. Compared to shuffled random sequences, dl and Dif binding motifs are significantly enriched in the AS gene set (dl: p = 3.33e -7 ; Dif: p = 7.70e -6 ) but not the non-AS gene set ( p > 0.05). In addition, the probabilities of the presence of dl and Dif binding motifs are significantly correlated with the numbers of C-terminal AS ( Fig 2E ; Spearman correlation; dl: ρ = 0.48, p < 2.2e -16 , dif: ρ = 0.48, p < 2.2e -16 ). These correlations are significantly higher than correlations using shuffled sequences, which have the same lengths as the true gene sequences (both comparisons: p < 0.0001). These results suggest that genes with more C-terminal AS are more likely to contain dl and Dif binding motifs, and this is not simply due to their longer sequence lengths.
We further asked how these dl and Dif binding motifs are distributed in sequences. We defined 6 region categories according to their strands (on coding or noncoding strands) and location (in promoter, exon or intron regions). Most of the best-hits were in the intron region of the coding strand for the AS gene set, while most of the best-hits were in the exon region of the noncoding strand for the non-AS gene set ( Fig 2F ). Best-hits of dl and Dif binding motifs are more likely to locate in the intron region of the coding strand of the AS gene set than that of the non-AS gene set ( Fig 2F ; dl: χ 2 = 50.40, p < 0.00001; dif: χ 2 = 55.38, p < 0.00001) and shuffled control sequences of the AS gene set (dl: χ 2 = 10.8984, p = 0.000962; dif: χ 2 = 7.9206, p = 0.004888). In addition, when we restricted the motif search to one of the six region categories, the correlations between the probabilities of the presence of dl and Dif binding motifs and the numbers of C-terminal AS were significantly higher than correlations using shuffled sequences in the intron region of the coding strand but not in the other five region categories (dl: z = 2.7, p = 0.0065; dif: z = 4.4, p < 0.0001; all other comparisons: p > 0.05). Therefore, we conclude that dl and Dif binding motifs tend to be enriched in the intron region of the coding strand within the AS gene set.
Parasitoid wasps employ fewer AS in serpin, but PpSerpin-1 is an expansion of the AS exon number
We further investigated the potential relationship between AS and the parasitic life strategy. Higher numbers of total serpin proteins/domains per species were found in parasitoid wasps than in nonparasitoid hymenopterans ( Fig 3A ; W = 370, p = 0.0026). Compared to nonparasitoid hymenopterans, more GD ( Fig 3B ; MWUT; W = 434, p = 5.5e -6 ) and domain duplication ( S2 Fig ; MWUT; average: W = 362.5, p = 2.1e -4 ; proportion: W = 350, p = 8.1e -4 ) but less AS ( Fig 3C and 3D ; MWUT; average: W = 96, p = 0.0022; proportion: W = 96, p = 0.0017) are utilized in parasitoid wasps to produce serpin protein/domain diversity. Consistent with this, gene expansions in parasitoid wasps are generally more recent ( S3A Fig ; MWUT; W = 11429, p = 3.0e -6 ), with a higher proportion of genus-specific and family-specific expansions ( Fig 3E ; genus-specific: χ 2 = 14.09, p = 1.74e -4 ; family-specific: χ 2 = 10.49, p = 1.20e -3 ) than those in nonparasitoid hymenopterans. Conversely, exon expansions in parasitoids are generally more ancient ( S3B Fig ; MWUT; W = 16338, p = 1.25e -4 ), with a lower proportion of lineage-specific exon expansions ( Fig 3F ; χ 2 = 20.14, p < 0.00001) than those in nonparasitoid hymenopterans. Patterns are similar when comparing parasitoid wasps with all nonparasitoid insects (not limited to Hymenoptera) (all comparisons: p < 0.05).
10.1371/journal.ppat.1011649.g003
Fig 3
Parasitoid wasps employ fewer AS events in serpin, but PpSerpin-1 shows an expansion of the C-terminal AS number.
(a-f) Comparison between parasitoid wasps and nonparasitoid hymenopterans of (a) total serpin protein per species, (b) gene copy number, (c) percentage of genes with C-terminal AS, (d) average C-terminal AS number, (e) distribution of gene expansion and (f) distribution of exon expansion. (g) Expansion of the C-terminal AS number in PpSerpin-1 . Black arrows indicate parasitoid wasps. The red and blue dots on branch terminals indicate the presence (extracellular) and absence (intracellular) of the signal peptide, respectively. Black dots on branch terminals indicate N-terminal AS with both extracellular and intracellular isoforms. The gradient color on branches indicates the estimated ancestral states of C-terminal AS numbers. The red bars indicate the log2 transformed numbers of C-terminal AS events. The gray bars indicate the log2 transformed numbers of serpin domains within a gene. *** p < 0.001, ** p < 0.01.
However, PpSerpin-1 shows an expansion of AS number compared to other parasitoids. The C-terminal AS numbers of the PpSerpin-1 and Nasonia vitripennis LOC100122505 genes are increased compared to their homologous genes in the Chalcidoidea wasps, i.e., Trichogramma pretiosum , Ceratosolen solmsi marchali , and Copidosoma floridanum ( Fig 3G ). For PpSerpin-1 alternative exons, H and X2 clustered together with the XP_008201831.1-specific exon of the N . vitripennis LOC100122505 gene as the closest outgroup, suggesting PpSerpin-1 -specific exon expansion after divergence from the common ancestor with the N . vitripennis LOC100122505 gene, which is estimated to have occurred approximately 19 million years ago (MYA) [ 46 ]. Sixteen out of 21 alternative exons show clear orthologous relationships with alternative exons of the N . vitripennis LOC100122505 gene, suggesting that most alternative exons of PpSerpin-1 existed prior to the divergence between Pteromalus and Nasonia .
We then estimated the pairwise substitution rates of orthologous exons between the PpSerpin-1 and N . vitripennis LOC100122505 genes. Compared to constitutive exons, alternative exons show higher nonsynonymous substitution rates ( dN ) ( S4B Fig ; MWUT; W = 15; p = 0.013) and lower synonymous substitution rates ( dS ) ( S4C Fig ; MWUT; W = 84; p = 0.0061), resulting in higher dN / dS values ( S4A Fig ; MWUT; W = 5; p = 5.1e -4 ). These results suggest positive selection diversifying protein sequences in alternative exons with lower substitution rates in synonymous sites, which can be important in the regulation of the AS process [ 1 , 2 , 47 ].
PpSerpin-1 is involved in the wasp’s immune response and recruited into both wasp venom and larval saliva
Next, we investigated the function of isoforms of PpSerpin-1 . First, isoform-specific RT–PCR confirmed the presence of all 21 C-terminal AS forms of PpSerpin-1 ( Fig 4A ). These isoforms were more likely to be upregulated by the gram-positive bacterium Micrococcus luteus and the fungus Beauveria bassiana than by PBS or the gram-negative bacterium Escherichia coli ( Fig 4B ; WSRT; all comparisons: p < 0.001).
10.1371/journal.ppat.1011649.g004
Fig 4
PpSerpin-1 isoforms are involved in the wasp’s immune response and parasitism.
(a) Isoform-specific RT–PCR of PpSerpin-1 isoforms. (b) Expression of PpSerpin-1 isoforms in response to injection of PBS, E . coli , M . luteus or B . bassiana . Expression levels were calculated from RNA-seq data. (c) Expression of PpSerpin-1 isoforms in different developmental stages and tissues. Black dots indicate proteomic identification in P . puparum larval saliva. Black stars indicate proteomic identification in venom of P . puparum female wasps. (d) Western blot detection in venom of P . puparum , Nasonia vitripennis , Muscidifurax raptor and M . uniraptor using PpSerpin-1 antibodies. (e) Western blot detection in larval saliva of P . puparum , Nasonia vitripennis , Trichomalopsis sarcophagae , Muscidifurax raptor and M . uniraptor using PpSerpin-1 antibodies.
In our previous study, PpSerpin-1O was isolated from P . puparum venom [ 22 ]. RNA-seq data showed that multiple PpSerpin-1 isoforms are expressed in the venom gland ( Fig 4C ). We also re-analyze the published P . puparum venom proteomic data [ 37 ] and identified isoforms B, G, O and P. By Western blotting, PpSerpin-1 and its homologous proteins were detected in the venoms of P . puparum and its relative species N . vitripennis , Muscidifurax raptor , and M . uniraptor ( Fig 4D ). Additionally, by examining published venom sequences, we also identified PpSerpin-1 homologous proteins in the venom of Trichomalopsis sarcophagae and Urolepis rufipes , exhibiting 90.1% and 91.6% identity to the constitutive sequence of PpSerpin-1 protein, respectively [ 17 , 35 ].
In addition, both RT–PCR and transcriptomic results showed that some PpSerpin-1 isoforms were highly expressed in the larval stage ( Fig 4A and 4C ), particularly in the larval salivary gland ( Fig 4C ). We therefore hypothesize that some isoforms have been recruited into P . pupuarum larval saliva, which has been demonstrated to suppress host melanization [ 48 ]. Consistent with this hypothesis, PpSerpin-1 proteins were detected by Western blot using PpSerpin-1 antibodies in PBS incubated with P . puparum larvae ( S5 Fig ) and in larval saliva of P . puparum ( Fig 4E ). Protein homologs of PpSerpin-1 were also detected in the larval saliva from relatives of P . puparum , i.e., N . vitripennis , Trichomalopsis sarcophagae , Muscidifurax raptor and M . uniraptor ( Fig 4E ). Moreover, we identified 11 isoforms of PpSerpin-1 , i.e., X1, B, E, G, H, I, L, M, N, O and P, in the larval saliva of P . puparum by the protomic approach ( Fig 4C ).
PpSerpin-1 isoforms inhibit host melanization immunity
To test the function of PpSerpin-1 in larval saliva of P . puparum , we demonstrated that P . puparum larval saliva inhibited the host’s hemolymph melanization ( Fig 5A ; Dunnett’s test; p = 0.0038), and this inhibitory effect could be eliminated by the antibody of PpSerpin-1 ( Fig 5A ; Dunnett’s test; p > 0.05). To further investigate the function of PpSerpin-1 in suppressing host melanization, recombinant proteins of each isoform were produced ( S6 Fig ). PpSerpin-1 isoforms A, G, O and P significantly inhibited hemolymph melanization of P . rapae ( Fig 5B ; Dunnett’s test; all comparisons: p < 0.001) in a dose-dependent manner ( S7 Fig ) and formed complexes with host hemolymph proteins in pull-down assays ( Fig 5C ). Pull-down assays were also performed for the remaining isoforms, which were identified in wasp venom or larval saliva. PpSerpin-1B and H formed complexes with host hemolymph proteins, although they showed no inhibition of host melanization ( Fig 5C ).
10.1371/journal.ppat.1011649.g005
Fig 5
PpSerpin-1 isoforms inhibit host melanization and form complexes with host hemolymph proteins.
(a) Effect of P . puparum larval saliva and PpSerpin-1 antibody on PO activity of host P . rapae hemolymph. (b) Effect of PpSerpin-1 isoforms on PO activity of host P . rapae hemolymph. (c) Pull-down assays against host hemolymph using PpSerpin-1A, G, O, P and other isoform proteins identified in venom or larval saliva of P . puparum . (d) Host hemolymph proteinases identified in pull-down samples of PpSerpin-1 isoforms. Pieris rapae hemolymph proteinases (PrHPs) are named based on their orthology with Manduca sexta hemolymph proteinases [ 49 ]. Pull-down samples of eGFP and PpSerpin-1M were used as negative controls. *** p < 0.001, ** p < 0.01.
To identify the targets of isoforms A, B, G, H, O, and P, we cut gel slices covering the complexes ranging from ~60 to 100 kDa and identified 59, 9, 56, 29, 42, and 86 host proteins, respectively ( S3 Table ). Among them, six identified proteins were serine proteases, i.e., PrPAP1, PrPAP3, PrHP8, PrHP17, PrHP and PrHP1 ( Fig 5D ). The nomenclature of Pieris rapae hemolymph proteases (PrHPs) was based on their orthology with Manduca sexta hemolymph proteases [ 49 ] ( S8 Fig ).
Among these protease candidates, PrPAP1 is P . rapae prophenoloxidase-activating proteinase 1 and is critical for hemolymph melanization. The presence of PrPAP1 in pull-down samples of PpSerpin-1A, O and P was confirmed by Western blotting using PrPAP1 antibodies ( Fig 6A ). In vitro inhibitory assays showed that isoforms A, O and P but not G significantly inhibited the activity of recombinant PrPAP1 ( Fig 6B ; Dunnett’s test, for A, O and P: p < 0.001; for G: p = 0.94). Consistent with this, isoforms A, O, and P but not G formed SDS–PAGE stable complexes with recombinant PrPAP1 (Figs 6C , 6D and S9 ). Presumably, two identified peptides of PrPAP1 in the pull-down sample of PpSerpin-1G may be leaks from the neighboring pull-down samples of PpSerpin1-A and O. The stoichiometries of inhibition (SIs) of PrPAP1 by PpSerpin1-A and P were 13.37 and 197.33, respectively ( Fig 6E and 6F ), which were larger than the previously reported SI of 2.3 by PpSerpin-1O [ 22 ]. These results demonstrate that PpSerpin-1 isoforms A, O, and P suppress host melanization immunity by inhibiting PrPAP1.
10.1371/journal.ppat.1011649.g006
Fig 6
PpSerpin-1 isoforms A, O, and P but not G inhibit host PrPAP1.
(a) Western blot detection of PrPAP1 in pull-down samples of PpSerpin-1A, G, O and P. (b) Effect of PpSerpin-1A, G, O and P on PrPAP1 activity. *** p < 0.001. (c) SDS-stable complex formation between PpSerpin-1A and PrPAP1. (d) SDS-stable complex formation between PpSerpin-1P and PrPAP1. (e) Stoichiometry of PrPAP1 inhibition by PpSerpin-1A. (f) Stoichiometry of PrPAP1 inhibition by PpSerpin-1P.
Discussion
In this study, we report the divergent features of AS and GD in the evolution of insect serpins and their potential associations with the parasitic life strategy. We also found that PpSerpin-1 shows a number expansion of AS exons, which is involved in the wasp’s immune response and recruited to both wasp venom and larval saliva to suppress host immunity.
Not all isoforms recruited into wasp venom or larval saliva inhibit host melanization. These isoforms may serve other functions, such as regulating the host’s Toll pathway or modulating host metabolism. Proteomic identification results from the pull-down assays of these isoforms can provide hints into their targets and functions, which will be an interesting future direction. Moreover, several isoforms identified in larval saliva failed to form complexes with host hemolymph proteins. These PpSerpin-1 isoform proteins may target proteins out of the host hemolymph, e.g., in host hemocytes or gonad cells, inside host guts, or regulate proteases from wasp larval saliva. Alternatively, the inconsistency between proteomic identification and functions of PpSerpin-1 isoforms may be simply due to leaky expression from the noise of AS regulation [ 50 ]. More investigations are needed to test these hypotheses.
In addition to AS, GD and domain duplication jointly contribute to the protein diversity of insect serpins. Among them, AS and GD are the major sources of protein diversity. We observed an inverse relationship between the production of AS variants and gene family size across different species within a serpin clade. This is similar to previous reports that the production of AS variants was inversely correlated with the size of the gene family within the same species [ 9 , 10 , 51 ]. These results suggest negative regulatory mechanisms between AS and GD in the production of protein diversity.
AS is an efficient mechanism to generate significant protein diversity without requiring duplication and divergence of the entire gene [ 2 , 3 , 8 ]. In serpin AS genes, C-terminal AS tends to occur at the end of the hinge region with an extra nucleotide G. The serpin hinge region is critical to the conformational change and thus the inhibitory function of serpin [ 30 , 31 , 33 , 34 ], while the extra nucleotide G may be important in the splicing process of AS. The splicing sites in AS serpins reflect the maximum reuse of sequence. Compared to non-AS genes, AS genes are more conserved in the constitutive region but more variable in the alternative region. In addition, in the comparison of PpSerpin-1 with its orthologous gene LOC100122505 in N . vitripennis , the alternative exons show significantly higher nonsynonymous substitution rates than the constitutive exons. Collectively, these results imply that AS allows for distinct regions to experience divergent evolutionary pressures, enabling rapid protein evolution in serpin RCL while preserving most of the functional backbone under strong conservation.
On the other hand, the regulation of AS is often complicated [ 52 ]. For pairwise substitution rates between PpSerpin-1 and its orthologous gene LOC100122505 in N . vitripennis , alternative exons show significantly lower synonymous substitution rates than constitutive exons, suggesting that synonymous sites may be critical in the regulation of AS and thus under higher selective pressure in alternative exons [ 1 , 2 , 47 ]. In addition, the binding motifs of two transcription factors, dl and Dif, are enriched in the AS gene set compared to the non-AS gene set, particularly preferring the intron region of the coding strand of AS genes. Genes with more C-terminal AS are more likely to contain dl and Dif binding motifs. Accumulated evidence suggests that transcription factors can be directly or indirectly involved in the regulation of AS, although the mechanisms are not fully understood [ 53 , 54 ]. As pre-mRNA splicing mainly takes place during transcription, extensive crosstalk has been reported between these two processes [ 53 ]. One possible explanation could be that some expanded serpin isoforms are involved in the Toll pathway. The expression of these isoforms might be regulated by Toll-pathway-related dl and Dif, probably through binding to the intron regions of serpin gene pre-mRNA.
To manipulate their hosts, parasitoid wasps often recruit effector genes, e.g., venom or larval salivary genes [ 14 , 15 ]. Serpin, a common component of parasitoid venom [ 17 , 20 – 23 , 35 – 43 ], has also been reported in teratocytes of parasitoid wasps for host regulation [ 55 , 56 ]. Here, we further extend serpin recruitment to wasp larval saliva. A previous study also reported that P . puparum larval saliva inhibits host immunity [ 48 ]. In light of this, the higher numbers of serpin protein/domains in parasitoid wasps may be related to the recruitment of host effector genes. Another example of an AS venom gene is LOC122512947 from Leptopilina heterotom with 2 N-terminal and 20 C-terminal AS isoforms [ 36 ]. In addition, Leptopilina boulardi LbSPNy (ACQ83466.1) was reported as a venom non-AS gene that inhibits host melanization [ 43 ]. LbSPNy forms a Leptopilina -specific non-AS gene cluster with LOC122509667, 122505241, 12200434, 122502279, 122503460, and 122502269 of L . heterotoma , suggesting that LbSPNy likely originated from GD. Extensive GD of the serpin gene was also reported in the venom of M . mediator [ 21 ]. Our findings further show that the increased number of serpins in parasitoids results from more GD but less AS. This is consistent with the general observation that venom genes in parasitoid wasps and other venomous animals are more likely to originate from GD than AS [ 14 , 57 ].
In PpSerpin-1 , we observed a number expansion of alternative exons compared to other parasitoid serpin genes. Reconstructing the accurate evolutionary history of these exons is challenging, particularly for ancient expanded exons, due to their accelerated protein evolution and short sequences (44~50 amino acids for PpSerpin-1 alternative regions). Phylogenetic tree analysis revealed that alternative exons of PpSerpin-1H and X2 form a gene-specific cluster, suggesting a recent Pteromalus -specific exon duplication. However, the majority (16 out of 21) of PpSerpin-1 alternative exons show clear orthologous relationships with those of the N . vitripennis LOC100122505 gene. One possible explanation is that these exons underwent lineage-specific exon duplications before the divergence of Pteromalus and Nasonia (~ 19 MYA) [ 46 ] and were retained in both species, probably serving conserved and essential functions. The expanded alternative exons in PpSerpin-1 may contribute to its ecological adaptation. Consistent with this hypothesis, we found that several isoforms of PpSerpin-1 are involved in the wasp’s immune response, have been recruited to both wasp venom and larval saliva, and suppress host immunity.
However, no obvious expressional specialization, especially to venom glands, was observed for PpSerpin-1 isoforms. We speculate that accurate expressional regulation is an obstacle to the recruitment of host effector genes by AS, particularly that PpSerpin-1 is likely involved in the regulation of wasp self-immunity. Several PpSerpin-1 isoforms can be upregulated by M . luteus and B . bassiana , suggesting that PpSerpin-1 may be involved in the wasp’s Toll pathway, which is preferentially activated by gram-positive bacteria and fungi in Drosophila [ 45 ]. In AS genes, differential expressional regulation between self and host manipulation functions may be difficult to evolve. On the other hand, expressional divergence of duplicated genes often occurs after location segregation of gene copies [ 58 ], presumably due to changes in cis-regulatory regions. The positional linkage of alternative exons makes the expressional regulation of AS more challenging and requires more complicated regulatory mechanisms. In particular, high expressional specialization should be required to avoid self-harm for toxic virulent genes. This may explain why AS is less employed than GD in the recruitment of parasitoid host effector genes.
In summary, we show that both AS and GD contribute to the evolution of insect serpin with differential features. We report that a parasitoid serpin gene has evolved through exon number expansion of AS and show its involvement in wasp’s immunity and recruitment into the wasp’s venom and larval saliva to manipulate host immunity.
Materials and methods
Insect rearing
Laboratory cultures of P . puparum , N . vitripennis , T . sarcophagae , M . raptor and M . uniraptor were maintained in Drosophila tubes at 25°C with a photoperiod of 14:10 h (light:dark) as previously described [ 17 , 37 ]. Pupae of Pieris rapae were used as hosts for P . puparum , and pupae of Sarchophaga bullata were used as hosts for N . vitripennis , T . sarcophagae , M . raptor and M . uniraptor . Once emerged, P . puparum adult wasps were fed ad libitum with 20% (v/v) honey solution to lengthen their life span.
Alternative isoform identification for the PpSerpin-1 gene
Using accumulated transcriptomic data [ 46 ], sequencing reads were filtered by Trimmomatic v0.38 [ 59 ], mapped to the P . puparum genome by TopHat v2.0.12 [ 60 ], and assembled into transcripts using Cufflinks v2.2.1 [ 60 ]. For exhaustive identification of PpSerpin-1 isoforms, we further manually identified isoforms using IGV browser v2.3.91 based on mapped reads [ 61 ]. The expression levels of PpSerpin-1 isoforms were estimated using Cufflink v2.2.1 [ 60 ].
Protein structure prediction
The protein structure of PpSerpin-1F was modeled using AlphaFold2 with default settings [ 62 , 63 ]. No structural template was used. All five AlphaFold2 models were tested. Model 4 was selected as the best prediction based on its highest pLDDT score of 88.7. Alignemnts for protein modeling were generated through MMSeqs2 [ 64 ] against the UniRef+Environmental database.
Sequence fetch and feature analyses
BLASTP [ 65 ] was performed using the constitutive protein sequence of the PpSerpin-1 gene against the Refseq_protein database on NCBI (accessed at Dec 2021). The organism was limited to “Insecta (taxid:50557)” with a maximum target sequence of 5000 and an e-value of 1e -5 . One species per genus was selected as a representative to reduce sampling bias. GenBank files were retrieved for each gene using EFetch v0.2.2 ( https://dataguide.nlm.nih.gov/edirect/efetch.html ). Isoform sequences and splicing positions were extracted from GenBank files for each gene using homemade python scripts. For each gene, the number of N-terminal AS events was approximately estimated by counting the number of different sequences in the first 20 aa. Similarly, the number of C-terminal AS events was approximately estimated by counting the number of different sequences in the last 50 aa. Signal peptides were predicted using SignalP v5.0 [ 66 ]. Serpin domains were annotated using the NCBI Batch Web CD-Search Tool ( https://www.ncbi.nlm.nih.gov/Structure/bwrpsb/bwrpsb.cgi ) against the “CDD 58235 PSSMs” database [ 67 ] (accessed at Jan, 2022).
Phylogenetic analyses
For gene phylogeny, the longest protein isoform was selected for each gene. Proteins were aligned using Mafft v7.310 [ 68 ]. The gene tree was constructed using IQTree v2.2.0 with an ultra bootstrap of 1000 [ 69 ]. The best-fit model (LG+R10) was automatically selected by built-in ModelFinder [ 70 ] in IQTree. For exon phylogeny, annotated serpin domains were extracted with a C-terminal extending 30 aa or to the end of the sequences if less than 30 aa. For genes with C-terminal AS, only nonredundant C-terminal sequences were included. Serpin domain sequences were aligned using Mafft v7.310 [ 68 ]. The tree was constructed using IQTree v2.2.0 [ 69 ]. To reduce sampling bias, one species per genus was selected as a representative for visualization and subsequent statistical analyses. Ancestral character estimation was performed for the C-terminal AS number of each internal node using the fastAnc function in the R package “phytools” v1.2.0 [ 71 ]. Trees were pruned by Newick Utilities v1.6 [ 72 ] and visualized on iTOL [ 73 ].
Expansion level analyses
For every pair of genes from the same species in the gene phylogeny, two genes are defined as species-specific expansion if all genes in the clade of the common ancestor of these two genes belong to the same species. If all genes in the clade of the common ancestor of these two genes from the same species belong to the same family, two genes are defined as family-specific expansion, and so on. Similarly, for every two alternative C-terminal exons from the same gene in the exon phylogeny, two exons are defined as gene-specific expansion if all exons in the clade of the common ancestor of these two exons belong to the same gene, and so on. Expansion levels were determined by homemade R scripts.
WebLogo
Consensus sequence logos were created using WebLogo v3.7.9 [ 74 ]. For protein alignments, conservation scores were estimated for each site using WebLogo v3.7.9 [ 74 ]. Conservation scores of the corresponding positions in the reference sequence were extracted for comparisons. For the constitutive region of AS genes, the longest protein isoforms of each gene were selected as representatives. For the C-terminal alternative region of AS genes, all nonredundant sequences (based on the last 50 aa at the C-terminus) were used. Conservation scores were mapped to the PpSerpin-1F structure using PyMOL v2.5.0 ( https://pymol.org ).
dN and dS estimation
For each orthologous exon pair between PpSerpin-1 and LOC100122505 (the ortholog of PpSerpin-1 in N . vitripennis ), protein sequences were aligned using Mafft v7.310 [ 68 ] and then reverse translated to codons using PAL2NAL v14 [ 75 ]. If a codon crossed the boundary of two exons, the entire codon was counted in the exon that contained more nucleotide bases of that codon. Pairwise dN and dS values were estimated using PAML v4.9j [ 76 ].
Motif scanning and enrichment analysis
Motif enrichment analyses were performed using SEA (Simple Enrichment Analysis) in MEME v5.5.0 [ 77 , 78 ]. Gene sequences with C-terminal AS were set as primary sequences, and shuffled sequences or gene sequences without C-terminal AS were set as control sequences. Drosophila melanogaster “CIS-BP 2.00 single species DNA” or “CISBP-RNA single species RNA” was set as the motif database. Fisher’s exact tests were performed if the primary and control sequences had the same average length (within 0.01%); otherwise, binomial tests were performed [ 78 ]. Sequences of pomotor, exon and intron regions were extracted from GenBank files. Promoter regions were approximated by using the upstream sequences of the translation start sites. Exon regions were approximated using coding sequences by masking other regions using “N”. For binding motifs of dl (M02712_2.00) and Dif (M03953_2.00), hits were detected with a p value threshold of 1 using FIMO in MEME [ 77 ]. Shuffled sequences were created by fasta-shuffle-letters in MEME [ 77 ].
Specific RT–PCR
Ptermoalus puparum embryos (<10 hr after parasitism), larvae (combined 2–3 day larvae after parasitism), yellow pupae (mixed with female and male pupae), adult females and males (combined 1–5 day adults after emergence) were collected and rinsed with PBS. Total RNA was extracted using TRIzol reagent (Invitrogen, USA) according to the manufacturer’s protocol and then reverse transcribed using TransScript One-step gDNA Removal and cDNA Synthesis SuperMix (TransGen, Beijing, China) with random primers. Isoform-specific primers were designed to span constitutive exon 7 and alternative exon 8 using PerlPrimer v1.1.21 [ 79 ] and are listed in S1 Table . PCR was performed using TransTaq HiFi DNA Polymerase (TransGen, Beijing, China) with 20–35 cycles of amplification. PCR products were analyzed by electrophoresis on a 1% (g/mL) agarose gel and confirmed by Sanger sequencing (Sangon Biotech, Shanghai, China).
Microbial stimulus of P . puparum
Freshly cultured M . luteus , E . coli , and B . bassiana were harvested by centrifugation and subsequently washed with sterile PBS three times. These microbes were suspended in PBS at a density of 5×10 7 cells/mL. Two-day old P . puparum female adults were selected and anesthetized using carbon dioxide. A sterilized acupuncture needle was immersed in the suspension containing either bacteria or fungi and then inserted into the inter-segmental space of the parasitoid wasp’s abdomen. Wasps that were pricked with sterile PBS were used as a control group. Female wasps were then collected at 6, 24, and 48 h after the microbial stimulus.
Wasp larval saliva collection
Saliva of wasp larvae was collected as previously described with minor modifications [ 48 ]. Briefly, wasp larvae were collected by opening host pupae 3 days after parasitism. Larvae were rinsed with PBS and electrically stimulated using an acupuncture device (Hwato, China) with a frequency of 50 Hz and a current of 2.5 mA. Secreted saliva drops at larval mothparts were transferred into PBS using pipette tips. The protein concentration was determined using a Pierce BCA Protein Assay Kit (Thermo Scientific, USA).
LC–MS/MS
Fifty micrograms of protein from wasp larval saliva was digested by trypsin using the filter-aided sample preparation (FASP) method. After desalting by a C18 cartridge, the digestion product was lyophilized and redissolved in 40 μl of 0.1% formic acid solution. LC–MS/MS was conducted on an Easy nLC HPLC system (Thermo Scientific, USA) with a flow rate of 300 nl/min followed by Q-Exactive (Thermo Finnigan, USA). The sample was loaded on a Thermo Scientific EASY column (5 μm, 2 cm × 100 μm, C18) and then separated on another Thermo Scientific EASY column (3 μm, 75 μm × 100 mm, C18). Buffer A was water with 0.1% formic acid, and buffer B was 84% acetonitrile with 0.1% formic acid. The columns were first equilibrated with 95% buffer A, then from 0% to 35% buffer B in 50 min, from 35% to 100% buffer B in 5 min, and finally 100% buffer B for 5 min. The charge-to-mass ratios of peptides and fractions of peptides were collected 20 times after every full scan. The resulting MS/MS spectra were searched against the P . puparum genome database using Mascot 2.2 in Proteome Discoverver. “Carbamidomethyl (C)” was set as a fixed modification. “Oxidation (M)” and “Acetyl (Protein N-term)” were set as variable modifications. The maximum number of missed cleavages was set as 2. FDR ≤ 0.01 was set to filter the protein identification. The same software and parameters were used for the reanalysis of P . puparum venom proteomic data [ 37 ]. This part was conducted by Shanghai Applied Protein Technology Co., Ltd. (Shanghai, China).
Western blot
Protein was separated by electrophoresis on 12% SDS–PAGE and transferred to a PVDF (polyvinylidene difluoride) membrane at 100 mA for 2 h using a Mini-PROTEAN Tetra system (Bio-Rad, USA). For detection of PpSerpin-1 isoform proteins or PrPAP1, antibodies against PrPAP1 and PpSerpin-1O [ 22 ] (diluted 1:1000) were used as primary antibodies, followed by HRP (horseradish peroxidase)-conjugated goat anti-rabbit IgG antibody (Sigma Aldrich, Germany; diluted 1:5000) as the secondary antibody. For detection of His-tagged proteins, THE His Tag mouse antibody (GenScript, Nanjin, China; diluted 1: 2000) was used as the primary antibody, followed by goat anti-mouse IgG antibody-HRP conjugate (Sigma Aldrich, Germany; diluted 1: 5000) as the secondary antibody. The membranes were detected using Pierce ECL Western Blotting Substrate ECL (Thermo Fisher, USA) and imaged by the Chemi Doc-It 600 Imaging System (UVP, Cambridge, UK).
Recombinant protein expression and purification
For recombinant expression of PpSerpin-1 isoforms, constitutive fragments without signal peptides and alternative fragments were separately amplified using TransTaq HiFi DNA Polymerase (TransGen, Beijing, China) and cloned into the pET-28a vector using the ClonExpress MultiS One Step Cloning Kit (Vazyme, Nanjing, China). Primers were designed using PerlPrimer V1.1.21 and are listed in S2 Table . The linear vector pET-28a was generated by digestion with BamHI and XhoI (TaKaRa, Dalian, China). Recombinant plasmids were then transfected into E . coli BL21(DE3) Chemically Competent Cell (TransGen Biotech, Beijing, China) and confirmed by Sanger sequencing (Sangon Biotech, Shanghai, China). E . coli cells were grown in autoinduction medium [ 80 ] containing 100 μg/μl kanamycin at 300 rpm and 20°C for 48 h and then harvested by centrifugation at 12000 × g for 20 min. Recombinant protein was extracted using BugBuster Protein Extraction Reagent (Thermo Scientific, USA) and purified using cOmplete His-Tag Purification Resin (Roche, Switzerland) and His-Bind Purification kit (Novagen, USA) according to the manufacturer’s protocol. The concentration of purified protein was determined using a Pierce BCA Protein Assay Kit (Thermo Scientific, USA).
Phenoloxidase activity assay
Plasma was harvested by cutting the hind legs of P . rapae larvae with scissors and diluted four times into ice-cold TBS buffer (20 mM Tris, 150 mM NaCl, pH = 7.6). Cell-free hemolymph was obtained by centrifugation at 4°C and 3000 × g for 10 min to remove hemocytes. To screen for the inhibitory activities of PpSerpin-1 isoforms on host melanization, 5 μl of recombinant protein (0.2 μg/μl) was mixed with 10 μl of diluted P . rapae hemolymph in a 384-well plate. For each sample, 5 μl of elicitor (0.1 μg/μl M . luteus ) and 5 μl of substrate solution (50 mM L-Dopa in PBS, pH = 7.5) were first mixed and added to another 384-well plate, which was fixed upside down on the sample plate. By centrifuging these two oppositely fixed 384-well plates, the PPO (prophenoloxidase) cascade was activated in each well simultaneously. Plates were measured at A470 and 25°C every 5 min for 2 h using a Varioskan Flash multimode reader (Thermo Scientific, USA). For dose-dependent assays, 5 μl of recombinant protein was mixed with 15 μl of diluted P . rapae hemolymph and 5 μl of elicitor (0.1 μg/μl M . luteus ). After incubation at 25°C for ~10 min, 800 μl of substrate solution (20 mM Dopa in PBS, pH = 6.5) was added. Samples (200 μl) were monitored at A470 in 96-well plates for 20 min using a Varioskan Flash multimode reader (Thermo Scientific, USA). One unit of PO activity was defined as 0.001 △A470/min.
PrPAP1 amidase activity assay
Recombinant PrPAP1Xa was secretively expressed using the Bac-to-Bac Baculovirus Expression System (Invitrogen, USA) as previously described [ 22 ]. Unexpectedly, PrPAP1Xa was activated for unknown reasons. For inhibitory assays, recombinant proteins of PpSerpin-1 isoforms were incubated with activated PrPAP1 at room temperature for 10 min. After adding 200 μl of 50 μM acetyl-Ile-Glu-Ala-Arg-p-nitroanilide (IEARpNA) in TBS (100 mM Tris, 100 mM NaCl, 5 mM CaCl 2 , pH = 8.0), residual amidase activities were measured at A405 for 20 min using a Varioskan Flash multimode reader (Thermo Scientific, USA). One unit of amidase activity was defined as 0.00001 △A405/min.
Pull-down assay
Recombinant protein of PpSerpin-1 isoforms (10 μg) was mixed with 1 ml diluted P . rapae hemolymph, 50 μl saturated PTU and 100 μl M . luteus (1 μg/μl) and incubated on a rotator overnight at 4°C. After centrifugation at 12000 g and 4°C for 20 min, the supernatant was further incubated with 25 μl of cOmplete His-tag purification resins (Roche, Switzerland) at 4°C for 2 h. After washing three times with 300 μl of washing buffer (1 M NaCl, 120 mM imidazole, 40 mM Tris-HCl, pH 7.9), proteins were eluted with 50 μl of elution buffer (1 M imidazole, 0.5 M NaCl, 20 mM Tris-HCl, pH 7.9) and subjected to SDS–PAGE followed by Lumitein Protein Gel Stain (Biotium, Hayward, CA, USA) and immunoblotting. To identify potential targets in pull-down complexes, gels were cut between 60 and 100 kDa. Gel slices were then in-gel digested by trypsin at 37°C for 20 h. After desalting and lyophilization, the enzymatic product was redissolved in 0.1% formic acid solution and subjected to LC–MS/MS. The parameters used were the same as for the above, except specifically mentioned. The gradient was 1 h. Annotated proteins from Pieris rapae genome assembly GCF_001856805.1 were set as the search database. “Oxidation (M)” was set as a variable modification. Protein identification was filtered by proteins with at least two peptides identified. Gel digestion and proteomic identification were conducted by Shanghai Applied Protein Technology Co., Ltd. (Shanghai, China).
Statistics
All statistical analyses were performed in R v4.1.2. Differences between Spearman correlations were tested using the Hittner2003 method in the R package “cocor” v1.1.4 [ 81 ]. To avoid biases that may arise from data non-independence due to shared evolutionary history, phylogenetic correction was performed using the independent contrasts method. Phylogenetic correction allows us to account for the evolutionary relationships among species and obtain more accurate estimates of trait correlations. For independent contrasts, the phylogenetic species tree was generated by phyloT v2 ( https://phylot.biobyte.de/ ). The branch length was estimated by compute.brlen in the R package “ape” v5.6.2 [ 82 ] using Grafen’s (1989) methods [ 83 ]. Independent contrasts were conducted using the “pic” function in the R package “ape” [ 82 ]. Correlations through origins were estimated for independent contrasts using the R package “picante” v1.8.2 [ 84 ].
Supporting information
S1 Data
Excel spreadsheet containing the underlying numerical data for Figs 1 – 6 in separate sheets.
(XLSX)
S1 Table
Primers used for isoform-specific RT-PCR.
(XLSX)
S2 Table
Primers used for recombinant expression vector construction.
(XLSX)
S3 Table
Proteomic identified proteins in pull-down gel slices.
(XLSX)
S1 Fig
Protein and codon alignments at the C-terminal alternative splicing site in the PpSerpin-1 gene.
The black arrow indicates the splicing site. Curly brackets indicate the hinge region.
(TIF)
S2 Fig
Comparison of multiple-domain serpin between parasitoid wasps and non-parasitoid hymenopterans.
SPN, serpin.
(TIF)
S3 Fig
Comparison of gene and exon expansion between parasitoid wasps and non-parasitoid hymenopterans.
Expansion levels were compared using a ranking method. 0 indicates gene-specific expansion; 1 indicates genus-specific expansion; 2 indicates family-specific expansion; 3 indicates order-specific expansion; 4 indicates not specific. Smaller rank numbers mean younger expansions.
(TIF)
S4 Fig
Comparison of dN/dS between shared and alternative exons.
Constitutive exons include exons 2–7, and alternative exons include 8B, 8C, 8D, 8E, 8F, 8G, 8H, 8I, 8J, 8K, 8X4, 8X5, 8L, 8N, 8O, and 8P.
(TIF)
S5 Fig
PpSerpin-1 protein secretion into PBS by P . puparum larvae.
(a) Western blot detection of PpSerpin-1 proteins in PBS incubated with P . puparum larvae. (b) Fluorescence intensity of larvae after 8 h of incubation in PBS. CellTox is a dye that detects cell integrity and can enter ruptured cells and bind to DNA to emit fluorescence. The stronger the fluorescence intensity, the more ruptured cells in the larvae. The larvae remained viable after incubation 8 h, and there was no significant difference compared to the spontaneous fluorescence without CellTox added. Larvae were heat-killed at 100°C for 10 min and used as the positive control. (c) Representative fluorescence images of P . puparum larvae after 8 h of incubation in PBS. * p < 0.05; ns.: p > 0.05.
(TIF)
S6 Fig
SDS–PAGE analyses of purified recombinant PpSerpin-1 isoform proteins.
Some isoform proteins show two bands, which might be truncated by-products during the expression process.
(TIF)
S7 Fig
Dose-dependent suppression of host PO activity by PpSerpin-1A, G, O and P isoform proteins.
(TIF)
S8 Fig
Phylogenetic tree of PrHPs in pull-down samples of PpSerpin-1 isoform proteins.
Sequences of MsHP and MsPAP were retrieved from [ 49 ] paper. Sequence, alignment and tree files can be downloaded at FigShare ( https://doi.org/10.6084/m9.figshare.21545598.v1 ).
(TIF)
S9 Fig
Interaction of Pr PAP1 with eGFP, G and O.
Black arrows indicate complexes formed for PpSerpin1 isoforms with PrPAP1.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Morphemes constitute the smallest units of meaning in an oral or written language, and in the case of English, include bases (e.g., <help>) that carry the main kernel of meaning in a word, and affixes (e.g. the prefix <un-> or the suffix <-ful>) that modify the meaning of the base (e.g., <unhelpful>). There is growing interest in the role that morphological knowledge plays in literacy acquisition in English. This interest is motivated by two general considerations. Firstly, English is a morphophonemic language with an orthographic system that evolved to represent both phonology and morphology. Indeed, as we describe below, the English spelling system prioritizes the consistent spellings of morphemes over phonemes. This provides a rationale for teaching morphology along with phonology and how they interrelate. Second, empirically, there is now ample evidence that morphological knowledge in young children predicts literacy outcomes [ 1 ], and growing evidence that morphological instruction improves literacy outcomes, especially for younger and less-abled readers [ 2 , 3 ]. Despite these theoretical and empirical considerations there is almost no guidance for how to teach morphology in the classroom, nor research into how to best teach English morphology to children in order to improve reading, writing, and vocabulary [ 4 , 5 ].
Here we compare two approaches of teaching morphology in the context of a memory experiment carried out with undergraduate students. The first was designed to mirror an affix-centric approach that focuses on teaching affixes one at a time. For example, students might be presented with a set of unrelated base words that all share a given affix in order to highlight the spelling, meaning, and grammatical role of the affix (e.g., presenting a list of words that all share the <-ing> affix such as <doing>, <going>, <talking>, <walking>, <playing>, etc.). This approach appears to be the most common guidance given to teachers [ 6 ]. The second was designed to mirror the base-centric approach used in Structured Word Inquiry or SWI [ 7 , 8 ]. SWI teaches words in the context of morphological families composed of words that share a common base (e.g., <do>, <doing>, <redo> and <done>). A common tool of SWI is the morphological matrix, as depicted in Fig 1 . We show that memory for morphologically complex words in a free recall task is better in both the affix- and base-centric morphological matrix conditions compared to a control condition that does not highlight the morphological composition of words, and importantly, demonstrate that studying words in a base-centric morphological matrix improves memory further still. We attribute these findings to the fact that memory, learning, and reasoning is best when information is encoded in an elaborative and organized manner [ 9 , 10 ].
10.1371/journal.pone.0262260.g001
Fig 1
An example of a morphological matrix.
Our use of undergraduate students to assess the promise of different forms of instruction with children is not uncommon [ 11 , 12 ]. The advantage of this approach is that it is more straightforward to assess the impact of different learning conditions in tightly controlled conditions. The obvious disadvantage, however, is that it is not certain that the results will extend to younger children in a classroom setting. Still, clear-cut findings obtained with adults can provide a strong motivation to explore similar interventions in children. And in this case, we take our findings to motivate further research into the morphological matrix as a teaching tool for reading instruction. Our findings may also help explain the recent disappointing results obtained in a large-scale study carried out with grade 5 students that adopted an affix-centric approach to morphological instruction [ 13 ].
The paper is organized as follows. First, we review the theoretical and empirical motivation for teaching morphology in order to improve various literacy outcomes including word naming, spelling, and vocabulary. Second, we briefly review different forms of morphological interventions, including affix-centric and various base-centric approaches that have been employed in intervention studies. Although there is evidence that morphological instruction is beneficial, there is no research to date directly comparing the efficacy of the different approaches. Third, we report two experiments that compare an affix-centric to a base-centric approach that uses morphological matrices. Finally, in the General Discussion, we consider the implications of our results for literacy instruction.
The theoretical motivation for teaching morphology
It is widely claimed that the English spelling system follows an “alphabetic principle” according to which letters represent meaning via sounds, and where the grapheme-to-phoneme system is riddled with irregularities [ 14 ]. But in fact, English has a morphophonemic spelling system that evolved to jointly reflect meaning (through influences of morphology and etymology) and pronunciation (through representations of phonemes). This provides a key motivation for teaching children the important role that morphology plays in explaining not only the meaning-spelling connections between words, but also the spelling-pronunciation connections informed by morphology. Since the pronunciation of morphemes in English shifts across words, it is not possible to have a one-to-one relationship between phonemes and graphemes (e.g., the grapheme <o> maps onto different phonemes in the words <do> and <does>, but maintains consistent spelling of the base morpheme <do>). Writing systems like English with many ways of spelling the same phonemes are described as having a “deep” orthography. This contrasts with languages with “shallow” orthographies such as Italian and Spanish where grapheme-phoneme correspondences are largely one-to-one (highly consistent) [ 15 ].
To illustrate the morphophonemic organization of English spellings, consider the morphological families formed from the bases <act>, <do> and <go> in Fig 1 . The important point to note is that the spelling of the base remains consistent across all members of the morphological families even when changes in pronunciation are observed (e.g., actor vs. action ; do vs. does ; go vs. gone ). Or consider the <-ed> suffix. In a word like “painted” the pronunciation of the <-ed> suffix is syllabic, and thus needs the <e> grapheme to represent the vowel phoneme of that syllable. However, the <-ed> suffix is not syllabic in “jumped” or “played” and countless other words. In such words there is no pronunciation associated with that <e> grapheme. Technically, the <e> grapheme in the <-ed> suffix of these words represents a “zero allophone” /∅/ of the /ə/ phoneme that is pronounced in words like “painted”. This “zeroing” of a phoneme frequently occurs and it makes sense in the context of a morphemic family. For example, the <t> grapheme represents the /t/ phoneme in the “print” or “printing”, but it is zeroed when the <-s> suffix is added to form the word “prints” or “footprints”. The <t> grapheme in the base <rupt> (meaning “break”) can represent the /t/ phoneme in “bankrupt” or “disrupt”, the /ʃ/ in “disruption”, the /tʃ/ in “rupture” and it is zeroed when we add the <-cy> suffix to produce the word “bankruptcy.” It is important to emphasize that the above examples of morphological constraints on spelling are not cherry-picked; they are the norm.
It should also be noted that there are two types of affixes, namely, derivational and inflectional. When derivational affixes are added to a base or a complex word, the result is considered a “new word,” or technically, a new lexical item (e.g., “help” vs “helpful”). By contrast, adding inflectional suffixes to a word results in what is considered different forms of the same word (e.g., “help” vs. “helps”). Technically, words like “help” and “helps” are considered the same lexical item. Inflectional suffixes mark grammatical relationships such as plural, past tense and possession, but unlike derivational suffixes, cannot alter the grammatical class of the words they construct [ 16 ]. This means that derivational affixes can dramatically change the meaning of the base (“cover” vs. “discover”) whereas the meaning shifts are minimal for inflectional affixes (e.g., “cover” vs. “covered”). Nevertheless, it is important to note that the phonological shifts that occur with derivations (“actor” vs. “action”) apply equally to inflections (“do” vs. “does”), as do spelling shifts. For example, the inflectional suffix <-ing> in “hopping” and “hoping” forces spelling changes in the base a result of suffixing conventions: i.e., doubling the <p> in the spelling <hopping> (hop(p) + ing—> hopping), and dropping the <e> in the spelling <hoping> (hope/ + ing—> hoping). An appreciation of the underlying logic of the English writing system has long motivated the hypothesis that reading instruction should emphasize both the phonological and meaningful regularities in word spellings [ 17 – 19 ]. Furthermore, as highlighted by Bowers and Bowers [ 8 ], this hypothesis is consistent with the finding that learning and memory is best when information can be encoded in a meaningful and organized manner [ 9 , 10 ]. For example, Bower et al. [ 9 ] carried out a memory experiment in which words were organized within a hierarchy that highlighted the meaningful relations amongst the words, as depicted in Fig 2 . Memory was much better in this condition compared to a condition that did not highlight these relations. In a similar way, morphological matrices (as depicted in Figs 1 and 3 ) highlight the meaning (and spelling) relations between words, and importantly, most words in English are members of morphological families. This raises the possibility that morphological matrices might be an effective way to learn the spelling, vocabulary, and pronunciation of most English words as well.
10.1371/journal.pone.0262260.g002
Fig 2
Hierarchy highlighting meaningful relations amongst words.
10.1371/journal.pone.0262260.g003
Fig 3
Morphological matrices highlighting membership in morphological families.
The empirical motivation for morphological instruction
The linguistic analysis of the English writing system shows why morphology (and morphological matrices) might be relevant to literacy instruction, and memory research provides some theoretical motivation for emphasizing the meaningful organization of words. But is there any empirical evidence that morphological knowledge and morphological instruction benefits literacy?
Research has shown that prior to starting school, children already possess rich morphological knowledge as reflected in their language production and comprehension [ 20 ], and that untaught morphological knowledge is a good predictor of word naming, spelling, and vocabulary. For example, Kirby, Geier, and Deacon [ 20 ] found that morphological knowledge uniquely predicted reading speed, accuracy, decoding and comprehension in a sample of 182 Grade 3 students even after controlling for various other aspects like verbal and nonverbal intelligence, phonological awareness, naming speed and orthographic processing. Similarly, Deacon, Kirby and Casselman-Bell [ 21 ] found that morphological knowledge measured in Grade 2 was able to explain approximately 8% of the variance of Grade 4 spelling even after controlling for other aspects like verbal and nonverbal intelligence, phonological awareness, verbal short-term memory, and rapid automized naming. In the case of vocabulary, McBride-Chang and colleagues [ 22 ] provided evidence that morphological knowledge was able to predict vocabulary learning in English-speaking kindergarten and Grade 2 students. They found that morphological measures uniquely contributed to an additional 10% of the variance in vocabulary in either group upon controlling statistically for other measures like those of reading, phonological awareness and naming speed. For a recent review of the impact of morphological awareness on literacy see Duncan [ 1 ].
With regards to instruction, it has long been argued that morphology should play a more prominent role [ 17 , 18 , 22 – 24 ], but only in the past decade have a sufficient number of morphological intervention studies been carried out to support meta-analyses and systematic reviews. Goodwin and Ahn [ 2 ] reported a meta-analysis that showed moderate and significant improvements on a range of outcomes for children with learning disabilities, and Goodwin and Ahn [ 3 ] reported a meta-analysis that showed moderate and significant improvements on a range of outcomes for children from the general population. Similar conclusions were made in systematic reviews [ 4 , 25 , 26 ]. Most recently, in a meta-analysis of spelling interventions with children with dyslexia, morphological instruction was found to be effective, and just as effective as phonics overall in the early years of instruction [ 27 ]. Indeed, they found that morphological and orthographic instruction became more effective for children with spelling difficulties, whereas phonics instruction became less effective for struggling spellers, consistent with previous research on phonics, as summarized by Bowers [ 28 ].
How is and how should morphology be taught?
Given these compelling reasons to include morphology as an ingredient to literacy instruction, it is striking how rarely morphology plays a role in reading, spelling, and vocabulary instruction in the classroom. This is highlighted by the fact that teachers know very little about morphology [ 29 , 30 ]. To the extent that morphology is taught in schools there appears to be an emphasis on affix centric approaches, although the teaching guidelines we can find are all extremely vague. For example, in the US, “The Common Core State Standards for English Language Arts & Literacy in History/Social Studies, Science, and Technical Subjects” [ 31 ] mentioned morphology twice, writing that Grade 4 and 5 students should “use combined knowledge of all letter-sound correspondences, syllabication patterns, and morphology (e.g., roots and affixes) to read accurately unfamiliar multisyllabic words in context and out of context” (p 17), and with regards to vocabulary instruction writes “Use the most frequently occurring inflections and affixes (e.g., -ed, -s, re-, un-, pre-, -ful, -less) as a clue to the meaning of an unknown word.” (p. 27). Similarly, in the UK National Curriculum for children in Key Stage 1–2 (ages 5–11), the main information regarding morphological instruction is found in tables in the document that provide some general guidance for affix centric instruction (e.g., page 13, 46–52). There is also brief mention of a more base-centric approach, with the following included in another table describing instruction for Year 3 students: “Word families based on common words, showing how words are related in form and leaning [for example, solve, solution, solver, dissolve, insoluble]” (p. 66). An affix-centric approach is likewise encouraged in countries such as Hong Kong, where teachers are advised to focus on affixes from Secondary levels 1–3 [ 32 ], and in Singapore, where it is recommended that the concept of affixes be introduced from Primary 1 onwards [ 33 ]. In all cases, English teachers are left to their own devices to teach morphology as they see fit, with only general guidance in affix-centric (and on occasion base-centric) instruction.
While less common than affix-centric approaches, there are also different forms of base-centric instruction reported in the literature. For example, “Words Their Way” [ 34 ] is a popular program which targets more advanced derivational morphology in upper elementary than many resources. This approach begins with activities for sorting words around affixes, then moves to sorting words around common bases. For example, students are asked to sort words around what they describe as a Latin Root <dic> for “speak, say” (such as <verdict>, <dictate>) or <vis> for “see” (such as <vision> or <revise>). Because the to-be-sorted words are presented to students this is called morphological ‘recognition tasks.’ There are also production base-centric tasks where children create new words that are part of a morphological family [ 35 ]. Structured word inquiry [ 7 , 8 ] involves both recognition and production: when students are asked to analyze the words in a matrix they are engaging morphological recognition and when they create their own matrices from a base, they are engaging in production. Promising results have been obtained with all of these approaches, including with young children in Grade 1 [ 36 , 37 ], but there is no evidence whether some base-centric approaches are more effective than others.
Even the more basic contrast between affix- vs. base-centric morphological instruction has received little attention. The research literature does include studies that have specifically focused on affixes and others that have focused on affixes and bases. For example, P. Bowers et al. [ 4 ] reviewed 22 intervention studies and found that whereas all targeted affixes only 8 targeted bases or stems. The authors did not attempt to compare the effectiveness of these two different approaches because of the variability of the studies, and neither did Goodwin and Ahn [ 3 ]. Interestingly, Reed [ 26 ] concluded that stronger effects were associated with instruction focused on root (base) words compared to affixes alone in a review of morphological interventions, but this was based on a small (and variable) set of studies making any conclusion insecure.
The present study
Here we report two experiments that directly compare an affix-centric approach to a base-centric approach using morphological matrices (as in Figs 1 and 3 ). In both experiments, participants studied a set of morphologically related words for a subsequent memory task when groups of words were organized by a common affix (using an affix-centric morphological matrix) or base (the base-centric morphological matrix used in SWI), or when randomly intermixed with no reference to morphology (control condition). Given that learning and memory is best when information can be encoded in a meaningful and organized manner [ 9 ], we predicted that memory performance would be better in the two morphological conditions compared to the control condition. Our critical prediction, however, is that organizing words by base in a base-centric morphological matrix will support better memory performance than organizing words by affix, and that this would extend to both familiar (e.g., mistake) as well as unfamiliar (e.g., misbalance) morphologically complex words.
Experiment 1
Method and results
Participants
A total of sixty-two English-speaking participants (mean age = 21.3 years, SD = 2.43 years; 30.6% were males) studying various majors at a large British university participated for either course credit or payment. Participants reported normal language abilities and normal or corrected-to-normal vision. No participants were excluded as all of them were able to successfully complete the task. Participation occurred with informed consent and the experimental protocols were approved by the Institutional Review Board.
This sample size can be justified on the assumption of a large effect size of d = 1.0 for the base-centric condition over the affix and control condition. Consistent with this, Bower et al. [ 9 ] reported memory performance over three times higher in the organized vs. disorganized memory condition with zero overlap between performance in the two groups (the effect sizes were so large that statistics were not even reported). In addition, in their review of generative learning strategies, Fiorella and Mayer [ 10 ] found that matrices (not morphological matrices) that organized to-be-learned information in a meaningful manner were the most effective method for improving learning in a classroom setting, with a median effect of d = 1.07 across eight studies (findings discussed in more detail in the General Discussion). Although there are important differences with our work, these findings lead us to predict large effect sizes. Based on G-Power analysis [ 38 ], an effect size of d = 1.0 with an alpha of 0.05 and power of 0.8, requires 17 participants per group. Accordingly, our sample size should be adequate to detect an effect size around this magnitude.
Design and materials
Participants were randomly assigned to one of three study conditions: A control condition (n = 21) in which words were presented in lists without any form of organization, an affix-centric condition (n = 20) in which words with different bases were grouped by a common affix in a matrix format, or a base-centric condition (n = 21) in which a morphological family of words that all shared a common free base were presented in a morphological matrix much like that used in SWI. We selected 40 morphologically complex words constructed from 10 base words and 10 affixes, half of which comprised either a base word combined with a prefix (e.g., un + cover → uncover), and the other half, a base word combined with a suffix (e.g., count + er → counter). These words were chosen from online dictionaries such as Lexico [ 39 ], Merriam-Webster [ 40 ], Collins [ 41 ] and online resources such as “Word Searcher” [ 42 ]. The words were selected by first determining some of the most commonly used prefixes and suffixes. This was followed by a selection process where base words were chosen based on whether they could form plausible morphologically complex word forms using the prefixes and suffixes that were selected in the previous step. In the base-centric study condition each base word was combined with two prefixes and two suffixes, whereas in the affix-centric condition these same affixes were combined with four bases (see S1 Appendix which contains all the appendices).
In the control condition, words were organized in lists of 4 randomly selected words from the pool of 40 words. Accordingly, all participants studied 10 lists each composed of four morphologically complex words, and all participants studied the same set of 40 words. Note, not all affixes were studied with all bases (e.g., the prefix <dis> was studied with the base <charge> but not <cover>), and accordingly, participants needed to remember which affixes were combined with which bases during the study phase. It is also the case that the spelling of some bases changed when adding an affix (e.g., charge/ + ing → charging), and that some of the morphologically complex words were highly familiar (e.g., mistake) and others not (e.g., misbalance). Regardless of whether such words are familiar or not, a general sense of what these words must mean can be deduced when the reader is familiar with the morphemes constructing the words. The word frequency of the 40 morphologically complex words according to the English Lexicon Project [ 43 ] are listed in S1 Appendix , which contains all the appendices. Each participant studied and recalled the 40 words twice.
Procedure
Participants sat in front of a computer and were tested in groups of 1–6 with stimuli presented using PsychoPy 3 [ 44 ]. Participants were instructed to remember the words presented on the screen. A practice session was conducted to familiarize participants with the general flow of the experiment. During this practice session, participants were given examples of how the stimuli would be presented in the experiment, as well as explicit instructions and a demonstration as to how to combine the affixes and bases in the affix- and base-centric conditions to form the to-be-remembered words. An example of this is shown in Fig 4 . For the base-centric matrices, participants were instructed to study the base combined with all the prefixes and the base combined with all the suffixes (e.g., discharge or charging), but not the base combined with a prefix and a suffix (e.g., discharging).
10.1371/journal.pone.0262260.g004
Fig 4
How the matrices were used to form the to-be-remembered word.
As the participants in the study were university-level students and experienced readers, the basic principle of dropping the final “e” in some of the words was not elaborated on, and in fact, such an error was absent in the responses made later on. Each participant was given 3 seconds to remember each word, with each stimulus list or matrix (composed of four words) presented for a total of 12 seconds. The order in which the lists and matrices were presented was randomized. After all lists or matrices were presented participants wrote down as many words that they could remember on a blank slip of paper. Participants were free to write down the words in any format that they chose, and while not explicitly told to do so, some chose to represent their answers in the matrices that had previously been presented. This slip of paper was then collected from the participants. The experiment was then repeated using the same materials and procedure without the practice session. The order in which the lists and matrices were presented were again randomized. This provided two sets of recall data in each condition, and it provided an assessment of whether the different methods of organizing words support greater or smaller differences with increased practice [ 9 ].
Results and discussion
The mean recall scores in the three study conditions and two recall tests are summarized in Fig 5 , with the y-axis being the proportion of words correctly recalled. A 3 x 2 between-within mixed ANOVA was conducted on the number of words correctly recalled, with Study Condition as the between factor (Control, Affix-centric and Base-centric) and Test Iteration as the within factor (Memory Test 1, Memory Test 2). Only words from the study phase that were correctly spelt were considered correct (erroneously recalled words and those that were misspelt were coded as incorrect). A main effect of test iteration was observed, with participants more likely to recall more words in the second compared to the first test, F1(1, 59) = 124.32, MSE = .336, p < .001; F2 (1,39) = 55.04, MSE = .315 p < .001. More importantly, we observed a main effect of study condition, F1(2, 59) = 16.37, MSE = .197, p < .001; F2 (2, 78) = 38.07, MSE = .463, p < .001, with both the affix-centric (24.2% recall) and base-centric (29.0% recall) conditions supporting better recall than the control condition (17.6% recall). We also observed a significant interaction between study condition and test iteration, F1(2, 59) = 6.29, MSE = .017, p < .005; for the item analyses, Mauchly’s test indicated a violation in the assumption of sphericity, χ 2 (2) = 14.136, p = 0.001, and so the Hyunh-Feldt correction was applied, F2(1.58, 59.51) = 8.199, MSE = .074, p < .005. Taken together, this reflected the greater memory performance in the base-centric condition in the second memory test. The planned comparison between the affix- and base-centric study conditions was significant, t1(39) = 2.17, p = .036; t2(39) = 2.57, p = .014, corresponding to a medium sized effect (Cohen’s d = 0.68), with a null effect observed in memory test 1, t1(39) = .679, p > .05; t2(39) = .761, p > .05 (Cohen’s d = 0.22), and a robust effect in memory test 2, t1(39) = 2.96, p = .005; t2(39) = 3.128, p < .005 (Cohen’s d = 0.95). Together these findings highlight the value of organizing words into their morphological components, and the added value of presenting words in a base-centric morphological matrix.
10.1371/journal.pone.0262260.g005
Fig 5
Mean recall scores for each group in each memory test for Experiment 1 (error bars indicate 1 standard error above and below the mean).
The pattern of errors that is summarized in Table 1 also highlights the impact of the morphological study conditions. We categorized errors into four sub-categories: Recombined Errors that reflected the correct recall of an affix and a base but recombining them into a non-studied word, partial errors that reflected the correct recall of an affix or a base but not both (i.e., a correctly recalled base or affix is paired with an affix or base outside of the study list), unrelated errors that reflected recalled words that included neither a studied based or affix, and spelling errors. For examples of some common errors, please refer to S1 Appendix . All error proportions were calculated with the total number of words presented as the denominator. As is clear from Table 1 , the rate of recombined errors was much greater in the two morphological study conditions, leading to somewhat higher error rates in the two morphological conditions. Note, these recombined errors reveal that the participants were using morphology to inform their memory responses. In the context of a memory experiment, the increase in recombined errors might be considered a negative. But in the context of assessing the overall impact of the morphological encoding conditions on memory and learning more generally (where all morphological forms should be learned, not just studied words), these errors might be considered further evidence that the two morphological study conditions are more effective (reflecting the learning of new morphologically complex words). In Experiment 2 we redesigned the study lists in an attempt to reduce the recombined errors in order to get a more straightforward assessment of the benefits of encoding words in the morphological conditions.
10.1371/journal.pone.0262260.t001
Table 1 Error type (%) across study conditions for Experiment 1.
Test 1
Condition
Overall
Recombined
Partial
Unrelated
Spelling
Control
10.71
3.87
0.50
6.35
0.00
Affix-Centric
13.66
12.04
0.12
1.50
0.00
Base-Centric
13.19
11.92
0.23
1.04
0.00
Test 2
Condition
Overall
Recombined
Partial
Unrelated
Spelling
Control
11.41
5.46
0.30
5.56
0.10
Affix-Centric
14.47
13.89
0.23
0.35
0.00
Base-Centric
16.44
14.35
0.35
1.74
0.00
Experiment 2
In order to reduce the incidence of errors we developed a new set of words in which eight bases and eight affixes were factorially combined (all affix-base combinations were studied), making a total of 64 words to recall. In this way, if a participant remembers a base and an affix correctly, he or she will not incorrectly combine them to produce a recombined error. Again, we anticipate that participants will use their morphological knowledge to improve memory performance in the two morphology conditions, and the critical question is whether the base-centric study condition improves memory performance more than an affix-centric condition.
Method and results
Participants
A total of sixty-four English-speaking participants (mean age = 24; 28% were males) studying various majors at a large British university participated for either course credit or payment. This new set of participants was drawn from the same pool of participants as Experiment 1. Participants reported normal language abilities and normal or corrected-to-normal vision. No participants were excluded as all of them were able to successfully complete the task. Participation occurred with informed consent and the experimental protocols were approved by the Institutional Review Board. Once again, participants were assigned to either the control condition (n = 20), the affix-centric condition (n = 22) or the base-centric condition (n = 22).
Design and materials
A total of 8 free base words and 8 affixes were chosen, and all the base words chosen were able to form legal permutations with all affixes. These words were chosen from the same online resources that were used in Experiment 1. This allowed us to factorially combine all bases and affixes during study so that participants could no longer make Recombined Errors as was observed in Experiment 1. In total, 64 morphologically complex words were formed, half of which comprised either a base word combined with a prefix (e.g., re + value → revalue), and the other half, a base word combined with a suffix (e.g., value + s → values). Once again, the affix and base sometimes composed a more familiar word (e.g., revalue) and in other cases an unfamiliar word (e.g., <disvalue>). These words were grouped either by affixes (for the affix-centric condition) or their bases (for the base-centric condition) to form lists or matrices of 8 words each. Lists of 8 randomly selected words from the pool of 64 words were also created for the control condition. In total, 8 lists or matrices were created for each condition, with each participant studying all 64 words. For full details of the lists, matrices and words used, please refer to S1 Appendix .
Procedure
As above, participants were seated in front of a computer, tested in groups of 1 to 6, and asked to remember words that were to be presented to them on the screen. A practice session similar to that which was conducted in Experiment 1 was conducted to familiarize participants with the general flow of the experiment. Each participant was given 3 seconds to remember each word, with each stimulus list or matrix (composed of eight words) presented for a total of 24 seconds. The order in which the lists and matrices were presented was randomized. After all lists or matrices had been presented, participants were asked to write down as many words that they could remember on a piece of paper, which was collected from the participants once they were done with this portion of the experiment. The experiment was then repeated using the same materials and procedures, without the practice session.
Results and discussion
As before, only words from the study phase that were correctly spelt were considered correct, and the mean recall scores across conditions are summarized in Fig 6 . A 3 x 2 between-within mixed ANOVA showed a main effect of test iteration, with participants performing better in the second compared to the first test, F1(1, 61) = 92.94, MSE = .984, p < .001; F2(1,63) = 331.22, MSE = 2.981, p < .001, and a main effect of study condition, F1(2, 61) = 41.609, MSE = 2.064, p < .001; F2 (2, 126) = 506.88, MSE = 6.659, p < .001, with both the affix-centric (58.7% recall), and base-centric (71.3% recall) condition supporting better recall compared to the control condition (28.0% recall). And again, a significant interaction was also found between test iteration and study condition, F1(2, 61) = 6.72, MSE = .071, p < .005; F2 (2, 126) = 24.75, MSE = .191, p < .001, reflecting the greater improvement in the base-centric compared to affix-centric conditions in the second memory test. The planned comparison between the affix- and base-centric conditions was again significant, t1(42) = 2.93, p = .005; t2(63) = 11.01, p < .001, corresponding to a large effect size (Cohen’s d = 0.88), with the contrast between the base- and affix-centric conditions failing to achieve significance in memory test 1 for the participant analyses, t1(42) = 1.67, p = .102 and significant in the item analyses, t2(63) = 5.78, p < .001 (Cohen’s d = 0.52), and robust in memory test 2 for both participant and item analyses, t1(42) = 2.559, p = 0.014; t2(63) = 11.76, p < .001 (Cohen’s d = 0.79).
10.1371/journal.pone.0262260.g006
Fig 6
Mean recall scores for each group in each memory test for Experiment 2 (error bars indicate 1 standard error above and below the mean).
Furthermore, as expected, the elimination of the Recombined Errors (due to fully crossing bases and affixes) resulted in much reduced overall error rates in the two morphological conditions. Indeed, the overall recall performance was much higher in Experiment 2 (52.5%) compared to Experiment 1 (28.3%), with recall rates most improved in the base-centric morphological matrix condition. The pattern of errors for experiment 2 is summarized in Table 2 .
10.1371/journal.pone.0262260.t002
Table 2 Error type (%) across study conditions for Experiment 2.
Test 1
Condition
Overall
Recombined
Partial
Unrelated
Spelling
Control
7.99
0.00
0.23
7.70
0.06
Affix-centric
4.33
0.00
0.36
3.05
0.92
Base-centric
2.58
0.00
1.14
1.20
0.24
Test 2
Condition
Overall
Recombined
Partial
Unrelated
Spelling
Control
8.28
0.00
1.10
6.77
0.41
Affix-centric
3.69
0.00
0.50
2.06
1.14
Base-centric
1.50
0.00
0.48
0.30
0.72
It is worth noting that we obtained the same pattern of results for the familiar and unfamiliar morphologically complex words in both in Experiments 1 and 2. The frequencies of the 40 morphologically complex words from Experiment 1 and 64 morphologically complex words from Experiment 2 were assessed using the English Lexicon Project [ 42 ] and were organized into a high-frequency category (43 words that all had a frequency of over 10 per million), low frequency category (38 words that all had a frequency between 10 and 0 per million), and zero frequency condition (32 words that did not occur amongst ~131 million words analyzed in the English Lexicon Project). Fig 7 summarizes the overall memory results in the control, affix, and base conditions across the two experiments for the three frequency conditions. The similar pattern of results obtained across the three frequency conditions shows that the memory advantage in the base-centric condition extends to low-frequency and novel words. This is consistent with the hypothesis that the base-centric morphological matrix is a useful tool for learning new words.
10.1371/journal.pone.0262260.g007
Fig 7
The overall memory results in terms of proportion of words recalled in the control, affix, and base conditions across both Experiments 1 and 2 for zero frequency, low frequency and high frequency words.
Another point worth noting is that the same pattern of results was obtained for inflections and derivations in both Experiments 1 and 2 (see Fig 8 ). This demonstrates the memory advantage in the base-centric condition extends to both inflectional and derivational words.
10.1371/journal.pone.0262260.g008
Fig 8
The overall memory results in terms of proportion of words recalled in the control, affix and base conditions across both Experiments 1 and 2 for inflections and derivations.
General discussion
In two experiments we found that memory for a set of morphologically related words was better when the study conditions highlighted the morphological composition of words compared to a control condition that ignored this structure. The critical finding, however, is that memory was best when words were studied in a base-centric manner using morphological matrices that organized words into morphological families compared to a closely matched condition in which the same set of words were organized by affixes in matrices. As noted above, morphological instruction plays a minor role (at best) in literacy instruction in schools at present, and when morphology is considered, affix-centric approaches tend to be adopted. Our findings are important because they motivate the introduction of a new tool for morphological instruction with relevance to teaching spelling, vocabulary, and word naming. More generally, our finding lends some support to a wide range of base-centric methods that have been developed over decades but have not been widely adopted in schools [ 35 ].
Why did the morphological matrices improve memory outcomes? We take our findings to be closely related to the earlier finding from Bower et al. [ 9 ] who observed improved memory when a set of semantically related words were organized in a semantic hierarchy. This organization helped participants generate a retrieval plan that in turn improved memory performance. For example, in the illustration depicted in Fig 2 , participants already knew many examples of gemstones and metals and could thus retrieve instances of gemstones and metals from background knowledge and then decide if the generated item was on the study list. The morphological matrices may have provided a similar benefit to memory retrieval strategies in our experiments given our participants likely already knew the meanings of the base words and affixes they had studied, and accordingly, they may have used their background morphological knowledge to generate plausible responses and then decide if the word was included in the study list.
Providing a context in which a learner can generate a retrieval plan may not only be useful for retrieval itself, but also the retrieval plan may improve the encoding of the to-be-remembered materials through “generative learning” that involves organizing and integrating new information with prior knowledge in order to improve the learning [ 10 ]. A classic finding in memory research is the “generation effect” in which generating a solution to a problem during the study phase of a memory experiment leads to better retention than being provided the answer to the problem at study [ 43 , 45 , 46 ]. For example, participants were better at remembering the word pair “rapid-fast” if they generated the word “fast” from the cue “rapid-f” during the study phase compared to simply studying the word pair “rapid-fast” [ 47 ]. Importantly, organized retrieval and generation effects may interact. For example, the memory advantage of organized vs. disorganized study condition in the above Bower et al. [ 9 ] study became larger following repeated memory tests, suggesting that an organized retrieval plan at memory test 1 improved the encoding of the words that in turn improved performance in memory test 2. We found a similar pattern here, with the advantage of the affix-centric and base-centric matrix conditions becoming stronger with repetition.
One common tool for generative learning that is especially relevant is a graphic organizer in which key concepts are arranged in meaningful format such as a matrix. Indeed, of all the generative learning strategies reviewed by Fiorella and Mayer [ 10 ], a matrix organizer was found to be the most effective with a median effect of d = 1.07 across eight studies. In the case of the morphological matrix, the content being arranged meaningfully just happens to be the orthographic structures of morphologically related words. Generative learning may help explain the improved performance of the base-centric compared to the affix-centric condition given that the base rather than the affix constitutes the core meaning of a word so that deeper semantic links can be developed when organizing words by base. It is also consistent with our finding that the memory advantage in the base-centric matrix condition extended to novel morphologically complex words, a case in which novel words can be organized and integrated with the familiar morphemes to make sense of them.
Of course, there is a large gulf between teaching morphology to children in the classroom and our memory experiment carried out on undergraduate students. Clearly further research is needed to determine whether the memory advantage we observed in the morphological base-centric matrix condition will translate into more effective morphological instruction in the classroom and under what conditions. Similarly, it is unclear from the current memory results whether the matrices would be best suited for teaching word naming, spelling, vocabulary, or some combination of these skills. But none of this undermines the main goal of the current research, namely, to provide a tightly controlled study that directly compares the efficacy of encoding words in an affix- and base-centric manner. We think our findings provide a strong motivation to explore the conditions under which morphological matrices are effective for instruction in children and highlight the value of base-centric instruction.
Indeed there are good reasons to hypothesize that our results are relevant to learning in children. Perhaps the most basic point is that the benefits observed in the morphological base-centric matrix condition reflect a fundamental property of memory and learning that applies to both adults and children; namely, it helps to study and recall information in a meaningful and organized manner. More directly relevant, there is some preliminary evidence that the morphological base-centric matrix does help literacy instruction in children. For example, Bowers and Kirby [ 7 ] reported that a Structured Word Inquiry study that included the morphological matrix improved vocabulary learning in grade 4 and 5 students. Similarly, Devonshire and Fluck [ 43 ] showed benefits in spelling for Year 4 and 5 students in England (ranging from age 7–9) following an intervention that included morphological matrices, and Devonshire et al. [ 36 ] reported benefits in spelling and naming aloud words in a Year 1 and 2 intervention (children aged between 5–7) using matrices. These studies all used a range of tools in addition to the morphological matrices, and the current study suggests that the matrix per se played a role in these results. It is interesting to contrast the study by Bowers and Kirby [ 7 ] with a recent large intervention by Foorman et al. [ 13 ] that adopted an affix-centric approach to teaching morphological knowledge. Both studies were concerned with improving vocabulary knowledge, and both studies targeted similar aged children, although the Foorman et. al study [ 13 ] included 200 minutes more instructional time and addressed far less content. Whereas Foorman et al. failed to observe any improvements in children inferring the meaning of novel words, Bowers & Kirby found children were better at defining words they had not been exposed as long as that word shared a base with a word they had been exposed to. The current study suggests that the use of the base-centric morphological matrix in the Bowers & Kirby study [ 7 ] may have contributed to the better outcomes.
In sum, our findings highlight that encoding words in a base-centric condition leads to better memory than a closely matched affix-centric condition. These findings suggest that that the morphological base-centric matrix may be a useful tool for morphological instruction, and support base-centric approaches to morphological instruction more generally. We hope our findings motivate future research that characterizes the conditions that morphological matrices are most useful for reading instruction and motivate more research into the Structured Word Inquiry method that uses the morphological base-centric matrix along with various other tools to not only teach about morphology, but also grapheme-phoneme correspondences in a meaningful context.
Supporting information
S1 Appendix
Contains all the appendices.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Bladder cancer (BC) accounts for 3.2–4% of all cancers worldwide and approximately 90% of BC are urothelial carcinoma (UC). Although UC of the bladder often has focal squamous differentiation, it is differentiated from squamous cell carcinoma (SCC) of bladder, which contains solely keratin-forming carcinoma cells. BC composed of mixed urothelial and squamous phenotypes is called UC with squamous differentiation (UC/SCC) [1] .
Among the known carcinogens of UC of the bladder, cigarette smoking is the most well-characterized risk factor. It is responsible for nearly half of BCs [2] . Other risk factors for BC include occupational exposure to a number of aromatic amines; infection-related carcinogenesis, such as schistosomal infection; and large doses of certain drugs, including cyclophosphamide and phenacetin [3] . However, a large number of BCs remain unexplained.
Recently, the role of viruses has received considerable attention as potential carcinogens, especially human papilloma virus (HPV). HPV is a small, circular, double-stranded DNA virus that infects stratified squamous epithelium and has been estimated linked to almost 10% of all cancers worldwide, especially a subset of SCCs of the vulva, penis, anus, and oropharynx [4] – [7] . Among the different HPV types, certain types, such as HPV-16 and HPV-18, have an established etiological role in development of anogenital cancers. Although HPV-16 and/or -18 genomic sequences were also identified in the urinary tract of female patients with recurrent and persistent urethritis and cystitis [8] , [9] , a number of studies have investigated the possibility that HPV infection is a risk factor contributing to UC of the bladder; however, their results have conflicted, so no definitive conclusions are possible [10] – [13] .
A related issue involves the possibility that p16 protein is overexpressed in UC of the bladder. The expression of p16 is well known to be associated with high-risk HPV infection in cervical cancer and head and neck cancer and it has been suggested as a qualified surrogate marker for the identification of biologically active HPV infection [14] – [16] . However, despite many recent studies focusing on the SCC component of UC of the bladder, it remains unclear whether overexpression of p16 protein is associated with oncogenesis of UC of bladder [5] , [10] , [17] – [19] .
Therefore, the aim of the present study was to determine the presence of HPV infection and to validate the possibility of p16 overexpression as a surrogate marker for HPV infection in UC/SCC of the bladder in Koreans smokers and non-smokers.
Materials and Methods
Following approval of the study by our Institutional Review Board, 47 patients were selected from the archives of the hospital from July 2001 to March 2011 (IRB No. NCCNCS-12-643). All patients provided verbal informed consent. The written consent was not obtained because the IRB of National Cancer Center approved to exempt the written consent procedure. We examined paraffin-embedded tissue samples obtained during transurethral resection of the bladder or cystectomy. The study group consisted of 35 patients with mixed UC/SCC of the bladder and the control group consisted of 12 patients with squamous metaplasia of the bladder. Of the study group patients, 33 had evidence of mixed UC/SCC and two had extensive SCC with a history of UC. The UC of the bladder specimens were classified according to the World Health Organization (WHO) and International Society of Urological pathology (ISUP) grading criteria and staged according to the TNM 2010 version. None of the patients had known or suspected histories of Schistosoma heamatobium infection, uterine cervix SCC, or head and neck SCC.
A representative SCC component of each paraffin block sample was selected based on the presence of adequate quality and quantity of extracted DNA. DNA was extracted using Qiagen BioRobot M48 workstation (Qiagen, CA) and the quality and quantity of DNA was measured using Nanodrop ND 1000 (NanoDrop Technologies, Wilmington, DE). The presence of HPV infection was analyzed using an HPV-DNA chip. HPV genotyping was performed as per the manufacturer's protocol using a polymerase chain reaction (PCR)-based DNA microarray system (Greencross, Gyeonggi, Korea), consisting of multiple probes of the HPV L1 sequence [20] . These probes included specific probes for high-risk HPV types (HPV-16, 18, 31, 33, 35, 39, 45, 51, 52, 53, 56, 58, 59, 66, and 68) and low-risk HPV types (HPV-6, 11, 34, 40, 42, 43, 44, 54, and 70). Consensus PCR products were hybridized to the probes on the chip, scanned (Scanner GX, PerkinElmer, MA), and then analyzed. PCR amplification was performed using 10 μL purified DNA.
Immunohistochemistry (IHC) staining of p16 protein was performed using specific E6H4 antibodies (DAKO, Carpinteria, CA). IHC p16 expression was scored using a semiquantitative composite scoring system as follows: (1) staining intensity, defined as 0 for negative, 1+ for weak, 2+ for moderate, and 3+ for strong; (2) positive area, defined as the (10 X) fraction of stained tumor cells in the entire tumor; and (3) expression score, defined as the staining intensity multiplied by the positive area. The highest possible score was 30. Overexpression of p16 was defined as a score >20 ( Figure 1 ).
10.1371/journal.pone.0093525.g001 Figure 1
Histological findings and expression patterns of p16.
A: The squamous cell carcinoma components are located beneath the urothelial lining. The tumor cells are strongly positive for p16 (x100). B: This case is a mixed papillary urothelial carcinoma (left upper) and squamous cell carcinoma. The squamous cell carcinoma component is strong positive for p16. C: Squamous metaplasia shows no expression of p16.
The relationships between the presence of HPV infection and clinicopathological parameters were assessed using Fisher's exact tests and Mann-Whitney tests. All statistical analyses were performed using STATA (release 9.2, STATA Inc, Tex, USA). Results were considered statistically significant if the two-sided p value was <0.05.
Results
Clinicopathological characteristics of the 35 UC/SCC study group patients and 12 squamous metaplasia control group patients are shown in Table 1 . No demographic differences were observed between the study and control groups, including age (71.2±7.7 vs. 66.5±10.1 years, p = 0.246), male to female ratio (82.9% vs. 66.7%, p = 0.251), and smoking history (62.9% vs. 41.7%, p = 0.311). In the study group, 4 (12.1%) patients had low-grade carcinoma and 29 (87.9%) had high-grade. For the remaining 2 patients, the available samples (obtained by transurethral resection of the bladder) contained only an SCC component, thereby prohibiting the assignment of a grade.
10.1371/journal.pone.0093525.t001 Table 1
Demographic and histopathological characteristics of the study group and control group.
Study group
Control group
p -value
No. of patients
35
12
Age (years)
71.2±7.7
66.5±10.1
0.247
Sex
Male
29 (82.9%)
8 (66.7%)
0.251
Female
6 (17.1%)
4 (33.3%)
Stage
NMIBC
8 (22.9%)
NA
NA
MIBC
17 (48.6%)
NA
NA
MBC
10 (28.6%)
NA
NA
Grade
Low
4 (12.1%)
NA
NA
High
29 (87.9%)
NA
NA
NA
2
NA
NA
Smoking history
0.311
Smoker or Ex- smoker
22 (62.9%)
5 (41.7%)
Non-smoker
13 (37.1%)
7 (58.3%)
NMIBC: non-muscle invasive bladder cancer. MIBC: muscle invasive bladder cancer. MBC: metastatic bladder cancer. NA: not applicable.
HPV DNA was detected in 6 of the 35 (17.1%) study group samples, and 1 of the 12 (8.3%) control group samples ( Table 2 ). All HPVs were type 18, except for 1 study group sample, which was type 35. Overexpression of p16 was detected in 16 (45.7%) study group samples and 1 (8.3%) control group sample ( Table 2 ). Both HPV positivity and p16 overexpression were present in only 3 (8.8%) study group samples and no control group sample.
10.1371/journal.pone.0093525.t002 Table 2
Detection of human papillomavirus and overexpression of p16 in the study group and control group.
Study group
Control group
p -value
No. of patients
35
12
HPV+
6 (17.1%)
1 (8.3%)
0.659
P16+
16 (45.7%)
1 (8.3%)
0.034
HPV+/P16+
3 (8.6%)
0 (0.0%)
0.295
HPV: human papillomavirus virus. HPV+: detection of HPV. P16+: overexpression of p16.
For the study group, there were no differences in age, sex, stage, or grade between HPV-negative and HPV-positive cases. Of the 29 HPV-negative cases, 20 (69.0%) were smokers or ex-smokers, whereas of the 6 HPV-positive cases, only 2 (33.3%) were smokers or ex-smokers. For smokers or ex-smokers, the percentage with HPV-negative BC was approximately 2-fold higher than the percentage with HPV-positive BC. By contrast, the BC samples of non-smokers demonstrated an approximately 2-fold higher percentage of HPV-positivity than HPV-negativity ( Table 3 ).
10.1371/journal.pone.0093525.t003 Table 3
Demographic and histopathological characteristics according to detection of human papillomavirus in the study group.
Study group (N = 35)
p -value
HPV- (n = 29)
HPV+ (n = 6)
Age (years)
71.4±7.9
70.0±6.5
0.718
Sex
0.973
Male
24 (82.8%)
5 (93.3%)
Female
5 (17.2%)
1 (16.7%)
Stage
0.220
NMIBC
6 (17.0%)
2 (5.8%)
MIBC
16 (45.7%)
1 (2.9%)
MBC
7 (20.0%)
3 (8.6%)
Grade
0.571
Low
3 (11.1%)
1 (16.7%)
High
24 (88.9%)
5 (83.3%)
NA
2
0
Smoking history
0.286
Positive+
20 (69.0%)
2 (33.3%)
Negative
9 (31.0%)
4 (66.7%)
NMIBC: non-muscle invasive bladder cancer. MIBC: muscle invasive bladder cancer. MBC: metastatic bladder cancer. NA: not applicable. + , Positive included with all of the present and ex-smokers.
Discussion
In addition to cigarette smoking, occupational exposure to aromatic amines, and use of specific drugs, such as cyclophosphamide and phenacetin, as known risk factors of BC, HPV infection has also been suggested as a potential causative agent of UC of the bladder [21] , [22] . The prevalence of HPV in UC of the bladder reported in previous studies has varied from 0% to 81.3% [23] – [25] . The disparity is probably due to sampling problems, contamination, sensitivity of the detection systems, and geographic variation [26] .
Youshya et al. [21] . reported that 60% of patients were positive for HPV L1 capsid protein expression by immunostaining, even though PCR using consensus GP5+/6+ primers failed to detect HPV DNA. The suggested reason for the disparity of HPV detection rates was the antibodies that were utilized and the sensitivity of analysis for detection of DNAs [27] . In this study, we utilized the HPV DNA chip microarray system. The HPV oligonucleotide microarray system is a newly developed biotechnology that can be applied in clinical practice for detection and genotyping of HPV. The HPV-DNA chip microarray has shown higher sensitivity for detection of HPV DNA than PCR alone in tissue block [28] , [29] .
The HPV has an affinity for squamous cells and its association with human SCC of the uterine cervix, head and neck, and anus has been well established [7] . In the current study, the detection rate of HPV DNA was 2-fold higher in the study group with mixed UC/SCC of the bladder (17.5%) than the control group with squamous dysplasia (8.3%). However, we could not confirm a statistically significant increase in risk of HPV infection in the study group, likely because of the relatively small sample size ( Table 2 ). Our findings are consistent with other studies in which HPV DNA was detected in both UC of the bladder and SCC of the bladder [30] .
Although HPV infection has been associated with BC in many studies, it is important to realize that such an association is not equivalent to causation. None of the previous studies have proven the existence of a direct link between HPV infection and carcinogenesis in BC, although the possibility of a causal role has been suggested because of the similarity to cervical and head and neck cancers, for which a causal relationship has been established [19] [31] , [32] . To address this issue, we tried to validate p16 overexpression as a surrogate marker for active HPV infection in BC; however, we were unable to detect a strong association between p16 overexpression and HPV infection in UC/SCC of the bladder( Table 2 ).
HPV encodes two oncoproteins: E6 and E7 [33] . E6 protein binds to wild type (wt) p53, thus triggering ubiquitin-mediated degradation of p53 or directly inactivating wt-p53 by complex formation [34] , [35] . E7 binds to pRb and releases the E2F transcription factor, which subsequently facilitates cell proliferation. P16 is a cyclin-dependent kinase (CDK) inhibitor that blocks CDK4-mediated phosphorylation of pRb and subsequent G1 to S progression. HPV-infected cells overexpress p16 in an attempt to compensate for the E7-induced loss of pRb. Because of these negative feedback mechanisms between pRb and p16, overexpression of p16 is now often used as a surrogate marker of HPV oncogene expression. In this study, p16 was overexpressed in 45.7% of UC/SCC samples; however, both HPV-positivity and p16 overexpression was present in only 8.6%. This finding thereby shows that p16 expression might not be a strong surrogate marker for evidence of HPV infection in UC/SCC, which is consistent with the results of Alexander et al. [10] .
Additional findings of interest in this study were the types of HPV and the higher percentage of non-smokers in the HPV-positive group than in the HPV-negative group. Almost all HPV-positive samples in the study group were type 18. The sole exception was a sample with HPV type 35, which has been considered a variant of HPV 16 and which has not been previously isolated from the bladder [36] . As it is well known that the HPV types 16 and 18 are categorized as high risky HPV for a leading cause of cancer among various HPV types [7] , [16] , [37] , [38] , our result with HPV types 18 and 35 might be supported our hypothesis of HPV infection in association with the carcinogenesis of bladder cancer. Although the increased likelihood of HPV-positivity in non-smokers did not reach statistical significance because of the small sample size, this finding suggests that HPV infection might further be associated with the development of mixed UC/SCC of urinary bladder, especially in non-smokers because of its twice greater the percentage of HPV positivity than that of smokers (p = 0.286, Table 3 ).
This study had some limitations. A retrospective study with a relatively small number of samples limited its statistical power and potential bias. The serology to HPV suggested by Badawi could not be performed because of the study design, even though the antibody response (immunoglobulin [Ig]G) to L1 capsids could be used as a marker of cumulative HPV exposures [9] . The control group had squamous metaplasia, whereas a control group with normal urothelium may have been more appropriate to show a strong association between HPV infection and carcinogenesis of UC/SCC. However, this study is significant in determination of the relationship of HPV infection with prevalence of BC as one of causative factors in carcinogenesis of bladder and this is the first paper describing the clinical importance of the possible association between HPV infection and bladder cancer in non-smokers.
Conclusion
Based on the findings of this study, HPV infection appears to be associated with the development of mixed UC/SCC of the urinary bladder, especially in non-smokers among Koreans. However, the role of HPV infection in the development of this type of cancer may not be as significant as for SCC of the uterine cervix or the head and neck. Additionally, p16 expression does not appear to be a strong surrogate marker for evidence of HPV infection in UC of the bladder with squamous differentiation.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Rift Valley fever (RVF) is a zoonotic disease that causes storm abortions in ruminants [ 1 – 3 ]. The disease leads to introduction of restrictions for international livestock trade from enzootic/endemic countries. The disease imposes a dual impact in that it exacerbates the poverty cycle in livestock-dependent communities, by causing substantial health costs and at the same time affecting negatively the livelihoods of the communities in many sub-Saharan countries where it is enzootic/endemic[ 4 , 5 ]. RVF was first reported in early 1930’s in the Eastern Rift Valley province of Kenya causing high rates of abortion in infected sheep [ 6 ]. Since then, the Rift Valley fever virus (RVFV) has been associated with several periodic disease epidemics and epizootics affecting human and animals in many regions of Africa. Although the virus is enzootic/endemic to sub-Saharan Africa, it has the potential for global spread and has already crossed significant natural geographic barriers such as the Indian Ocean, the Sahara Desert and the Red Sea to reach naive ecologies [ 7 ]. Outside Africa, RVF outbreaks were first reported in Saudi Arabia [ 8 ] and Yemen [ 9 ] in 2000. This northward spread of RVFV suggests the possibility of the virus being introduced into Europe and North America where several species of mosquitoes competent for viral transmission exist [ 10 ].
Recent spatial and temporal analysis of RVF in Tanzania showed that RVF-like disease was reported for the first time in 1930 concurrently with the outbreak in Kenya, with a further ten outbreaks being reported between 1947 and 2007 [ 7 ]. In 2006/2007, there was a massive outbreak with a total of 684 human cases and 234 deaths reported in Kenya, 114 cases with 51 deaths in Somalia and 264 cases with 109 deaths reported in Tanzania [ 11 ]. In Tanzania, the 2006/2007 RVF outbreak was widely spread to more than ten regions in the northern, eastern-central and southern parts of the country [ 7 ].
In RVFV enzootic/endemic regions, outbreaks occur with 3 to 17 year intervals [ 12 ], which is an average inter-epizootic/inter-epidemic (IE) interval of 7.9 years [ 7 ]. The RVFV maintenance between the long IE periods is not fully understood. Although it has been widely hypothesized that the virus is maintained via transovarially infected Aedes mosquito eggs [ 12 , 13 ], serological evidence suggests that the virus could be maintained through IE circulation in domestic ruminants, wild animals and humans [ 9 , 14 – 17 ]
Evidence of RVFV circulating in Tanzania during an IE period has been shown previously in the Kilombero river valley, where various livestock species had RVFV antibodies (cattle 11.03%; sheep 11.86% and goats 11.37%, respectively) [ 14 ], although Kilombero river valley was among the sites that experienced the 2006/07 RVF outbreak [ 7 ]. Additional surveys in areas without previous history of RVF outbreaks [ 18 ] or clinical cases in humans [ 9 ] show related prevalence rates in livestock and humans respectively. In addition, the role of mammals as maintenance hosts for RVFV remains largely unknown [ 19 ]. Furthermore, the detection of antibodies in areas where no clinical disease has been reported [ 9 , 18 ] raises the question of whether the disease is overlooked due to lack-of effective surveillance systems, or whether there are strains of RVFV with low pathogenicity. This study aimed at determining the involvement of non-vaccinated cattle in the IE maintenance and transmission of RVFV in areas with no history of RVF outbreaks in Tanzania.
Materials and methods
Study areas
This study was conducted from June, 2014 to October, 2015 in the Kyela and Morogoro districts, Tanzania. During the 2006/2007 RVF outbreak in Tanzania, ten regions were affected [ 20 ]. The Morogoro region was one of the ten regions, but only two districts, the Kilombero and Ulanga districts were affected, while the Morogoro district was not affected [ 7 ]. The Mbeya region, where the Kyela district is located, was not affected during the 2006/2007 RVF outbreak in Tanzania [ 7 , 20 ].
Kyela is one of the districts in Mbeya region, located in south-western part of the country. Most of the Kyela district is lowland situated in the Great Rift Valley at 505m above sea level, in the flood plains of Lake Nyasa. It receives heavy rains, of about 2000-3000mm per annum and floods are common in March through May. The district has a warm and humid climate, with a mean daily temperature of 23°C. Together with Lake Nyasa, the district also has four large rivers, (Songwe, Mbaka, Lufilyo, and Kiwira), and many streams (Mkalizi, Kampala, Mgaya, Chiji, Kandete, Masukila, Njisi, and Kubanga). Agriculture dominates livelihoods and economic activities of the Kyela district. In addition to rain fed paddy farming, other crops include banana and cocoa cultivation. Other livelihood activities include livestock farming and fishing. Because of the water logging condition, few sheep and goats are kept in the district. Few (1–5) cattle are kept per household by tethering in communal grazing areas during the day and on the doorsteps of their houses at night for fear of theft, providing an animal reservoir of RVFV in proximity of humans.
The Morogoro district is located within the Morogoro region, 200 km east of Dar es Salaam. The annual average rainfall for Morogoro ranges between 500 and 1800 mm with temperatures between 18°C to 28°C. The main occupation of the inhabitants include crop cultivation and livestock keeping and a number of livestock species are kept including cattle, goats, sheep, pigs, camels, donkeys and horses. The livestock production in Morogoro is organised under commercial and traditional sectors. The livestock production systems are pastoralisim, agro-pastoralism and small scale intensive system which is becoming popular as land shortage force many livestock keepers to intensify their production. In the latter system, mainly crossbred animals are kept, and cut and carry system of feeding is practicied.
Selection of districts and wards
The sampling process involved a two-stage purposive selection of districts and wards based on the findings of the past studies (2) reporting status of RVF outbreaks in Tanzania. The number of wards was not based on statistical considerations, but on logistic and resource availability.
Based on the above, the Kyela and Morogoro districts with no previous history of RVF outbreaks were selected. In both districts, all veterinary officers were consulted to identify wards within each district considered to be at highest risk of RVF occurrence. Criteria used included areas subject to regular flooding, ecological features suitable for mosquito breeding, relatively high concentration of domestic ruminants, proximity to rivers, ponds and lakes. The wards within the districts that were identified with most of these epidemiological characteristics were selected for the study.
Within the selected wards, all households keeping domestic ruminants and not having a history of vaccination against RVF were identified using local official veterinary records. The spatial and temporal patterns of RVF outbreaks in Tanzania; 1930 to 2007 [ 7 ], showed that no previous outbreak had occurred in the two study districts. It was further confirmed during the study where, in each district, the veterinary offices were asked for any occurrence of RVF disease and/or outbreak, and history of livestock RVFV vaccination. Furthermore, information on the animal movements into the selected wards was retrieved from the veterinary officials, household heads and herdsmen. Additional information on the sources of replacement heifers was also requested. Wards without inward migration of animals from other areas were selected for the study.
Sample collection
The local breed of zebu cattle ( Bos indicus ) and the crosses with exotic breed ( Bos taurus ) were sampled by collecting 5 ml of blood from the jugular vein into plain vacuitaner tubes. The criteria for selection of animals included a history of non-vaccinated status against RVFV, animals born after the 2006/2007 outbreak, calves above 6 months of age, and owners consent to using the animals for study. Herd and individual animal epidemiological data were obtained from the household head and herders as well as through clinical examination. The data collected included the breed, sex and age and feeding practices. In addition, a history of animal movements into the herd and whether the animals were born within the herd or introduced (moved) into the herd from another district was recorded. Sampling was based on only those herds with restricted animals without history of movements to high RVF risk areas.
The blood samples were collected in vacutainer tubes without an anti-coagulant, labeled and stored in a cooler box with ice packs while in the field. Before blood collection, animals were restrained into the crush or by use of ropes and halters. The blood was allowed to coagulate before serum was separated into a 1.5 ml cryovial tube, labeled and stored in a cool box with ice packs until transfer to the laboratory for analysis. Serum samples were stored at -80°C until analysis.
Age and breed determination
Individual animal age was estimated from epidemiological data collected from household heads and herders, and where possible, by review of available records on date of birth and dentition. Records were available from farms keeping crossbred dairy cattle. Dentition was used in determining the age of cattle divided into young or adult, depending on the eruption of the permanent incisors [ 21 ]. All cattle that had at least a permanent middle incisor were categorised as adult, while those without were categorised as young. Normally, the permanent incisors in cattle erupt at about 18 months of age and by 24 months they are fully developed. To exclude sampling young animals less than six months old, age was estimated by asking the head of the household, herd boys and other members of the household for the month, season and year of birth. Also, we performed physical observation of animal size and asked if they still were suckling. To avoid sampling animals present during the 2006/2007 RVF outbreak, animals that had initial wear on their incisor teeth (5 to 6 years old) and those which had noticeable wear (7 to 8 years old) were excluded from the study.
Breed types were recorded as local or cross-breeds, depending on the body colouration, presence of a hump and horns. Local breeds were shorthorn humped zebu with various body colouration. Sex of the animal and test results were provided in the same data set.
Multi-species competition enzyme linked immunosorbent assay (cELISA)
Each serum sample was analysed with the commercial Innovative Diagnostic (ID.vet) Screen RVF competition multispecies ELISA (cELISA) (IDVet, Montpelier, France). The commercial cELISA is based on the recombinant RVFV nucleoprotein and detects both RVFV IgM and IgG antibodies. The cELISA was carried out according to the manufacturer’s instruction. Briefly, 50 μl of the dilution buffer was dispensed into each well of a labelled ELISA plate pre-coated with recombinant RVFV nucleoprotein. Then, 50 μl of the internal positive (freeze-dried RVFV IgG positive bovine serum supplied by the manufacturer) and the internal negative control (supplied by the manufacturer) were added in duplicates. To the remaining wells, 50 μl of each sample was added. After mixing samples and controls with the TST dilution buffer (50 mM Tris/150 mM NaCl/0.1% Tween 20, pH 8.0), we incubated at 37°C for 1 hour. The wells were washed three times with washing buffer using a well plate washer (Thermo Scientific Wellwash Microplate Washer, Waltham, MA USA). Next, 100 μl of antinucleoprotein peroxidase (HRP) conjugate was added to the wells and the contents of the plate were incubated at room temperature for 30 min, followed by washing three times with 300 μl of wash solution as before to remove excess conjugate. Then, 100 μl of substrate solution 3,3',5,5'-tetramethylbenzidine (TMB) was added to each well and the plate was incubated at room temperature for 15 min in the dark. To terminate the reaction 100 μl of 2n Sulphuric acid (2NH 2 SO 4 ) stop solution was added to each well. The presence of antibodies to RVFV was detected by lack of a colour change, whereas absence of antibodies to RVFV was detected by a change in substrate colour to blue. The contents of the wells of the microplate were read at a wavelength of 450 nm by a microplate absorbance reader (Molecular Devices, CA, USA).
For each cELISA experiment, duplicate internal controls were incorporated. The optical densities (ODs) of the control were detected at 450 nm. To verify the reliability and validity of the results obtained from each cELISA test, the average of the ODs of the two negative controls (NCs) was > 0.7 while the average of the two positive controls divided by the average OD of the NCs was > 0.3. For each sample, the competition percentage was calculated by dividing the OD of the sample by the average OD of the negative control multiplied by 100 ([ODsample/ODNC] x 100). A sample was considered positive if the value obtained from the formula was ≤ 40%. Any sample with a value of > 50% was considered to be negative, whereas values ranging from 40–50% were considered to be doubtful.
RVFV IgM antibody capture ELISA
The IgM ELISA test was employed for cELISA positive samples only. These samples were analysed with the commercial ID Screen RVF IgM Capture kit (IDvet, MOntpelier, France) according to the manufacturer’s instruction. Briefly, 40 μl of the diluent buffer was dispensed into each well of a labelled microwell plate pre-coated with anti-bovine-ovine-caprine IgM polyclonal antibodies. Then, 10 μl of the internal positive control (freeze-dried anti-RVFV recombinant NP bovine serum supplied by the manufacturer) and the internal negative control (supplied by the manufacturer) were added in duplicates. To the remaining wells, serum samples were added in duplicate and the plate with all samples was incubated at 37°C for 1 hour. The microplate wells were then washed three times with 300 μl by a microplate washer as above. Next, 50 μl of RVFV nucleoprotein or diluent buffer was added and incubated at 37°C for 1 hour. The wells were washed three times followed by addition of 50 μl of anti-RVFV nucleoprotein horseradish peroxidase (HRP) conjugate solution to each well and incubation for 1 hour at 37°C. Again, the wells were washed three times as above and 100 μl of the substrate solution, TMB, was added to each well and then incubated for 15 min at room temperature in the dark. Then, 100 μl of stop solution was added to terminate the reaction.
The presence of IgM antibodies to RVFV was detected by appearance of blue colouration, which became yellow after addition of the stop solution. The contents of the wells of the microplate were analysed at 450 nm by a microplate absorbance reader (Molecular Devices, CA, USA).
For each IgM antibody capture ELISA experiment duplicate internal controls were incorporated. The optical densities (ODs) obtained from the samples at 450 nm were validated in accordance with the manufacturer’s instructions as follows:
The net OD was calculated: net OD = OD even well -OD odd well
The plate was valid if the mean value of the net positive control OD was greater than 0.35 and the ratio of the mean values of the net positive and negative control (absolute value of ODs) is greater than 3 (net OD PC /net OD NC > 3
Interpretation of antibody detection results
For each sample, the percentage of the ratio of sample and positive control (s/p%) was calculated.
S / P % = net OD sample / net OD postive control
Samples presenting a S/P percentage (S/P%):
Less than or equal to 40% were negative
Between 40% and 50% were doubtful
Greater than or equal to 50% were positive
Plaque reduction neutralization test 80% (PRNT 80 )
All samples that were positive for RVFV antibodies by the cELISA kit were analyzed by PRNT 80 . The PRNT 80 protocol used was adopted as previously described [ 22 ]. The RVFV MP-12 vaccine strain, propagated in Vero-E6 cells, was used in the PRNT assay.
Each PRNT assay included the test sera, and a known RVFV antibody positive serum sample and a RVFV antibody negative serum sample from cattle. Each serum sample was diluted in Hanks’ Balanced Salt Solution (HBSS) supplemented with one % each of HEPES, penicillin and streptomycin and heat-inactivated fetal bovine serum (FBS). The dilutions sera samples were made in 96 well plates beginning with a 1:5 dilution in the first wells followed by 4-fold serial dilutions of 1:20, 1:80, 1:320, 1:1280, and 1:5120 in each of subsequent wells. Each diluted serum sample was then mixed with an equal volume of 60–80 plaque-forming units (PFU) of MP-12 vaccine virus. The quantification of PFU was confirmed by a plaque assay based on testing a mixture of equal volumes of the 60–80 PFU and HBSS to confirm that the final virus dose ranged from 30–40 PFUs. The antibody positive control consisted of a mixture of equal volume of 60–80 PFU and a 1:10 dilution of antibody positive cattle serum. The antibody negative control consisted of a mixture of equal volume of 60–80 PFU and a 1:10 dilution of RVFV antibody negative cattle serum. The virus/serum dilution mixtures were incubated at 37°C in the absence of CO 2 for one hour. Next, 50 μl of the virus/serum dilution mixtures were inoculated onto each of two Vero E6 cell monolayer cultures propagated in 24-well tissue culture plates and incubated for one hour at 37°C and 5% CO 2 . Virus mixed with the antibody-positive control serum, was inoculated onto twenty separate Vero E6 cultures. Virus mixed with antibody-negative control serum mixture was inoculated onto four Vero E6 cultures. After incubation for one hour at 37°C with 5% CO 2, each cell culture was overlaid with 0.5 ml of a Seakem agarose (1%) with an equal volume of 2X Eagle’s Basal Medium with Earle’s salts (EBME) supplemented with 8% FBS and one % penicillin/streptomycin, and Glutamine+8g/l HEPES. After two more days of incubation at 37°C with 5% CO 2 , each culture was overlaid with 0.5 ml of a mixture of an equal volume of agarose (1%) and 2X EBME supplemented with 5% neutral red, 8% FBS, and penicillin and streptomycin (1%) and Glutamine + 8g/l HEPES and incubated overnight at 37°C with 5% CO 2 . The PFUs were counted and recorded for both the controls and cattle serum samples. An 80% reduction in the number of PFUs was used as the endpoint for antibody virus-neutralization titers (PRNT80). Wells with too high number of PFUs, that were impossible to count at that dilution, were recorded as TNTC (too numerous to count).
Data analysis
The data were entered into a Microsoft Excel spreadsheet and imported into STATA version 12 (Statacorp, College Station, TX, USA) for cleaning and statistical analysis. Descriptive statistics was carried out followed by univariable analysis to assess initial association between potential risk factors and the outcome variable defined by RVFV seropositivity. The mixed effects logistic regression modelling was used to investigate the association between various potential risk factors and the outcome variable defined by RVFV seropositivity. The models included districts, age, breed, sex and the type of holding. The analysis was conducted in two steps. The statistically significant variables were included in a mixed effects multivariable logistic regression analysis based on a forward variable selection approach, utilising the likelihood ratio statistic and a p -value ≤ 0.05. Because of the differences in the sample sizes and agro-ecological features between Kyela and Morogoro districts, RVFV seropositivity was compared among the wards within the respective district. The Chi-square test was used to compare the RVFV seropositivity among the wards, by using the Rstudio statistical software at p -value ≤ 0.05.
Ethics statement
During blood collection from cattle, the research team adhered to the generally acceptable ethical standards and strictly followed existing national and international guidelines for minimizing pain and stress to the animals. The study purpose was explained to cattle owners prior to sample collection and upon agreeing to allow samples to be collected from their animals, they provided a written consent form. The study protocol was approved by Research and Publication Committee, College of Veterinary and Biomedical Sciences, Sokoine University of Agriculture, Tanzania.
The protocol/permit number assigned by the Institutional Animal Care and Use Committee IACUC/ethics committee Protocol No. SUA/FVM/R.19 of 17 th March 2014.
National or international regulations/guidelines to which animal care and use protocol adhered to: Public health Service Policy on Humane Care and Use of Laboratory Animals and Animal Welfare Regulations.
Results
A total of 356 cattle serum samples were analysed for presence of RVFV antibodies, of which 147 samples were from the Kyela district and 209 samples were from the Morogoro district. The overall seropositivity by cELISA was 29.2% (104/356) and a seroprevalence of 32% (47/147) and 27% (57/209) were recorded among animals in Kyela and Morogoro districts respectively. Animals older than 2 years were more likely to be seropositive than animals younger than 2 years (OR = 0.19; p = 0.000) ( Table 1 ). Likewise, zebu cattle were more likely to be seropositive than crosses, (OR = 2.5; p = 0.000). There were no significant differences between the districts (OR = 0.98; p = 1.0) and the type of holding (OR = 1.25; p = 0.34).
10.1371/journal.pntd.0006931.t001
Table 1 Potential risk factors related to RVFV seroprevalence in cattle in two districts of Kyela and Morogoro, Tanzania.
Variable
Level
% cELISA positive (n)
Odds ratio ( OR )
95% Confidence Interval (CI)
p-value
District
Kyela (n = 147)
32.0(47)
0.8
0.50–1.27
0.34
Morogoro (n = 209)
27.3(57)
Age
Adult (n = 234)
38.9(91)
0.19
0.1–0.35
<0.001 *
Variable
Level
% cELISA positive (n)
Odds ratio ( OR )
95% Confidence Interval (CI)
p-value
District
Kyela (n = 147)
32.0(47)
0.8
0.50–1.27
0.34
Morogoro (n = 209)
27.3(57)
Age
Adult (n = 234)
38.9(91)
0.19
0.1–0.35
<0.001 *
Young (n = 122)
10.7(13)
Breed
Cross (n = 180)
20.0(36)
2.5
1.57–4.05
<0.001 *
Zebu (n = 176)
38.6(68)
Sex
Female (n = 273)
29.3(80)
0.98
0.57–1.69
0.95
Male (n = 83)
28.9(24)
Holding
Farm (n = 209)
27.3(57)
1.25
0.79–1.99
0.34
Open (n = 147)
32.0(47)
*Statistical significant differences between the groups
In the Morogoro district, the Mikese ward had the highest RVFV seroprevalence at 55.3% (21/38) followed by Magadu at 32% (33/103) and Mazimbu 4.4% (3/68). In the Kyela district, the RVFV seroprevalence was 35.7% (25/70), 40.5% 15/37) and 17.5% (7/40) in Bujonde, Kajujumele, and Katumba Songwe wards respectively ( Fig 1 , Table 2 ). There were statistical significant differences in RVFV seroprevalence between the wards in Morogoro district ( p < 0.001) ( Table 2 ). However, this was not the case for wards in the Kyela district ( p = 0.06) ( Table 2 ).
10.1371/journal.pntd.0006931.g001
Fig 1
Seroprevalence of RVFV in cattle from different wards in Kyela and Morogoro districts.
10.1371/journal.pntd.0006931.t002
Table 2 Distribution of RVFV seroprevalence in different wards in Kyela and Morogoro districts, Tanzania.
District
Ward
Samples tested (n)
cELISA positives (n)
% cELISA positive (n)
Chi-square
Df
p-value
Kyela
Bujonde
70
25
35.7
5.55
2
0.06
Kajunjumele
37
15
40.5
Katumba Songwe
40
7
17.5
Morogoro
Magadu
103
33
32.0
34.11
2
< 0.001 *
Mazimbu
68
3
4.4
Mikese
38
21
55.3
*Statistical significant differences between the groups
To specifically detect RVFV IgM antibodies, the RVFV antibody positive samples analysed by the cELISA method were subjected to an IgM capture ELISA. Of the 104 analyzed samples, 30 (29%) were positive for RVFV IgM antibodies. In total 8.4% (30/356) of all cattle sampled in the two districts had RVFV IgM antibodies. When segregated by districts, the IgM antibody seroprevalence was 2.0% (3/147) and 12.9% (27/209) in Kyela and Morogoro districts respectively. In the Morogoro district, the RVFV IgM seroprevalence was 11.7% (12/103), 1.5% (1/68) and 36.8% (14/38) for Magadu, Mazimbu and Mikese wards respectively while in Kyela district, the IgM seroprevalence was 2.9% (2/70), 0% (0/37) and 2.5% (1/40) in Bujonde, Kajunjumele and Katumba Songwe wards respectively ( Table 3 ).
10.1371/journal.pntd.0006931.t003
Table 3 Distribution of RVFV IgM seropositivity and neutralizing RVFV antibodies in Kyela and Morogoro districts, Tanzania.
Region
Variable
Level
Total samples (n)
IgM positives (n)
% IgM positive
% PRNT 80 positive of cELISA positives (n) *
Mbeya
District
Kyela
147
3
2.0
100 (n = 47)
Wards
Bujonde
70
2
2.9
100 (n = 25)
Kajunjumele
37
0
0
100 (n = 15)
Katumba Songwe
40
1
2.5
100 (n = 7)
Morogoro
District
Morogoro
209
27
12.9
81 (n = 57)
Wards
Magadu
103
12
11.7
85 (n = 33)
Mazimbu
68
1
1.5
0 (n = 3)
Mikese
38
14
36.8
86 (n = 21)
*The plaque reduction neutralization (PRNT) assay detected neutralizing RVFV antibodies
Positive samples by cELISA were also analyzed for presence of RVFV neutralising antibody by the PRNT 80 assay and 89% (93/104) of all cELISA-positive samples were PRNT-positive. All (47/47) cELISA positive samples from the Kyela district contained RVFV neutralising antibody, while 81% (46/57) of the samples from Morogoro district had neutralising antibody. Antibody titres ranged from 1:10 to 1:10240 and above ( S1 Table ). Some ELISA positive cattle samples collected from wards in Morogoro district gave titers below 1:10 which was considered negative.
Discussion
We found IgG and/or IgM antibodies to RVFV in 29.2% of cattle sampled during 2014–2015 in Tanzania, from two districts with no RVF outbreaks. All samples were collected during an inter-epizootic/inter-epidemic (IE) period from animals born after the large RVF outbreak in East Africa 2006/2007. The finding of both IgG and IgM positive cattle suggests both long-term persistence of RVFV antibodies and a low level of recent circulation of RVFV. In previous studies from Kilombero, Tanzania and Ijara Kenya the seroprevalence in livestock born after the 2006/2007 outbreak was only 5.5% and 13.1% respectively, while in a study in Tanzania from 2013 in the Kajunjumele ward in the Kyela region, 25.8% had RVFV IgG antibodies [ 14 , 23 , 24 ]. Interestingly, we detected a RVFV seroprevalence of 40.5% (15/37) in the Kajunjumele ward from samples collected 2014–2015, but none of the fifteen seropositive animals were RVFV IgM positive. This suggested that new infections have occurred in the Kajunjumele ward between the previous study, ending August 2013, and our study, starting June 2014. This should then have occurred at least 6–8 weeks before our sampling, since IgM antibodies only persist for that time (25), but no reported animal or human cases were reported from that region during the period. On the other hand, cattle samples collected from Tanzania during the 2006/2007 outbreak had a seroprevalence of 38.7% [ 25 ]. The variation in seroprevalence could be explained by time of sampling, new infections, slaughter, removal of seropositive animals, methods used to analyze the samples, as well as the agro-ecological conditions of the study sites.
The results reported in this study indicated that domestic cattle from the two studied districts have been exposed to RVFV infection during the IE period and could function as virus amplifiers, although the two study districts have no previous history of RVF outbreaks. The cELISA method detected both IgG and IgM RVFV specific antibodies. The RVFV IgG antibodies are believed to persist in animals for life following infection, and therefore its detection provides a reliable index of previous exposure to RVFV (5, 6), but does not indicate when the animals were infected. The detection of RVFV IgM antibodies indicated that the virus was actively circulating sub-clinically in the both the Kyela and Morogoro districts during the time of sampling, although mainly in the Morogoro district. This is supported by the fact that IgM antibodies persist for only 6 to 8 weeks after initial infection [ 26 ], disappears in 50% of infected animals after 45 days, and are absent in almost 100% of infected animals by 120 days post infection [ 27 ]. In the Morogoro district, the Mikese ward had the highest RVFV IgM seroprevalence followed by Magadu and Mazimbu, indicating that RVFV infections have recently occurred in the region and especially in the Mikese ward with 14 RVFV IgM positive cattle out of 38 analyzed. Other studies have also detected RVFV activity in cattle and humans in areas where the disease has never been reported before [ 12 , 15 , 19 , 28 – 30 ].
To summarize, we detected RVFV IgM antibodies in all study wards except Kajunjumele, with Morogoro district having a relatively high IgM seroprevalence compared to Kyela. These findings indicated the presence of active RVFV infection at the time of sampling, during the dry season, at least in the wards examined. Despite the small number of wards and animals tested for IgM, these findings clearly demonstrated the circulation of RVFV during IE periods in non-outbreak areas. It is not clear why the circulating RVFV in these areas did not lead into clinical disease, and the possible mechanisms for the virus maintenance remain to be elucidated. However, possible explanations could be circulation of non-virulent strains of RVFV in these areas or misdiagnosis excluding RVF for other febrile conditions with similar clinical features of fever and abortions. A limitation of the present study was that we unfortunately did not attempt to isolate RVFV from the IgM-positive animals, due to biosafety issues. RVFV is classified as a biosafety level-3 agent and demands biosecurity measures not available during the study. Furthermore, we did not perform any RT-PCR analysis to detect virus RNA.
The relatively high RVFV general seroprevalence recorded among cattle in Kyela (32%) and Morogoro (27%) districts could in some part be attributed to the physical characteristics of the respective district. The study site in Kyela is a low-lying area, close to Lake Nyasa, with many swamps and rivers and is subjected to regular flooding during the rainy seasons [ 9 ]. The ecology of low altitude and proximity to perennial water bodies were found to be associated with RVFV seropositivity in ruminant herds in Senegal and Madagascar, as well as in humans in Gabon and Tanzania [ 9 , 31 – 33 ]. Such an ecology provides good breeding habitats for mosquito vectors involved in the transmission of the RVFV. Furthermore, a large part of the study site in Kyela is used for wetland paddy cultivation with frequent water logging, suitable for mosquito breeding. In the Kyela district, cattle are usually grazed by tethering in open grassland, communal grazing land, near the wetlands and paddy farms, thus increasing the risk of acquiring RVFV from mosquitoes. On the other hand, the Kyela district is characterized by the abundance of banana and cocoa plantations. One study in Ngorongoro district, Tanzania, trapped more Aedes aegypti (a vector for RVFV) in banana and maize farms than in other habitats [ 34 ]. Thus, the banana and cocoa plantations in Kyela could form additional breeding habitats for mosquitoes that may transmit the virus to the animals and thus the observed high seroprevalence in this area.
Other factors that may contribute to the observed high RVFV seroprevalence include activities that facilitate animal movements such as livestock trade, moving animals to areas with green pastures during the drought season, lending animals among the community members and payment of dowry. Animal movements from high-risk areas could introduce RVFV into naïve animals in new areas [ 16 , 35 ] Thus, it is important to carry out studies also in areas found to have high RVFV seroprevalence to better understand the role of animal movements in the dispersal of RVFV and/or its vectors. Such data will be essential for formulation of RVFV control strategies.
The observed RVFV transmission hotspots during the sampling period in Magadu and Mikese point to locally existing factors playing a major role in RVFV maintenance and transmission dynamics. Although, entomological surveys were not conducted, the existence of suitable mosquito breeding habitats was evident. The presence of water in farms throughout the year provides suitable habitats for the breeding of RVFV mosquito vectors. Persistent water in aquaculture ponds and waste lagoons close to animal bans and grazing fields at Magadu farm may serve as important breeding habitats for the Aedes mosquito species. The presence of old machinery like tractors, discarded combined harvesters, old automobile tires and water storage containers may serve as water holding places thus providing harbour and breeding habitats for mosquitoes and continuous low-level transmission of RVFV to vertebrate hosts.
Older cattle (>2 years old) were found to be at a higher risk of having RVFV antibodies than younger cattle (OR = 0.19, 95% CI (0.1–0.35)). These findings agree with reports from studies which found higher seroprevalence in older animals [ 14 , 36 – 38 ]. The exotic breeds and their crosses are more susceptible to RVFV infection than local breed which are resistant and well adapted to the environment (12,13). However, in this study indigenous zebu breed appeared more likely to be RVFV seropositive than crosses (OR = 2.5, 95% CI (1.57–4.05). Sampled crossbred animals were from dairy farms with frequent use of acaricides to control tick infestation. The acaricides use could possibly prevent cattle from mosquito bites as well (Mnyone, 2018, personal communication) thus reducing the RVFV transmission in these animals. No significant difference between sex or the type of holding was observed.
Seroepidemiological studies provides the antibody status in respective species in each area, for the period during which the surveillance was carried out. Thus, continuous surveillance of the antibody prevalence in susceptible species is therefore highly recommended in epizootic/endemic areas.
Conclusion and recommendations
The results demonstrated widespread prevalence of RVFV antibody among cattle during an inter-epizootic period in regions without previously reported RVF outbreaks. Therefore, it is important for animal health officers in these areas to be aware of the current RVFV circulation so that preventive measures such as vaccination could be implemented. It would be interesting to perform further studies in other similar areas with no history of RVF outbreaks, since it is likely that undetected low-level RVFV is occurring in many places. This would help to identify and target RVF hot spots by control measures, aiming for prevention of RVFV transmission to animals and humans.
Future studies could for instance focus on a more comprehensive and inclusive surveillance to identify and characterize RVFV reservoirs and vectors during the IE periods. Longitudinal investigations leading to a better understanding of ongoing RVFV circulation will lead to a better understanding of the IE virus maintenance.
Supporting information
S1 Table
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
The incorrect reference is given in the S4 Fig caption. Please view the figure with the correct caption below.
References 22–35 are missing. Please view these references below.
22. Messerli MA, Danuser G, Robinson KR. Pulsative influxes of H + , K + and Ca 2+ lag growth pulses of Lilium longiflorum pollen tubes. J Cell Sci 1999; 112: 1497–1509.
23. Messerli MA, Creton R, Jaffe LF, Robinson KR. Periodic increases in elongation rate precede increases in cytosolic Ca2+ during pollen tube growth. Dev Biol 2000; 222: 84–98.
24. Holdaway-Clarke TL, Feijó JA, Hackett GR, Kunkel JG, Hepler PK. Pollen tube growth and the intracellular cytosolic calcium gradient oscillate in phase while extracellular calcium influx is delayed. Plant Cell 1997; 9: 1999–2010.
25. Cardenas L, McKenna ST, Kunkel JG, Hepler PK. NAD(P)H oscillates in pollen tubes and is correlated with tip growth. Plant Physiol 2006; 142: 1460–1468.
26. Cardenas L, Lovy-Wheeler A, Kunkel JG, Hepler PK. Pollen tube growth oscillations and intercellular calcium levels are reversibly modulated by actin polymerization. Plant Physiol 2008; 146: 1611–1621.
27. Kroeger JH, Daher FB, Grant M, Geitmann A. Microfilament orientation constrains vesicle flow and spatial distribution in growing pollen tubes. Biophys J 2009; 97: 1822–1831.
28. Michard E, Alves F, Feijo JA. The role of ion flues in polarized cell growth and morphogenesis: the pollen tube as an experimental paradigm. Int J Dev Biol 2009; 53: 1609–1622.
29. Liu J. Piette BMAG, Deeks MJ, Franklin-Tong VE, Hussey PJ. A compartmental model analysis of integrative and self-regulatory ion dynamics in pollen tube growth. PLoS ONE 2010; 5: e13157.
30. Chavarria-Krauser A,Yejie D. A model of plasma membrane flow and cytosis regulation in growing pollen tubes. J Theor Biol 2011; 285: 10–24.
31. Kroeger JH, Zerzour R, Geitmann A. Regulator or driving force? The role of turgor pressure in oscillatory plant cell growth. PLoS ONE 2011; 6: e1854. doi: 10.1371/journal.pone.0018549 .
32. Kochian LV, Shaff JE, Kühtreiber WM, Jaffe LF, Lucas WJ. Use of an extracellular, ion-selective, vibrating microelectrode system for the quantification of K+, H+, and Ca2+ fluxes in maize roots and maize suspension cells. Planta 1992; 12: 601–610.
33. Shabala SN, Newman IA, Morris J. Oscillations in H+ and Ca2+ ion fluxes around the elongation region of corn roots and effects of external pH. Plant Physiol 1997; 113: 111–118.
34. Shabala S, Newman I, Whittington J, Juswono U. Protoplast ion fluxes: their measurement and variation with time, position and osmoticum. Planta 1998; 204: 146–152.
35. Shipway A, Shipway S. Solenoid properties. Inductance and magnetic B-field. 2008. http://www.calctool.org/CALC/phys/electromagnetism/solenoid (Date access: 15/05/2014).
Supporting Information
S4 Fig
The solenoid that was used for measurements.
The elongating zone of the coleoptile is shown in green. The figure is based on the Shipway and Shipway [35] solenoid properties calculator.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Similar to high-income countries, low- and middle-income countries-initiated processes of nutritional and food environment transition in the 1990s [ 1 ]. These transitions are characterized by an increase in the consumption of processed foods, edible oils, and sugary beverages, as well as a greater tendency to eat outside the home and an increased availability of ultra-processed products. All these changes have come at the expense of healthy and traditional diets. Simultaneously, the population in these countries has gradually reduced physical activity and increased sedentary behavior [ 1 – 4 ].
These transitions have led to a gradual increase in overweight and obesity (OW/OB) in all age groups, with a special increase in childhood OW/OB [ 2 – 5 ]. This global increase follows a pattern known as the "obesity transition" [ 6 ]. This pattern is characterized by a gradual shift in the burden of OW/OB from high-income to low- and middle-income countries, from wealthy households to poor ones, from urban to rural areas, and from adults to children. This changes affect several countries in Latin America [ 6 ].
In Ecuador, the prevalence of childhood OW/OB surged by almost 5 percentage points from 2012 to 2018, reaching 35.4% [ 7 ]. This alarming increase predominantly impacts urban areas, males, those with mixed and white ethnic backgrounds, and wealthier households. Importantly, this trend extends beyond, affecting middle- and low-income households and rural populations, highlighting the urgency for a thorough exploration of its determinants [ 7 ].
While there is extensive knowledge about the social, environmental, and clinical determinants of excess malnutrition in school-age children, particularly in high-income countries [ 8 ], few countries in the region have studied these determinants [ 9 , 10 ]. In low- and middle-income countries such as Ecuador, there is a lack of national information regarding the determinants of childhood obesity in school-age children, except for some localized studies [ 11 – 13 ]. The existing knowledge gap precludes the development of effective public health policies. Despite the bulk of scientific evidence, the Ecuadorian government’s efforts are not focused on improving school environments, increasing taxes on sugary beverages, or improving the labelling of processed and ultra-processed products; rather, it has reduced taxes on ultra-processed food [ 14 ]. Therefore, initiatives with proper implementation, supervision, and robust evaluations are necessary to demonstrate their impact and cost-effectiveness in school-age populations. These insights will serve as a compass for evidence-based public policies and interventions, which are crucial for combating childhood obesity in Ecuador.
Childhood obesity is a precursor to adult cardiovascular diseases and cancer [ 15 – 17 ]. Understanding its determinants is vital for making informed public health policy decisions. This study aims to identify the independent factors associated with obesity and overweight in Ecuadorian school-age children (5–11 years). By delving into obesogenic environments and contextual sociodemographic conditions, this research offers valuable territorial insights.
Materials and methods
Study design
This cross-sectional study involved a secondary analysis, utilizing data from the 2018 National Health and Nutrition Survey of Ecuador. To ensure methodological rigor and transparency in both the study design and dissemination of findings, compliance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) [ 18 ] guidelines was scrupulously maintained, as detailed in S1 Table .
Population and sample
In our study, we included data from children encompassing sociodemographic information, anthropometric measurements, dietary habits at home and school, and physical activity status. Additionally, we integrated information about whether the children’s households identified, understood, and used the nutritional traffic light labelling system for processed foods and beverages that was implemented in Ecuador. These questions were incorporated into the National Health and Nutrition Survey of 2018 (see flowchart in Fig 1 ). We included study subjects who: (i) are ≥5 years old or ≤11 years old; (ii) had complete anthropometric information, and (iii) had complete data regarding age, ethnicity, economic quintile, schooling characteristics, dietary habits, and physical activity.
10.1371/journal.pone.0296538.g001
Fig 1
Flowchart showing the study population and selection of study participants.
The ENSANUT 2018 survey
The Ecuadorian National Health and Nutrition Survey 2018 (ENSANUT 2018, for its acronym in Spanish) was a cross-sectional study conducted in 2018 that involved nationally representative samples from the Ecuadorian population [ 19 ].
In the ENSANUT 2018 study, a two-stage sampling strategy was employed to secure a representative sample of the Ecuadorian populace. Initially, Primary Sampling Units (PSU) were chosen through stratified sampling, incorporating proportional probability to size. Subsequently, an average of 18 households per PSU were randomly selected for investigation. Within these households, specific demographic groups were identified. For households with children aged 5 to 11 years, a qualified child informant was selected for interview and asked to complete a specialized questionnaire. The abovementioned sampling approach ensured the data quality and representativeness of the study. Further information on the methodology, datasets, and findings of ENSANUT 2018 is available at: https://www.ecuadorencifras.gob.ec/institucional/home/ .
Measurements
We used the information that the survey collected about sex, age of the child, ethnicity, education of the children, economic quintile, regular class attendance, geographical regions of Ecuador, receiving the human developing bonus (BDH, for its acronym in Spanish), number of people in the household, disposal of excreta, physical activity, perception of consumption of vegetables, consumption of fast food, days per week of school food consumption, buying food at school, consumption of the food provided by the school, recognizing, understanding, and using the nutritional traffic light labelling of processed foods; and, consumption of processed foods with a red label. The Ecuadorian Nutritional Traffic Light Labelling system uses three colours—red, yellow, and green—to indicate the levels of sugar, fat, and salt in processed foods. Red signifies high concentrations, yellow denotes medium levels, and green indicates low content [ 20 ]. This system empowers consumers to make healthier food choices.
Main outcome
The main outcome variable was, overweight or obesity. The World Health Organization (WHO) macro program Stata (WHO AnthroPlus) was used to establish the nutritional status of children based on WHO 2007 standards for the classification of children as overweight or obese between 5 and 19 years. Overweight was defined as a Body-Mass-Index (BMI)-for-age greater than 1 standard deviation above the WHO Growth Reference median; and obesity as a greater than 2 standard deviations above the WHO Growth Reference median [ 21 ].
Statistical analyses and sample considerations
A priori, we calculated that a sample of 9759 individuals is enough to estimate, with 95% confidence and an accuracy of +/- 1 percentage units, a population percentage that will predictably be around 35.38% [ 7 ]. The percentage of necessary replacements is expected to be 10%.
In the analysis conducted, the ’svy’ function of Stata 16.1 (StataCorp. 2019. Stata Statistical Software: Release 16. College Station, TX: StataCorp LLC.) was employed, which is adept at tailoring calculations to the specific structure of our survey data. Adjusted percentages for categorical data and means with standard errors for continuous data are computed by this function, taking into account the sample design. Furthermore, it has been ensured that the ’svy’ function facilitates the fitting of complex statistical models, conforming the results to the survey settings predefined by svyset. Following this, we compared the characteristics of children who are not overweight or obese with those who are, ensuring a methodical approach for a balanced and representative analysis. To assess the differences between groups, we employed Pearson’s chi-squared test for categorical variables and the z test for numerical variables.
Then, multilevel logistic regression models were employed to discern the associations between independent variables and the prevalence of overweight and obesity (OW/OB). These models facilitated the estimation of both unadjusted and adjusted odds ratios (OR and aOR, respectively), confidence intervals (95%CI) of ORs, and their corresponding p-values for each independent variable, offering insights into the likelihood of OW/OB presence relative to each variable or their respective categories. To enhance the precision of our analysis, we incorporated geographical regions as a level within our models. This was imperative due to the observed variance in OW/OB prevalence among different regions. By doing so, we could account for regional disparities that may influence the health outcome. Additionally, we integrated expansion factors in our estimations to ensure congruence with the stratified sampling design and the primary sampling units. This stratification was an essential step in managing the inherent variability and correlations within our sampled groups, thereby ensuring that our estimates remained robust and representative of the broader population.
We started by creating a saturated model that considered all possible variables. Next, we removed any variable that did not show a strong enough link to the outcomes we were interested in; specifically, with a p-value of 0.05 or higher [ 22 ], indicating a less than 5% chance that the factor was meaningfully related to the outcome. After trimming down the model in this way, we were left with a parsimonious model that only included the most relevant variables. Finally, we compared the saturated model with parsimonious one and selected the best model based on the likelihood ratio test that measure how well each model predicts our outcomes of interest. The final model was stratified by sex. Given the small number of missing data (there were missing values in <1% of the whole database), we employed complete case analysis to estimate statistical associations.
To test for potential effect modification, we performed several secondary analyses to assess the sensitivity of our estimates with our assumptions regarding biases, and to test for model misspecifications. We ran the final model excluding: (i) children categorized in the highest income quintile, (ii) children whose parents received the BDH, and (iii) children within the upper third of the highest number of people per household.
Ethical issues
The research protocol was thoroughly reviewed and approved by the Ethics Subcommittee for Research in Human Beings of the Faculty of Medicine of the Pontifical Catholic University of Ecuador under the code SB-CEISH-POS-691. The committee determined that informed consent was not required for this study.
Results
A comprehensive demographic and health snapshot of children aged 5 to 11 years is presented in Table 1 . In the study sample comprising 10807 children aged 5 to 11 from Ecuador, it was found that the mean age of the children was 8.0 years (standard error [SE]: 0.03). About ethnicity, Mestizo (mixed ethnic background) children were predominant, making up 80.8% of the sample. When stratified by economic quintiles, approximately a quarter of the children (25.4%) belonged to the lowest income quintile. Educationally, it was noted that the vast majority (98.3%) were currently attending elementary school. Regarding dietary habits, a significant proportion of the sample (64.3%) reported consuming food provided by the school, and nearly half (48.4%) indicated a reduction in the consumption of processed foods with a red label. Notably, 36.0% of the children were identified as being overweight or obese, according to the Body Mass Index (BMI) adjusted for age and sex, using the WHO Growth References.
10.1371/journal.pone.0296538.t001
Table 1 Description of the sample.
Variable
Whole sample (n = 10807)
Weighted percentage or mean (SE)
No.
Sex (Male)
5541
50.3
Age of the children
10807
8.0 (0.03)
Ethnicity
Ethnicity (Indigenous)
1252
7.2
Ethnicity (Afroecuadorian)
459
4.8
Ethnicity (Mestizo-mixed ethnic background)
8536
80.8
Ethnicity (White)
141
1.3
Ethnicity (Montubio or other)
419
5.9
Economic quintiles by income a
Economic quintile by income (1st quintile)
2865
25.4
Economic quintile by income (2nd quintile)
2344
23.6
Economic quintile by income (3rd quintile)
2096
20.4
Economic quintile by income (4th quintile)
1767
16.7
Economic quintile by income (5th quintile)
1608
13.9
Education of children (Elementary school ongoing)
10618
98.3
Regular class attendance (Yes)
10764
99.6
The head of the household receives the BDH (Yes)
455
4.4
Number of people in the household
10807
5.0 (0.03)
Inadequate disposal of excreta (Yes)
2570
24.1
Regular physical activity (Yes)
1389
13.5
Perception of low consumption of vegetables (Yes)
5218
47.1
Days per week of consumption in fast food restaurants
10807
0.9 (0.02)
Days per week of school food consumption
7930
4.4 (0.03)
The child buys food at school (Yes)
7271
64.1
The child eats the food provided by the school (Yes)
7103
64.3
Family members recognize, understand, and use the labelling of processed foods (Yes)
6635
65.4
In the family, they reduced the consumption of processed foods with a red label (Yes)
4256
48.4
Overweight or obesity b
3931
36.0
BDH = Human Development Voucher, by its Spanish spelling.
SE = Standard error.
a Income quintiles are calculated at the household level using monetary labour income per capita, first calculating the total income for each income earner. This total income includes earnings from work, income from investments, transfers, and other benefits, such as cash social transfers. Once we add all these up, we obtain the total household income. Then, we determine the average income per person (per capita income) by dividing the total household income by the number of people in each household. Subsequently, the population is systematically arranged on the basis of the per capita income variable. The calculation of the quintiles was performed by dividing the population into five equal groups, known as quintiles. The first quintile includes the percentage of households with the lowest income, the second quintile includes the next percentage, and so on until the fifth quintile, which includes the percentage of households with the highest income.
b Overweight and obesity were determined by calculating the Body Mass Index (BMI), adjusted for age and sex, according to the WHO growth references. Overweight is BMI-for-age greater than 1 standard deviation above the WHO Growth Reference median; and obesity is greater than 2 standard deviations above the WHO Growth Reference median.
We found distinct differences between non-overweight or non-obese children and their overweight or obese counterparts ( S2 Table ). In our bivariate crude comparison of characteristics between non-overweight/non-obese children (n = 6876) and those identified as overweight or obese (n = 3931), we observed several notable differences. Among these, the average age of children classified as overweight or obese was marginally higher, at 8.3 years, compared to 7.9 years for their non-overweight peers. This age discrepancy was statistically significant, as evidenced by a p-value of less than 0.001 in the z test. Ethnic distribution showed that 81.7% of the overweight or obese groups were Mestizo (mixed ethnic background), compared with 80.3% in the non-overweight group (p-value = 0.602, Pearson’s chi-squared). Economic stratification revealed that 20.6% of children were from the lowest income quintile among the overweight or obese compared to the 28.2% among those non-overweight (p-value < 0.001, Pearson’s chi-squared). Furthermore, a greater proportion of overweight or obese children (65.7%) consumed food provided by the school, in contrast to 63.2% of the non-overweight children (p-value = 0.124, Pearson’s chi-squared). Finally, a difference was detected in the perception of reduced consumption of processed foods with a red label, with 50.4% of overweight or obese children indicating consumption, compared to 47.2% of their non-overweight peers (p-value = 0.097, Pearson’s chi-squared).
After running multivariate logistic regression models, we found that, several factors were significantly associated with childhood overweight and obesity ( Table 2 ). According to the final adjusted model, several variables exhibited statistically significant associations between several explanatory variables and being overweight or obese. Notably, male children exhibited a higher likelihood of being OW/OB, with 1.26 times increased adjusted odds (95% CI: 1.20 to 1.33) compared to female children. In addition, for every yearly increase in a child’s age, the odds of being overweight or obese increased by 1.10 times (95% CI: 1.09 to 1.10). When broken down by ethnicity, compared to Indigenous children, the Afroecuadorian ethnicity presented a slightly elevated but not statistically significant odds of 1.12 (95% CI: 0.99 to 1.26), while Mestizo children showed 1.14 times increased odds (95% CI: 1.04 to 1.25). White children and those from Montubio (mixed ethnic background of coastal Ecuador) or other ethnicities did not demonstrate statistically significant differences in this model. When considering economic quintiles by income, children in the 2nd quintile demonstrated 1.17 times higher odds (95% CI: 1.07 to 1.31), those in the 3rd quintile showed 1.33 times (95% CI: 1.11 to 1.59), in the 4th quintile it was 1.39 times (95% CI: 1.18 to 1.65), and in the 5th quintile, the odds were 1.39 times higher (95% CI: 1.29 to 1.51) compared to those in the 1st quintile. An increase in the number of household members corresponded to a slight reduction in odds by 0.93 times for each additional person (95% CI: 0.91 to 0.95). Moreover, children with inadequate disposal of excreta exhibited 0.82 times lower odds of being overweight or obese (95% CI: 0.76 to 0.90). Similarly, regular physical activity was associated with reduced odds, at 0.79 times (95% CI: 0.75 to 0.82). Interestingly, children from families that recognized and used processed food labels exhibited a higher likelihood of being overweight or obese, with an adjusted odds ratio (aOR) of 1.14 (95% CI: 1.02 to 1.26). Conversely, the consumption of food provided by schools was linked with a non-significant reduction in the risk of overweight or obesity, with an aOR of 0.93 (95% CI: 0.82 to 1.06).
10.1371/journal.pone.0296538.t002
Table 2 Crude and adjusted Odds Ratios of overweight or obesity from each explanatory variable using multilevel and logistic regression models.
Variable
Crude models
Adjusted models
OR (IC95%)
p-value
Saturated
p-value
Parsimonious
p-value
aOR (IC95%)
aOR (IC95%)
Male sex (female is the ref.)
1.24 (1.18 to 1.31)
<0.001
1.23 (1.08 to 1.40)
0.002
1.26 (1.20 to 1.33)
<0.001
Age of the child (per each increase in one year)
1.10 (1.08 to 1.12)
<0.001
1.09 (1.08 to 1.10)
<0.001
1.10 (1.09 to 1.10)
<0.01
Ethnicity
Ethnicity (Indigenous is the ref.)
1
-
1
-
1
-
Ethnicity (Afroecuadorian)
1.09 (0.90 to 1.32)
0.381
0.99 (0.86 to 1.15)
0.917
1.12 (0.99 to 1.26)
0.062
Ethnicity (Mestizo—mixed ethnic background)
1.25 (1.06 to 1.49)
0.009
1.06 (0.88 to 1.27)
0.560
1.14 (1.04 to 1.25)
0.004
Ethnicity (White)
1.39 (1.05 to 1.85)
0.021
1.17 (0.80 to 1.71)
0.412
1.26 (0.95 to 1.67)
0.114
Ethnicity (Montubio or other)
1.01 (0.83 to 1.24)
0.903
0.82 (0.76 to 0.89)
<0.001
0.97 (0.83 to 1.14)
0.718
Basic education of the children (no formal education is the ref.)
2.52 (0.69 to 9.18)
0.160
0.82 (0.50 to 1.35)
0.433
Economic quintiles by income a
Economic quintile by income (1st quintile is the ref.)
1
-
1
-
1
-
Economic quintile by income (2nd quintile)
1.21 (1.14 to 1.29)
<0.001
1.13 (1.07 to 1.19)
<0.001
1.17 (1.12 to 1.21)
<0.001
Economic quintile by income (3rd quintile)
1.47 (1.24 to 1.75)
<0.001
1.32 (1.05 to 1.64)
0.015
1.33 (1.11 to 1.59)
0.002
Economic quintile by income (4th quintile)
1.59 (1.36 to 1.86)
<0.001
1.34 (1.10 to 1.63)
0.004
1.39 (1.18 to 1.65)
<0.001
Economic quintile by income (5th quintile)
1.67 (1.50 to 1.85)
<0.001
1.41 (1.19 to 1.67)
<0.001
1.39 (1.29 to 1.51)
<0.001
p-for-trend
1.15 (1.10 to 1.19)
<0.001
1.09 (1.05 to 1.14)
<0.001
1.09 (1.05 to 1.14)
<0.001
Regular class attendance (otherwise is the ref)
1.48 (0.94 to 2.34)
0.091
0.67 (0.26 to 1.71)
0.403
-
-
The head of the household receives the BDH (otherwise is the ref.)
0.77 (0.74 to 0.81)
<0.001
0.97 (0.86 to 1.08)
0.563
-
-
Number of people in the household (per each extra person)
0.92 (0.91 to 0.93)
<0.001
0.94 (0.91 to 0.97)
<0.001
0.93 (0.91 to 0.95)
<0.001
Inadequate disposal of excreta (otherwise is the ref.)
0.73 (0.68 to 0.78)
<0.001
0.87 (0.75 to 1.00)
0.058
0.82 (0.76 to 0.90)
<0.001
Regular physical activity (otherwise is the ref.)
0.76 (0.75 to 0.78)
<0.001
0.77 (0.76 to 0.78)
<0.001
0.79 (0.75 to 0.82)
<0.001
Perception of low consumption of vegetables (otherwise is the ref.)
0.91 (0.79 to 1.05)
0.196
1.01 (0.91 to 1.13)
0.824
-
-
Days per week of consumption in fast food restaurants (per each extra day of consumption)
1.05 (1.03 to 1.07)
<0.001
1.01 (0.96 to 1.06)
0.810
-
-
Days per week of school food consumption (per each extra day of consumption)
0.98 (0.94 to 1.02)
0.318
0.99 (0.93 to 1.06)
0.751
-
-
The child buys food at school (otherwise is the ref.)
1.17 (0.95 to 1.44)
0.139
1.10 (0.90 to 1.36)
0.349
-
-
Consumption of food provided by the school (otherwise is the ref.)
0.83 (0.73 to 0.95)
0.007
0.82 (0.74 to 0.90)
<0.001
0.93 (0.82 to 1.06)
0.288
Family members recognize, understand, and use the labelling of processed foods (otherwise is the ref.)
1.24 (1.14 to 1.35)
<0.001
1.14 (1.09 to 1.19)
<0.001
1.14 (1.02 to 1.26)
0.019
In the family, they reduced the consumption of processed foods with a red label (otherwise is the ref.)
1.14 (1.10 to 1.18)
<0.001
1.08 (0.94 to 1.25)
0.260
-
-
BDH = Human Development Voucher, by its Spanish spelling
a Income quintiles are calculated at the household level using monetary labour income per capita, first calculating the total income for each income earner. This total income includes earnings from work, income from investments, transfers, and other benefits, such as cash social transfers. Once we add all these up, we obtain the total household income. Then, we determine the average income per person (per capita income) by dividing the total household income by the number of people in each household. Subsequently, the population is systematically arranged on the basis of the per capita income variable. The calculation of the quintiles was performed by dividing the population into five equal groups, known as quintiles. The first quintile includes the percentage of households with the lowest income, the second quintile includes the next percentage, and so on until the fifth quintile, which includes the percentage of households with the highest income.
In analysing the determinants of overweight and obesity in children, notable differences emerged between genders when running the final parsimonious model ( Table 3 ). For each incremental year in age, a significant association with overweight or obesity was noted in both genders, with the odds of being overweight or obese increasing by an adjusted odds ratio (aOR) of 1.09 (95% CI: 1.07 to 1.11) for women and 1.10 (95% CI: 1.09 to 1.11) for men. Among the ethnicities, White ethnicity was associated with the highest risk in women, with an aOR of 1.57 (95% CI: 1.34 to 1.83). With respect to economic quintiles, women in the 5th quintile exhibited the greatest risk, with aOR of 1.38 (95% CI: 1.13 to 1.70). In relation to other determining factors, it was observed that inadequate disposal of excreta had significant associations with overweight or obesity in both genders. For women, the risk was reduced with an adjusted odds ratio (aOR) of 0.86 (95% CI: 0.80 to 0.92), whereas for men, the reduction in risk was slightly more pronounced with an aOR of 0.81 (95% CI: 0.74 to 0.89). Regular physical activity appeared to be protective against overweight or obesity. Women who engaged in regular physical activity presented a reduced risk, as indicated by an aOR of 0.84 (95% CI: 0.80 to 0.88), whereas men benefitted slightly more from such activity, displaying an aOR of 0.76 (95% CI: 0.70 to 0.83). Notably, in households where processed food labelling was recognized, understood, and utilized, the risk of overweight or obesity increased in both women and men. This association, for women, was statistically significant with an aOR of 1.16 (95% CI: 1.09 to 1.24), and for men, it was not significant (aOR = 1.11; 95% CI: 0.95 to 1.29).
10.1371/journal.pone.0296538.t003
Table 3 Adjusted Odds Ratios of overweight or obesity from each explanatory variable using the parsimonious logistic regression model of Table 2 between women and men.
Variable
Parsimonious model in women
p-value
Parsimonious model in men
p-value
aOR (IC95%)
aOR (IC95%)
Age of the child (per each increase in one year)
1.09 (1.07 to 1.11)
<0.001
1.10 (1.09 to 1.11)
<0.001
Ethnicity
Ethnicity (Indigenous is the ref.)
1
-
1
-
Ethnicity (Afroecuadorian)
1.19 (0.82 to 1.72)
0.374
1.07 (0.93 to 1.23)
0.325
Ethnicity (Mestizo–mixed ethnic background)
1.32 (1.09 to 1.60)
0.005
1.01 (0.97 to 1.06)
0.653
Ethnicity (White)
1.57 (1.34 to 1.83)
<0.001
1.05 (0.64 to 1.70)
0.858
Ethnicity (Montubio or other)
0.94 (0.67 to 1.32)
0.716
0.99 (0.97 to 1.00)
0.196
Economic quintiles by income a
Economic quintile by income (1st quintile is the ref.)
1
-
1
-
Economic quintile by income (2nd quintile)
1.13 (1.10 to 1.17)
<0.001
1.19 (1.14 to 1.24)
<0.001
Economic quintile by income (3rd quintile)
1.27 (0.88 to 1.83)
0.203
1.38 (1.32 to 1.44)
<0.001
Economic quintile by income (4th quintile)
1.31 (1.07 to 1.60)
0.009
1.48 (1.26 to 1.73)
<0.001
Economic quintile by income (5th quintile)
1.38 (1.13 to 1.70)
0.002
1.40 (1.33 to 1.48)
<0.001
p-for-trend
1.09 (1.01 to 1.17)
0.021
1.10 (1.09 to 1.12)
<0.001
Number of people in the household (per each extra person)
0.92 (0.90 to 0.95)
<0.001
0.93 (0.88 to 1.00)
0.041
Inadequate disposal of excreta (otherwise is the ref.)
0.86 (0.80 to 0.92)
<0.001
0.81 (0.74 to 0.89)
<0.001
Regular physical activity (otherwise is the ref.)
0.84 (0.80 to 0.88)
<0.001
0.76 (0.69 to 0.83)
<0.001
Consumption of food provided by the school (otherwise is the ref.)
1.01 (0.98 to 1.05)
0.499
0.86 (0.69 to 1.09)
<0.001
Family members recognize, understand, and use the labelling of processed foods (otherwise is the ref.)
1.16 (1.09 to 1.24)
<0.001
1.11 (0.95 to 1.29)
0.179
BDH = Human Development Voucher, by its Spanish spelling
a Income quintiles are calculated at the household level using monetary labour income per capita, first calculating the total income for each income earner. This total income includes earnings from work, income from investments, transfers, and other benefits, such as cash social transfers. Once we add all these up, we obtain the total household income. Then, we determine the average income per person (per capita income) by dividing the total household income by the number of people in each household. Subsequently, the population is systematically arranged on the basis of the per capita income variable. The calculation of the quintiles was performed by dividing the population into five equal groups, known as quintiles. The first quintile includes the percentage of households with the lowest income, the second quintile includes the next percentage, and so on until the fifth quintile, which includes the percentage of households with the highest income.
After conducting an analysis excluding children in the highest quintile, children with parents receiving the BDH, and those residing in households with a larger number of individuals, the observed associations maintained similar trends. However, the relationships were less statistically significant, as detailed in S3 Table .
Discussion
Principle findings
Increasing age, male gender, mestizo (mixed ethnic background) ethnicity, higher economic quintiles, inadequate disposal of excreta, and lack of physical activity are factors associated with a higher likelihood of overweight or obesity in children aged 5 to 11 years in Ecuador. The impact of consuming school-provided meals was inconclusive. Children from families with a higher number of individuals in the household and with families that recognize and use processed food labels exhibited a higher likelihood of being overweight or obese.
Comparison with the literature
The prevalence of OW/OB in school children places the country in the eighth position in the Americas, following countries such as Mexico, Chile, Panama, and the United States, and surpassing the prevalences in Colombia and Peru [ 23 ]. International data reported in 2016 ranked the country 15th in the same region [ 24 ]. This highlights the drastic increase in prevalence in the absence of effective public policies, compared with other countries that have taken strong measures against childhood overweight and obesity [ 25 , 26 ].
The study reveals that children who purchase food at school are at a greater risk of being overweight or obese than those who consume meals provided by the school, but this difference is not statistically significant. Despite that, it is important to mention that prior research from the ENSANUT 2018 survey, indicated a link between school foods, particularly those sold in stores (73%), and elevated BMI [ 27 ]. A plausible explanation is that 60% of these stores provide unhealthy products labelled with a “red traffic light”. The lack of significant results may stem from the Ecuadorian school feeding program, which relies on processed and packaged products mandated to bear nutritional labelling (traffic light system) [ 27 ]. These products contain high levels of sugar, salt, and fats similar to the ultra-processed items distributed by food industries in formal markets [ 27 ].
The study underscores the crucial role of the school food environment in addressing childhood obesity. Despite robust evidence [ 28 ], many Latin American schools, including those in Ecuador [ 29 ], lack effective public health policies. Practical recommendations include prohibiting the advertisement and sale of processed foods in and around schools, imposing higher taxes on sugar-rich, ultra-processed products, and promoting the consumption of fresh fruits and vegetables, along with education on food labelling [ 30 ].
In a broader epidemiological context, the gradual increase in the prevalence of OW/OB in the population of this study, along with a population of adults who have been overweight and obese for decades, coupled with a distribution of excess malnutrition across all socioeconomic strata in both children and adults [ 7 ], places Ecuador in a second stage of transitioning to obesity [ 6 ]. It is noteworthy that the richest quintile of households has a lower prevalence of OW/OB than the quintile immediately below, which could indicate that Ecuador is entering the third stage of transitioning to obesity, where excess malnutrition decreases in wealthier households and becomes concentrated in socioeconomically vulnerable households [ 6 ]. Additionally, the observation that children residing in homes with inadequate excreta disposal are less likely to be OW/OB underscores the presence of a socioeconomic gradient in obesity and overweight prevalence in Ecuador. This characteristic is compounded by the fact that the country has not overcome malnutrition in early childhood [ 7 ], further complicating the issue of overall malnutrition, resulting in a country experiencing a double burden of malnutrition (malnutrition due to deficiency and excess).
The prevalence patterns of OW/OB in Ecuadorian school children are similar to those in other low and middle-income countries in the region. Similar to a study of school children in Mexico, there is a higher prevalence of OW/OB in boys than in girls, in the non-indigenous population, and in higher-income households, and OW/OB also increases with age [ 31 , 32 ]. However, the results of this study differ from patterns seen in high-income countries in the region such as Chile, where OW/OB is more prevalent in socioeconomically vulnerable families [ 33 ]. It is challenging to compare the determinants found in this study with those in other countries in the region because of to the scarcity of published studies [ 34 ].
The fact that children from households self-identified as mestizo have higher rates of OW/OB may be explained by the fact that these children tend to live in urban environments where they are more exposed to obesogenic environments such as advertisements for sugary beverages, ultra-processed foods, and junk food. This would also explain why indigenous children are protected from obesity and overweight, in addition to socioeconomic vulnerabilities in this population, along with household size, which is more strongly linked to malnutrition due to deficiency rather than excess in the country [ 13 , 35 ]. Therefore, further research is needed on the circumstances of social vulnerability and food insecurity in households of school children with OW/OB to explain why certain negative socioeconomic factors seem to protect against OW/OB.
Physical activity among schoolchildren acts as a protective factor against OW/OB [ 36 , 37 ]. However, the percentage of children in the sample who engage in regular physical activity is very low, which aligns with previously published findings [ 38 ]. There is limited research linking physical activity to excess malnutrition in this group. Therefore, further research on this topic is necessary [ 39 ].
Regarding gender differences, our findings support the current understanding of gender-based dietary segregation, both in Ecuador and globally. In Ecuador, prevailing dietary customs, shaped by a patriarchal and sexist societal framework, often prioritize feeding boys. This may partly account for why physical activity appears to benefit boys more significantly in reducing the likelihood of being overweight or obese, as compared to girls, where such activities do not seem to diminish the probability of overweight or obesity as effectively [ 40 ]. This is exacerbated by the inclination to maintain purely aesthetic standards, which results in food restrictions for girls, especially as they enter adolescence [ 41 ]. Additionally, there are existing and differential unhealthy exposures based on socioeconomic strata [ 41 ].
Our analysis uncovers a 14% increase in obesity and overweight risk among individuals interacting with the nutritional traffic light labelling system, casting light on its ineffectiveness in reducing the intake of processed and ultra-processed foods. This finding not only challenges the system’s foundational goal but also implies a significant, albeit counterintuitive, relationship between frequent exposure to the labelling system and an elevated prevalence of obesity and overweight. Such an association suggests that engagement with the system, rather than mitigating risks, may inadvertently contribute to higher obesity rates. Additionally, our research highlights a stark gender disparity in the impact of the labelling system, with women experiencing a markedly higher risk of obesity and overweight than men. This gender-based difference underscores the unique vulnerability of women, potentially linked to either limited [ 42 ] or less intense physical activity opportunities [ 43 ], thus exacerbating their risk in the face of processed food consumption.
Given Ecuador’s extensive history of childhood chronic malnutrition, a pivotal risk factor for overweight, obesity, and non-communicable diseases, coupled with the severe impacts of climate change linked to environmental factors, it is imperative to focus research and public policies on understanding the global syndemic of malnutrition. This approach should incorporate triple-duty actions, considering the distinctive aspects of diverse life stages [ 44 ].
Our findings suggest that public health policies should place greater focus on improving the quality of available foods within schools to mitigate the risks of childhood overweight and obesity.
In the context of Ecuador and the wider region, our study underscores the criticality of scrutinizing policy missteps that deregulate the marketing of unhealthy foods and other harmful products, thereby hindering the enhancement of food environments. Remarkably, the President of Ecuador enacted Presidential Decree No. 645 on January 10, 2023, which aims to slash taxes on known health hazards, such as alcohol, tobacco, sugary beverages, and firearms, in stark contrast to advancing public health policies [ 14 ]. This decree is poised to adversely affect collective health. This underscores a significant lacuna in public health strategies, potentially exacerbating the obesity and overweight crisis among school children. This situation urgently demands the attention of authorities and policymakers to realign regulations with public health principles, ensuring that economic interests and conflicts of interest do not undermine the development of sound public health policies [ 45 , 46 ]. In this context, it becomes imperative for Ecuadorian authorities to implement and strengthen comprehensive policies that address both the quality of food offered within schools and those sold in their vicinity.
The congruence between the primary and sensitivity analyses bolsters the robustness of our findings. Although the consumption of food purchased at school and food provided by the school did not attain statistical significance in the model adjusted for expansion factors, the overall pattern of results remained consistent. Importantly, the existing scientific literature supports our findings concerning these variables. It is also crucial to emphasize that there are additional food environments associated with formal markets and marketing targeted at children, which were not considered in this study but significantly contribute to childhood OW/OB. Consequently, it is imperative that national health and nutrition surveys not only possess sensitivity to growth retardation but also to excess malnutrition. While the study delves into various sociodemographic factors, some potentially relevant variables, such as cultural practices or parental education, are not thoroughly examined. It is worth mentioning that not all information from different forms can be cross-referenced for analysis because of variations in the design and objectives of the national survey.
We believe that our findings necessitate authorities to contemplate public policies aimed at reducing the burden of OW/OB from an early age, thereby diminishing future disabilities and deaths, which evidently result in heavier economic costs and social burdens.
The study calls for comprehensive policy actions to curb the marketing and availability of unhealthy foods, including strict regulation of advertising and sales practices for sugary beverages and other unhealthy products. It critiques policy missteps like Presidential Decree 645, which reduce taxes on health hazards, potentially exacerbating the obesity crisis by making such products more affordable. Highlighting the need for regulations to adhere to public health principles, the research advocates for a multifaceted strategy to improve food quality in and around schools, emphasizing the importance of evidence-based policymaking and prioritizing public health to effectively tackle childhood overweight and obesity in Ecuador.
Strengths and limitations
The foremost strength of our study lies in its utilization of the nationally representative ENSANUT 2018 survey, which significantly enhances the credibility of our analysis and offers a genuine reflection of the child population in Ecuador. By adopting WHO criteria for evaluating overweight and obesity, we not only ensure the accuracy of our measurements but also improve the comparability of our findings regarding their prevalence and determinants with those of other nations and contexts. Additionally, our use of multilevel modelling techniques in the statistical analysis deepens our insight, adeptly addressing the complexity of the data structure, such as the distribution of children across various regions. This approach allows for a nuanced exploration of the collective impact of diverse factors on the rates of overweight and obesity. Moreover, the extensive scope of the ENSANUT 2018 survey, encompassing a broad array of demographic, socioeconomic, and health-related factors reflective of conditions common to numerous countries beyond Ecuador, supports the external validity of our study. This suggests that our findings have the potential to be applicable in other nations with analogous contexts. Consequently, our research provides critical epidemiological insights into childhood overweight and obesity in Ecuador, while also serving as an invaluable tool for guiding public health strategies in other countries confronting similar health issues.
Given the reliance of this study on secondary analysis of existing survey data, it is crucial to acknowledge the limitations posed by the variables available from the survey, which may not capture the full spectrum of contextual influences on childhood obesity. These include cultural, environmental, and policy-related factors that significantly affect obesity prevalence among children. Additionally, the potential exists for other important, yet unmeasured or unconsidered, factors to contribute, such as genetic predispositions, past medical conditions (e.g., hypothyroidism), medication use, detailed eating behaviours, parental practices, and community-level interventions. Moreover, while our study establishes correlations between certain demographic and socioeconomic factors and childhood obesity, the observational cross-sectional design precludes the determination of causality.
Conclusions
Increasing age, male gender, mestizo (mixed ethnic background) ethnicity, higher economic quintiles, inadequate disposal of excreta, and lack of physical activity are factors associated with a higher likelihood of overweight or obesity in children aged 5 to 11 years in Ecuador. The impact of consuming school-provided meals was inconclusive. Children from families with that recognize and use processed food labels exhibited a higher likelihood of being overweight or obese; this indicates that the nutritional traffic light labelling system, contrary to its intended purpose, is linked to a 14% heightened risk of obesity and overweight, especially among women, highlighting its limited efficacy and underscoring the urgency to reinforce this public health strategy. Our findings underscore the need for a critical reassessment of Ecuador’s public health policies, emphasizing the improvement of school food quality and stricter regulation of unhealthy product marketing to mitigate childhood overweight and obesity risks through school-based dietary interventions.
Supporting information
S1 Data
(CSV)
S1 Table
STROBE statement—Reporting checklist for cross sectional study.
(DOCX)
S2 Table
Characteristics comparison between non-overweight/non-obese and overweight/obese children.
Each category’s percentages highlight the proportional differences between the two groups.
(DOCX)
S3 Table
Adjusted Odds Ratios of overweight or obesity from each explanatory variable using the parsimonious logistic regression model of Table 2 after excluding: (i) people categorized in the lowest income quintile, (ii) people categorized in the highest income quintile, (iii) people who receive the BDH, (iv) people within the upper third of the highest number of people per household.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
About 88 years ago, Moore and his collaborators have provided the evidence for the first time that the scrotum is a local thermo-regulator and the temperature environment of the testes, which is below that of the general body temperature, is essential for the occurrence of normal spermatogenesis [1] . Retention of testes inside the abdomen, namely cryptorchidism, often observed in human birth defect regarding male genitalia, can result in disruption of spermatogenesis and impairment of steroidogenesis [2] , [3] . Similarly in mice, precoital testicular heating not only reduced the number of successful matings, but also produced a transient retardation in embryo growth [4] , [5] . Besides scrotum, a second thermoregulatory system is located in the spermatic cord where there is a counter-current heat exchange between incoming arterial blood and outgoing venous blood [6] , [7] . The pathological dilation of testicular veins and pampiniform plexus, usually called varicocele, can elevate intrascrotal temperatures and thus is frequently associated with decreased conception rates among infertile couple [8] , [9] . In short, accumulated studies have conclusively documented the adverse effects of hyperthermia on the normal adult testis among different species. Heat stress can result in disruption of the seminiferous epithelium, accumulation of lipid in Sertoli cells (Ser), local dilations of the intercellular spaces between Ser junctions, and increased apoptotic rate [10] . Based on histopathology, primary spermatocytes are the most susceptible cell type [11] . Although the physiological and cellular responses to heat treatment of the testes have been well documented, the molecular mechanisms through which these responses are directed remain largely unknown.
Metastasis-associated protein 1 (MTA1), a ubiquitously expressed chromatin modifier, plays an intergral role in nucleosome remodeling and histone deacetylase (NuRD) complexes [12] , [13] . Soon after its identification, it became apparent that, in addition to its well-proven correlation with metastatic potential, MTA1 is also involved in the regulation of non-histone proteins by modifying the acetylation status of crucial target genes [14] . For example, during DNA double-strand break (DSB) repair in response to IR, MTA1 could directly stabilize the p53 protein by inhibiting its ubiquitination by E3 ligases, and therefore regulates p53-dependent function in DNA repair [15] . Interestingly, MTA1 could also inhibit p53-induced apoptosis by deacetylating p53, resulting in a more metastatic state in human cancer cells [16] . One possible explanation for these conflicting observations is that MTA1 may serve distinct roles in different physiopathological systems in response to different stimulations.
Our laboratory has a long-standing interest in the potential involvement of MTA1 during spermatogenesis. MTA1 protein is gradually increased in the testis since 14 days postnatal and reaches the maximum in adults, reversely correlating to the appearance of p53 in murine testis [17] , [18] . Moreover, we have demonstrated that overexpression of MTA1 in vitro could remarkably elevate the capability of spermatogenic tumor cells against heat-induced apoptosis [19] . To better understanding the physiological function of MTA1 during normal spermatogenesis, we addressed here the distinctive interaction and regulatory roles between MTA1 and p53. Our data indicate that endogenous MTA1 can serve as a transient protector of primary spermatocytes against heat-stress induced apoptosis and therefore might act as a novel coregulator of p53 in maintenance of cellular integrity during early phase after hyperthermal stimulation in testicular germ cells.
Results
Tetraploid accumulation of MTA1 is not required for meiotic division but enhanced during acute phase of heat-stress induced apoptosis
MTA1 is mainly expressed in tetraploid primary spermatocytes and is slightly expressed in Ser of mouse testis, which suggests a potential involvement in meiosis [17] . To determine whether the predominant expression level of MTA1 is the prerequisite for the meiotic divisions of mouse spermatocytes, we treated isolated primary spermatocytes ( Figure S1 ) in culture with okadaic acid (OA). This serine/threonine phosphatase inhibitor triggers entry of pachytene spermatocytes into meiotic divisions, which can be visualized by the overwhelming expression of phosphoH3 ( Figure S2 ) [20] . Surprisingly, brief introduction of mid- to late pachytene spermatocytes to proceed to metaphase I was unable to stimulate the expression level of MTA1 at either transcriptional or translational levels ( Fig. 1A and 1B ). Unchanged nuclear localization of MTA1 after OA treatment also excludes the potential of being a ribonucleoprotein (RNPs) to favor mRNA translation during meiosis ( Fig. 1C ) [21] . MTA1 has been shown able to regulate the p53-dependent transcription for supplying nucleotides to repair damaged DNA in response to IR [15] . This raises the possibility that MTA1 might also play as a DNA damage responsive protein in coordination with p53 under hyperthermal condition. To preliminarily explore this assumption, we incubated primary spermatocytes at 43°C to induce heat stress and then analyzed the expression levels of MTA1 and of p53. There was a mild increase of MTA1 expression 0.25 h right after heat stress. This upregulation sustained for about 1 h before a dramatic decrease was observed at 2 h after heat stress ( Fig. 1D ). We also confirmed this expression profile at the transcriptional level ( Figure S3 ). Consistent with the previous reports [22] , [23] , p53 expression gradually increased upon heat stimulation (first appearance of p53 expression was detected 0.25 h after heat stress). Acetylated status and functionality of p53 were also elevated from 0.5 h after heat stress, as evidenced by the expression level of Ac-p53 and p21. These data collectively suggested that there is a reverse correlation between the expression level of MTA1 and the activity of p53 induced by acetylation modification in tetraploid primary spermatocytes during acute phase of heat stimulation.
10.1371/journal.pone.0026013.g001
Figure 1
Expression of MTA1 was not required for meiotic division but was enhanced during acute phase of heat-stress induced apoptosis.
A Expression of MTA1 in isolated primary spermatocytes after OA treatment was examined using RT-PCR (left panel). PCR products were then quantified by SYBR green intercalation in real-time PCR (right panel). 18S was served as an internal control for each PCR amplification. Data are expressed as mean±SEM (n = 3; p>0.05 vs. control). B Western analysis of MTA1 expression in isolated primary spermatocytes after OA treatment. The blot was reprobed with a β-actin (42 kDa) antibody to confirm equal loading. C Subcellular localization of MTA1 in isolated primary spermatocytes after OA treatment was determined by immunofluorescent staining. DAPI was used to stain nuclei of cells. Bar = 10 µm D Expression levels of MTA1, p53, Ac-p53 and p21 in primary spermatocytes were revealed at different time-points after heat stress by western blotting. β-actin was used as internal control.
MTA1 protects spermatocytes from temperature induced apoptosis by upregulation of HDAC2-mediated decetylation of p53
Next, we try to figure out whether the endogenous MTA1 can perform certain roles in spermatogenic cells during apoptotic process induced by deregulated temperature. To address this question, we took advantage of a temperature sensitive spermatocyte-like cell line, namely GC-2spd (ts), which allows for full p53 nuclear localization at 32°C and partial p53 nuclear localization at 37°C but none at 39°C. The activity of the protein is determinant on the correct folding of p53, which is inhibited at the higher temperatures [24] . Moreover, these cells were derived from spermatocytes, the germ cell type very sensitive to heat-induced apoptosis [25] . Thus, GC-2spd (ts) affords experimentally an appropriately functional p53+/+ or p53 -/- in vitro model. After inhibiting about 78% expression level of MTA1 /MTA1 (When compared to the internal control 18S , the relative expression level of MTA1 after RNAi was 0.2243±0.07, p = 0.0126) as demonstrated by real-time PCR and western blot analyses in Figure S4B, S4C , we subjected the GC-2spd (ts) into apoptotic permission temperature (32°C). The decrease of MTA1 expression notably increased the apoptotic susceptivity as histologically demonstrated by the elevated positive signals in TUNEL staining ( Fig 2A ). ELISA methodology also confirmed that the apoptotic rate of KD cells was significantly higher than that of control cells after heat stress ( Fig. 2B ). This apoptotic rate was trichostatin A (TSA)-sensitive and could be elevated when we achieved less expression of MTA1 by using higher level of siRNA ( Fig. 2B , Figure S4D ), confirming the indispensable role of MTA1 upon heat stress. Among the different candidates influencing the apoptotic process during spermatogenesis, p53 is believed to be a guardian of genome integrity [26] . MTA1 could inhibit p53-induced apoptosis by deacetylating p53 in cancer cells [16] . Therefore, we then focused our study on the potential relationship between these two molecules during heat stress induced apoptosis. As shown in Fig. 2C , the elevation of MTA1 expression induced by incubation at 32°C was not detected in knockdown (KD) groups. Surprisingly, inhibition of MTA1 expression could only affect the deacetylation status and functionality of p53 (as evidenced by the increasing expression level of Ac-p53 and of p21, respectively), instead of total p53 level. This indicated that endogenous MTA1 may exert effects on the activity of p53 at the post-translational level, probably through deacetylation modification. To further confirm that the p53 gene chromatin is a direct target of MTA1, we performed ChIP assays in GC-2spd (ts) cultured at 32°C ( Fig. 2D ). Both MTA1 and HDAC2 could be recruited to p53 promoter in a TSA-sensitive manner, suggesting the possible involvement of nucleosome remodeling and histone deacetylase complex in the noted MTA1 regulation of p53 transcription. After we knocked down the MTA1 expression, the endogenous association between p53 and HDAC2 induced by permission temperature was greatly reduced, reversely consistent with the evoked activity of p53 (as demonstrated by the increased expression level of p21) instead of total p53 level ( Fig. 2E ). Taken together, these findings identify MTA1 as a post-translational regulator of p53 through deacetylation modification upon heat stimulation.
10.1371/journal.pone.0026013.g002
Figure 2
Impairment of the endogenous MTA1 in GC-2spd (ts) upregulated the acetylation of p53 by diminishing the recruitment of HDAC2 and led to an increase of apoptosis after temperature switch.
A TUNEL staining of GC-2spd (ts) culture at 32°C for 6 h revealed a higher apoptotic wave due to the interference of MTA1 expression. Inserted phase images showed an equal cell density in two groups. Bar = 20 µm B Higher apoptotic rate in MTA1 RNAi GC-2spd (ts) was confirmed by ELISA methodology. C Inhibition of MTA1 expression only affected the deacetylation status and functionality of p53 instead of total p53 level. D Recruitment of MTA1 and HDAC2 to p53 promoter after heat stress was in a TSA-sensitive manner as shown by CHIP assay. E Endogenous association between p53 and HDAC2 induced by permission temperature (32°C) in GC-2spd (ts) was significantly decreased after suppression of MTA1 expression. Expression levels of Ac-p53 and p21 were also monitored in D and E .
In vivo suppression of MTA1 increases heat-stress induced apoptotic rate and thereafter impairs the spermatogenic differentiation
The mammalian spermatogenesis carries out an amazing biological process. Significant knowledge about the molecular mechanisms of this process has been gained mainly by using genetically engineered animals, which however, are usually time-consuming and labor-intensive [27] . In vivo electroporation of the mammalian testis could reduce this burden because a gene of interest can be easily interrupted in target cells with adequate electric shock, and their behavior can be assayed directly [28] . To this end, we decided to use RNAi approach to knock down MTA1 gene expression in vivo to further elucidate its biological function. As shown in Fig. 3A and 3B , MTA1 expression in the mouse testis was suppressed by 48.06±18.83% after 48 h by RNAi treatment as compared with the control group. This inhibition could be maintained until 72 h after treatment regardless of body metabolism (data not shown). Immunohistochemical examination revealed two cell types, namely pachytene spermatocytes (pachy) and Ser, which were deprived of MTA1 expression to the maximum extent ( Figure 3C , Table 1 , Figure S5 ). In line with the in vitro data, suppression of MTA1 expression failed to induce notable morphological difference under normal condition ( Fig. 3D , HE staining) (To be noted, the tissue damages resulted from the injection of the plasmid and the electroporation manipulation have been taken into consideration when carrying out the histological examination). Instead, we found more degenerated seminiferous tubules (marked by asterisks in HE staining or by a pound key in TUNEL staining of Fig. 3D ) in KD group 48 h after a single, transient scrotal heat stress stimulation. Further quantitative evaluation confirmed that the increased ratio of degenerated tubules in KD group was statistically significant ( Fig. 3E ). This degeneration was likely due to the upregulated apoptotic frequency as evidenced by TUNEL assay ( Fig. 3D ) and increased level of cleaved caspase-3 ( Fig. 3F ), respectively. Because the single testicular heating has been established as a reversible approach for impairment of spermatogenesis [29] , we were then keen to examine whether this elevated apoptosis beard any biological effects. We explored the expression levels of genes known to be sequentially tuned in spermatocytes development. Proacrosin mRNAs are known to be expressed in mid-pachytene spermatocytes, and Sprm-1 and cyclin A1 mRNAs appear at the end of prophase of meiosis. Expression of the Gapd-s gene only begins in spermatids [30] . In accord with the deregulation of MTA1 and the increased apoptotic rate in KD tubules, transcripts from the Proacrosin , Sprm-1 and cyclin A1 were practically decreased compared to the relatively steady level of Gapd-s mRNA ( Fig. 3G ). These data together suggested that the protective effect of MTA1 against heat stress induced apoptosis in primary spermatocyte was required for normal differentiation.
10.1371/journal.pone.0026013.g003
Figure 3
Increased apoptosis in a single, transient scrotal heat stress model after in vivo interference of MTA1 expression impaired the differentiation of primary spermatocytes.
A Relative expression level of MTA1/18S was analyzed using real-time PCR 48 h after siRNA treatment. Data are expressed as mean±SEM (n = 3; * p<0.05 vs. control). B Reduced expression of MTA1 was also confirmed by western blotting analysis. β-actin was used as internal control for equal loading. C Immunolocalization of MTA1 in mouse testicular sections treated with control siRNA and MTA1 siRNA. Abbreviations: Ser, Sertoli cell; pachy, pachytene spermatocyte; Rsd, round spermatid; Esd, elongated spermatid. Bar = 25 µm D In vivo suppression of MTA1 had no effect on the histology of testis but resulted in more damaged seminiferous tubules and higher apoptosis 48 h after transient heat stress as shown by HE and TUNEL staining, respectively. Arrows indicated the apoptotic germ cells. Bar = 50 µm E Western blotting analysis of caspase-3 expression in testes treated with control siRNA and MTA1 siRNA after transient heat stress. F The differentiating markers of primary spermatocytes not those of round spermatids were deregulated in MTA1 siRNA-treated testis 48 after hyperthermal stimulation as demonstrated by RT-PCR. 18S was served as an internal control.
10.1371/journal.pone.0026013.t001 Table 1
Summary of the immunohistochemical analysis of MTA1 in mouse and human testes.
Group
Region/structure (Positive cell type)
Sub-cellular localization
Immunoreactivity
Positive cell number
CT mouse
Spermatogonium
Nuclei
-/2+ (moderate)
[2+] 25–50%
Spermatocyte
Nuclei
3+ (strong)
4+ (>75%)
Spermatid (round spermatid)
Nuclei
[-/+] faint
[-/+] <10%
Spermatozoon
-
-
-
Sertoli cell
Nuclei
1+ (weak)
[1+] 10–25%
Leydig cell
-
-
-
KD Mouse
Spermatogonium
Nuclei
1+ (weak)
[-/+] <10%
Spermatocyte
Nuclei
1+ (weak)
[3+] 50–75%
Spermatid (round spermatid)
Nuclei
[-/+] faint
[-/+] <10%
Spermatozoon
-
-
-
Sertoli cell
Nuclei
[-/+] faint
[-/+] <10%
Leydig cell
-
-
-
Nor Human
Spermatogonium
Nuclei
1+ (weak)
2+ (25–50%)
Spermatocyte
Nuclei
3+ (strong)
3+ (50–75%)
Spermatid (round spermatid)
Nuclei
3+ (strong)
4+ (>75%)
Spermatozoon
-
-
-
Sertoli cell
Nuclei
-/2+ (moderate)
1+ (10–25%)
Leydig cell
-
-
-
Var Human
Spermatogonium
-
-
-
Spermatocyte
-
-
-
Spermatid (round spermatid)
Nuclei
2+ (moderate)
-/+ (<10%)
Spermatozoon
-
-
-
Sertoli cell
-
-
-
Leydig cell
-
-
-
Score for positive cell number: [-/+] <10%, [1+] 10–25%, [2+] 25–50%, [3+] 50–75%, [4+] >75%.
Score for signal intensity: [-/+] faint, [1+] weak, [2+] moderate, [3+] strong.
Impaired expression of MTA1 in hyperthermia related infertile testes
Varicocele is the pathological dilation of testicular veins and pampiniform plexus. The facts that varicoceles are found in a higher percentage among males attending the infertility clinics and that treatment of varicoceles is associated with increased spontaneous conception rates among infertile couples strongly suggest a deleterious effect of varicocele on male reproduction [31] . To date, the pathogenesis of varicocele is far from clear. However, emerging evidences have pointed out the elevated intrascrotal temperature due to retrograde flow in the affected veins as the top ranked aetiology involved [32] , [33] . To better understand the pathological relevance of MTA1, therefore, we extended our investigation into the potential involvement of MTA1 in arrested spermatogenesis at the round spermatid level of varicocele. We observed a relatively lower level of MTA1 expression in pathological testis lysates ( Fig. 4A ). Subsequent immunohistochemical staining revealed a more detailed expression pattern. MTA1 beard the most intensive staining in the nuclear of round spermatid (Rsd) and of pachy in normal group, with a modest signal in the nuclear of some Ser. Weak staining was also observed in the nuclear of spermatogonia (spg). In contrast, elongate spermatids (Esd) and interstitial compartment did not exhibit any MTA1-specific immunostaining ( Fig. 4B ). Overall, the expression level of MTA1 in varicocele group (Var) was relatively lower at the low magnification. The moderate staining could only be detected in partial Rsd. Primary spermatocytes and Ser were deprived of positive signals. Replacement of the primary antibody with normal goat IgG (inserted window in Fig. 4B ) abolished the immunostaining, confirming the specificity of immunohistochemical outcome. The details of immunohistochemical results were demonstrated in Table 1 . Histochemical and biochemical assays have proved that MTA1 has a dominant expression in primary spermatocytes in both human and mouse testes [34] , which led us to hypothesize that MTA1 might play a more conservative role in deciding the fate of tetraploid spermatocytes. To further verify the deteriorating effect of the impaired MTA1 expression on the status of primary spermatocytes, we went back to check the expression profile of MTA1 in the above-mentioned mouse transient scrotal heat stress model. An increase in TUNEL-positive apoptotic spermatocytes was not significantly obvious until 24 h after heat stress ( Figure S7A ). By contrast, the expression level of MTA1 was decreased as early as 8 h after hyperthermal operation as compared to that of control group. In contrast to the MTA1 expression, transient heat stress did not affect the expression of other genes, such as vasa and pgk-2 , which are predominantly expressed in primary spermatocytes and/or round spermatids [35] , [36] , giving strong evidence that the reduction of MTA1 was not due to a decrease in the number of cells ( Figure S7B ). As shown in Figure S7C , most primary spermatocyes at 8 h after heat stress, which had not yet undergone apoptosis as indicated by the normal morphology of chromatin, were deprived of the predominant nuclear staining of MTA1. These results suggested that the decrease of MTA1 expression level could be detected even earlier than the beginning of massive elimination of primary spermatocyes induced by hyperthermia.
10.1371/journal.pone.0026013.g004
Figure 4
Attenuated expression of MTA1 in spermatogenesis-arrested testis (at the round spermatid level) of human varicocele patients.
A Western analysis of MTA1 protein in normal (Nor) and varicocele (Var) testes. The blot was reprobed with β-actin antibody to confirm equal loading. B Immunohistochemical analysis of MTA1 expression in human pathological testes. Replacement of the primary antibody with normal goat IgG (NC) abolished the immunostaining in the tissues, confirming the specificity of the assay. Bar = 20 µm .
Discussion
Given the normal testicular function is temperature dependent, the testes are usually kept between 2 and 8°C below core body temperature by virtue of being held outside the body cavity in the scrotum [37] . However, the temperature of scrotum can be easily affected by numerous external factors such as posture, clothing, lifestyle, occupation and season [38] , [39] . During the complicated process of spermatogenesis, the DNA of the spermatocytes in meiosis is most vulnerable to the introduction of a range of errors [40] . To this end, multifactorial protective mechanisms are warranted in spermatocytes to ensure the maintenance of cellular integrity upon stress conditions, such as DNA damage and hyperthermia. In fact, the expression of a number of DNA repair genes such Ogg1 (involved in base excision repair), Xpg (involved in nucleotide excision repair) and Rad54 (involved in double-strand break repair) are all down-regulated following heat stress at 43°C in murine testis [41] . Emerging literatures also establish MTA1 to be a valid DNA-damage responsive protein [42] . Brief induction of G2/M progression of primary spermatocytes by OA treatment could not upregulate the expression of MTA1, indicating that MTA1 is not necessarily tuned in the pathway of meiotic division triggered by OA. Actually, depletion of MTA1 can result in a defect of G2/M checkpoint in human cancer cells on the premise of UV radiation [43] . Thus, MTA1 may participate in checkpoint pathway under certain pathological conditions. In line with this assumption, our in vitro data showed an attenuated expression profile of MTA1 in primary spermatocytes within 2 h after hyperthermal stimulation, raising the possibility that it may help to maintain the optimum DNA-repair activity in germ cells exposed to heat stress during the acute phase.
In our previous study on the androgen-regulated expression of MTA1 in mouse epididymis, we found an unexpectedly sharp increase of the MTA1 mRNA level at the 5 th day of androgen supplementary treatment in the castrated animals, which was contrasted to the relatively lower circulated androgen level at this time-point [44] . This observation raised the possibility that MTA1 expression was actually modulated by androgen signaling but it may also be transiently evoked by certain pathological conditions. After all, MTA1 is a DNA-damage response gene and castration is an intrinsic stress condition. Similarly in the present study, heat stress could be another stimulator to the upregulation of MTA1 expression during the very early phase. This may explain why MTA1 expression was maintained at relatively high level from 0.25 h to 1 h after heat stress, regardless of a reverse correlation between the expression level of MTA1 and the activity of p53 induced by acetylation modification in tetraploid primary spermatocytes thereafter.
Next, we determined the biological effect of biochemical ablation of MTA1 in spermatocyte-like cells. Disruption of MTA1 expression could introduce remarkable increase of apoptotic rate in GC-2spd (ts) cultured at p53-active temperature. Interestingly, this endogenous inhibition did not impair the total level of p53, which was observed in our prior MTA1-overexpressing system [19] . Two possibilities were suggested for this discrepancy. Firstly, there are not necessarily relevant results capable of being obtained when introducing suppression of endogenous gene or overexpression of exogenous gene. For example, knockdown of candidate dyslexia susceptibility gene (CDSG) homologs in cerebral cortical progenitor cells results in acute disturbances of neocortical migration, while overexpression of CDSG does not have any effect [45] . Secondly, overexpression of exogenous MTA1 could induce at least 4-fold increase of target gene while in the in vitro knockdown GC-2spd (ts), we only detected about 50% reduction of MTA1 expression. It is possible that this overwhelming overexpression might activate other pathway to compromise the expression level of p53. These possibilities are being addressed in ongoing experiments.
Nevertheless, we did find the evidence that suppression of endogenous MTA1 in GC-2spd (ts) could reduce the recruitment of HDAC2 into p53 promoter, thus leading to a less deacetylated status of p53 and a parallel elevated acetylation of p53 ( Fig. 2C, 2D ). The acetylation of p53 is generally believed to be synonymous with the activation of the tumor suppressor, so we reasoned that the inhibition of MTA1 expression could result in the increased p53 activity upon heat stimulation, as evidenced by the increased expression level of p21 and thereafter a higher apoptotic rate in primary spermatocytes after heat stress. MTA1 may serve as a negative coregulator of p53 in spermatocytes. After heat shock, p53-mediated spontaneous testicular apoptosis is helpful to remove defective germ cells during initial phase and is subsequently followed by Fas-dependent apoptosis [23] . However, the p53-governed effect should not be exaggerated. For example, adenovirus-mediated p53 gene overexpression to rodent spermatocytes significantly impairs spermatogenesis [46] . Our data also demonstrated a deteriorating effect of this overdosed apoptotic wave on the spermatocyte differentiation after heat treatment. Given the facts that p53 is responsible for the initial phase of germ cell apoptosis induced by hyperthermia and MTA1 expression was also maintained at a relatively high level during acute phase in primary spermatocytes after heat stress, we hypothesize that MTA1 acts as a “buffer solution” to ensure the p53 activity below the threshold within the acute phase after hyperthermal stimulation. This apoptotic balance could be achieved by elevating the deacetylation level of p53 ( Fig. 5 ).
10.1371/journal.pone.0026013.g005
Figure 5
Summary diagram of the possible mechanisms related to MTA1 function contributing to transient protection of pachytene spermatocytes following heat stress.
As expected, inhibition of in vivo MTA1 expression induced a significant increase of apoptosis upon hyperthermal treatment at 48 h after 43°C hyperthermal stimulation. This elevated apoptotic rate finally resulted in an impairment of late development of spermatocytes. Meiosis in male mice occurs over a 2 week period [47] . Therefore, the relatively transitory experimental duration in our model (about 2 days after heat treatment) was not enough for any change of spermatids to occur. Besides primary spermatocytes, Ser was the other major cell type affected by the interference. This notion was supported by the observation that the expression of a number of genes essential for the complex interaction between germ cells and Ser was slightly disrupted in MTA1 KD testis after heat shock ( Figure S6 ) [48] . A single exposure of the rodent testis to 43°C can only produce specific damage limited to the spermatocytes [49] , [50] . Therefore, this disruption of Ser function is probably a subsequent result due to the apoptosis of germ cells. After all, Ser fate is also determined by the status of contacting germ cells [51] . Moreover, elevated temperature in adult mouse testis results in a complex stress response, including induction of genes associated with oxidative stress and hypoxia [20] . MTA1 exists in a biochemically distinct protein complex which comprises HDAC1/2, RbAp46/48, and MBD3 components. MTA1 induces the deacetylation of HIF-1alpha by increasing the expression of HDAC1 [42] .In this context, MTA1 also bear potential to be a novel histone modifier participating in the indirect regulation of Ser activity.
Due to the limitation of knockdown assay, we could not reveal a long-term effect of suppression of MTA1 on testicular function. We tried to answer this question from a devious angle. The deteriorating effects of hyperthermia have been described in many male infertility-related diseases such as cryptorchidism, varicocele and chronic fever [52] . The influence of varicoceles on testicular function is variable, leaving it apparently unaltered in some cases, and causing partial or total arrest of spermatogenesis in others. It is suggested that an infertile man with varicocele might have testicular damage reflected by abnormal histologic patterns, such as maturation arrest, before becoming azoospermic [53] . Immunoblotting analysis revealed that the expression of MTA1 was decreased in the testes of varicocele patients compared with that in normal group. In line with our in vivo model, pachytene spermatocytes and Ser were two main cell types devoid of MTA1 immunolocalization. These data suggest that the decrease in MTA1 expression may be related to the pathogenesis of infertility in patients with varicocele.
DNA arrays have been employed to link gene expression to mechanisms of heat toxicity on murine testis. The overall pattern of gene expression was one of cellular shutdown. This is intuitively understandable, wherein an acute environmental insult causes cellular activity to cease and certain protective mechanisms are therefore guaranteed to initiate a defensive or reparative response [41] . To this end, further analysis of MTA1 function in the testes exposed to various exogenous thermal factors would help elucidate the underlying molecular mechanisms leading to male infertility. The fact that decrease of MTA1 could be detected even earlier than the beginning of the damage of testicular cells after heat stress also warrants its diagnostic potential for male infertility induced by hyperthermia.
Materials and Methods
Ethics Statement
The Ethics Committee for Animal Experiments of the Fourth Military Medical University approved all animal work (Permit number: 10001) and the experimental protocols strictly complied with the institutional guidelines and the criteria outlined in the “Guide for Care and Use of Laboratory Animals”. All surgery was performed under sodium pentobarbital anesthesia, and all efforts were made to minimize suffering. For human samples, the written informed consents have been obtained from all participants. The use of the human tissue in this study was approved by the Human Research Committee of the Fourth Military Medical University for Approval of Research Involving Human Subjects. The protocol employed strictly conformed to the standards set by The 2008 Revised Declaration of Helsinki .
Tissue samples and cells
Adult male mice (C57BL/6) were obtained from the Animal Research Center of the Fourth Military Medical University, Xi'an (China) and maintained on a 12-hour light: 12-hour dark in a 20–25°C environment. They were allowed to acclimatize for at least 1 week before the experiment. To induce the transient heat stress, males were subjected to a single heat stress as described elsewhere with little modification [54] . Briefly, each animal was anesthetized using sodium pentobarbital (Sigma, 5 mg/100 g), and the lower third of the body (hind legs, tail, and scrotum) was submerged in a water bath of 43°C for 20 min. Animals were then dried and returned to their cages. Control animals, designated as 0 h, were anesthetized and left at room temperature. They were sacrificed at 8 h, 24 or 48 h after hyperthermal exposure. For biochemical analysis, tissues were stored at −80°C until use. For histological studies, murine testes were fixed in Bouin's solution for about 24 hours. After dehydration, the testes were embedded in paraffin and further processed into 5-μm-thick sections for morphological examination. To assess testicular degeneration after heat stress, hematoxylin-eosin (H&E) stained transverse testis sections were evaluated for the presence of ‘normal tubules’ if the general architecture showed normal spermatogenesis with all the layers of spermatogenic cells associated with each of the XII stages of spermatogenesis, or ‘degenerated’ if they showed depletion of all or some of germ cell layers usually associated with vacuolization within the seminiferous tubule. One hundred seminiferous tubules in a randomly selected section of the testis from five mice in each group were selected to quantify the relative percentage of degenerated tubules. This quantification was performed blindly and by two independent investigators.
Normal human testes were obtained from three victims of traffic accidents aged 25–40 years as described in previous publication [44] . Two biopsy specimens of histologically arrested testis (at the round spermatid level) were obtained from two patients, 29 and 37 years old, with varicocele at the male infertility clinic. Tissues were processed as described above.
Spermatogenic cell line GC-2spd (ts) was obtained from American Type Culture Collection (Rockville, MD) and maintained in Dulbecco's modified Eagle's medium (Gibco, Grand Island, NY) supplemented with 10% fetal calf serum (Life Technologies, Inc., CA), 1% nonessential amino acids solution (Gibco), and 100 lU/ml penicillin at 39°C in a humidified atmosphere of 10% CO 2 in air. For temperature-shift experiments, cells were maintained at 32°C. After 6 h, cells were harvested and either fixed using 10% neutral buffered formalin for 10 minutes or stored at -80°C until use.
Testicular cells were prepared as previously reported [34] . Briefly, testes were decapsulated and digested for 15 min in 0.25% (w/v) collagenase (type IX, Sigma) at room temperature with constant shaking. Seminiferous tubules were then cut in pieces using a sterile blade and further digested in minimum essential medium containing 1 mg/ml trypsin for 30 min at 30 °C. Digestion was stopped by adding 10% fetal calf serum and the released germ cells were collected after sedimentation (10 min at room temperature) of tissue debris. Germ cells were centrifuged for 10 min at 1,500 rpm at 4°C and the pellet resuspended in 20 ml of elutriation medium (120.1 mM NaCl, 4.8 mM KCl, 25.2 mM NaHCO 3 , 1.2 mM KH 2 PO 4 , 1.2 mM MgSO 4 (7H 2 O), 1.3 mM CaCl 2 , 11 mM glucose, 1× essential amino acid (Life Technologies, Inc.), penicillin, streptomycin, 0.5% bovine serum albumin). Homogeneity of cell populations ranging between 80 and 85% (pachytene spermatocytes) and 95% (round spermatids) was routinely monitored morphologically. Primary pachytene spermatocytes were cultured in minimum essential medium (Gibco) and supplemented with 0.5% BSA (Sigma-Aldrich), 1 mM sodium pyruvate, and 2 mM lactate at 32°C in a humidified atmosphere containing 95% air and 5% CO 2 as described previously [55] . Cells were treated overnight with 10 µM U0126 (EMD) before the addition of DMSO (Sangon Biotech Co., Ltd, Shanghai, China) or 0.5 µM OA (Invitrogen, Beijing, China) for 4–6 h to induce metaphase entry. For heat stress, primary cells were exposure at 43°C for 1 h and then cultured at 32.5°C for 23 h followed by harvest from 0 h to 2.5 h. For isolation of Ser after in vivo siRNA experiment, after being dispersed (but not fragmented) in 0.1% collagenase (type IV) and 0.04% DNase I (all from Sigma) in 1× Hanks fluid (pH 7.4) at 34°C for 10–15 min with constant shaking (100 oscillations/min), the seminiferous tubules were incubated in 1× Hanks fluid (pH 7.4) containing 0.04% DNase I, 0.05% hyaluronidase, and 0.5% trypsin for at least 10 min at 34°C with agitation. The fragmented tubules were allowed to settle and cells were subsequently centrifuged at 900 rpm and at 700 rpm (each for 3 min). Cells in the supernatant were collected and cultured (in DMEM/F12 medium containing 5% FCS) overnight. Sertoli cells attached to the bottom and acquired an irregular shape, whereas the germ cells did not attach and could easily be removed by repeated washing. The purity of isolated testicular cells could reach 85% (pachy) and 90% (Ser), respectively. Further confirmation of the cell types were carried out using RT-PCR analysis of specific cell markers as described in the following part.
In vitro and in vivo siRNA treatment
In vitro siRNA treatment was performed as previous report [56] . In brief, GC-2spd (ts) was plated in six-well culture plates in DMEM media without antibiotics. The cells were allowed to grow to 50–60% confluence before being transfected with siRNA against MTA1 (sc-35982, Santa Cruz Biotechnology, Santa Cruz, CA) or with a control siRNA (sc-37007, Santa Cruz Biotechnology, Santa Cruz, CA). Subsequent incubation of cells in transfection medium along with the transfection reagent was strictly followed the Santa protocol. The cells were incubated with the siRNA mixture for a period of 48 h before being subjected to either mRNA and protein analyses or heat stress treatment.
For in vivo siRNA experiment, adult mice were anesthetized as described above. Testes were pulled out from scrotum, and about 30 µl of plasmid DNA solution (sc-35982, Santa Cruz Biotechnology, Santa Cruz, CA) was injected into the rete testis using glass capillaries under a dissecting microscope as previously described [57] . Electric pulses were delivered using BTX ECM-830 Electro Square Porator (Waukegan, IL, USA). Square electric pulses were applied four times, and again four times in the reverse direction at 50 V for 50 ms for each pulse. The testes were then returned to the scrotum and skin was closed with sutures. The total RNA and protein of testis were collected for assessment 48 h after injection. All efforts were made during the whole process to minimize animal suffering.
RT-PCR and QRT-PCR
Total RNA was extracted from cells or mouse testis using RNeasy Mini Kit (QIAGEN Inc., Valencia, CA, USA) according to the manufacturer's instructions. Routine DNase (Applied Biosystems/Ambion, Austin, TX, USA) treatment (1 U DNaseI per µg total RNA) was performed before reverse transcription. First-strand cDNA was synthesized using 1 µg RNA with Superscript III (Rnase H-Reverse Transcriptase; Invitrogen), according to the manufacturer's instructions and PCR was set up according to Promega's reverse transcription system protocol. The details of primers used in this study were listed in Table 2 . The amplification of 18S and GAPDH were served as internal controls. All PCR reactions for all samples were repeated at least three times. PCR products were then quantified by SYBR green intercalation using the MiniOpticon™ system (Bio-Rad Laboratories, Inc., Hercules, CA, USA). Standard curves were constructed for MTA1 (specific target) and 18S (internal control) by plotting values of CT (the cycle at which the fluorescence signal exceeds background) versus log cDNA input (in nanograms). Accordingly, CT values from each experimental sample were then used to calculate the amount of MTA1 and 18S mRNAs relative to the standard. For each sample, results in terms of MTA1 expression levels were normalized to those of the internal control 18S .
10.1371/journal.pone.0026013.t002 Table 2
Primer sequences for RT-PCR or real-time PCR of target genes.
Gene
GenBank access number
Primer sequence
product length ( bp )
18S
M10098.1
F: CTCGCCGCGCTCTACCTACCTA
R: ATGAGCCATTCGCAGTTTCACTGTA
120
MTA1
AF288137.1
F: CAGTGTCGCCTCTGCGCATC
R: TCCACTGCTCCGAGCTGGAA
141
Gapd-s
M60978.1
F: CTGGCCAAGCCTGCTTCTTAC
R: CAAGGAGGGGCCTTTTGTGTTAC
286
Proacrosin
NM013455.3
F: TCCCCAAATACCCCACACCTG
R: CCCACCACTGTCCCCCTG
217
Sprm-1
NM029315.1
F: TGCGACCCCTGCTGAAAATG
R: GAGGCGCCCAGCAATATAGC
203
cyclin A1
NM007628.3
F: ATTGCAGCTTGTCGGGACAG
R: TGGTGGTTGGAACGGTCAGA
179
GAPDH
M32599.1
F: GGGTGAGGCCGGTGCTGAGT
R: TGACCCGTTTGGCTCCACCCT
98
Vasa
NM001145885.1
F: AGGAAGCAGAGATATTGGCGAGTCT
R: ACCTCTGTTTCCAAAGCCCTTTCCT
91
pgk-2
NM031190.2
F: AGGCCACCTCCAATGGCTGTGT
R: TGTGCGCAGGAAACAGGAAGCAA
210
Gata-1
NM_008089.1
F: GTGAACTGTGGAGCAACG
R: TTGACAGTTAGTGCATTGGG
174
Prm-1
NM_013637.4
F: ACTCCTGCGTGAGAATTTTAC
R: TTATTGACAGGTGGCATTGTT
113
Primers for CHIP assay and for Sertoli function analysis after in vivo RNAi were referenced at Literature 59 and 48, respectively.
Western blot analysis
Protein samples were prepared in ice-cold RIPA buffer (Tris-HCl 50 mM, NaCl 150 mM, Triton X-100 1% vol/vol, sodium deoxycholate 1% wt/vol, and SDS 0.1% wt/vol pH 7.5) supplemented with complete proteinase-inhibitor cocktail tablets (Roche Diagnostic, Mannheim, Germany). Protein was separated on 8–15% SDS/PAGE and transferred to nitrocellulose membrane (Millipore, Bedford, MA, USA). Membranes were then incubated with primary antibodies including anti-MTA1 (Santa Cruz biotechnology, CA, USA; dilution 1∶500), anti-β-actin (Santa Cruz biotechnology, CA, USA; dilution 1∶2000), anti-p53 (Santa Cruz biotechnology, CA, USA; dilution 1∶1000), anti-acetyl-p53 (Millipore, MA, USA; dilution 1∶1000) and anti-p21 (Santa Cruz biotechnology, CA, USA; dilution 1∶1000) in blocking solution overnight at 4 o C. Positive signals were finally detected by using an ECL kit (Amersham Biosciences, Buckinghamshire, UK).
Immunofluorescence and immunohistochemistry
GC-2spd (ts) were fixed in 4% paraformaldehyde and washed three times with PBS. Cells were permeabilized with 0.1% Triton X-100 for 10 min and then incubated for 1 h in 0.5% BSA. Cells were washed three times with PBS and incubated overnight at 4°C with antibody against MTA1 (1∶200) followed by 1 h of incubation with FITC-labeled anti-goat IgG (dilution 1∶800; Sigma). Nuclear were visualize by 10-minute staining of DAPI (dilution 1∶2000; Sigma). Slides were finally analyzed by microscopy using an inverted microscope (Axio Imager M1 microscope, Zeiss).
Streptavidin-biotin complex (SABC) immunohistochemical method was conducted as previously described [17] . In brief, the sections were exposed to 0.3% hydrogen peroxide in methanol for 30 min to destroy endogenous peroxides activity after deparaffinization and rehydration. The slides were then incubated with the goat anti-MTA1 antibody (Santa Cruz Biotechnology, Santa Cruz, CA, USA; 1∶150 dilution) diluted in PBS, at 4°C overnight in a moist box. Biotinylated rabbit anti-goat IgG (1∶800 dilution; Sigma) was incubated on the sections for 1 h at room temperature and detected with streptavidinperoxidase complex. Peroxidases were detected with 0.7mg/ml 3-3′-diaminobenzidine tetrahydrochloride (Sigma, St. Louis, MO, USA) in 1.6mg/ml urea hydrogen peroxide. The sections were subsequently counterstained with hematoxylin for 40 sec. Control slides were incubated with a nonimmune serum instead of primary antibody. The immunohistochemical staining for MTA1 protein in mouse and human testes was evaluated by scanning the entire tissue specimen under low-power magnification (×100) and then confirmed under high-power magnification (×400) by two pathologists. 50 areas (×400) were randomly selected and totally 10 sections of each sample were included in the quantitative analysis. The signal intensity was stratified as strong staining (+++), moderate staining (++), weak staining (+), no staining (-). For semi-quantitative evaluation of positive cell number, at least 1000 cells in each cell type was counted and stratified as [-/+] <10%, [+] 10–25%, [++] 25–50%, [+++] 50–75% or [++++] >75% positive.
In situ end-labeling of fragmented DNA
Apoptotic cells were identified by the TUNEL technique (in situ end-labeling of fragmented DNA), using In Situ Cell Death Detection Kit, POD. (Roche Applied Science, Mannheim, Germany) following instructions of the manufacturer, on fixed cells or paraffin-embedded tissue sections.
Quantification of the apoptotic cells
An apoptosis ELISA kit (Roche Diagnostics, Mannheim, Germany) was used to quantitatively measure cytoplasmic histone-associated DNA fragments as previously reported [58] .
Chromatin immunoprecipitation
Germ cells were incubated with 1% formaldehyde in medium for the last 10 min of culture, washed in cold PBS and harvested. Nuclei were isolated by lysing cells in hypotonic buffer (5 mM Pipes pH 8.0, 85 mM KCl and NP-40 0.5%). Nuclei were then re-suspended, lysed in a buffer containing 1% SDS, 10 mM EDTA and 50 mM Tris/HCl pH 8.1 and sonicated with 8 pulses (1−, 90% Amplitude), clarified on ProteinA/agarose/salmon sperm DNA (Millipore) and used (100 µg of DNA/sample) for immunoprecipitation with 2 µg of MTA1, HDAC2 antibodies or control IgG (Santa Cruz Biotechnology, Santa Cruz, CA, USA). Primers used for ChIP were listed in Table 2 [59] .
Immunoprecipitation analysis
Immunoprecipitation analysis was performed as reported before [60] . In brief, cell lysates were obtained using RIPA buffer containing a complete proteinase-inhibitor cocktail tablets (Roche Diagnostic, Mannheim, Germany) and centrifugation at 5000 g at 4°C for 10 min. The lysates were incubated with mouse anti-p53 antibody at 4°C overnight. On the following day, protein G-Sepharose (Pierce, Rockford, IL, USA) was added into lystates and the compound was incubated at 4°C for another 2 h. Immunocomplexes were finally eluted from the sepharose beads by boiling in Laemmli sample buffer and subjected to SDS-PAGE of immunoblotting analysis with rabbit anti-HDAC2 antibody (Abcam, Shatin, N.T., Hong Kong, China).
Statistical analysis
Experiments were repeated at least three times, and one representative from at least three similar results is presented. The significance of the results was determined by using the one-way ANOVA parametric test. Statistical differences were considered significant at P<0.05. Data were presented as the mean±SD.
Supporting Information
Figure S1
Identification of isolated primary pachytene spermatocytes by RT-PCR analysis. Gata-1, Prm-1 and cyclin A1 were employed as specific markers for Ser, Rsd and Pachy, respectively.
(TIF)
Figure S2
Isolated pachytene spermatocytes were stimulated to enter meiotic divisions by treatment of OA. Control (DMSO) or treated (OA) spermatocytes were analyzed by immunoblotting ( A ) or immunofluorescence ( B ) assays with the anti-phosphoH3 (p-H3) antibody, Bar = 10 µm . Appearance of phosphorylated H3 indicates that spermatocytes have progressed to the M phase of the first meiotic division. C Upreguation of phospho-Erks was only detected by western blot in primary spermatocytes treated with OA.
(TIF)
Figure S3
MTA1/18S relative expression level in primary spermatocytes at different time-points after exposure to heat stress was examined by real-time PCR. Data are expressed as mean±SEM (n = 3; * p<0.05, ** p<0.01 vs. 0 h).
(TIF)
Figure S4
Knockdown of MTA1 expression in GC-2spd (ts) by RNAi. A RT-PCR analysis of MTA1 expression after in vitro RNAi. 18S was served as an internal control. B PCR products were then quantified by SYBR green intercalation in real-time PCR. Data are expressed as mean±SEM (n = 3; * p<0.05 vs. control). C Western analysis of MTA1 protein in GC-2spd (ts) after RNAi treatment. β-actin was used to confirm equal loading. Three separate experiments were repeated and one representative result was presented. D Dose-dependent interference effect of siRNA against MTA1 was revealed by western blot analysis. CT, group treated with control siRNA; KD, group treated with MTA1 siRNA.
(TIF)
Figure S5
Effect of in vivo RNAi on the expression of MTA1 was further confirmed by RT-PCR analysis on isolated Ser and pachy from siRNA treated testes. A Identification of isolated Ser and pachy by RT-PCR analysis. B Transcriptional expression of MTA1 in isolated Ser and pachy.
(TIF)
Figure S6
48 h after in vivo knockdown of MTA1, the mouse testis was subjected to transient heat stress as described in Materials and methods . Changes in testicular expression of functional genes of Ser were then evaluated using real-time PCR. Data are expressed as mean±SEM (n = 3; * p<0.05 vs. control).
(TIF)
Figure S7
MTA1 expression was impaired in the testis of a mouse transient scrotal heat stress model. A Murine testicular sections collected at 8 h (n = 5) or 24 h (n = 5) after hyperthermia exposure were subjected to TUNEL staining. Apoptotic activity was quantified as the number of cells positive for TUNEL staining within 50 seminiferous tubules. # p>0.05 or ** p<0.01 vs. 0 h group. B Effect of heat stresses on the expression of MTA1 , pgk-2 , vasa and GAPDH at different time-points was elucidated at the transcriptional level. C, control group; H, hyperthermia-treated group. C Immunolocalization of MTA1 in the testicular sections at 0 h and 8 h after hyperthermia exposure. Replacement of the primary antibody with normal goat IgG was served as negative control (NC). Arrows indicate pachytene spermatocytes. Bar = 25 µm .
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Conclusive rabies diagnosis can only be achieved by appropriate laboratory testing. Clinical and epidemiological diagnosis is challenging and leads to under-reporting [ 1 , 2 , 3 ]. The Direct Fluorescent Antibody test (DFA) for detection of rabies virus antigen remains as the gold standard test for laboratory diagnosis of rabies in post-mortem brain tissues [ 3 ].
In addition to its essential role in the confirmation of clinical rabies, laboratory diagnosis supports rabies surveillance and the clinical management of patients exposed to a potentially rabid animal. In the case of disease surveillance in rabies endemic countries, laboratory diagnosis of suspected rabid animals is important for assessing the distribution and prevalence of the disease in its major reservoir hosts. In the case of human exposures to potentially rabid animals, the decision to administer post-exposure prophylaxis (PEP) depends on the results of observation and/or laboratory testing of the animal involved [ 2 ]. Laboratory diagnosis can avert the financial losses incurred by the unnecessary application of PEP in the case of negative results, and trigger adequate human case patient management and adequate occupational health risk assessments in the case of positive results [ 4 , 5 ].
Due to a concerted elimination effort by the Latin America and Caribbean (LAC) countries, cases of human rabies transmitted by dogs have been reduced to a small number in well-defined areas in the region [ 6 , 7 ]. As part of the regional strategies towards dog-associated rabies elimination, it is necessary to standardize rabies diagnostic testing across LAC to improve quality, enhance laboratory diagnostic capability and allow multinational inter-laboratory comparisons and analyses of regional surveillance data [ 8 , 9 ]. To address this need, the Action Plan for the Elimination of Dog-Transmitted Rabies in LAC [ 6 ] seeks the implementation of an Inter-American Network of Rabies Diagnostic Laboratories (REDILAR- from the Spanish acronym, Red Interamericana de Laboratorios de Diagnostico de Rabia) to facilitate rapid and reliable diagnosis, promote regional training opportunities and develop a regional laboratory quality assurance program. To this effect, and following recommendations from the 14th Regional Meeting of National Rabies Program Managers, (REDIPRA 14 from the Spanish acronym) [ 10 ] a proficiency testing exercise was carried out within the LAC countries. This report describes an Inter-laboratory proficiency testing (IPT) exercise for rabies diagnosis by DFA, and a baseline assessment of laboratory practices and infrastructure among rabies diagnostic laboratories in LAC. Specifically, the objectives of this IPT were to assess the diagnostic capacity of each laboratory to perform the DFA test and identify specific laboratory needs for training and equipment. In regards the assessment laboratory practices, the objective was to identify differences in the laboratory protocols that could explain discrepant laboratory results and provide baseline knowledge for regional standardization of protocols.
Materials and methods
Participating laboratories
Thirty national or regional (as in regions or states within the countries) rabies reference laboratories, affiliated with the ministries of health (MoH) and/or agriculture (MoAg) agreed to participate in the exercise in response to an invitation from the Pan American Foot-and-Mouth Disease Center (PANAFTOSA in Spanish) of the Pan American Health Organization (PAHO) located in Rio de Janeiro, Brazil. By targeting MoH and MoAg affiliated laboratories, the exercise aimed to comprehensively capture rabies diagnosis practices, on human and animal samples. Across LAC countries, MoH laboratories mostly perform diagnosis on human and domestic animals samples, whereas MoAg laboratories focus on livestock and wildlife submissions. Participation in the IPT exercise and baseline operations assessment was voluntary. Thirteen laboratories were affiliated with the MoH and seventeen with the MoAg. The 30 participating laboratories were located in the following countries: Argentina (MOH/MoAg), Bolivia (MOH/MoAg), Brazil (MOH/MoAg), Canada (MoAg), Chile (MOH), Colombia (MOH/MoAg), Costa Rica (MoAg), Cuba (MoAg), Ecuador (MoAg), Dominican Republic (MoAg), El Salvador (MoAg), Guatemala (MOH/MoAg), Haiti (MoAg), Honduras (MoAg), Mexico (MOH/MoAg), Nicaragua (MOH), Panama (MoAg), Paraguay (MOH/MoAg), Peru (MOH), Trinidad and Tobago (MoAg) and Uruguay(MoAg). The laboratories were randomly coded as L1 to L30.
Source of panel material and panel composition
Brain tissue impression slides used in the exercise were prepared at the Poxvirus and Rabies Branch of the US Centers for Disease Control and Prevention (CDC) and represented samples received for routine rabies diagnosis or typing with variations in antigen load and distribution. All tissues were obtained from naturally occurring rabies cases in major host reservoir species without passage in a secondary susceptible animal. No animals were used or harmed in any way for this study.
Each PT panel constituted by 20 samples included 17 test samples and 3 controls. Positive tissues contained major rabies virus variants circulating in the Americas. Negative samples, encompassed tissues demonstrating complete absence of rabies virus antigen and artifacts, and some with atypical golden lipofuscin and atypical nonspecific fluorescent bacteria or fluorescent artifacts. Lipofuscin within CNS tissues is commonly detected in cerebellum of cats and brain tissues from older animals, and appears as golden "pseudo-inclusions" with diffuse outline when DFA slides are observed with FITC filters exciting in blue light wavelengths. The lipofuscin will be observed in tissues which are stained by immunofluorescence as well as in tissues without conjugate applied. Often inexperienced individuals reading slides, particularly those working with conjugates at less than optimal dilution or deficient microscope equipment, may not be able to observe rabies positive virus antigen as sparkling 4+ apple-green fluorescence, which in turn prevents distinguishing artifacts.
The samples with lipofuscin were purposely chosen among CDC archived samples to demonstrate the golden atypical objects easily distinguished by competent laboratorians.
Details of the rabies virus variants included in the IPT panel are listed in Table 1 by both predominant host reservoir species, and antigenic type as indicated by the use of a panel of eight monoclonal antibodies (Mabs) directed to the rabies virus N protein [ 11 ].
10.1371/journal.pntd.0005427.t001
Table 1 Rabies panel composition and rabies virus variants used in this study.
ID
SAMPLE
AG VARIANT
PASSAGED ON
ORIGIN
COMMENTS
INTENSITY/
AG DISTRIBUTION
1
Strong Positive Control
(Dog Variant)
Not Passaged
Canis lupus familiaris
Naturally infected animal
4+/4+
2
Weak Positive Control
3
Not Passaged
Bos taurus
Naturally infected
4+/2+
3
Negative Control
N/A
N/A
Negative healthy (Sheep)
Non-rabid animal
0/0
4
PT 1
1 (Dog Variant)
Not Passaged
Canis lupus familiaris
Brain stem or cerebellum
4+/3-4+
5
PT 2
(Negative)
N/A
Felis catus
Cat cerebellum containing lipofuschin
0/0 Golden autofluorescence
6
PT 3
7 (AZ Gray Fox)
Not passaged
Urocyon cinereoargenteus
Brain stem or cerebellum
4+/2-3+
7
PT 4
Negative Dog
N/A
Canis lupus familiaris
Very clear negative
0/0
8
PT 5
3 (Vampire)
Bovine
Bos taurus
Brain stem or Cerebellum
4+/2+
9
PT 6
4 (Tadarida brasiliensis)
Not Passaged
Tadarida brasiliensis
Brain from naturally infected
4+/3+
10
PT 7
(Negative)
N/A
fox Infected with Listeria monocytogenes
brain tissue containing bacteria
Non-specific staining
11
PT 8
Negative skunk
NA
Mephitis mephitis
DFA Negative sample
0/0
12
PT 9
South Central US North Central MX
Not Passaged
Mephitis mephitis
Brain stem or Cerebellum
4+/2+
13
PT 10
6 (Lasiurus cinereus)
Fox
Lasiurus cinereus
Brain from naturally infected
4+/4+
14
PT 11
Negative
Non-rabid
Molossus molossus
Non rabid animal
0/0
15
PT 12
Western Eptesicus fuscus
Not passaged
Eptesicus fuscus
Brain from naturally infected
4+/3-4+
16
PT 13
Negative
Non-rabid
Ursus arctos
Brain stem or cerebellum
0/0 Golden autofluorescence
17
PT 14
Negative
Non-rabid
Sus domesticus
Brain stem or cerebellum
Non-specific staining
18
PT 15
Negative
Non-rabid
mouse
Sample infected with group G streptococcus
0/0
19
PT 16
Mongoose (PuertoRico)
Not passaged
Herpestes javanicus
Brain stem or cerebellum
4+/4+
20
PT 17
Negative
Non-passaged
Sheep
Brain stem or Cerebellum
0/0 Negative
All rabies positive brain tissues were inactivated via gamma-irradiation with 5 x 10 6 rads using a Gamma Cell Irradiator (Gammacell 220 Excel, MDS Nordion, Canada) to exclude any infectious agent in order to facilitate safe importation procedures. To confirm complete inactivation of rabies virus, isolation was attempted in mouse neuroblastoma (MNA) cell culture. Brain homogenates (10% w/v) were prepared in Magna NA Lyser tissue homogenizer tubes containing 1.4-mm (diameter) ceramic beads (Roche Applied Science, Penzberg, Germany), using 1.0 mL of MEM-10 as a diluent. For virus recovery, 100μL of test inoculum was added to 1mL of MEM-10 containing 5x10 6 mouse neuroblastoma cells (MNA) in a T-25 tissue culture flask (Corning, NY). Tissue culture flasks were incubated at 0.5% CO 2 at 37°C for 72 hours prior to passaging. All cultures were sub-passaged a minimum of three times. For infectivity assessments, Teflon-coated four well slides were seeded with 30uL of MEM-10 containing 0.5 x 10 6 cells per mL, and incubated in a humid chamber at 0.5% CO 2 at 37°C for 24 hours. The slides were then rinsed with phosphate-buffered saline (PBS), and fixed in cold acetone at -20°C for one hour. Cells were checked for the presence of RABV antigens by the DFA test, using optimal working dilutions of FITC-labeled anti-RABV mAb conjugate (Fujirebio Diagnostics, Inc., Malvern, PA, USA) after each passage. This testing method was performed in duplicate by multiple, qualified testing personnel at the CDC in Atlanta, Georgia.
After tissue inactivation and safety testing procedures concluded, tissues were re-tested by DFA to confirm antigen was not damaged. From all selected inactivated tissues, a total of 600 touch impression slides were prepared according to the standard protocol for the postmortem diagnosis of rabies by the DFA test [ 12 ]. All slides were acetone fixed at -20°C overnight and stored at -80°C until packaging and shipment from CDC to PANAFTOSA. After the panels arrived to PANAFTOSA in Brazil, the material was distributed to all the participating laboratories. Each shipment contained a document with instructions, and the results reporting sheet. The shipping time from dispatch to receipt by the laboratories was recorded. Laboratories were requested to submit the reporting sheet within two months of receipt of the samples.
Panel testing
The slides were designed to cover different situations faced during routine testing in the participating laboratories. The participants were asked to test the samples using the standard protocol in their laboratory, record their results (positivity, intensity, and distribution of the fluorescence staining) in the reporting sheet, and the microscopic condition and impression quality of the tissues (Good, Acceptable, or Deficient) as evaluated by the laboratory operator. The microscopic condition refers to the integrity of the cells facilitating fluorescence interpretation (cytoplasmic or nuclear) under the microscope and the tissue impression quality refers to how well the tissue impressions were made and if there was sufficient material (tissue) to read the slide and reach a diagnosis as assessed by an overall visual inspection.
In addition, for the laboratory practices assessment, a questionnaire to study the variability of diagnostic techniques, available resources, quality control, and safety procedures was sent electronically to the participating laboratories in parallel to the panel
Data synthesis and analysis
The sensitivity, specificity, and Cohen's kappa coefficient to measure the inter-rater agreement values were calculated for each laboratory. Samples for which no result was returned were considered missing and were not included in the analysis. Inconclusive samples (as classified by the laboratory on the reporting sheet) were excluded from the sensitivity and specificity calculations, but were considered for the kappa and concordance calculations. The level of concordance was calculated by laboratory and by sample as the number of results in agreement with those of the CDC over the total number of results; inconclusive results were included in the denominator.
Exploratory data analyses were carried out to investigate the relationship between the laboratories concordance results and laboratory practices as captured by the laboratory practices assessment. One-way ANOVA and Mann-Whitney U tests, the latter if the distributional assumptions underlying the parametric test were not satisfied, were used to compare groups of respondents to the questionnaire. An alpha of < 0.05 was considered statistically significant.
The overall microscopic condition and impression quality of the tissues for each laboratory panel and for each of the samples across the laboratories was calculated as the percent samples reported as deficient. In the case of two samples for which the operators from the same laboratory evaluated the quality of a slide differently (i.e. Good/Deficient or Adequate/Deficient), the quality was coded as Deficient for the purposes of the analysis. The data analysis was performed in R (i386 3.1.2) [ 13 ].
Quality control verification
To assess potential shipping problems during transportation of PT panels to all participating laboratories, the stability of rabies antigen and reproducibility of the DFA test results was determined a priori . Two complete panel sets were maintained under 5 different temperatures (-80°C, -20°C and 4°C, room temperature and 37°C) as well as 3 different storage time (3, 7 and 14 days) to mimic potential transport and storage conditions. Thus, a great total of 600 slides with two tissue impressions were tested for all combinations of these conditions. At all time points and storage conditions one complete panel set was run using FDI Fujirebio Diagnostics Inc. cat# 800–092 and the other with EMD Millipore Light Diagnostics cat#5100). A control conjugate EMD Millipore Cat#5102 was used (non-rabies FITC labeled antibody of the same isotype as the FITC labelled monoclonal) to assess the specificity of the reaction. Both the fluorescent intensity and the antigen distribution were reported for each case. A total of 36 test results were reported and analyzed for each of the 20 samples in each panel tested. The result of this testing was compared against the baseline value provided by the CDC which was the result of samples stored at -80°C. There were sufficient sample slides for all planned storage variations except sample# 12 (WEF variant) whose tissue was very limited in amount available due to the small size of a bat brain.
Results
Direct fluorescent antibody test panel testing results
Of the 30 laboratories, from 21 countries, that accepted the invitation to participate in the exercise, 23 laboratories (9 affiliated to MoH and 14 affiliated to MoAg), from 18 countries, received the panel and completed the proficiency testing exercise. Seven laboratories, two in Argentina, two in Guatemala, one in Cuba, one in El Salvador and one in Paraguay, did not receive their samples due to shipping or customs problems. All the participating laboratories reported their results within two months after receiving the panel.
The agreement between the laboratory results and those of the CDC, as measured by the sensitivity, specificity, concordance and kappa values are shown in Table 2 . Two laboratories correctly identified all samples tested (sensitivity and specificity of 1.0). However, 30% (7/23) of all laboratories reported at least one false positive and 83% (19/23) of all laboratories reported at least one false negative sample. The average sensitivity was 76% with a range of 40% to 100%. The average specificity was 88% with a range of 22% to 100%. While a majority of the laboratories had low false positive rates, there were considerable differences in the sensitivity ( Fig 1 ). The mean concordance was 81% with a range of 50% to 100% and the mean kappa score was 0.56 with a range of 0.02 to 1.00.
10.1371/journal.pntd.0005427.g001
Fig 1
Scatterplot of Sensitivity by Specificity.
The sensitivity (y-axis), or the true positive rates and the specificity (x-axis), or the true negative rates, for each laboratory are shown with an open circle. A laboratory correctly identifying all samples would be found in the upper right corner.
10.1371/journal.pntd.0005427.t002
Table 2 Results by Laboratory and agreement relative to CDC expected results.
Laboratory
Missing
Inconclusive
a
b
c
d
Sensitivity
Specificity
Concordance
Kappa
L01
0
0
8
0
3
9
0.73
1.00
0.85
0.71
L06
0
0
9
5
2
4
0.82
0.44
0.65
0.27
L07
0
2
9
0
1
8
0.90
1.00
0.94
0.73
L08
0
2
6
0
5
7
0.55
1.00
0.72
0.38
L09
1
0
4
0
6
9
0.40
1.00
0.68
0.39
L10
0
0
10
0
1
9
0.91
1.00
0.95
0.90
L11
0
0
7
0
4
9
0.64
1.00
0.80
0.61
L12
0
1
6
0
4
9
0.60
1.00
0.79
0.54
L13
0
1
11
0
0
8
1.00
1.00
1.00
0.90
L14
0
1
7
0
3
9
0.70
1.00
0.84
0.63
L15
0
0
9
0
2
9
0.82
1.00
0.90
0.80
L16
0
1
8
0
2
9
0.80
1.00
0.89
0.72
L17
2
0
8
2
3
5
0.73
0.71
0.72
0.43
L18
0
0
8
2
3
7
0.73
0.78
0.75
0.50
L20
0
1
11
2
0
6
1.00
0.75
0.89
0.70
L21
3
7
4
2
0
4
1.00
0.67
0.80
0.25
L22
0
1
7
0
3
9
0.70
1.00
0.84
0.63
L23
0
0
5
0
6
9
0.45
1.00
0.70
0.43
L24
0
2
10
0
1
7
0.91
1.00
0.94
0.72
L25
4
0
4
3
5
4
0.44
0.57
0.50
0.02
L27
0
0
11
0
0
9
1.00
1.00
1.00
1.00
L28
0
0
10
7
1
2
0.91
0.22
0.60
0.14
L29
0
2
7
0
3
8
0.70
1.00
0.83
0.55
Missing: Samples with no result due to shipping and/or receiving error–excluded from analysis.
Inconclusive: Samples with inconclusive result as classified by laboratory, unable to determine a positive or negative results. Excluded from the sensitivity and specificity calculations, but included in the kappa calculations.
a: True Positive
b: False Positive
c: False Negative
d: True Negative
Concordance: percent of laboratory results in agreement the CDC
Kappa: Cohen's kappa coefficient measures inter-rater agreement considering agreement occurring by chance.
The inter-laboratory agreement by sample is presented in Table 3 . The Texas Grey Fox (S01), the Mongoose/Dog Variant (S02), and the Vampire Bat weak positive control (S19), had the lowest level of laboratory concordance with, respectively, only 8/23 (30%), 9/23 (39%) and 9/20 (45%, 3 missing) laboratories correctly identifying the sample. Four of the nine negative samples in the panel had a concordance of less than 0.8, two between 0.8 and 0.9, and only three negative samples had a high level of concordance (0.91, 0.96 and 1.0).
10.1371/journal.pntd.0005427.t003
Table 3 Agreement between Laboratory and CDC values, by Sample.
Laboratory Results
Sample
Ag Variant
CDC I/D
Pos
Neg
Indeter-mined
Missing
Concordance (%)
S01
Texas Gray Fox
4+/3+
8
12
3
0
30
S02
Mongoose/ Dog Variant
4+/3+
9
13
1
0
39
S03
Negative
0/0
3
19
1
0
83
S04
AZ Gray Fox (AV7)
3+/4+
18
4
1
83
S05
Negative
0/0
4
16
3
0
74
S06
Lasionycteris noctivagans
4+/4+
19
4
0
87
S07
Tadarida brasiliensis (AV4)
4+/4+
18
0
0
78
S08
Negative
0/0
4
17
2
0
74
S09
Hognose skunk
4+/4+
22
1
0
0
96
S10
South Central Skunk US North Central MX
4+/1+
19
3
1
0
83
S11
Negative
0/0
5
17
1
0
74
S12
Western Eptesicus fuscus
4+/4+
15
8
0
0
65
S13
Negative
0/0
3
15
3
2
76
S14
Negative
0/0
3
19
0
1
86
S15
Negative
0/0
3
20
0
0
91
S16
Mongoose (PuertoRico)
4+/4+
20
2
1
0
91
S17
Negative
0/0
1
22
0
0
96
S18 Control 1
LA Dog Variant
4+/4+
19
2
0
2
95
S19 Control 2
3 (Vampire)
4+/2+
9
7
4
3
45
S20 Control 3
Negative
0/0
0
21
0
2
100
CDC I/D: CDC determined values for Staining Intensity and Antigen Distribution
Pos: Number of laboratories reporting the sample positive for rabies
Neg: Number of laboratories reporting the sample negative for rabies
Inconclusive: Samples with inconclusive result as classified by laboratory, unable to determine a positive or negative results. Excluded from the sensitivity and specificity calculations, but included in the kappa and concordance calculations. Concordance: percent of laboratory results in agreement the CDC
Logistics and quality control questionnaire
The mean shipping time for laboratories to receive their samples was 11.87 days, ranging from 1 to 65 days (n = 22). Twenty laboratories (87%) reported on the tissue quality of the samples; two of the 20 reported on the quality of samples 1–18, omitting the control sample quality. The mean percentage of samples with deficient microscopic condition and impression quality, as reported by each laboratory, was 13% (range: 0–68%) and 18% (range: 0–63%), respectively. The three control samples (S18-S20) had the highest proportion of samples reported as Deficient (5/18, 5/18 and 7/18, respectively) when considering impression quality. No statistically significant correlation was found between shipping time and the percent of tissues with deficient impression (Pearson’s r = 0.06, p-value = 0.77) or microscopic condition (Pearson’s r = 0.13, p = 0.53, respectively).
Quality control verification
Four months into the exercise and due to perceived problems with the slides reported by several participating laboratories, we assessed the stability of the rabies antigen and the reproducibility of the DFA test results. Duplicate sets of slides were removed from the -80°C storage at the CDC and tested for stability under different storage conditions and time. The complete panel was maintained under 5 different temperatures (-80°C, -20°C and 4°C, room temperature and 37°C) as well as 3 different storage times (3, 7 and 14 days) to mimic potential transport and storage conditions. All combinations of these conditions were run using two different anti-rabies conjugates (FDI Fujirebio Diagnostics Inc. cat# 800–092 and EMD Millipore Light Diagnostics cat#5100), and one control conjugate EMD Millipore Cat#5102 (non-rabies FITC labeled antibody of the same isotype as the FITC labelled monoclonal) to assess the specificity of the reaction. Both the fluorescent intensity and the antigen distribution were reported for each case. A total of 30 test results (replicas) were reported and analyzed for each of the 20 samples of the panel. The result of this testing was compared against the baseline value provided by the CDC. There were sufficient sample slides for all planned storage variations except sample# 12 (WEF variant).
Results indicate that weak positive control (S19) had false negative results at day 3, 7 and 14 at both -80 and 4°C. Sample 1 also had a negative result after 3 days in storage at -80°C but was positive at other times and storage temperatures. Inconsistent results for these two samples indicate this sample had uneven amount of antigen distributed across the tissue, which is consistent with a low viral load.
Laboratory practices technical questionnaire
The twenty-three participating laboratories answered the practices questionnaire on laboratory testing, biosafety and quality control (Tables 4 and 5 ). All laboratories performed DFA and had a dedicated rabies laboratory, which was restricted to authorized personnel in 21 laboratories (91%). The laboratories varied in capacity with 52% testing 0–10 samples per week; one laboratory, affiliated to MoAg reported an average of 120+ samples per week. Only 13 (56%) laboratories responded that the darkroom used for reading DFA slides was for rabies only. While 22 of 23 laboratories reported having a biological safety cabinet, only 65% are checked annually and 74% are for rabies only.
10.1371/journal.pntd.0005427.t004
Table 4 Selected questions of the rabies laboratory practices questionnaire and results.
Survey Item and Response
Freq.
Percent
Sensitivity(Mean)
Test Stat
p-value
Number of rabies samples tested per week
-
0.07
0.93
0–10
12
52
0.78
-
-
11–30
5
22
0.78
-
-
>30
6
26
0.75
-
-
Number of people performing tests in the lab
-
2.5
0.11
1–2
8
35
0.86
-
-
3–4
11
48
0.69
-
-
4+
4
17
0.81
-
-
Has a formal quality program
-
44.5
0.82
Yes
17
74
0.78
-
-
No
6
26
0.74
-
-
Has a quality control person with direct access to top level management
-
35.5
1.00
Yes
19
83
0.78
-
-
No
4
17
0.75
-
-
Has established, written lab quality manual
-
24.5
0.65
Yes
21
91
0.76
-
-
No
2
9
0.85
-
-
Rabies lab is restricted to the (rabies) testing personnel
-
26.0
0.53
Yes
21
91
0.76
-
-
No
2
9
0.86
-
-
Biological safety cabinet systems annually verified
-
48.5
1.00
Yes
15
65
0.78
-
-
No
7
30
0.76
-
-
Missing
1
4
Biological safety cabinet restricted for rabies testing
-
25.5
0.24
Yes
17
74
0.80
-
-
No
5
22
0.68
-
-
Missing
1
4
Has a fluorescence microscope
-
-
-
Yes
23
100
-
-
-
No
0
0
-
-
-
Age of scope (years)
-
1.63
0.22
0–5
10
44
0.80
-
-
6–15
8
35
0.68
-
-
>15
5
22
0.85
-
-
Optical surfaces cleaned after every use
-
44.5
0.49
Yes
18
78
0.76
-
-
No
5
22
0.84
-
-
10.1371/journal.pntd.0005427.t005
Table 5 Selected questions of the rabies laboratory practices questionnaire and results.
Survey Item and Response
Freq.
Percent
Sensitivity(Mean)
Test Stat
p-value
Service and maintenance records available
-
8.0
0.19
Yes
21
91
0.79
-
-
No
2
9
0.56
-
-
Has calibration programs for the laboratory instruments
-
51.0
0.94
Yes
16
70
0.78
-
-
No
7
30
0.76
-
-
Has an official (institutional) DFA protocol
-
58.5
0.06
Yes
19
83
0.74
-
-
No
4
17
0.92
-
-
Approval and/or rejection criteria of samples established and written
-
25.5
0.39
Yes
4
17
0.85
-
-
No
19
83
0.76
-
-
# of people who read slides per test
-
71.5
0.49
1
11
48
0.75
-
-
2–3
12
52
0.79
-
-
Performs confirmatory test(s) on weak or inconclusive samples
-
46.5
0.09
Yes
20
87
0.74
-
-
No
3
13
0.91
-
-
Participates in a proficiency program for rabies
-
54.0
0.94
Yes
8
35
0.78
-
-
No
15
65
0.77
-
-
Routinely checks VNA titer of all rabies lab employees
-
54.5
0.95
Yes
15
65
0.78
-
-
No
8
35
0.76
-
-
Area of brain used for samples (labs may report >1 area)
-
-
-
Hippocampus
17
74
-
-
-
Cerebellum
16
70
-
-
-
Cortex
11
48
-
-
-
Brain Stem (medulla oblongata)
11
48
-
-
-
Pons
2
9
-
-
-
Thalamus
1
4
-
-
-
Spinal cord
1
4
-
-
-
Olfactory Trigone
1
4
-
-
-
# of Anti-Rabies Conjugates Used
-
5.0
0.11
1
20
87
0.75
-
-
2
3
13
0.96
-
-
Uses rabies negative control FITC labeled Conjugate
-
52.5
0.45
Yes
5
22
0.72
-
-
No
18
78
0.79
-
-
Slide Incubation Time (Minutes)
-
20.5
0.47
30 or less
20
87
0.76
-
-
>30
3
13
0.84
-
-
While all laboratories reported having a fluorescence microscope, the manufacturer brand is highly variable. The age of the microscope varied from less than one year (43%) to greater than 20 years (14%); with 30% of scopes older than 11 years. The light source of the microscope is also highly variable with 82% of the labs using Mercury lamps (HBO 100W or 50W), while only 13% using LED and 4% using halogen as a light source. For routine scanning of the DFA slides, 4 laboratories (17%) use a magnifying power of 40X or higher with oil immersion, 7 laboratories (30%) use a combination of 20X and 40X dry format, and 11 (48%) of the laboratories use 40X in a dry format. One laboratory did not know what type of lens was used to scan the slides. Eighteen (78%) laboratories clean the optical lenses after each use, 21 (91%) have service records and 16 (70%) have calibration programs.
All laboratories reported having written procedures for sample processing, including transportation, reception, handling, protection storage and safe disposal. However, only 19 (82%) laboratories had an official DFA protocol and only 4 (17%) had written sample approval and or rejection criteria.
We found considerable variation in the general laboratory protocols including the area of the brain used for routine diagnosis, the type of sample used (impressions vs. brain smears), the fixation protocol (temperature, time and dilutions for slides) and the conjugate procedures (staining and incubation time). Regarding the selected area of the brain routinely tested, most laboratories selected from more than one location: 17 laboratories (74%) select the hippocampus, 16 (70%) the cerebellum, 11 (48%) the cortex and the 11 (48%) brain stem. In addition, a smaller proportion of the labs select the Pons, Thalamus, Spinal cord or olfactory trigone (n = 3, 13%).
The vast majority of laboratories use brain impressions (n = 22, 96%). Only one laboratory uses brain smears. Nineteen (82%) of the responding laboratories fix the DFA slides. The fixing solution is also variable with 17 (74%) of laboratories fixing in acetone. Of the18 laboratories that reported on fixing times, 13 (72%) employed between 15 to 60 minutes, 4 (22%) from 1 to 4 hrs, and 1 left the slides overnight. Of the 20 laboratories that reported on fixation temperatures, 16 (69%) reported temperatures of -20°C, and the remaining three reported temperatures of 80°C, 4°C, and room temperature, respectively.
The conjugate is a critical reagent in DFA; 20 (87%) laboratories use a single conjugate for rabies diagnosis, while 3 laboratories use more than one. Sixty-five percent of responders using a commercial source and 35% using a conjugated produced by a regional, national or local rabies laboratory. Eighteen (78%) laboratories do not use a rabies negative control FITC labeled conjugate. Eighteen (78%) laboratories incubate the stained slides for 30 min while 5 laboratories incubate the slides for 15, 20, 40, 45 and 60 minutes, respectively.
One laboratory incubates the stained slides at room temperature while the rest of laboratories (22, 95%) incubate at 37°C. Differences were also seen in the solutions used for washing the slides, the time and washing steps and the mounting media used to coverslip: 17% of the laboratories use 2x 3–5 min PBS rinses and 13% use as mounting media 90% glycerol in carbonate-bicarbonate buffer pH 9.0. Eleven (48%) laboratories use only one dedicated slide reader, 10 (43%) use 2 and 2 laboratories have 3 readers per slide. Although 20 (86%) laboratories perform a confirmatory testing on weak or inconclusive results, 3 (13%) do not have this capacity. The most common confirmatory test reported was the mouse inoculation (n = 15, 65%) followed by repeat DFA with using an un-labelled monoclonal to assess the specificity of the fluorescence (n = 4, 17%). Only one laboratory uses cell culture and two laboratories (9%) use RT-PCR.
Sixty-five percent of the laboratories [ 15 ] reported having at least three dedicated laboratory employees for rabies diagnostics. While all the laboratories require pre-exposure prophylaxis for their employees, 5 (22%) check the employee titer biannually, 7 (30%) annually, 3 (9%) biennially and 9 (39%) reported that they don’t check or did not have a specific time frame for checking employee titers.
Only 8 laboratories (35%) reported participating in a proficiency testing program for rabies, different from this one, 17 (74%) had a formal quality control program, and 19 (82.6%) had a person responsible for quality control with access to top level management.
Discussion
The level of concordance between the 23 participating laboratories and the CDC panel showed large variability. Two laboratories had 100% concordance, while 91% of the labs had at least one discordant sample, with a total of 26 false positive and 61 false negative results among all laboratories. This level of concordance is lower than those found in similar inter-laboratory exercises of DFA diagnosis of rabies elsewhere. In an inter-laboratory exercise of Middle and Eastern European countries in 2001 [ 11 ], 3 of the 16 (19%) participating national laboratories produced false-positive results. During the EURL/ANSES annual inter-laboratory trials from 2009 to 2014, which included 5 laboratories from the Americas, the percent of laboratories with at least one discordant sample ranged from 8% to 20% [ 15 ]. One possible cause of the lower level of concordance relative to other studies, may be related to the sample stability during shipping and conditions of storage of the panel. However, we did not find a significant association between the shipping time and the laboratory-reported microscopic and tissue impression quality. This assumes that shipping time captures adequately unobserved variables such as shipping and storage temperature. At the times and temperatures of shipping and/or storage evaluated in this study, the CDC found no impact on the degree of positivity of samples, with the exception of the weak positive control. Only 4 panels had shipping times longer than the 14 day limit of the stability test (2 labs with 15 and 16 days, respectively). In addition, we tested the level of concordance for the subset of PT panels delivered within the 14-day limit of the stability test, which encompasses 81% of all the panels tested across LAC, and found no significant increase in the concordance among laboratories.
For the current exercise, a technical questionnaire was completed in parallel to the panel in order to identify variations in methodology, equipment and human resources that could potentially explain discordant results to help implement standardized rabies diagnosis protocols across the region. While no significant associations were seen between differences in survey responses and laboratory performance, this may be due to the small sample size, as even small methodological changes have been shown to affect the sensitivity of the DFA test including the mounting medium [ 16 – 17 ], laboratorian’s expertise level, number of employees reading the slides, the anti-rabies conjugate utilized, specificity controls used and working performance of fluorescence microscope [ 18 , 19 ].
Forty-percent of the laboratories have only one person reading the slides. While no difference in sensitivity was found between laboratories with one or two readers, the OIE manual for DFA recommends two readers per slide [ 3 ] and in the EURL/ANSES study, laboratories using two readers had a higher concordance during the DFA than laboratories with one reader [ 18 ].
Half of the participating laboratories tested less than 10 samples per week. While the sensitivity did not vary by laboratory work load in this study, laboratories with small throughput may have reduced test interpretation expertise. In the Barrat 2001 inter-laboratory technical questionnaire, low laboratory throughput was generally associated with a lower frequency in the use of fresh (recently prepared, replaced for new one from the bottle or reagents far from expiring) reagents including fixatives, buffers and solutions, which in turn will affect the analytical sensitivity and quality of the test [ 14 ].
The type of conjugate may also affect the sensitivity of the DFA test (monoclonal cocktail versus polyclonal, in-house made versus commercial). For the current exercise, laboratories used commercial (65%) or in-house (35%) conjugates. A study of 12 rabies reference laboratories in Europe demonstrated that the variability of conjugates could potentially lead to discordant results and influence assay sensitivity [ 19 ]. The OIE manual states that the conjugates should be fully validated for specificity and sensitivity before use, including their ability to detect lyssaviruses other than rabies [ 3 ]; it is unclear whether such validation has been completed for the in-house conjugates used in the LAC region. It was recommended by members of the regional laboratory network during a briefing on the results of this exercise that PAHO should provide a standard set of conjugates for future exercises to reduce the potential variability in results associated with conjugate type.
The maintenance, type and quality of fluorescence microscope including quality and type optical lens, magnification and numerical aperture could also affect the DFA results. All labs reported having a fluorescence microscope. However, the manufacturer brand, age, light source and magnifying power was highly variable, which may have affected the overall concordance results among laboratories.
The PT panel was composed of 20 samples with samples of diverse origin. Potentially, the variety of samples and virus variants caused difficulty for some laboratories. The Texas Grey Fox (S01), the Mongoose/Dog Variant (S02), and the Vampire Bat weak positive control (S19) had the lowest level of laboratory concordance. In a two-year review of the 2009–2010 EURL/ANSES international inter-laboratory trials all errors were associated with bat strains (EBLV-1, EBLV-2 and ABLV); the authors commented that these strains do not provide the same type of fluorescence than the conventional RABV strains, which could lead to false negative results [ 20 ].
Among the participating laboratories, 28% did not have a written DFA protocol and 82% did not have a written sample approval or rejection criteria. There was variation in the general protocol for sample selection and processing, including the area of brain routinely tested, the type of sample (impressions vs. brain smears), and the fixation protocol, including temperature, time and dilutions for the slides. Details in execution of the immune-staining procedure, including incubation times, incubation temperatures, number of washes as well as mounting medium used also varied greatly. Variations in slide preparation and in the DFA protocols were noted in the 2001 WHO inter-laboratory study [ 14 ], particularly in the use of heat and/or acetone for fixation and fixation time and temperature. Variation in methodology including rinsing and washing times, fixation procedures, type of conjugate used and the staining time and temperature was also noted in the 2009–2010 EURL/ANSES technical questionnaires; nevertheless, in this study, the number of persons reading the same slide was the only factor that significantly affected the proficiency test results [ 18 ]. Standardized protocols for DFA exist from the OIE Terrestrial Manual, the CDC, the EURL and the WHO [ 12 , 21 , 3 ]. The LAC region would benefit from the use of standardized DFA protocols.
Three of the 23 laboratories did not have the capacity to perform a confirmatory testing on weak or inconclusive results. The most common confirmatory test reported was the mouse inoculation; only two laboratories use RT-PCR. Although, the WHO does not currently recommend RT-PCR as a primary rabies routine diagnostic tool, a number of laboratories in LAC are in the process of implementing RT-PCR as a confirmatory test as ascertained during the post-exercise briefing. In an European ring test study, RT-PCR showed a high level of variation between laboratories with cross-contamination as a potential issue [ 4 ]. Concerns about the potential of RT-PCR to identify imported strains were raised by Fischer et al., 2013. Given the expressed interest of LAC laboratories to implement RT-PCR, a region-wide RT-PCR harmonization exercise is recommended due to the test’s requirements for strict quality control protocols and high level of experience and expertise for accurate diagnosis [ 2 ].
In contrast to the acetone-fixed slides provided in this exercise, the three inter-laboratory studies cited in this discussion used freeze-dried viable virus homogenates obtained from animals inoculated intra-cerebrally [ 14 , 15 , 20 ]. The use of fully inactivated gamma radiated homogenates in future iterations of this exercise would also allow for each lab to test its slide or smear preparation protocol as commented in Barrat et al.,2007. Furthermore, homogenizing the material in their own laboratory settings may provide more consistency across the panels than impressions from an infected brain [ 14 ].
The WHO recommends the preventive immunization of all staff handling infected or suspect material with the titer checked every 6 months. While all of the laboratories participating in this exercise required pre-exposure prophylaxis for employees, only 21.7% of the labs reported that employee titers are checked in compliance of WHO-recommendations. Employee safety should be a primary concern for all rabies laboratories and a regional effort to increase the frequency of employee titer checks should be implemented.
As a first iteration of a regional DFA inter-laboratory exercise, this effort provided useful insights into the lab performance and protocols, as well as suggested improvements for future rounds of the exercise. There is a clear need to increase participation in regional exercises on rabies diagnosis across LAC, since only 34.8% of the 23 laboratories reported previous participation in diagnostic test exercises of this nature. A regional laboratory network (REDILAR) is part of the REDIPRA plan for elimination of canine-rabies in the Americas [ 6 ]. Noting the heterogeneity in the DFA protocols described in the technical surveys conducted in this study, and the level of discordant results, an effective REDILAR network is critical to develop a rabies diagnosis and reporting harmonization scheme concerted with an annual testing exercise in the way it has been implemented for Europe [ 21 , 22 ].
Results of this exercise were presented before the Regional Meeting of the Rabies Directors of the Americas in 2015 (REDIPRA 15), which end up in an official resolution instructing the REDILAR coordinated by PANAFTOSA, to define a framework for the harmonization of rabies diagnosis and to implement annual exercises. This resolution demonstrates the commitment by the national authorities, PAHO, and WHO Collaborating Centers for Rabies Research to support the REDILAR laboratory network in improving the sensitivity of rabies diagnosis. Sensitivity, specificity as well as timely rabies diagnosis and surveillance will be of increasing relevance in the race towards elimination of human rabies transmitted by dogs in the Region.
Supporting information
S1 Table
Results of the Rabies Panel Testing by laboratory and sample.
(XLSX)
S2 Table
Responses to the laboratory practices assessment by laboratory.
(XLSX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Overview
With the advent of deep sequencing technologies and the ability to analyze whole genome sequences and transcriptomes, there has been a growing interest in exploring putative functions of the very large fraction of the genome that is commonly referred to as “junk DNA.” Whereas this is an issue of considerable importance in genome biology, there is an unfortunate tendency for researchers and science writers to proclaim the demise of junk DNA on a regular basis without properly addressing some of the fundamental issues that first led to the rise of the concept. In this review, we provide an overview of the major arguments that have been presented in support of the notion that a large portion of most eukaryotic genomes lacks an organism-level function. Some of these are based on observations or basic genetic principles that are decades old, whereas others stem from new knowledge regarding molecular processes such as transcription and gene regulation.
Introduction
The search for function in the genome
It has been known for several decades that only a small fraction of the human genome is made up of protein-coding sequences and that at least some noncoding DNA has important biological functions. In addition to coding exons, the genome contains sequences that are transcribed into functional RNA molecules (e.g., tRNA, rRNA, and snRNA), regulatory regions that control gene expression (e.g., promoters, silencers, and enhancers), origins of replication, and repeats that play structural roles at the chromosomal level (e.g., telomeres and centromeres).
New discoveries regarding potentially important sequences amongst the nonprotein-coding majority of the genome are becoming more prevalent. By far the best-known effort to identify functional regions in the human genome is the recently completed Encyclopaedia of DNA Elements (ENCODE) project [1] , whose authors made the remarkable claim that a “biochemical function” could be assigned to 80% of the human genome [2] . Reports that ENCODE had refuted the existence of large amounts of junk DNA in the human genome received considerable media attention [3] , [4] . Criticisms that these claims were based on an extremely loose definition of “function” soon followed [5] – [8] (for a discussion of the relevant function concepts, see [9] ), and debate continues regarding the most appropriate interpretation of the ENCODE results. Nevertheless, the excitement and subsequent backlash served to illustrate the widespread interest among scientists and nonspecialists in determining how much of the human genome is functionally significant at the organism level.
The origin of “junk DNA”
Although the term “junk DNA” was already in use as early as the 1960s [10] – [12] , the term's origin is usually attributed to Susumu Ohno [13] . As Ohno pointed out, gene duplication can alleviate the constraint imposed by natural selection on changes to important gene regions by allowing one copy to maintain the original function as the other undergoes mutation. Rarely, these mutations will turn out to be beneficial, and a new gene may arise (“neofunctionalization”) [14] . Most of the time, however, one copy sustains a mutation that eliminates its ability to encode a functional protein, turning it into a pseudogene. These sequences are what Ohno initially referred to as “junk” [13] , although the term was quickly extended to include many types of noncoding DNA [15] . Today, “junk DNA” is often used in the broad sense of referring to any DNA sequence that does not play a functional role in development, physiology, or some other organism-level capacity. This broader sense of the term is at the centre of most current debate about the quantity—or even the existence—of “junk DNA” in the genomes of humans and other organisms.
It has now become something of a cliché to begin both media stories and journal articles with the simplistic claim that most or all noncoding DNA was “long dismissed as useless junk.” The implication, of course, is that current research is revealing function in much of the supposed junk that was unwisely ignored as biologically uninteresting by past investigators. Yet, it is simply not true that potential functions for noncoding DNA were ignored until recently. In fact, various early commenters considered the notion that large swaths of the genome were nonfunctional to be “repugnant” [10] , [16] , and possible functions were discussed each time a new type of nonprotein-coding sequence was identified (including pseudogenes, transposable elements, satellite DNA, and introns; for a compilation of relevant literature, see [17] ).
Importantly, the concept of junk DNA was not based on ignorance about genomes. On the contrary, the term reflected known details about genome size variability, the mechanism of gene duplication and mutational degradation, and population genetics theory. Moreover, each of these observations and theoretical considerations remains valid. In this review, we examine several lines of evidence—both empirical and conceptual—that support the notion that a substantial percentage of the DNA in many eukaryotic genomes lacks an organism-level function and that the junk DNA concept remains viable post-ENCODE.
Genome Size and “The Onion Test”
There are several key points to be understood regarding genome size diversity among eukaryotes and its relationship to the concept of junk DNA. First, genome size varies enormously among species [18] , [19] : at least 7,000-fold among animals and 350-fold even within vertebrates. Second, genome size varies independently of intuitive notions of organism complexity or presumed number of protein-coding genes ( Figure 1 ). For example, a human genome contains eight times more DNA than that of a pufferfish but is 40 times smaller than that of a lungfish. Third, organisms that have very large genomes are not few in number or outliers—for example, of the >200 salamander genomes analyzed thus far, all are between four and 35 times larger than the human genome [18] . Fourth, even closely related species with very similar biological properties and the same ploidy level can differ significantly in genome size.
10.1371/journal.pgen.1004351.g001 Figure 1
Summary of haploid nuclear DNA contents (“genome sizes”) for various groups of eukaryotes.
This graph is based on data for about 10,000 species [18] , [19] . There is a wide range in genome sizes even among developmentally similar species, and there is no correspondence between genome size and general organism complexity. Humans, which have an average-sized genome for a mammal, are indicated by a star. Note the logarithmic scale.
These observations pose an important challenge to any claim that most eukaryotic DNA is functional at the organism level. This logic is perhaps best illustrated by invoking “the onion test” [20] . The domestic onion, Allium cepa , is a diploid plant (2n = 16) with a haploid genome size of roughly 16 billion base pairs (16 Gbp), or about five times larger than humans. Although any number of species with large genomes could be chosen for such a comparison, the onion test simply asks: if most eukaryotic DNA is functional at the organism level, be it for gene regulation, protection against mutations, maintenance of chromosome structure, or any other such role, then why does an onion require five times more of it than a human? Importantly, the comparison is not restricted to onions versus humans. It could as easily be between pufferfish and lungfish, which differ by ∼350-fold, or members of the genus Allium , which have more than a 4-fold range in genome size that is not the result of polyploidy [21] .
In summary, the notion that the majority of eukaryotic noncoding DNA is functional is very difficult to reconcile with the massive diversity in genome size observed among species, including among some closely related taxa. The onion test is merely a restatement of this issue, which has been well known to genome biologists for many decades [18] .
Genome Composition
Another important consideration is the composition of eukaryotic genomes. Far from being composed of mysterious “dark matter,” the characteristics of the sequences constituting 98% or so of the human genome that is nonprotein-coding are generally well understood.
Transposable elements
By far the dominant type of nongenic DNA are transposable elements (TEs), including various well-described retroelements such as Short and Long Interspersed Nuclear Elements (SINEs and LINEs), endogenous retroviruses, and cut-and-paste DNA transposons. Because of their capacity to increase in copy number, transposable elements have long been described as “parasitic” or “selfish” [22] , [23] . However, the vast majority of these elements are inactive in humans, due to a very large fraction being highly degraded by mutation. Due to this degeneracy, estimates of the proportion of the human genome occupied by TEs has varied widely, between one-half and two-thirds [24] , [25] . Larger genomes, such as those of salamanders and lungfishes, almost certainly contain an even more enormous quantity of transposable element DNA [26] , [27] .
Many examples have been found in which TEs have taken on regulatory or other functional roles in the genome [28] . In recognition of the more complex interactions between transposable elements and their hosts, Kidwell and Lisch proposed an expansion of the “parasitism” framework where each TE can be classified along a spectrum from parasitism to mutualism [29] . Nevertheless, there is evidence of organism-level function for only a tiny minority of TE sequences. It is therefore not obvious that functional explanations can be extrapolated from a small number of specific examples to all TEs within the genome.
Highly repetitive DNA
Another large fraction of the genome consists of highly repetitive DNA. These regions are extremely variable even amongst individuals of the same population (hence their use as “DNA fingerprints”) and can expand or contract through processes such as unequal crossing over or replication slippage. Many repeats are thought to be derived from truncated TEs, but others consist of tandem arrays of di- and trinucleotides [30] . As with TEs, some highly repetitive sequences play a role in gene regulation (for example, [31] ). Others, such as telomeric- and centromeric-associated repeats [32] , [33] , play critical roles in chromosomal maintenance. Despite this, there is currently no evidence that the majority of highly repetitive elements are functional.
Introns
According to Gencode v17, about 40% of the human genome is comprised of intronic regions; however, this figure is likely an overestimate as it includes all annotated events. It is also important to note that a large fraction of TEs and repetitive elements are found in introns. Although introns can increase the diversity of protein products by modulating alternative splicing, it is also clear that the vast majority of intronic sequence evolves in an unconstrained way, accumulating mutations at about the same rate as neutral regions. Although the median intron size in humans is ∼1.5 kb [30] , data suggest that most of the constrained sequence is confined to the first and last 150 nucleotides [34] .
Pseudogenes
The human genome is also home to a large number of pseudogenes. Estimates of the total number range from 12,600 to 19,700 [35] . These include both “classical” pseudogenes (direct duplicates, of the sort imagined by Ohno [13] ) and “processed” pseudogenes, which are reverse transcribed from mRNA [36] . Once again, although some pseudogenes have been co-opted for organism-level function (for example see [37] ), most are simply evolving without selective constraints on their sequences and likely have no function [38] .
Conserved sequences
Several analyses of sequence conservation between humans and other mammals have found that about 5% of the genome is conserved [1] , [39] – [42] . It is possible that an additional 4% of the human genome is under lineage-specific selection pressure [39] ; however, this estimate appears to be somewhat questionable [43] , [44] (also see [45] ). Ignoring these problems, the idea that 9% of the human genome shows signs of functionality is actually consistent with the results of ENCODE and other large-scale genome analyses.
Besides protein-coding sequences (including associated untranslated regions), which make up 1.5%–2.5% of the human genome [24] , data from ENCODE suggest that conserved long noncoding RNAs (lncRNAs) are generated from about 9,000 loci that add up to less than an additional 0.4% [46] , [47] . Thus, even if a vast new untapped world of functional noncoding RNA is discovered, this will probably be transcribed from a small fraction of the human genome.
At first blush, sequences that are bound by transcription factors (TFs) appear to be very abundant, making up about 8.5% of the genome according to ENCODE [2] . This number, however, is an estimate of regions that are hypersensitive to DNase I treatment due to the displacement of nucleosomes by TFs. As pointed out by others [6] , these regions are annotated as being several hundreds of nucleotides long and are thus much larger than the actual size of individual TF-binding motifs, which are typically 10 bp in length [48] . By ENCODE's own estimates, less than half of the nucleotide bases in these DNase I hypersensitivity regions contain actual TF recognition motifs [2] , and only 60% of these are under purifying selection [49] . Others have found that weak and transient TF-binding events are routinely identified by chromatin IP experiments despite the fact that they do not significantly contribute to gene expression [50] – [53] and are poorly conserved [53] . Given that experiments performed in a diverse number of eukaryotic systems have found only a small correlation between TF-binding events and mRNA expression [54] , [51] , it appears that in most cases only a fraction of TF-binding sites significantly impacts local gene expression.
In summary, most of the major constituents of the genome have been well characterized. The majority of human DNA consists of repetitive, mutationally degraded sequences. There are unambiguous examples of nonprotein-coding sequences of various types having been co-opted for organism-level functions in gene regulation, chromosome structure, and other roles, but at present evidence from the published literature suggests that these represent a small minority of the human genome.
Evolutionary Forces
To understand the current state of the human genome, we need to examine how it evolved, and as Michael Lynch once wrote, “Nothing in evolution makes sense except in the light of population genetics” [55] . Unfortunately, concepts that have been generated by this field have not been widely recognized in other domains of the life sciences. In particular, what is underappreciated by many nonevolution specialists is that much of molecular evolution in eukaryotes is primarily the result of genetic drift, or the fixation of neutral mutations. This view has been widely appreciated by molecular evolutionary biologists for the past 35 years.
The nearly neutral theory of molecular evolution
An important development in the understanding of how various evolutionary forces shape eukaryotic genes and genomes came with the theories developed by Kimura, Ohta, King, and Jukes. They demonstrated that alleles that were slightly beneficial or deleterious behaved like neutral alleles, provided that the absolute value of their selection coefficient was smaller than the inverse of the “effective” population size [56] – [59] . In other words, it is important to keep in mind population size when thinking about whether deleterious mutations are subjected to purifying selection.
It is also important to realize that the “effective” population size is dependent on many factors and is typically much lower than the total number of individuals in a species [55] . For humans it has been estimated that the historical effective population size is approximately 10,000, and this is on the low side in comparison to most metazoans [60] . Given the overall low figures for multicellular organisms in general, we would expect that natural selection would be powerless to stop the accumulation of certain genomic alterations over the entirety of metazoan evolution. One type of mutation that fits this description is intergenic insertions, be they transposable elements, pseudogenes, or random sequence [55] . The creation and loss of TF-binding motifs or cryptic transcriptional start sites in these same intergenic regions will equally be invisible to natural selection, provided that these do not drastically alter the expression of any nearby genes or cause the production of stable toxic transcripts. Thus, a central tenet of the nearly neutral theory of molecular evolution is that extraneous DNA sequences can be present within genomes, provided that they do not significantly impact the fitness of the organism.
Genetic load
It has long been appreciated that there is a limit to the number of deleterious mutations that an organism can sustain per generation [61] , [62] . The presence of these mutations is usually not harmful, because diploid organisms generally require only one functional copy of any given gene. However, if the rate at which these mutations are generated is higher than the rate at which natural selection can weed them out, then the collective genomes of the organisms in the species will suffer a meltdown as the total number of deleterious alleles increases with each generation [63] . This rate is approximately one deleterious mutation per generation. In this context it becomes clear that the overall mutation rate would place an upper limit to the amount of functional DNA. Currently, the rate of mutation in humans is estimated to be anywhere from 70–150 mutations per generation [64] , [65] . By this line of reasoning, we would estimate that, at most, only 1% of the nucleotides in the genome are essential for viability in a strict sequence-specific way. However, more recent computational models have demonstrated that genomes could sustain multiple slightly deleterious mutations per generation [66] . Using statistical methods, it has been estimated that humans sustain 2.1–10 deleterious mutations per generation [66] – [68] . These data would suggest that at most 10% of the human genome exhibits detectable organism-level function and conversely that at least 90% of the genome consists of junk DNA. These figures agree with measurements of genome conservation (∼9%, see above) and are incompatible with the view that 80% of the genome is functional in the sense implied by ENCODE. It remains possible that large amounts of noncoding DNA play structural or other roles independent of nucleotide sequence, but it far from obvious how this would be reconciled with “the onion test.”
The evolution of the nucleus
When dealing with the evolution of any lineage, one must also keep in mind unique events, also known as historical contingencies, which constrain and shape subsequent evolutionary trajectories [69] . One of these key events in our own ancestry was the evolution of the eukaryotic nucleus. A further examination of why the nucleus evolved and how this altered cellular function may generate significant insights into the current shape of the eukaryotic genome.
One important event in early eukaryotic evolution was the development of a symbiotic relationship between the α-proteobacteria progenitor of mitochondria and an archaebacteria-like host [70] , [71] . As with most endosymbiotically derived organelles [72] , DNA was transferred from mitochondria to the host. In this way, Group II introns, which are still found in both mitochondria and α-proteobacteria [73] , invaded the host genome. Group II introns are parasitic DNA fragments that replicate when they are transcribed, typically as part of a larger transcript. The intron then folds into a catalytic ribozyme that splices itself out of the precursor transcript and then reinserts itself at a new genomic locus by reversing the splicing reaction. Importantly, functional fragments of Group II introns can splice out inactive versions in a trans-splicing reaction [74] , [75] . As described elsewhere, it is likely that Group II introns proliferated and evolved into two populations: inactivated copies that could be nonetheless spliced out in trans, and active fragments that promoted splicing of the former group. This latter group eventually evolved into the spliceosomal snRNAs [75] – [77] . This idea is supported by not only structural, catalytic, and functional similarities between Group II introns and snRNAs [78] , [79] but also by the fact that expression of the U5 snRNA rescues the splicing of Group II introns that lack the corresponding U5-like region [80] .
It is likely that the proliferation of trans-splicing triggered the spatial segregation of RNA processing (the nucleoplasm) from the translation machinery (the cytoplasm) [77] . This subdivision ensured that mRNAs were properly spliced before they encountered the translation machinery. Not only would this segregation prevent translating ribosomes from interfering with the splicing reaction (and vice versa) but would also prevent the translation of incompletely processed mRNAs, which often encode toxic proteins [81] , [82] . Importantly, the segregation of translation from both transcription and RNA processing provided an opportunity for nuclear quality-control processes to eliminate misprocessed and spurious transcripts that did not meet the minimal requirements of “mRNA identity” (see below) before these RNAs ever encountered a ribosome. This in turn permitted intergenic DNA and cryptic transcriptional start sites to proliferate with minimal cost to the fitness of the organism. It should also be noted that the increase in ATP regeneration due to mitochondrial-derived metabolic pathways provided the surplus energy that is required to support an expansion not only in genome size and membranes [83] , [84] but also wasteful transcription. Thus, by several independent mechanisms, the acquisition of mitochondria likely allowed the expansion of nonfunctional intergenic DNA and the evolution of a noisy transcriptional system.
Gene Expression in Eukaryotes
Eukaryotic transcription is inherently noisy
One of the most widely discussed discoveries of the past decade of transcriptome analysis is that much of the metazoan genome is transcribed at some level (although this, too, was already recognized in rough outline in the 1970s [15] ). When nascent transcripts from mouse have been analyzed by deep sequencing, the total number of reads that map to intergenic loci is almost equivalent to the number mapping to exonic regions ( Figure 2A , reproduced from reference [85] ). This is consistent with the observation that a large fraction of the cellular pool of RNA Polymerase II is associated with intergenic regions [86] and that transcription can be initiated at random sequences (see Figure S4 in [87] ) and nucleosome-free regions [88] , [89] . Strikingly, when one examines the steady state level of polyadenylated RNA, very little maps to intergenic regions ( Figure 2A, 2B , the latter reproduced from reference [46] ; also see [85] , [90] – [92] ). In fact, when one eliminates the ∼9,000 transcript species that are thought to be derived from conserved lncRNA, then most of the annotated noncoding polyadenylated RNAs are present at levels below one copy per cell and are found exclusively in the nucleus ( Figure 2B ). The situation is no better in the unpolyadenylated pool, in which the amount of lncRNA and intergenic RNA is practically insignificant, especially in the cytoplasmic pool ( Figure 2B ). In aggregate, these data indicate that the majority of intergenic RNAs are degraded almost immediately after transcription. Consistent with this idea, the level of intergenic transcripts increase when RNA degradation machinery is inhibited [93] – [101] . Although pervasive transcription has been used as an argument against junk DNA [3] , [4] , it is in fact entirely in line with the idea that intergenic regions are evolving under little-to-no constraint, especially when one considers that this intergenic transcription is unstable.
10.1371/journal.pgen.1004351.g002 Figure 2
Levels of protein-coding and intergenic RNAs in mammalian cells.
(A) Analysis of nascent and total poly(A)+ RNA levels from mouse liver nuclei. Nascent (i.e., polymerase-associated) RNA and poly(A)+ RNA were isolated from mouse liver nuclei and analyzed by high-throughput sequencing. Individual reads were categorized by their source. Exonic and intronic are from known referenced genes (i.e., “RefSeq” genes), while intergenic originate from nonreferenced loci (i.e., “non-RefSeq”) in the mouse genome. Reproduced from [85] . (B) Empirical Cumulative Distribution Function (ECDF) of transcript expression in each cell compartment as determined by the ENCODE consortia. Results for RNA that either contain (“polyA+”) or lack (“polyA−”) a poly(A)-tail in the nucleus and cytosolic fractions are shown. Each human cell line that was analyzed is represented by three lines, one for each pool of RNA (red for protein-coding RNAs, blue for lncRNAs [“noncoding”], and green for intergenic transcripts [“novel intergenic”]). The lines indicate the cumulative fraction of RNAs in a given pool (y-axis) that are expressed at levels that are equal or less than the reads per kilobase per million mapped reads (RPKM) on the x-axis. Total numbers in each pool are as follows: reference protein coding genes: 20,679, loci producing lncRNAs: 9,277, and regions producing intergenic transcripts: 41,204. Transcripts with expression levels of 0 RPKM were adjusted to an artificial value of 10 −6 RPKM so that the onset of each graph represents the fraction of nonexpressed genes or loci. Note that 1–4 RPKM is approximately equivalent to one copy per tissue culture cell [46] , [129] . Using this figure, one can easily deduce that the vast majority of intergenic transcripts are present at levels less than one copy per cell. Reproduced with permission from [46] .
Identifying mRNA from intergenic transcription
A common theme that has emerged from the study of mRNA synthesis is that various steps in RNA synthesis and processing are biochemically coupled. In other words, cellular machineries that participate in one biochemical activity also promote subsequent steps. For example, during the splicing of the 5′most intron, the spliceosome collaborates with the 5′cap binding complex to deposit nuclear export factors onto the 5′end of the processed transcript [102] , [103] , and this helps to explain why splicing enhances the nuclear export of mRNA [104] – [106] . Countless other examples of coupling exist (for reviews, see [107] – [111] ).
The ultimate goal of these coupling reactions is to sort protein-coding RNAs (i.e. mRNA) from intergenic transcripts [111] , [112] . Given that, on average, protein-coding genes have eight introns [30] , while the majority of annotated ENCODE intergenic transcripts tend not to be spliced [46] , introns help distinguish these two populations and thus serve as “mRNA identity” markers. These mRNA identity features activate coupling reactions, which in turn promote the further processing, nuclear export, and translation of a particular transcript. Likewise, other classes of functional RNAs (e.g., tRNAs and snRNAs) have their own identity elements [113] . In contrast, transcripts that lack identity elements are targeted for degradation. In agreement with this model, intronless RNA molecules that have a random sequence are poorly exported from the nucleus and have a very short half-life [114] , [115] . In contrast, intronless mRNAs have specialized motifs that promote their nuclear export [105] , [116] – [119] .
In light of the fact that many functional lncRNAs serve a role in regulating chromatin structure or transcription, it is not surprising that most localise to the nucleoplasm [46] . One would predict that lncRNAs contain a differential set of identity elements that not only serve to prevent their decay but also retain them in the nucleus. This would especially be critical for lncRNAs that are spliced. Despite this, the elements that regulate the localization and stability of these RNAs have received little attention, but can be informed by the view that they may have their own identity markers.
It is also important to point out that eukaryotes have other mechanisms that either degrade aberrant mRNAs (e.g., nonsense-mediated decay) or limit the amount of intergenic transcription (e.g., heterochromatin). Nevertheless, eukaryotes appear to have evolved an intricate network of coupling reactions that are required to cope with a large burden of junk RNA. These findings are consistent with the idea that eukaryotic genomes are filled with junk DNA that is transcribed at a low level.
An alternative view of transcription and conservation?
In an attempt to counter the argument that sequence conservation is a prerequisite for functionality, it has been recently proposed that certain transcriptional events may serve some role in regulating cellular function, despite the fact that the sequence of the transcriptional product is unconstrained [120] . Indeed, this view is in line with the findings that the transcription of certain yeast genes is inhibited as a consequence of the production of cryptic unstable transcripts originating from upstream and/or downstream promoters (for a review see [121] ). Other examples have linked the generation of cryptic unstable transcripts to chromatin modifications [101] , [122] , DNA methylation [123] , and DNA stability [124] . However, it remains unclear whether the majority of unstable noncoding RNAs have any effect on DNA or chromatin, let alone contribute to the fitness of the organism. In the cases where cryptic unstable transcriptional events impact gene expression, they usually consist of short transcripts that are synthesized from regions around the transcriptional start sites or within the gene itself [121] . Indeed most of the available data are consistent with the fact that transcriptional start sites are promiscuous, often generating bidirectional transcription [100] , [101] , and that subsequent coupling processes, such as the interaction between promoter-associated complexes and 3′end processing factors, are required to enforce proper transcriptional directionality [125] . Other unstable transcripts function to promote or maintain heterochromatin formation in the vicinity of the transcriptional site, likely because these regions produce toxic transcripts [122] . Although this form of transcription has a function (viz., to maintain a repressive state), it is not clear that the elimination of these regions would have any effect on the organism [8] . The transcription of other short unstable transcripts, mostly produced from enhancer regions, has been shown to promote gene expression [126] ; however, again these “enhancer RNAs” are transcribed from a small fraction of the total genome [127] . As stated by others [128] , it is imperative that those who claim that the vast majority of intergenic transcription is functional test their hypotheses. In the absence of this evidence, the declaration that we are in the midst of a paradigm shift with regards to eukaryotic genomes and gene expression [120] seems premature.
Concluding Remarks
For decades, there has been considerable interest in determining what role, if any, the majority of the DNA in eukaryotic genomes plays in organismal development and physiology. The ENCODE data are only the most recent contribution to a long-standing research program that has sought to address this issue. However, evidence casting doubt that most of the human genome possesses a functional role has existed for some time. This is not to say that none of the nonprotein-coding majority of the genome is functional—examples of functional noncoding sequences have been known for more than half a century, and even the earliest proponents of “junk DNA” and “selfish DNA” predicted that further examples would be found. Nevertheless, they also pointed out that evolutionary considerations, information regarding genome size diversity, and knowledge about the origins and features of genomic components do not support the notion that all of the DNA must have a function by virtue of its mere existence. Nothing in the recent research or commentary on the subject has challenged these observations.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
The enormous complexity of the human nervous system represents a barrier for modern drug discovery. Advances in combinatorial chemistry and high-throughput screening (HTS) technologies mean that high-throughput target-based screening assays are frequently used to identify potential drug candidates [1] . This creates a gap between simplified target-based screenings and in vivo animal models of human central nervous system (CNS) diseases. Consequently, candidate compounds have a high dropout rate in this transition. The development of in vitro phenotypic assays can help bridge this gap by combining moderate throughput with enough biological context to capture some of the molecular mechanisms of the disease [2] . Given the demands and the difficulty of working with animal models of CNS diseases, in vitro phenotypic assays have potential utility as screens for compounds before these are progressed to in vivo testing [1] .
The field of epilepsy is perhaps one of the best examples of the benefit of using animal models for screening compounds. Although many modern antiepileptic drugs (AEDs) have been discovered using target-based approaches, much of the success of this field comes from the early development of animal seizure models [3] – [5] . However, as the field moves from a primary focus of controlling seizures to the development of drugs able to address aspects of disease pathophysiology [6] – [8] , this requires the adoption of much resource- and time-consuming chronic animal models that can longer sustain the testing of even moderate numbers of compounds [9] . In vitro models of epilepsy largely rely on electrophysiological measurements of epileptiform activity in brain slices or dissociated neuronal cultures [10] , [11] . Such readouts impose significant throughput limitations, making these assays more suitable for mode of action studies than for compound identification or secondary screening. Consequently, there is a need for new in vitro functional epilepsy assays able to provide a medium throughput while still preserving sufficient biological context to allow for the identification of compounds with new modes of action.
An alternative approach for measuring neuronal activity is the use of fluorescent probes to monitor the fluctuations of intracellular calcium levels that accompany neuronal depolarization [12] . Image-based measurement of intracellular calcium levels relies on single neuron analysis and has therefore similar throughput limitations as electrophysiological recordings. Fluorescent plate reader-based calcium assays, on the contrary, are amenable to medium or high throughput screening [12] – [14] , but can only measure fluctuations of intracellular calcium levels when these occur in a large number of cells. Interestingly, such synchronized intracellular calcium oscillations are reported to occur in high density primary cultures of the hippocampus and cortex, either in a spontaneous fashion or in response to triggering factors such as incubation in low magnesium buffer [15] – [18] . Dual intracellular recordings from multiple neurons and calcium imaging have established that calcium oscillations are strictly associated with bursts of action potentials in these neurons representing epileptiform activity [11] , [19] , [20] . The same phenomenon has also been reported in hippocampal slices, where it is known that each calcium oscillation corresponds to an individual ictal event experienced by most neurons in the network in a synchronized manner [21] . Despite the high potential of such medium-throughout fluorescent assays, they have only been described in a small number of publications and have not yet been characterized as a predictive tool for AED discovery.
Here we describe a robust and simple fluorescence-based calcium assay to measure epileptiform network activity using rat primary cortical cultures in a 96-well format. We have adapted two well-known cellular models of epileptiform activity to this assay format –low magnesium and 4-aminopyridine models –, and evaluated the contribution of sodium and calcium voltage-gated ion channels and AMPA, NMDA and GABA receptors to epileptiform activity in both models. We have also evaluated their translatability using a panel of AEDs with a variety of modes of action. The in vitro phenotypic assay we describe here has a great potential as a functional screen for antiepileptic activity to help identifying compounds with new modes of action and pathways with previously unknown contribution to epilepsy.
Results
High density cortical cultures develop spontaneous calcium oscillations
Cortical and hippocampal primary cultures from mouse or rat tissue have been described to develop spontaneous intracellular calcium oscillations when cultured at high density. These oscillations occur throughout the entire population of primary neurons in a pattern that resembles neuronal hyper-synchrony during epileptiform activity [11] , [15] – [20] . To determine the extent to which neuronal calcium oscillations can be used as an in vitro model of epileptiform activity, we started by reproducing these observations in our laboratory. We loaded primary cultures from rat cortex with the fluorescence calcium indicator Fluo-4 and observed their spontaneous behavior using a fluorescent plate reader (Flexstation). We evaluated cultures at different stages of development (6–20 days in vitro ) and at different densities by monitoring their activity over 10-minute periods with data acquisition every 0.8 seconds. In early experiments we determined that this short, sub-second, acquisition timing was needed not to miss the peak of the oscillations.
In high density neuronal cultures (50,000 cells per well in 96-well plates), the spontaneous development of intracellular calcium oscillations occurred around days 9–10 ( Figure 1 ). The pattern of the oscillations among the wells was variable, with some showing no oscillations and others producing a limited number of oscillations during the 10-minute recording ( Figure 1B and D ). After 13 days in vitro the calcium oscillations pattern reached a plateau and stabilized at around 15 oscillations per 10-minute recording ( Figure 1C and D ). Cultures younger than 9 days failed to show any calcium oscillation ( Figure 1A and D ).
10.1371/journal.pone.0084755.g001 Figure 1
High-density cortical cultures develop spontaneous intracellular calcium oscillations.
Primary cultures from rat cortex were loaded with the intracellular calcium reporter dye Fluo-4 and imaged in a fluorescent plate reader. Representative images of 10-minute recordings are used for A-C. (A) Up to 9 days in vitro (DIV), the cellular population intracellular calcium level remained constant, with no peak observed during 10-minute recordings. (B) around days 9 and 10, fast elevations in intracellular calcium levels could be detected in some wells. (C-D) After 11-12 DIV the phenotype of calcium oscillations becomes established, with cultures displaying more frequent and rhythmic oscillations. (D) Graph showing the onset of calcium oscillations quantified as number of oscillations per 10-minute recording after following cultures from 3 separate batches of neurons (2 wells per batch and day, minimum 4 wells per time point) over a period of 8 days, from 6 to 14 DIV, and during their third week in vitro . Asterisk denotes sample statistically different from zero (one sample t-test, p<0.05).
While the number of oscillations per well was variable, their frequency was constant for any given well, allowing for pharmacological studies aimed at modulating this frequency.
Cultures plated at a slightly lower density (40,000 cells per well) had a more variable onset of oscillations but reached a similar plateau. Cultures of 20,000 cells per well or lower failed to produce any oscillations during a period of 20 days or produced a few oscillations in random wells (not shown). We therefore decided to work with cultures at 13–15 days in vitro and 50,000 cells per well for all subsequent pharmacological studies.
Calcium oscillations are affected by compounds that modulate the activity of voltage- and ligand-gated ion channels
In order to determine to how closely synchronized calcium oscillations mimic epileptiform activity, we explored their modulation by low magnesium and 4-aminopyridine (4-AP), two established methods for inducing epileptiform activity in tissue slices [10] , [22] , [23] .
We first evaluated the impact of varying concentrations of magnesium in the recording buffer on spontaneous calcium oscillations ( Figure 2A ). When high density neuronal cultures were exposed to low magnesium buffer (0.1 mM MgCl 2 ), their spontaneous calcium oscillations increased in frequency and became more stable, with most oscillations reaching the same peak fluorescence level. Conversely, when the cultures where exposed to high magnesium concentrations, cultures displayed very few or no oscillations during the 10-minute recording ( Figure 2A ). These observations are consistent with the ability of extracellular magnesium levels to reduce calcium influx by blocking the NMDA receptor channel [10] .
10.1371/journal.pone.0084755.g002 Figure 2
Low magnesium-enhanced calcium oscillations are modulated by GABA and glutamate receptors, and ion channels.
(A) Representative recordings of cultures imaged in the presence of normal magnesium levels in the buffer (1 mM, center), high magnesium buffer (3 mM) or low magnesium buffer (0.1 mM). While high magnesium in the buffer arrested or severely reduced the number of oscillations, incubation in 0.1 mM magnesium led to more numerous and stable (similar amplitude) oscillations. (B-D) Response of the low magnesium-enhanced calcium oscillations to GABA receptor agonism (B), NMDA receptor inhibition (MK-501, C) and AMPA receptor inhibition (CNQX, D) injected 4 minutes into the recording (time exposed to the compound denoted by blue bar). (E-G) Response of the calcium oscillations to ion channel inhibitors.
We next evaluated the sensitivity of the calcium oscillations to modulation of the four main molecular targets of AEDs, namely GABA receptors, glutamate receptors, and voltage-gated sodium and calcium channels [24] , [25] . For these experiments we recorded the oscillation pattern in each well for a baseline period of 4 minutes, followed by a further 6 minutes after injecting the different compounds into the wells. We also used low magnesium buffer to stabilize the variability in the pattern and frequency of the spontaneous oscillations. Increasing GABAergic tone using 30 µM GABA ( Figure 2B ) and reducing glutamatergic neurotransmission using selective NMDA and AMPA inhibitors (MK-801 and CNQX at 25µM, respectively; Figure 2C-D ) resulted in a decrease or suppression in the appearance of calcium oscillations. Similarly, inhibition of sodium and L-type calcium channels using 3 nM TTX and 30 µM nimodipine, respectively, suppressed the calcium oscillations ( Figure 2E-F ). Inhibition of N-type calcium channels by omega-conotoxin GVIA, even at high concentrations (8µM) had little impact on the oscillations, Figure 2G ). These findings are consistent with the ability of magnesium to modulate the oscillations frequency as reported previously [15] .
We also explored the modulation of calcium oscillations by 4-AP, widely reported in the literature to trigger epileptiform activity in non-epileptic tissues ( [10] , [22] , [23] Figure 3 ). In our assay, application of 4-AP resulted in an increase in basal intracellular calcium levels in a dose-response, but transient manner. As the calcium levels gradually returned to baseline, we observed a concomitant increase in the frequency of calcium oscillations that remained for the duration of the 10-minute recording ( Figure 3A ).
10.1371/journal.pone.0084755.g003 Figure 3
4-AP-enhanced calcium oscillations are modulated by GABA and glutamate receptors, and ion channels.
(A) Representative recordings of cultures exposed to vehicle or 4-AP 4 minutes into the recording (time exposed to the compound denoted by blue bar is 6 minutes). 4-AP induced a dose-response transient elevation of intracellular calcium baseline accompanied by an increase in frequency. (B-D) Response of the 4-AP-enhanced calcium oscillations to GABA receptor agonism (B), NMDA receptor inhibition (MK-501, C) and AMPA receptor inhibition (CNQX, D). (E-G) Response of the calcium oscillations to ion channel inhibitors.
Similarly to the low magnesium model, 4-AP-enhanced calcium oscillations were sensitive to the same compounds that reduced or inhibited calcium oscillations in the low magnesium model, namely GABA, MK-801, CNQX, TTX and nimodipine, while they were insensitive to omega-conotoxin GVIA ( Figure 3B -G). 4-AP-enhanced calcium oscillations had a tendency to be more resistant than low magnesium-enhanced calcium oscillations, with doses of compounds that completely suppressed low magnesium-enhanced calcium oscillations producing only partial reduction in the model. Nevertheless, all the modes of action tested suppressed both low magnesium- and 4-AP-enhanced calcium oscillations at higher compound concentration with the exception of omega-conotoxin GVIA ( Figure 2 - 3 ). Collectively, these data confirm the involvement of voltage-and ligand-gated ion channels known to control neuronal excitability in the generation of synchronized calcium oscillations in vitro and suggest this assay could mimic the mechanisms involved in generating epileptiform activity.
Neuronal calcium oscillations respond to a variety of antiepileptic drugs
In order to be used as a secondary screen and predictive tool for the discovery of antiepileptic compounds, the epileptiform-like phenotype displayed by the high density cortical cultures must be relevant to the human disease. To better understand the translatability of this in vitro phenotypic assay, we evaluated the ability of a panel of approved AEDs with a variety of modes of action to reduce the expression of calcium oscillations ( Figures 4 and 5 ). Because low magnesium-induced calcium oscillations were more sensitive to compounds targeting ion channels and receptors than those enhanced by 4-AP, we profiled the panel of AEDs in this model.
10.1371/journal.pone.0084755.g004 Figure 4
Changes in Calcium oscillations in response to application of classic antiepileptic drugs.
(A-G) Representative recordings of cultures exposed to antiepileptic drugs that work by enhancing GABA receptor activity (A-B), inhibiting sodium channels (C-D) and inhibiting calcium channels (E-G). While other AEDs showed activity in the model, ethosuximide turned from inactive to excitatory at doses higher than 500µM. Time exposed to the compound denoted by blue bar is 6 minutes.
10.1371/journal.pone.0084755.g005 Figure 5
Changes in calcium oscillations in response to application of antiepileptic drugs working through a variety of modes of action.
(A-E) Representative recordings of cultures exposed to an agonist of cannabinoid CB1 receptors (A), a potassium channel opener (B), showing activity in this model Topiramate was inactive at the highest dose that could be tested (C). Levetiracetam also failed to show activity in this model at concentrations up to 5 mM (D-E) while valproate showed partial activity at concentrations up to 2 mM (G-H) and suppression of calcium oscillations at 5 mM (I). Time exposed to the compound denoted by blue bar is 6 minutes.
We first evaluated AEDs known to act through classic modes of action, namely by activating GABA receptors or inhibiting sodium and calcium voltage-gated channels ( Figure 4 ) [25] . Administration of GABA A and GABA B receptor agonists (30µM diazepam and 30µM baclofen, respectively) suppressed the formation of calcium oscillations ( Figure 4A–B ). The sodium channel inhibitors carbamazepine and phenytoin were also able to suppress the formation of calcium oscillations (100 and 30µM respectively; Figure 4C-D ). In contrast, gabapentin was only able to reduce the generation of calcium oscillations when used at high concentrations (2 mM, Figure E-F). In the case of a second AED also thought to have efficacy by modulating calcium channels, ethosuximide, there was a trend towards exacerbating the calcium oscillations when applied at high doses, in a similar manner to 4-AP ( Figure 4G ) and no activity at lower doses. These observations suggest some of the main AED targets, but not all, are able to modulate calcium oscillations in this assay.
A number of AEDs or compounds with anticonvulsant properties in rodents are believed to act through other, non-classic, modes of action [24] . Application of the cannabinoid receptor agonist WIN 55212-2 resulted in a gradually decrease on calcium oscillations at 10µM, the highest dose that could be tested due to solubility limitations ( Figure 5A ) suggesting CB1 receptors are expressed in the mixed cortical cultures and contribute to regulating neuronal excitability. The potassium channel opener retigabine also had efficacy in the oscillation assay at low µM concentrations(3 µM; Figure 5B ). In contrast, the calcium oscillation assay was not sensitive to other approved AEDs such as topiramate and levetiracetam at the highest dose that could be tested for each compound ( Figure 5C and D-F ), and in the case of valproate, it only showed activity at very high concentrations (mM range, Figure 5G-I ). These results are summarized in Table 1 . Overall these data suggest the fluorescence-based calcium oscillations assay captures several signaling pathways that are involved in controlling neuronal excitability including a number of targets for known antiepileptic drugs.
10.1371/journal.pone.0084755.t001 Table 1
Summary of the compounds and antiepileptic drugs evaluated in the calcium oscillations assay.
Compound
Mechanism
Modulation of calcium oscillations
GABA
GABA receptor agonist
+
Diazepam
GABA A receptor agonist
+
Baclofen
GABA B receptor agonist
+
TTX
Na + channel inhibitor
+
Carbamazepine
Na + channel inhibitor
+
Phenytoin
Na + channel inhibitor
+
Nimodipine
L-type Ca +2 channel inhibitor
+
Conotoxin GVIA
N-type Ca +2 channel inhibitor
−
Ethosuximide
T-type Ca +2 channel inhibitor
−
Gabapentin
Alpha2delta-1 Ca +2 channel inhibitor
+
MK-801
NMDA receptor inhibitor
+
CNQX
AMPA/KA receptor inhibitor
+
WIN 55212-2
Cannabinoid receptor agonist
+
Retigabine
K + channel opener
+
Levetiracetam
SV2A ligand
−
Topiramate
Multiple mechanisms
−
Valproate
Multiple mechanisms
+
Discussion
As the epilepsy field moves from a primary focus on controlling seizures to addressing disease pathophysiology there is a need for new in vitro functional assays able to balance throughput and translatability. Here we describe a phenotypic cell-based assay using rat cortical cultures that captures several of the mechanisms involved in epileptiform activity, including those through which numerous approved AEDs exert their therapeutic activity. Because of its throughput and translatability, the calcium oscillations assay has the potential to be used for screening compounds with potential antiepileptic activity before in vivo testing.
Studies reported in the literature have shown that hippocampal and cortical cultures develop spontaneous intracellular calcium oscillations that resemble epileptiform activity [15] , [16] , [18] , [19] , [26] . A well-defined phenotype is the single most important consideration when designing a phenotypic assay or screen [2] . Our data confirm and extend previous observations that glutamate and GABA receptors are involved in the formation of the spontaneous calcium oscillations [15] , [17] – [20] . We observed enhanced sensitivity to NMDA receptor blockade in the low magnesium model compared with the 4-AP model, with doses of MK-801 able to suppress calcium oscillations in the former providing only partial suppression in the latter. Interestingly we observed the reverse pattern in the case of the AMPA and kainate receptor inhibitor CNQX, which had preferential activity in the 4-AP model. These observations are consistent with previous descriptions of the neuronal oscillations model using calcium imaging by Wang and Gruenstein [17] . In their study, Wang and Gruenstein also observed the necessity of L-type, but not N-type, calcium channels for the continuation of calcium oscillations in cortical cultures that we observed in our plate reader assay [17] . Also supporting our observations, synchronized calcium oscillations are described to be sensitive to the sodium channel blocker TTX and to the L-type calcium channel blocker nimodipine [15] , [19] , [20] . Therefore, we confirm in our study that a majority of targets associated with the mode of action of classic AEDs appear to be involved in the generation of calcium oscillations in cortical cultures.
To evaluate the translatability of this assay, we tested the ability of a panel of AEDs with a diverse set of modes of actions to suppress the formation of calcium oscillations. While there is in vitro data for all of these drugs, many have been profiled in rat hippocampal slices by electrophysiology and not in primary culture models of epileptiform activity. Compounds that were active in our assay have also been reported to have activity either in similar primary culture assays (WIN 55212-2, baclofen [11] , [18] , [27] ) or in the low magnesium model in slices (carbamazepine, phenytoin and retigabine [28] – [30] ). Interestingly ethosuximide, which failed to inhibit calcium oscillations in our assay, was also inactive up to 1 mM in a similar assay using hippocampal neurons [11] , confirming our observations.
Some AEDs with proven efficacy for certain types of epilepsy were unable to suppress the phenomenon of calcium oscillations in this assay. In the case of the synaptic vesicle 2A protein ligand levetiracetam, previous studies using tissue slices have shown that its activity might depend on the specific model tested, being reported to reduce epileptiform activity in human neocortical slices exposed to the GABA receptor inhibitor bicuculline but not when epileptiform activity was induced by low magnesium [31] . While the activity of levetiracetam in the first model was reported at 100–500µM, concentrations up to 1 mM were inactive in the low magnesium slice model [31] , consistent with our observations using primary cortical neurons with concentrations up to 5 mM. Topiramate has also been shown to be active in hippocampal primary neurons where epileptiform burst firing events are induced following transient exposure to a medium with no magnesium at doses that were inactive in our assay [33] . However a potential activity of topiramate in this model cannot be ruled out because the compound solubility limited our ability to evaluate concentrations higher than 100µM. Indeed two other AEDs that were inactive in this assay at micromolar concentrations were able to suppress the calcium oscillations when tested at concentrations above 1 mM. This was the case of gabapentin, which has been shown to be also inactive in the low magnesium model using human neocortical slices at micromolar concentration [32] , but to be partly active in a combined low magnesium high potassium model using rat hippocampal slices [28] , suggesting its activity might also depend on the specific model being tested. Valproate has been described to have modest activity at 1 mM concentration in the low magnesium model using rat hippocampal slices [28] , [29] . In our assay, valproate only produced a small reduction in the calcium oscillations peak amplitude at 1 mM but progressed to complete efficacy when tested at 5 mM. Because slice studies measure epileptiform activity by electrophysiology it is possible that such a read out identifies the anti-epileptiform activity of some compounds before it can be measured in the fluorescent assay. Alternatively, the efficacy of gabapentin and valproate in this assay at high, potentially non-physiological, concentrations might be mediated by receptors other than those driving their antiepileptic activity in vivo and in more comprehensive in vitro systems such as slice assays. Overall, our head to head comparison of a panel of AEDs indicates the fluorescence-based calcium oscillations assay captures several signaling pathways that are involved in controlling neuronal excitability including a number of targets for known antiepileptic drugs and is comparable to other in vitro epilepsy assays.
For those pathways that are conserved in the mixed cortical cultures, this assay has significant potential as a predictive tool for AED discovery and evaluation. Not only does it offers a fast response and medium throughput, but it can also be used to obtain early in vitro proof of concept with compounds that are not suitable for in vivo testing due to poor pharmacokinetic properties or via genetic modulation such as target knockdown [34] . The nature of the mixed cortical cultures also offers the possibility of evaluating the role of glial proteins in the generation of epileptiform activity [35] . For all of these pathways, in vitro phenotypic assays can help bridge the gap between primary target-based screening and in vivo testing and be used to characterize and preselect the compounds that will progress through the drug discovery cascade.
Nevertheless, the calcium oscillations assay shows some degree of drug refractoriness, in particular to those AEDs that are believed to have complex modes of action ( [24] , Table 1 ). A number of reasons might explain this observation. First, some of the targets of these AEDs might not be expressed in the cultures or not contribute substantially to the epileptiform activity. This might be the case of the N-type calcium channels. Second, some important ligands or factors that are needed to engage those targets might be missing in these cultures, thereby making modulators of those targets ineffective in the assay. Last, some important aspects of epilepsy are not cell autonomous and might require an appropriate circuit connectivity that is lacking in the dissociated cultures. For these modes of action, the calcium oscillations model might not sufficiently capture the complexity of the disease and highlights the need to use different models when evaluating the potential of a candidate drug [36] .
Some degree of drug refractoriness is also observed in animal models of epilepsy, where some models or subgroups of animals show a poor response to AEDs [37] . The specific molecular and cellular mechanisms that lead to pharmacoresistant forms of epilepsy remain largely unknown [6] , [7] . This lack of knowledge represents a barrier for the development of target-based drug screening for refractory epilepsy. The phenotypic calcium oscillations assay that we describe here offers an interesting opportunity by allowing the primary screening of compounds with potential efficacy in pharmacoresistant epilepsy without previous information about their biochemical targets. Whether phenotypic screens using this assay can successfully identify compounds with efficacy in rodent models and patients with pharmacoresistant epilepsy remains to be determined.
In addition to being used as primary and secondary screens, phenotypic assays can be used to better understand the mechanism behind their disease-relevant phenotype and shed light on the pathology that they model. In the case of the calcium oscillations assay, future research on the process leading to the onset of the oscillations in vitro might also provide useful information for understanding the mechanisms underlying epileptogenesis. Overall, the fluorescent-based phenotypic assay that we have characterized has the potential to be used as a predictive tool for AED discovery and to bridge the gap between target-based screening and in vivo testing. Given the demands and difficulties of working with animal models of CNS diseases, such in vitro phenotypic assay can bring great value to the discovery of AEDs.
Materials and Methods
Ethics statement
All experiments involving animals were approved by the UCB ethical committee for animal experimentation, in accordance with the European Directive 2010/63/EU on the protection of animals used for scientific purpose and with the Belgian law on the use of laboratory animals.
Reagents
Fluo-4 non wash was purchased from Invitrogen. Hank's balanced salt solution (HBSS), Hepes, CaCl 2 and MgCl 2 from Lonza. Diazepam was from Fragon; phenytoin from Fluka; conotoxin GVIA and TTX from Alomone Labs; gabapentin from TCI Chemicals; and levetiracetam and retigabine were synthetized at UCB. All other reagents were from Sigma Aldrich.
Primary neurons
Cortical cultures were prepared from rat embryos (OFA) at 17–18 days postcoitum and were cultured on poly-L-lysine coated 96-well plates (Greiner) following a protocol adapted from Mingorance-Le Meur and O’Connor (2009). Brain cortex was collected in ice-cold PBS and dissociated in 0.25% trypsin (Invitrogen) in PBS for 20 minutes in the incubator. Trypsin activity was then stopped by adding serum-containing plating medium consisting of neurobasal, 1% penicillin-streptomycin, 2% B27, 0.25% glutamax and 10% normal horse serum (Invitrogen). After inactivation the tissue was collected with a centrifugation step, resuspended in 1 ml of plating medium and viable cells were counted using a cell counter before being plated at 50,000 cells/well in plating medium. Four hours after plating, the medium was changed to serum free maintenance medium based on neurobasal-B27. Three days after plating, half of the medium was renewed and the cultures were maintained with two full media changes per week afterwards. For all pharmacological studies, cultures were used after 14 days in vitro DIV.
Calcium imaging
Incubation with Fluo-4 was performed according to recommendations from the manufacturer (Invitrogen). Briefly, a vial of Fluo-4 was resuspended in HBSS-Hepes 20 mM containing 3 mg/ml BSA and 100µl probenecid and the cultures were incubated in 80µl of Fluo-4 per well for 1 hour at 37°C. Plates were then taken out of the incubator and allowed to reach room temperature before being imaged. For imaging, the Fluo-4-containing HBSS was replaced by 180µl of either normal magnesium HBSS, or calcium and magnesium-free HBSS supplemented with 0.1 mM MgCl 2 and 2 mM CaCl 2 . Cultures were imaged in this medium.
For calcium imaging, Flexstation 2 and 3 plate readers running SoftMax pro at set at 24°C were used. Fluo-4 fluorescent was imaged with an excitation wavelength of 485 nm, emission of 525 and a cut off at 515. Recording time was 600 seconds with intervals of 0.8 seconds, photomultipliers set at high and 6 reads per well. This allowed for recording of 4 wells (in one column) at a time. For compound injection, a compound transfer protocol was used with a 20µl volume of compound added to the wells at 240 seconds. Compounds were first dissolved in HBSS or DMSO and then diluted to 10-fold final concentration in HBSS for injection. Matching vehicles (HBSS or 0.1 to 0.3% final DMSO in HBSS) were used to monitor a potential injection effect in the calcium oscillations read out. The final compound dose was chosen based on the literature and for those compounds with no activity on the calcium oscillations, higher doses were also tested.
Data analysis
Fluorescence intensity measurements every 0.8 seconds were transferred to GraphPad Prism for presentation. Effectiveness of the test compounds was assessed in a qualitative way by determining efficacy, partial efficacy or lack of efficacy based on changes after compound addition compared with the baseline before injection. For the first figure showing the onset of calcium oscillations, the number of oscillations per 10 minutes recording was manually counted, then analyzed in GraphPad Prism using a one-sample t-test to determine when they become different from 0.
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Protein disulfide isomerase (PDI) is a member of the thioredoxin superfamily with an abb'xa'c structural organization that consists of two catalytic domains (a & a′) separated by two non-catalytic domains (b & b′) and an short x linker, along with an acidic C-terminal c extension [1] – [3] . It is mainly located in the endoplasmic reticulum (ER) where it exhibits linked but independent oxidoreductase and chaperone activities. These activities allow it to facilitate the proper folding of nascent secretory proteins as well as the disposal of terminally misfolded proteins through the quality control mechanism of ER-associated degradation (ERAD). The structure and function of PDI is regulated by its redox status: it is a dynamic, flexible molecule which assumes a compact conformation in the reduced state and a more open conformation in the oxidized state [4] – [6] . PDI thus acts as a redox-dependent chaperone in its interactions with certain substrate proteins [7] – [10] .
The chaperone function of PDI is defined by its ability to prevent the aggregation of misfolded proteins [11] – [14] . The importance of this activity is highlighted by the link between PDI dysfunction and neurodegeneration: a S-nitrosylated form of PDI that cannot prevent protein aggregation is found in the brains of individuals with Parkinson's or Alzheimer's disease [15] . PDI also prevents the aggregation of α-synuclein which occurs in Parkinson's disease [16] , [17] . PDI can even prevent protein aggregation when added to the substrate 40 minutes after aggregation has begun [18] . This strongly suggests the chaperone function of PDI involves something other than simply binding and masking the solvent-exposed hydrophobic amino acid residues of a misfolded protein. However, the molecular mechanism of PDI chaperone function remains unknown.
PDI also plays a key role in cholera intoxication. Cholera toxin (CT) is an AB 5 toxin that consists of a catalytic A1 subunit, an A2 linker, and a cell-binding B pentamer ( Fig. S1 ) [19] . It moves by vesicle carriers from the cell surface to the ER where the A1 subunit dissociates from the rest of the toxin [20] . The free A1 subunit then shifts to a disordered conformation which allows it to exploit the ERAD system for export to the cytosol [21] – [24] . The translocated pool of CTA1 interacts with host factors in the cytosol to regain an active conformation, and it avoids proteasomal degradation long enough to effectively modify its Gsα target [24] – [27] .
CTA1 is anchored to CTA2 by a single disulfide bond and numerous non-covalent interactions. Reduction of the A1/A2 disulfide bond can occur at the resident redox state of the ER [28] , yet reduction alone is not sufficient for holotoxin disassembly [29] : PDI must displace the reduced A1 subunit from the rest of the toxin [7] , [8] . This process is essential for intoxication, as PDI-deficient cells are completely resistant to CT [7] .
PDI was originally thought to actively unfold the holotoxin-associated CTA1 subunit and to thereby displace CTA1 from the rest of the toxin [8] . This model was based upon the results of a protease sensitivity assay that only provided an indirect measure of protein structure. An alternative explanation for the “unfoldase” activity of PDI was suggested by our later work which demonstrated the intrinsic instability of CTA1 will allow it to spontaneously unfold upon its separation from CTA2/CTB 5 at physiological temperature [24] . Thus, PDI could trigger toxin unfolding simply by removing CTA1 from the CT holotoxin. Our recent biophysical analysis provided experimental support for this alternative model and demonstrated that PDI does not unfold CTA1 [7] . We also found that PDI exhibits conformation-dependent interactions with CTA1: PDI recognizes the folded conformations of CTA1 present at low temperatures and in the CT holotoxin, but it does not bind to the disordered, 37°C conformation of free CTA1 [7] . Consistent with previous reports [8] , we also noted only the reduced form of PDI will interact with CT and CTA1. This interaction did not appear to involve the oxidoreductase activity of PDI, as PDI did not form mixed disulfides with CTA1 and could bind to cysteine-free CTA1 deletion constructs [7] , [8] .
In this work we employed a biophysical and biochemical approach to define the structural basis for PDI-mediated disassembly of the CT holotoxin. Using isotope-edited Fourier transform infrared (FTIR) spectroscopy and circular dichroism (CD), we have demonstrated that PDI unfolds upon contact with CTA1. The substrate-induced unfolding of PDI provides a molecular explanation for holotoxin disassembly: the expanded hydrodynamic radius of unfolded PDI would act as a lever to dislodge reduced CTA1 from its non-covalent association with the rest of the toxin. In support of this model, we found the displacement of reduced CTA1 from CTA2/CTB 5 does not occur when PDI is locked in a folded conformation or when PDI chaperone function is disrupted by ribostamycin treatment. Additional drug treatments with bacitracin indicated the oxidoreductase activity of PDI is not required for holotoxin disassembly. Consistent with our model, the substrate-induced unfolding of PDI is blocked by ribostamycin but not bacitracin. Two other ER-localized oxidoreductases (ERp57 and ERp72) did not unfold in the presence of CTA1 and did not displace reduced CTA1 from its holotoxin. Substrate-induced unfolding thus appears to be a unique property of PDI that is linked to its chaperone function and could explain its ability to disrupt protein aggregation.
Results
PDI unfolds upon contact with CTA1
Far-UV CD spectroscopy was used to assess conformational changes in reduced PDI and CTA1 upon their interaction at 10°C and neutral pH ( Fig. 1 ). The spectrum of PDI alone was dominated by two components around 208 and 221 nm that can be ascribed to ππ* and nπ* backbone α-helical electronic transitions, respectively. The spectrum of CTA1 displayed a major component around 221 nm and a shoulder between 210 and 216 nm, consistent with its α/β secondary structure identified by X-ray crystallography [30] , [31] . When PDI and CTA1 were combined in an equimolar ratio, the resulting spectrum was different from the sum of the two individual spectra of PDI and CTA1, suggesting that protein-protein interactions result in conformational changes in either CTA1, PDI, or both proteins. The spectral difference (i.e., [CTA1+PDI] - [CTA1] - [PDI]) revealed two well defined peaks at 207 and 223 nm, as well as a deep minimum just above 190 nm, indicating a significant loss in the α-helical structure and a gain in the unordered structure. Loss of the PDI “double minima” α-helical signature from the spectrum of PDI+CTA1 combined sample implies that significant conformational changes occur in PDI. However, the spectral overlap of CD signals generated by both proteins prevents unambiguous assignment of the structural changes to one or the other protein. To resolve individual conformational changes in PDI and CTA1 upon their interaction, we used isotope-edited FTIR spectroscopy as described below.
10.1371/journal.ppat.1003925.g001 Figure 1
Loss of PDI structure in the presence of CTA1.
Far-UV CD spectra of PDI (green), CTA1 (blue), and both proteins in the same sample at a 1∶1 molar ratio (red) are shown. The black dotted line was obtained by spectral subtraction of the individual CTA1 and PDI spectra from the spectrum of both proteins together. All measurements were taken at 10°C in pH 7.0 buffer containing 1 mM GSH.
Isotope-edited FTIR spectroscopy allows the conformation of a protein to be monitored in the presence of a second, 13 C-labeled protein [32] – [34] . 13 C labeling does not alter the conformation of a protein. However, the heavier nuclear mass of the stable 13 C isotope generates a spectral downshift which allows the FTIR spectrum of a 13 C-labeled protein to be resolved from the spectrum of an unlabeled protein. The structures of both unlabeled and labeled proteins can thus be determined with isotope-edited FTIR spectroscopy. This technique allowed us, for the first time, to specifically and directly monitor the conformation of CTA1 in the presence of PDI. Our studies demonstrated that PDI does not unfold CTA1 [7] .
Here, we used isotope-edited FTIR spectroscopy to examine the impact of toxin binding on the structure of PDI ( Fig. 2 ). Our data indicated that PDI unfolds after it contacts CTA1. In the absence of CTA1, the 10°C structure of PDI exhibited a folded conformation with 37% α-helical and 42% β-sheet content ( Fig. 2A , Table 1 ). These percentages were consistent with the secondary structural content predicted from the crystal structure of PDI [35] . In the presence of CTA1 at 10°C, reduced PDI lost substantial α-helical and β-sheet content ( Fig. 2B ). The percentage of irregular PDI structure rose from 12% in the absence of CTA1 to 42% in the presence of CTA1 ( Table 1 ). The toxin-induced loss of PDI structure did not occur in the absence of GSH ( Fig. 2C–D , Table 1 ), which confirmed the specificity of our data: only reduced PDI can bind to CTA1 [7] , [8] .
10.1371/journal.ppat.1003925.g002 Figure 2
Structure of PDI in the absence or presence of CTA1.
Curve fitting (left panels) and second derivatives (right panels) for the FTIR spectrum of PDI recorded in the absence ( A , C , E , G ) or presence ( B , D , F , H ) of 13 C-labeled CTA1 are shown. For curve fitting, the dotted line represents the sum of all deconvoluted components (solid lines) from the measured spectrum (dashed line). Unless otherwise noted, all experiments were performed with sodium borate buffer (pH 7.0) containing 1 mM GSH. ( A , B ) PDI structure at 10°C. ( C , D ) PDI structure at 10°C in the absence of reductant. ( E , F ) PDI structure at 37°C. ( G , H ) PDI structure at 37°C in pH 6.5 buffer.
10.1371/journal.ppat.1003925.t001 Table 1
Toxin-induced unfolding of PDI.
% of Protein Structure
Condition
α-Helix
β-Sheet
Irregular
Other
PDI
10°C
37±4
42±2
12±3
9±2
10°C+CTA1
11±1
29±3
42±4
18±2
10°C, no GSH
31±4
45±3
8±2
15±4
10°C, no GSH+CTA1
44±2
41±1
7±3
8±4
10°C EDC treatment
32±2
39±2
16±1
13±4
10°C EDC treatment+CTA1
33±4
49±1
12±3
5±5
10°C bacitracin treatment
41±2
34±3
23±1
2±4
10°C bacitracin treatment+CTA1
16±4
33±2
37±1
14±2
10°C ribostamycin treatment
36±3
41±1
18±4
6±1
10°C ribostamycin treatment+CTA1
44±1
36±2
9±2
10±3
37°C
30±3
42±1
24±4
4±1
37°C+CTA1
25±2
44±3
18±4
14±1
37°C, pH 6.5
30±1
43±3
18±2
9±2
37°C+CTA1, pH 6.5
12±4
26±3
40±4
21±1
10°C+CTA1, 37°C measurement
35±3
39±3
13±1
12±2
37°C+CT holotoxin, 1 minute
27±3
22±4
40±1
10±1
37°C+CT holotoxin, 25 minutes
30±4
38±1
17±2
16±±4
ERp57
10°C
49±2
39±4
7±3
5±1
10°C+CTA1
46±3
36±4
11±2
8±3
ERp72
10°C
43±2
20±2
31±2
6±1
10°C+CTA1
48±4
21±3
22±4
9±2
Deconvolution of the conformation-sensitive amide I bands from FTIR data presented in Fig. 2 , 3 , 4 , 6 , 8 , S2 , and S3 was used to calculate the percentages of PDI, ERp57, or ERp72 structure under the stated conditions. Unless otherwise indicated, all conditions included GSH in pH 7.0 buffer. The averages ± standard deviations from three or four separate curve fitting iterations are shown.
An additional set of FTIR experiments were performed at 37°C to further validate our findings. Reduced PDI does not bind to CTA1 at 37°C in pH 7.0 buffer [7] and did not exhibit an increase in irregular structure when incubated with CTA1 under this condition ( Fig. 2E–F , Table 1 ). However, reduced PDI can bind to CTA1 at 37°C in pH 6.5 buffer [7] . Acidified medium stabilizes the CTA1 polypeptide, allowing it to retain a substantial amount of its secondary structure at physiological temperature [21] . Thus, as expected, PDI shifted to a disordered conformation with 40% irregular structure when mixed with CTA1 at 37°C in pH 6.5 buffer ( Fig. 2G–H , Table 1 ). These collective observations demonstrated the unfolding of PDI resulted from its physical interaction with CTA1.
PDI unfolding is reversible
The substrate-induced unfolding of PDI was a reversible event. When CTA1 was added to reduced PDI at 10°C, the folded conformation of PDI shifted to a disordered state ( Fig. 2A–B , Table 1 ). However, PDI returned to a folded conformation upon warming the PDI/CTA1 complex to 37°C ( Fig. 3 , Table 1 ). CTA1 unfolds at 37°C, and this unfolding event displaces its PDI binding partner [7] . The displacement of reduced PDI from CTA1 at physiological temperature thus allowed PDI to regain a folded structure.
10.1371/journal.ppat.1003925.g003 Figure 3
Refolding of disordered PDI.
PDI was placed at 10°C in sodium borate buffer (pH 7.0) containing 13 C-labeled CTA1 and 1 mM GSH. The temperature was then raised to 37°C for 60 min. ( A ) The FTIR spectra of PDI+CTA1 were recorded at 10°C (solid line) and at 37°C (dotted line). ( B ) Curve fitting (left panel) and second derivatives (right panel) for the FTIR spectrum of PDI at 37°C are shown. For curve fitting, the dotted line represents the sum of all deconvoluted components (solid lines) from the measured spectrum (dashed line).
An interaction with the CT holotoxin also resulted in the unfolding and refolding of PDI ( Fig. 4 , Table 1 ). For this experiment, 13 C-labeled PDI was mixed with the CT holotoxin at 37°C and pH 7.0. The holotoxin-associated CTA1 subunit maintains a folded conformation at 37°C and neutral pH [36] , so reduced PDI can bind to holotoxin-associated CTA1 under physiological conditions. However, the spontaneous unfolding of CTA1 which occurs after holotoxin disassembly at 37°C results in the displacement of its PDI binding partner [7] . This process allowed us to monitor both the unfolding of PDI upon its interaction with the CT holotoxin and the refolding of PDI after its release from the dissociated CTA1 subunit. One minute after exposure to the CT holotoxin, reduced PDI exhibited a disordered conformation containing 40% irregular structure ( Fig. 4A , Table 1 ). This was similar to the percentage of irregular structure in reduced PDI upon its binding to free CTA1 at 10°C and neutral pH or at 37°C and pH 6.5. After holotoxin disassembly, PDI could no longer interact with CTA1 and consequently returned to a folded conformation within 25 minutes of exposure to the CT holotoxin ( Fig. 4B , Table 1 ). As seen from Table 1 , the conformational changes in PDI between 1 and 25 minutes of its interaction with the CT holotoxin at 37°C involve an increase in the β-sheet fraction by 16% and an increase in the α-helical structure by only 3%, implying that the refolding of PDI begins with a gain of α-helical structure which is followed by a gain of β-sheet structure. The PDI-mediated displacement of reduced CTA1 from the CT holotoxin and subsequent release of PDI from the dissociated, unfolded CTA1 polypeptide are both extremely rapid events ( [7] , see also Fig. 5 ). Our data suggest these events had already occurred within 1 minute of combining PDI with CT, and the process of PDI refolding was already underway at our first time point. Technical limitations prevented the measurement of PDI structure before 1 minute of incubation with CT. Nonetheless, the data from Figures 2 – 4 collectively demonstrated that PDI unfolding occurs upon binding to either free CTA1 or holotoxin-associated CTA1, and the results further indicated that PDI will return to a folded state after the release of its bound substrate.
10.1371/journal.ppat.1003925.g004 Figure 4
Unfolding and refolding of PDI upon interaction with the CT holotoxin.
13 C-labeled PDI was placed at 37°C in sodium borate buffer (pH 7.0) containing 1 mM GSH and the CT holotoxin. FTIR spectra were then recorded after 1 min ( A ) or 25 min ( B ). Curve fitting (left panels) and second derivatives (right panels) for the FTIR spectrum of PDI are shown. For curve fitting, the dotted line represents the sum of all deconvoluted components (solid lines) from the measured spectrum (dashed line).
10.1371/journal.ppat.1003925.g005 Figure 5
PDI unfolding but not oxidoreductase activity is required for disassembly of the CT holotoxin.
SPR was used to monitor the real-time PDI-mediated disassembly of CT. A baseline measurement corresponding to the mass of the sensor-bound CT holotoxin established the 0 MicroRIU signal. The time course was then initiated with perfusion of PDI ( A ), EDC-treated PDI ( B ), or bacitracin-treated PDI ( C ) over the CT-coated sensor. The perfusion buffer contained either 30 mM GSH (left panels) or 1 mM GSH (right panels); non-reducing SDS-PAGE with Coomassie staining found the CTA1/CTA2 disulfide bond was reduced at 30 mM GSH but not 1 mM GSH ( inset , right panel of A ). PDI was removed from the perfusion buffer at time intervals denoted by asterisks and was replaced with sequential additions of anti-PDI, anti-CTA1, and anti-CTB antibodies as indicated by the arrowheads. One of two representative experiments is shown for each condition.
PDI unfolding is required for holotoxin disassembly
The unfolding of PDI provides a molecular basis for the PDI-mediated disassembly of CT: unfolding will expand the hydrodynamic radius of PDI, thus acting as a wedge to dislodge reduced CTA1 from its non-covalent association with CTA2/CTB 5 . To test this model, we locked PDI in a folded conformation by treating it with 400 mM EDC. Previous work has shown this “zero-length” intramolecular cross-linker will activate carboxylic side chains for reaction with nearby primary amines on lysine residues [37] , [38] . A standard cross-linking molecule is absent from this process, so the reactive side chains can only act on intramolecular targets. SDS-PAGE and size exclusion chromatography were used to confirm the absence of PDI dimers or oligomers after EDC treatment ( Fig. S2A –B). Isotope-edited FTIR spectroscopy further demonstrated that EDC-treated PDI did not unfold in the presence of CTA1 ( Fig. S2C–D , Table 1 ): the secondary structure content of EDC-treated PDI in either the absence or presence of CTA1 was similar to the secondary structure content of the isolated, untreated PDI polypeptide.
To examine the functional role of PDI unfolding in toxin disassembly, untreated PDI and EDC-treated PDI were perfused over a surface plasmon resonance (SPR) sensor coated with the CT holotoxin ( Fig. 5 ). A functional interaction between PDI and the toxin will result in the displacement of CTA1 from the SPR sensor and a corresponding drop in the refractive index unit (RIU) below the baseline value corresponding to the mass of the initial sensor-bound holotoxin [7] . When PDI was perfused over the CT-coated sensor under reducing conditions, we detected a rapid rise in RIU which was indicative of PDI binding to the toxin. This was followed by a drop in the RIU signal to a point below the initial baseline value ( Fig. 5A ). Identical results were obtained with either 30 mM GSH (left panel) or 1 mM GSH (right panel) in the perfusion buffer; non-reducing SDS-PAGE with Coomassie staining demonstrated the CTA1/CTA2 disulfide bond is reduced at 30 mM GSH but not 1 mM GSH ( Fig. 5A , right panel inset). The loss of signal around 200 seconds occurred even though PDI was still present in the perfusion buffer, suggesting that both PDI and CTA1 had been lost from the sensor. Sequential perfusions of anti-PDI, anti-CTA1, and anti-CTB antibodies over the PDI-treated slide confirmed this interpretation: only the anti-CTB antibody gave a positive response ( Fig. 5A ). Perfusion of an anti-KDEL antibody over a PDI-treated sensor also gave a positive response (not shown), which indicated the KDEL-containing CTA2 subunit remained associated with CTB 5 after the release of CTA1. This observation was consistent with previous reports [7] , [39] , and it demonstrated that PDI specifically removes CTA1 from the sensor-bound CTA2/CTB 5 complex.
EDC-treated PDI bound tightly to the CT holotoxin under reducing conditions but did not displace CTA1 from CTA2/CTB 5 ( Fig. 5B ). Identical results were obtained with either 30 mM GSH (left panel) or 1 mM GSH (right panel) in the perfusion buffer. Positive signals for the anti-PDI, anti-CTA1, and anti-CTB antibodies demonstrated that EDC-treated PDI remained associated with the intact CT holotoxin after removal from the perfusion buffer. The locked conformation of PDI thus exhibited a high affinity interaction with CT, but it lacked the mechanism required to separate CTA1 from the rest of the toxin. Our FTIR data strongly suggested this missing mechanism involves the toxin-induced unfolding of PDI.
Intramolecular cross-linking could disrupt the enzymatic activity of PDI, although previous work has suggested an oxidoreductase function is not required for PDI to displace CTA1 from CTA2/CTB 5 . To confirm this observation, bacitracin-treated PDI was perfused over a CT-coated SPR sensor under reducing conditions. Bacitracin is a peptide antibiotic that inhibits the reductive activity of PDI [40] , but it did not inhibit the toxin-induced unfolding of PDI as assessed by isotope-edited FTIR spectroscopy ( Fig. S3A–B , Table 1 ). Bacitracin-treated PDI could displace reduced CTA1 from its non-covalent association with CTA2/CTB 5 ( Fig. 5C , left panel), but it could not separate CTA1 from the rest of the toxin when the CTA1/CTA2 disulfide bond was intact ( Fig. 5C , right panel). In contrast, untreated PDI could mediate the disassembly of a CT holotoxin with an intact disulfide bond [7] , [8] ( Fig. 5A , right panel). These results indicated that the oxidoreductase activity of PDI could cleave the CTA1/CTA2 disulfide bond and that this activity was inhibited by bacitracin. Thus, bacitracin-treated PDI did not require an enzymatic function to dislodge reduced CTA1 from its non-covalent association with the rest of the toxin. By extension, the inability of EDC-treated PDI to displace reduced CTA1 from its holotoxin ( Fig. 5B , left panel) could not be attributed to a loss of enzymatic function.
The toxin-induced unfolding of PDI is related to its chaperone function
PDI can, independently of its oxidoreductase function, act as a chaperone to prevent the aggregation of misfolded proteins [12] , [13] . To determine whether the toxin-induced unfolding of PDI was related to its role as a chaperone, we used S-nitrosylation and ribostamycin to disrupt the chaperone activity of PDI [15] , [41] . Nitrosylated PDI and ribostamycin-treated PDI were then used in our toxin disassembly assay. As shown in Fig. 6A , nitrosylated PDI could not bind to the CT holotoxin. The inability of nitrosylated PDI to prevent protein aggregation [15] , [42] thus appears to result from the loss of substrate binding.
10.1371/journal.ppat.1003925.g006 Figure 6
The chaperone activity of PDI is required for disassembly of the CT holotoxin.
( A , B ) A baseline SPR measurement corresponding to the mass of the sensor-bound CT holotoxin established the 0 MicroRIU signal. The time course was then initiated with perfusion of S-nitrosylated PDI ( A ) or ribostamycin-treated PDI ( B ) over the CT-coated sensor in buffer containing 30 mM GSH. In ( B ), PDI was removed from the perfusion buffer after 600 sec and replaced with sequential additions of anti-PDI, anti-CTA1, and anti-CTB antibodies as indicated by the arrowheads. One of two representative experiments is shown for each condition. ( C , D ) Curve fitting (left panels) and second derivatives (right panels) for the FTIR spectrum of ribostamycin-treated PDI recorded in the absence ( C ) or presence ( D ) of 13 C-labeled CTA1 are shown. For curve fitting, the dotted line represents the sum of all deconvoluted components (solid lines) from the measured spectrum (dashed line).
Ribostamycin-treated PDI could bind to the CT holotoxin but could not separate reduced CTA1 from CTA2/CTB 5 ( Fig. 6B ). Removal of ribostamycin-treated PDI from the perfusion buffer resulted in a rapid drop in RIU to the initial baseline value corresponding to the mass of the CT holotoxin. Anti-PDI, anti-CTA1, and anti-CTB antibody controls confirmed that ribostamycin-treated PDI had dissociated from the intact CT holotoxin. Furthermore, isotope-edited FTIR spectroscopy demonstrated that ribostamycin-treated PDI did not unfold in the presence of CTA1 ( Fig. 6C–D , Table 1 ). The loss of chaperone activity for ribostamycin-treated PDI thus corresponded to an inhibition of both PDI unfolding and holotoxin disassembly.
Toxin resistance results from the loss of PDI chaperone function
Cells treated with 50 µM ribostamycin were almost completely resistant to CT ( Fig. 7A ). However, no protective effect was observed in cells transfected with a plasmid encoding a CTA1 construct that is co-translationally targeted to the ER before dislocation back into the cytosol ( Fig. 7B ). This expression system mimics the translocation events occurring after holotoxin disassembly [43] and was used to ensure ribostamycin treatment did not affect the intoxication process downstream of PDI-mediated toxin disassembly. We also found that the CTA1/CTA2 disulfide bond could be reduced in ribostamycin-treated cells ( Fig. 7C ). This result indicated that ribostamycin does not inhibit toxin transport to the ER, as reduction of the CTA1/CTA2 disulfide bond takes place in the ER [28] , [44] . In this experiment, brefeldin A (BfA) was used a positive control to demonstrate that CTA1/CTA2 reduction does not occur when toxin transport to the ER is blocked [45] . Ribostamycin did not block CT transport to the ER or reduction of the CTA1/CTA2 disulfide bond, yet only a minimal quantity of CTA1 could be detected in the cytosol of ribostamycin-treated cells ( Fig. 7D ). This was consistent with the toxin-resistant phenotype of ribostamycin-treated cells and indicated ribostamycin prevents the in vivo displacement of reduced CTA1 from CTA2/CTB 5 . Collectively, our data demonstrated that ribostamycin does not disrupt (i) toxin trafficking to the ER; (ii) reduction of the CTA1/CTA2 disulfide bond in the ER; (iii) translocation of the free CTA1 subunit from the ER to the cytosol; or (iv) CTA1 activity in the cytosol. The block of intoxication in ribostamycin-treated cells was therefore most likely due to the inhibition of PDI unfolding which facilitates holotoxin disassembly ( Fig. 6 ). This result demonstrated the critical role of PDI unfolding in the CT intoxication process.
10.1371/journal.ppat.1003925.g007 Figure 7
The chaperone activity of PDI is required for CT intoxication.
( A ) Untreated and ribostamycin-treated CHO cells were challenged with the stated concentrations of CT for 2 hr before cAMP levels were quantified. The averages ± ranges of 2 independent experiments with triplicate samples are shown. ( B ) CHO cells were transfected with a plasmid encoding a CTA1 construct that is co-translationally inserted into the ER. Dislocation of this CTA1 construct back to the cytosol of untreated or ribostamycin-treated cells was detected by the rise in intracellular cAMP at 4 hr post-transfection. Cells transfected with an empty plasmid (Mock) were used to establish the resting levels of cAMP. Data are presented as the averages ± standard deviations of three replicate samples per condition. One of three representative experiments is shown. ( C ) CHO cells were pulse-labeled at 4°C with 1 µg/mL of CT. Untreated or drug-treated cells were then chased in toxin-free medium for 2 hr at 37°C. Membrane fractions from digitonin-permeabilized cells were resolved by non-reducing SDS-PAGE and probed by Western blot with an anti-CTA antibody. ( D ) Untreated or ribostamycin-treated CHO cells were pulse-labeled at 4°C with 1 µg/mL of CT and then chased in toxin-free medium for 2 hr at 37°C. Cytosolic fractions from digitonin-permeabilized cells were then perfused over an SPR sensor coated with an anti-CTA1 monoclonal antibody. Known quantities of CTA were perfused over the sensor as positive controls, and the cytosolic fraction from unintoxicated cells was perfused over the sensor as a negative control. At the end of each perfusion, bound ligand was stripped from the sensor slide.
Toxin-induced unfolding is a unique property of PDI
Reduced CTA1 could not be separated from its holotoxin by ERp57 ( Fig. 8A ) or ERp72 ( Fig. 8B ), two ER-localized oxidoreductases with an overall domain structure similar to PDI [1] , [2] . ERp57 and ERp72 bound to the CT holotoxin under reducing conditions, but they did not dislodge CTA1 from CTA2/CTB 5 . Indeed, as demonstrated with our antibody controls, ERp57 and ERp72 remained stably associated with the intact CT holotoxin. Additional SPR experiments documented direct binding of ERp57 and ERp72 to the CTA1 subunit ( Fig. S4 ). ERp57 usually binds to glycosylated substrates in a complex with calnexin or calreticulin, but a direct interaction between ERp57 and its substrate has been reported as well [46] . Isotope-edited FTIR spectroscopy further demonstrated that ERp57 and ERp72 do not unfold upon contact with CTA1 ( Fig. 8C–F , Table 1 ). The toxin-induced unfolding of PDI and the PDI-mediated displacement of CTA1 from CTA2/CTB 5 thus appear to be unique, linked properties of PDI that are not shared by other oxidoreductases. This is consistent with the inability of CT to affect PDI-deficient cells [7] : if other resident ER oxidoreductases could perform the same function as PDI, then PDI-deficient cell lines would not be resistant to CT.
10.1371/journal.ppat.1003925.g008 Figure 8
ERp57 and ERp72 do not disassemble the CT holotoxin and do not unfold upon contact with CTA1.
( A , B ) A baseline SPR measurement corresponding to the mass of the sensor-bound CT holotoxin established the 0 MicroRIU signal. The time course was then initiated with perfusion of ERp57 ( A ) or ERp72 ( B ) over the CT-coated sensor in buffer containing 30 mM GSH. After 300 sec (ERp57) or 350 sec (ERp72), ligand was removed from the perfusion buffer and replaced with sequential additions of anti-ERp57/anti-ERp72, anti-CTA1, and anti-CTB antibodies as indicated by the arrowheads. One of two representative experiments is shown for each condition. ( C–F ): Curve fitting (left panels) and second derivatives (right panels) for the FTIR spectrum of ERp57 ( C , D ) or ERp72 ( E , F ) recorded in the absence ( C , E ) or presence ( D , F ) of 13 C-labeled CTA1 are shown. For curve fitting, the dotted line represents the sum of all deconvoluted components (solid lines) from the measured spectrum (dashed line).
Discussion
CT moves from the cell surface to the ER as an intact AB holotoxin. The CTA1/CTA2 disulfide bond is reduced in the ER, but this is insufficient for holotoxin disassembly: PDI must displace reduced CTA1 from the rest of the toxin. Our biophysical analysis has provided a structural explanation for this event. We have shown by isotope-edited FTIR spectroscopy and CD that PDI unfolds upon contact with CTA1. A real-time holotoxin disassembly assay demonstrated that the displacement of reduced CTA1 from CTA2/CTB 5 does not occur when PDI is locked in a folded conformation or when the substrate-induced unfolding of PDI is blocked due to the loss of its chaperone function. However, the oxidoreductase activity of PDI was not required for this event. The toxin-induced unfolding of PDI provides a molecular basis for holotoxin disassembly: the expanded hydrodynamic radius of unfolded PDI would act as a wedge to physically displace reduced CTA1 from the rest of the toxin. ERp57 and ERp72 did not unfold in the presence of CTA1 and did not displace reduced CTA1 from its holotoxin. Substrate-induced unfolding thus appears to be a unique property of PDI.
PDI does not directly interact with CTA2 or the CTB pentamer; it only recognizes the folded conformation of the CTA1 subunit [7] . In order for PDI to dislodge CTA1 from the CT holotoxin, it must bind to a region of CTA1 near the CTA2/CTB 5 interface. The expanded hydrodynamic radius of PDI resulting from its toxin-induced unfolding would then push against two components of the holotoxin and thereby dislodge the A1 subunit from its non-covalent association with the rest of the toxin. PDI was originally thought to interact with the C-terminal hydrophobic A1 3 subdomain of CTA1 [47] , which is distal to the CTA2/CTB 5 interface ( Fig. S1 ). However, binding assays with CTA1 deletion constructs have demonstrated the A1 3 subdomain is not required for PDI-CTA1 interaction [7] . Binding instead occurred in a region of CTA1 (residues 1–133) that is, in part, proximal to CTA2/CTB 5 . The exact location of the PDI binding site on CTA1 remains to be determined, and this information is important for further elucidation of the CT disassembly mechanism. However, the current data are consistent with our model for the physical, PDI-mediated displacement of reduced CTA1 from its holotoxin.
PDI binds to the folded conformations of CTA1 that are present at low temperature and in the CT holotoxin [7] . This induces the partial unfolding of PDI ( Figs. 1 – 2 , Fig. 4 , Table 1 ), but disordered PDI still remains associated with CTA1 [7] . The modular structure of PDI, which consists of a rigid b′ substrate binding domain flanked by other more flexible domains [2] , [4] , [48] , likely accounts for the ability of partially disordered PDI to remain associated with its folded CTA1 partner. As shown in our recent publication, PDI is only displaced from CTA1 when the toxin unfolds [7] . CT disassembly thus appears to involve the following events: (i) the CTA1/CTA2 disulfide bond is reduced at the resident redox state of the ER [28] , but CTA1 remains associated with CTA2/CTB 5 through non-covalent interactions [29] ; (ii) reduced PDI binds to holotoxin-associated CTA1 [7] , [8] ; (iii) the substrate-induced unfolding of PDI results in the separation of CTA1 from CTA2/CTB 5 (this work); and (iv) the dissociated CTA1 subunit spontaneously unfolds at 37°C [24] , which consequently displaces its PDI binding partner [7] . PDI regains its native conformation after release from CTA1 ( Figs. 3 – 4 ). PDI and other oxidoreductases may assist reduction of the CT disulfide bond ( Fig. 5A , right panel) [28] , [44] as observed for other ER-translocating AB toxins [49] – [51] . However, the essential and specific role of PDI in holotoxin disassembly appears to be the physical separation of reduced CTA1 from its holotoxin.
The toxin-induced unfolding of PDI suggests a molecular mechanism for its role as a chaperone that prevents protein aggregation: by unfolding in the presence of an aggregation-prone substrate, PDI would act as a lever to dislodge individual proteins from the forming aggregate. With this model, treatments that block the chaperone activity of PDI should prevent the substrate-induced unfolding of PDI. S-nitrosylation and ribostamycin represent two such conditions, as it has already been shown that S-nitrosylated PDI and ribostamycin-treated PDI can no longer prevent protein aggregation [15] , [41] . S-nitrosylation and ribostamycin treatment also blocked the PDI-mediated disassembly of CT. In the case of S-nitrosylation, the inhibition of toxin disassembly resulted from the loss of substrate binding. Ribostamycin-treated PDI could bind to CT, but it did not undergo substrate-induced unfolding. Many chaperones prevent protein aggregation by a simple physical mechanism that involves binding and masking the exposed hydrophobic amino acid residues of a disordered protein. Given that ribostamycin-treated PDI could still bind to CTA1, it is unlikely that ribostamycin disrupts the chaperone function of PDI through an inhibition of substrate binding. This strongly suggests the chaperone function of PDI involves an activity in addition to substrate binding; we propose this activity is linked to the substrate-induced unfolding of PDI. Consistent with this model, we also demonstrated that EDC-treated PDI binds to CTA1 but does not unfold in the presence of CTA1 and does not displace reduced CTA1 from CTA2/CTB 5 .
Bacitracin inhibited the enzymatic activity of PDI, but it did not affect the toxin-induced unfolding of PDI and did not prevent the displacement of reduced CTA1 from CTA2/CTB 5 . This again suggested that the substrate-induced unfolding of PDI is linked to its chaperone function, as unfolding was blocked by ribostamycin (an inhibitor of PDI chaperone function) but not bacitracin (an inhibitor of PDI oxidoreductase activity). Since holotoxin disassembly does not require the oxidoreductase function of PDI, the inhibitory effects of ribostamycin and EDC cannot be attributed to the potential disruption of PDI enzyme activity.
Our work has established a new functional property of PDI that is linked to its role as a chaperone. This property of substrate-induced unfolding would not have evolved for the benefit of a bacterial pathogen. There must therefore be a normal, physiological role for PDI unfolding. We propose the action of PDI in CT disassembly is related to its established role as a chaperone that prevents protein aggregation: in both cases, the expanded hydrodynamic radius of unfolded PDI would act as a wedge to disrupt non-covalent macromolecular complexes. This event could also be related to the structural changes that occur when PDI expands from a compact, reduced conformation to its more open, oxidized state. Thus, in addition to elucidating the molecular details of PDI-mediated toxin disassembly, our data provide a possible mechanistic basis for the known but structurally uncharacterized chaperone function of PDI.
Materials and Methods
Materials
Ribostamycin, PDI, BfA, GM1, GSH, CTA, S-nitrosoglutathione, and anti-CTB antibodies were purchased from Sigma-Aldrich (St. Louis, MO). Bacitracin was purchased from Calbiochem (La Jolla, CA), CT was from List Biological Laboratories (Campbell, CA), and phosphate-buffered saline (pH 7.4) with 0.05% Tween 20 (PBST) was from Medicago (Uppsala, Sweden). ERp57, the anti-ERp57 antibody, and the anti-ERp72 antibody were from Abcam (Cambridge, MA). ERp72 and the anti-PDI antibody were purchased from Enzo Life Sciences (Farmingdale, NY). The anti-CTA1 monoclonal antibody 35C2 [52] was a generous gift from Dr. Randall K. Holmes (University of Colorado School of Medicine). The pOLR130 plasmid encoding mature human PDI with an N-terminal His 6 tag [53] was generously provided by Dr. Lloyd Ruddock (University of Oulu, Finland). Uniformly 13 C-labeled 13 C 6 - D -glucose and D 2 O were purchased from Cambridge Isotope Laboratories (Andover, MA). Uniformly 13 C-labeled CTA1-His 6 was produced as described in [7] and purified as described in [22] .
13 C labeling and purification of His 6 -PDI
Escherichia coli strain BL21 pLysS transformed with pOLR130 was inoculated into 5 mL M9 minimal media containing 100 µg/mL of ampicillin and was grown at 37°C with shaking. The culture was then expanded in 200 mL M9 minimal media supplemented with 100 µg/mL of ampicillin and uniformly 13 C-labeled 13 C 6 -D-glucose as the sole metabolic carbon source. Incubation at 37°C with shaking continued until the culture reached an O.D. 600 of 0.2–0.3. The culture was then induced with 1 mM IPTG for 4 hr at 37°C, followed by centrifugation of the cells at 3,500× g and 4°C for 20 min. The supernatant was discarded, and the cells were resuspended in extraction buffer containing 20 mM Tris-HCl (pH 7.0), 300 mM NaCl, 0.1% sodium deoxycholate, 100 µg/mL of lysozyme, and 0.1 µL/mL of DNAse. Following three freeze/thaw cycles oscillating between −80°C and 37°C, the lysed cells were centrifuged at 13,800× g and 4°C for 30 min. The supernatant was collected, syringe filtered to remove any remaining cellular debris, and supplemented with 10 µL/mL of His-PIC (Sigma-Aldrich). For every 5 mL of crude lysate, 1 mL Talon resin beads were prepared as described by the manufacturer (Clontech, Mountain View, CA). After equilibration of the resin with extraction buffer, lysate was added to the resin and agitated at room temperature for 45 min. Unbound proteins in the supernatant were removed after a 5 min room temperature spin at 700× g, and the pelleted resin was resuspended in wash buffer containing 20 mM Tris-HCl (pH 7.0), 600 mM NaCl, and 0.1% Triton X-100. Following 15 min of agitation at room temperature, the resin was centrifuged at 700× g and room temperature for 5 min. The supernatant was removed, and the resin was washed two more times with wash buffer. The washed resin was then resuspended in 20 mM Tris-HCl (pH 7.0) with 600 mM NaCl, transferred to a gravity flow column, and allowed to settle. After washing the resin bed with 5 mL 20 mM Tris-HCl (pH 7.0) containing 600 mM NaCl, His 6 -PDI was eluted from the column using 2 mL quantities of increasing imidazole concentrations (10, 15, 20, 25, 35, 40, and 50 mM). Collected fractions of 0.5 mL volume were stored at −20°C until needed. Eluted fractions as well as samples from each step of the purification process were visualized by SDS-PAGE and Coomassie staining to verify the purity of His 6 -PDI. Samples of the 13 C-labeled protein were dialyzed in sodium borate buffer with decreasing concentrations of NaCl, lyophilized, and stored at −80°C before reconstitution in a D 2 0-based sodium borate buffer for use in FTIR spectroscopy.
Isotope-edited FTIR spectroscopy
As previously described in [7] , FTIR spectra for PDI, ERp57, or ERp72 were collected using a Jasco 4200 FTIR spectrometer at 0.964 cm −1 spectral resolution and a set resolution of 1 cm −1 . Samples were prepared in a D 2 O-based 10 mM sodium borate pH* 6.6 buffer that corresponds to pH 7.0. Where indicated, a D 2 O-based 10 mM sodium borate pH* 6.1 buffer that corresponds to pH 6.5 was used for measurements at acidic pH. The buffers also contained 100 mM NaCl and, unless otherwise noted, 1 mM GSH. The 70 µL samples contained either unlabeled oxidoreductase (40 µg) or a 1∶1 molar ratio of unlabeled oxidoreductase (90 µg) and uniformly 13 C-labeled CTA1. Studies involving PDI and the CT holotoxin used 20 µg of 13 C-labeled His 6 -PDI and 25 µg of CT; readings for this experiment were taken from 1–25 minutes after mixing the two proteins. Absorbance spectra were determined using the matched buffer as a reference and were corrected by subtraction of water vapor contribution, smoothing, and baseline correction in the amide I region. Wavenumbers were assigned to specific protein secondary structures for unlabeled PDI as detailed in [54] : 1655±5 cm −1 , α-helix; 1630±10 cm −1 , β-sheet; 1644±4 cm −1 , irregular. For 13 C-labeled PDI, wavenumber assignments were shifted 40–50 cm −1 . Deconvolution of protein secondary structures was performed as previously described [7] . The percentages of oxidoreductase secondary structure calculated by FTIR spectroscopy were consistent with the secondary structure content predicted from crystal structures for PDI (PDB 2B5E and PDB 4EKZ), ERp57 (PDB 3F8U), and ERp72 (reconstructed from PDBs 2DJ3, 2DJ2, 2DJ1, and 3EC3).
Far-UV CD
Experiments were conducted with 15 µM CTA1 and equimolar PDI in 10 mM borate buffer (pH 7.0) containing 100 mM NaCl and 1 mM GSH. Measurements were recorded at 10°C using a 0.1 mm optical path-length quartz cuvette and a J-810 spectrofluoropolarimeter equipped with a PFD-425S Peltier temperature controller (Jasco Corp., Tokyo, Japan). Samples were equilibrated at 10°C for 4 min before measurement, and each spectra was averaged from 5 scans.
SPR
To monitor disassembly of the CT holotoxin, a gold plated Reichert (Depew, NY) SPR sensor slide was coated with ganglioside GM1 and subsequently appended with CT as previously described [22] . PBST was perfused over the sensor for 10 min at 37°C with a 41 µL/min flow rate to generate a baseline RIU signal corresponding to the mass of the bound holotoxin. A PBST solution containing GSH (1 or 30 mM) and PDI, ERp57, or ERp72 (all at 100 nM final concentration) was then perfused over the sensor at 37°C and a flow rate of 41 µL/min. After removal of the oxidoreductase from the perfusion buffer, sequential additions of antibodies were perfused over the sensor at the following dilutions: anti-PDI antibody, 1∶10,000; anti-ERp57 antibody, 1∶500; anti-ERp72 antibody, 1∶500; anti-CTA1 monoclonal antibody, 1∶500; anti-CTB antibody, 1∶15,000. To detect the physical association of CTA1 with ERp57 or ERp72, each oxidoreductase was perfused at 10°C over a SPR sensor appended with CTA1-His 6 as previously described for ARF6-CTA1 interactions [24] . All experiments were performed with a Reichert SR7000 SPR refractometer.
Drug treatments
S-nitrosylated PDI was generated by treatment with S-Nitrosoglutathione for 20 min at room temperature. S-Nitrosoglutathione was prepared by incubating 5 mM reduced GSH with 5 mM sodium nitrate for 1 hr at room temperature. Other in vitro drug treatments involved exposing PDI for 30 min at room temperature to 0.2 mM EDC, 50 µM ribostamycin, or 0.1 mM bacitracin. A PDI stock concentration of 1.6 mg/mL was used for all treatments.
CT intoxication assay
CHO cells grown to 80% confluency in a 24-well plate were incubated in serum-free medium with 1, 10, or 100 ng/mL of CT for 2 hr at 37°C. Toxin-challenged cells were either left untreated or were co-incubated with 50 µM ribostamycin. The cAMP content of intoxicated and unintoxicated control cells was quantified with a commercial kit (Perkin-Elmer, Boston, MA) following the manufacturer's instructions. Values obtained from unintoxicated cells were background subtracted from the results with intoxicated cells, and the experimental data were then expressed as percentages of the maximal cAMP response obtained from cells exposed to 100 ng/mL of CT in the absence of ribostamycin. Triplicate samples were used for each condition.
CTA1 transfection/intoxication assay
CHO cells grown to 75% confluency in 6-well plates were exposed for 3 hr to a mixture of pcDNA3.1/ssCTA1 [55] and Lipofectamine according to manufacturer's instructions (Invitrogen, Carlsbad, CA). At 4 hr post-transfection, cAMP levels in untreated and ribostamycin-treated cells were quantified with a commercial kit (Perkin-Elmer). Resting levels of cAMP from cells mock transfected with the empty pcDNA3.1 vector were also recorded.
CT transport assay
CHO cells grown to 80% confluency in 6-well plates were pulse-labeled for 30 min at 4°C in serum-free medium containing 1 µg/mL of CT. Unbound toxin was removed by washing with PBS, after which the cells were returned to toxin- and serum-free medium for a 2 hr incubation at 37°C. Chase conditions included untreated cells, cells co-incubated with 50 µM ribostamycin, and cells co-incubated with 5 µg/mL of BfA. Membrane fractions from digitonin-permeabilized cells were collected at the end of the pulse and the end of the chase conditions. Samples were resolved by non-reducing SDS-PAGE with 15% polyacrylamide gels and probed by Western blot with an anti-CTA antibody as described in [23] .
CTA1 translocation assay
As described above for the CT transport assay, untreated and ribostamycin-treated CHO cells were pulsed-labeled with CT at 4°C and chased at 37°C for 2 hr. Cytosolic fractions from digitonin-permeabilized cells were then perfused over an SPR sensor coated with the anti-CTA1 35C2 monoclonal antibody [52] . A detailed protocol for this SPR-based translocation assay has been provided in [56] .
UniProtKB/Swiss-Prot accession numbers
CT, P01555 and P01556; PDI, P17967and P07237; ERp57, P30101; ERp72, P08003.
Supporting Information
Figure S1
CT structure. The catalytic 21 kDa CTA1 subunit (blue) is anchored to a CTA2 linker (red) by numerous non-covalent interactions and a single disulfide bond connecting the C-terminal A1 3 subdomain of CTA1 (light blue) to the N-terminus of CTA2. The 5 kDa CTA2 subunit extends into the central pore of the ring-like CTB homopentamer (grey) and thus maintains extensive non-covalent contacts with CTB. A KDEL tetrapeptide is found at the C-terminus of CTA2. Separation of CTA1 from CTA2/CTB 5 is required for the ER-to-cytosol translocation of CTA1 and optimal activation of its latent enzymatic activity. The ribbon diagram was derived from PDB 1S5F.
(TIF)
Figure S2
Impact of EDC on the structure of PDI. ( A ) PDI was treated with the stated concentrations of EDC for 30 min at room temperature before resolution on a non-reducing SDS-PAGE gel. One of two representative experiments is shown. ( B ) Untreated PDI (solid line) and PDI treated with 400 mM EDC (dashed line) were subjected to gel filtration with a Superdex G-75 column on an AKTA purifier. Each sample was eluted at 4°C in a buffer of 150 mM KCl and 25 mM Tris (pH 7.4) at a rate of 1 mL/min. Sample elution was detected by absorbance at 280 nm. Molecular mass standards of 150 kDa, 66 kDa, and 29 kDa eluted at 27 mL, 37 mL, and 52 mL, respectively. ( C , D ) PDI treated with 400 mM EDC for 30 min at room temperature was placed at 10°C in sodium borate buffer (pH 7.0) containing GSH. Curve fitting (left panels) and second derivatives (right panels) for the FTIR spectrum of EDC-treated PDI recorded in the absence ( C ) or presence ( D ) of 13 C-labeled CTA1 are shown.
(TIF)
Figure S3
Impact of bacitracin on the structure of PDI. ( A , B ) Curve fitting (left panels) and second derivatives (right panels) for the FTIR spectrum of bactitracin-treated PDI recorded in the absence ( A ) or presence ( B ) of 13 C-labeled CTA1 are shown. For all curve fitting, the dotted line represents the sum of all deconvoluted components (solid lines) from the measured spectrum (dashed line).
(TIF)
Figure S4
ERp57 and ERp72 bind to the CTA1 subunit at 10°C. ERp57 ( A ) or ERp72 ( B ) was perfused over a CTA1-coated SPR sensor slide in buffer containing 1 mM GSH. Arrowheads denote when the oxidoreductase was removed from the perfusion buffer. One of two representative experiments is shown for each condition.
(TIF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Hyperkyphosis, the excessive anterior curvature of the thoracic spine, affects an estimated 20–40 percent of older people [ 1 – 3 ]. Despite being relatively common in the population with no standard diagnoses or treatments available, hyperkyphosis is associated with a number of serious health complications, including physical disability [ 4 ], increased risk of fractures [ 5 ], and even premature mortality [ 6 ]. As hyperkyphosis progresses with advancing age, there is a need for better understanding of the extent to which it may affect daily functioning needed for healthy aging.
Hyperkyphosis is also associated with restricted pulmonary function [ 7 ] and posture [ 8 ], which may impact sleep quality possibly by increasing sleep fragmentation. Additionally, available treatments for hyperkyphosis, including physical therapy, therapeutic exercises, use of bracing, and surgery [ 9 ], may all impact sleep as well. Surprisingly, there has only been one study examining the association between hyperkyphosis and sleep quality [ 10 ], and this study was cross-sectional. Wankie and colleagues showed that older women with hyperkyphosis had worse self-reported sleep quality whereas no association was found in men [ 10 ]. While there was no association in men, the measure of sleep was subjective, and it is possible that the impact of hyperkyphosis in men might be better captured by objectively measured sleep dysfunction. Additionally, given the importance of sleep for healthy aging, it is important to explore whether relationships of hyperkyphosis with sleep be examined longitudinally to determine whether individuals with hyperkyphosis are susceptible to developing sleep problems over time.
To address these gaps, we analyzed the longitudinal association between hyperkyphosis and future sleep quality among older men participating in the Osteoporotic Fractures in Men Study (MrOS), using both subjective and objective sleep quality measures. The specific aims of our analysis were to determine whether there were differences in a) subjective and b) objective measurements of sleep quality between those with hyperkyphosis and those without. We hypothesized that hyperkyphosis will be associated with impairments in both subjective and objective sleep quality in older men.
Methods
Data source
Data for this study came from the Osteoporotic Fractures in Men Study (MrOS), a multi-center longitudinal observational cohort study of older men. From 2000–2002, 5,994 community-dwelling men age 65+ years were recruited from six sites in the US: Birmingham, Alabama; Minneapolis, Minnesota; Palo Alto, California; Monongahela Valley, Pennsylvania; Portland, Oregon; and San Diego, California. Eligibility requirements for study participation included walking without assistance and no history of bilateral hip replacement. From 2003–2005, a total of 3,135 MrOS participants were recruited to participate in the MrOS Sleep Study, an ancillary study which sought to characterize subjective and objective sleep quality among participants. For MrOS Sleep, participants were screened for nightly (or near nightly) use of mechanical devices (e.g., CPAP or BiPAP for sleep apnea, supplemental oxygen therapy), and were excluded from participation if they were unable or unwilling to forego use of such devices during the night of the sleep study. A second assessment in the sleep study was completed between 2009–2012, in which a subset of 1,055 original MrOS Sleep Study participants (recruitment goal was 1,000 subjects) repeated wrist actigraphy. To date, MrOS parent study has had four visits spaced 3–4 years apart. Our study used hyperkyphosis data from the third MrOS parent study visit (fielded from March 2007-March 2009) and MrOS Sleep visit 2 (fielded from November 2009-March 2012). Mean follow-up time between these two visits is approximately 2.9 years (SD = 0.77). The design and conduct of MrOS and MrOS Sleep have been previously reported [ 11 , 12 ]. All sites obtained their own institutional review board approval prior to completion of study (Sutter Health Institutional Review Board, University of California, San Francisco Human Research Protection Program, University of Alabama at Birmingham Institutional Review Board for Human Use, Human Research Protection Program at the University of Minnesota, Stanford University Institutional Review Board, University of Pittsburgh Institutional Review Board, Oregon Health & Science University Institutional Review Board, and the University of California, San Diego Human Research Protections Program). Additionally, data are publicly available at the following website: https://mrosdata.sfcc-cpmc.net/ .
Participants
Participants in the MrOS Sleep Study who completed blocks measurement of kyphosis (described below) during MrOS visit 3, and had wrist actigraphy recorded at the MrOS Sleep visit 2, were included in analyses. Of the 1,044 subjects who provided actigraphy data for MrOS Sleep visit 2, 754 were among the 4,681 subjects who had also completed MrOS visit 3 and provided data for this analysis (See Fig 1 ). All MrOS study participants provided informed consent at each of the clinical centers to participate in the study. Institutional review board approval was obtained by institutions at each site.
10.1371/journal.pone.0228638.g001
Fig 1
Flow diagram of analysis sample generation.
Measures
Hyperkyphosis
Measurement of hyperkyphosis was completed using the Rancho Bernardo block method [ 1 ]. Wooden blocks of 1.7 cm thickness were placed under the participant’s head to achieve a neutral spine position while lying supine on a DXA table. The number of blocks needed to yield a neutral spine (e.g., head is not hyperextended or hyperflexed) was recorded. Interrater reliability of the block method for measuring hyperkyphosis has ranged from 0.85 to 1.00 Spearman correlation across clinical sites. This study defined presence of hyperkyphosis as >3 blocks.
Subjective sleep quality
Self-reported sleep quality measures included the Pittsburgh Sleep Quality Index (PSQI) [ 13 ] and the Epworth Sleepiness Scale (ESS) [ 14 ]. The PSQI measures self-reported sleep quality and sleep disturbances over the past month, and consists of a global score (with scores ranging from 0 to 21; scores >5 are considered as poor sleep quality), as well as seven component scores including subjective quality, latency, duration, efficiency, sleep disturbance, sleep medication usage, and daytime dysfunction (scores range from 0 to 3 for each component). The ESS is a questionnaire that measures subjective daytime sleepiness with scores ranging from 0 to 24 (scores >10 are considered as excessive daytime sleepiness). For the PSQI and ESS, higher scores indicate worse symptomatology.
Objective sleep quality
Objective sleep quality was measured through the use of wrist actigraphy, a validated method [ 15 ] to estimate sleep/wake patterns based upon movement detected from a wrist worn device called an actigraph. Specifically, the actigraph model used was the Actiwatch 2 (Respironics, Inc., Bend, Oregon). Subjects were asked to wear the actigraph around the non-dominant wrist for five consecutive days, only taking the device off for bathing and while swimming. Participants on average wore the watch for a mean of 5.0 days (SD = 0.73, Range = 1–14 days). Movement data are recorded continuously throughout the day and summed in 1-minute intervals (epochs) with a 20 count threshold for determination of sleep/wake categorization. Concurrently with wrist actigraphy, participants were asked to self-report the times they went into and out of bed, when they turned lights out, as well as any periods in which the actigraph was not worn. Trained scorers identify intervals during which the device was removed, and mark the night-time sleep period with assistance from this diary. Action W-2 software was used to classify each epoch as sleep or wake, and objective sleep variables were derived [ 16 ]. Variables estimated from wrist actigraphy during the night-time sleep period include total sleep time (TST; duration in minutes spent asleep), wake after sleep onset (WASO; time in minutes awake while in bed), sleep efficiency (SE; percentage of time in bed spent asleep), and sleep latency (SL; time to fall asleep after lights out). All variables were averaged over all nights of data collection.
Sleep apnea
In addition to wrist actigraphy, subjects underwent a single night, unattended in-home polysomnography session. Methods have been previously described [ 17 ]. AHI was defined as number of apneas or hypopneas per hour of sleep with at least 3% oxygen desaturation. An apnea was defined as a complete cessation of breathing, where as a hypopnea was a partial (at least 50% reduction in airflow), each lasting at least 10 seconds.
Covariates
We also examined a number of covariates including baseline age, height, weight, marital status, self-reported health status, physical activity as measured by the Physical Activity Scale for the Elderly (PASE) [ 18 ], smoking status, history of alcohol use, and selected medical conditions (including sleep apnea, depression, arthritis, cardiovascular disease, diabetes). Additionally, via reports from clinics, MrOS collected data on prescription medications used of which were coded using the Iowa Drug Information Service (IDIS) Drug Vocabulary (College of Pharmacy, University of Iowa, Iowa City, IA) [ 19 ]. We assessed sleep medications, antidepressants, and benzodiazepines.
Analyses
Baseline characteristics of participants with vs. without hyperkyphosis were compared using chi-squared tests for binary/categorical variables and F-tests for continuous variables. In order to compare those with and without hyperkyphosis, we used linear regression models to analyze the associations between baseline hyperkyphosis and future subjective (PSQI total and component scores and ESS) and objective (TST, WASO, SE, and SL from wrist actigraphy; RDI at 3% desaturation from overnight in-home polysomnography) sleep outcomes. We further used logistic regression to determine the association between hyperkyphosis and dichotomous sleep outcomes, based on established cut points to define those with poor sleep. We conducted both bivariate and multivariate analyses, adjusting for significant baseline characteristics associated with hyperkyphosis at p <0.10 and study site. All analyses were completed in Stata SE version 15 (Stata Corp, College Station, TX).
Results
Subject characteristics by hyperkyphosis group are displayed in Table 1 . Hyperkyphosis subjects were on average 79.5 years old (SD = 5.30), had a height of 174.9 cm (SD = 7.84) and weight of 85.3 kg (SD = 13.93). Ninety percent were Caucasian, 71.7% were married, and 88.3% were in excellent or good health status. The mean PASE score was 128.0 (SD = 64.08). Fewer than 1% were current smokers, and subjects reported an average of 1.8 drinks per week (SD = 1.60). In terms of health conditions, 16.2% reported sleep apnea, 11.7% depression, 6.2% arthritis, 73.1% cardiovascular disease, and 17.2% diabetes. For medication use, just under a fifth reported use of sleep medications and similar percentages were reported for antidepressant use. Four-percent reported use of benzodiazepines. Compared to the non-hyperkyphosis group, the hyperkyphosis group was older, taller and heavier, less likely to be married, less physically active and more likely to have cardiovascular disease ( Table 1 ).
10.1371/journal.pone.0228638.t001
Table 1 Baseline characteristics.
Hyperkyphosis
Comparison
Normal n = 609 n (%) 1
Hyperkyphotic n = 145 n (%) 1
p -value 2
Age (years, mean ± SD)
77.6 (4.53)
79.5 (5.30)
<0.001
Height (cm, mean ± SD)
173.5 (6.68)
174.9 (7.84)
0.027
Weight (kg, mean ± SD)
81.0 (12.73)
85.3 (13.93)
<0.001
Race/Ethnicity
0.105
White
508 (83.4)
130 (89.7)
African American
32 (5.3)
2 (1.4)
Asian
41 (6.7)
5 (3.5)
Hispanic
16 (2.6)
6 (4.1)
Other
12 (2.0)
2 (1.4)
Marital Status
0.029
Married
496 (81.4)
104 (71.7)
Widowed
63 (10.3)
29 (20.0)
Separated
2 (0.3)
1 (0.7)
Divorced
29 (4.8)
7 (4.8)
Single, never married
19 (3.1)
4 (2.8)
Self-Reported Health Status
0.814
Fair/poor or very poor
67 (11.0)
17 (11.7)
Excellent or good
540 (89.0)
128 (88.3)
Physical Activity (PASE score, mean ± SD)
145.8 (67.36)
128.0 (64.08)
0.004
Smoking Status
0.732
No
263 (43.3)
65 (44.8)
Past
335 (55.2)
79 (54.5)
Current
9 (1.5)
1 (0.7)
Alcohol Use (drinks/wk, mean ± SD)
1.9 (1.71)
1.8 (1.60)
0.480
Selected Medical Conditions 3
Sleep apnea
78 (13.2)
23 (16.2)
0.352
Depression
44 (7.2)
17 (11.7)
0.074
Arthritis
44 (7.2)
9 (6.2)
0.666
Cardiovascular disease
375 (61.6)
106 (73.1)
0.009
Diabetes
89 (14.6)
25 (17.2)
0.427
Selected medication use 3
Sleep medication
82 (13.5)
24 (16.6)
0.340
Antidepressant
57 (9.4)
21 (14.5)
0.070
Benzodiazepine
23 (3.8)
5 (3.5)
0.848
Study site
<0.001
Birmingham, AL
101 (16.6)
22 (15.2)
Minneapolis, MN
85 (14.0)
19 (13.1)
Palo Alto, CA
79 (13.0)
50 (34.5)
Monongahela Valley, PA
111 (18.2)
11 (7.6)
Portland, OR
101 (16.6)
30 (20.7)
San Diego, CA
132 (21.7)
13 (9.0)
1 All cells report column percentages unless otherwise specified
2 p -value corresponds to chi-square tests for binary/categorical variables, and F-test for continuous variables
3 Column percentages for medical conditions and medications do not add to 100%; each condition/medication is measured dichotomously
No significant differences between the hyperkyphosis and non-hyperkyphosis groups were observed for all subjective (PSQI, ESS) and objective (wrist actigraphy and polysomnography) sleep measures in both unadjusted and adjusted models (Tables 2 and 3 ). In examining hyperkyphosis as a continuous measure (e.g., number of blocks with scores from 0–8 blocks), we noted no statistically significant correlations between severity of hyperkyphosis and sleep measures (PSQI: B = 0.05, 95% CI = -0.11, 0.22; ESS: B = 0.10, 95% CI = -0.11, 0.31; TST: B = 2.31, 95% CI = -1.23, 5.84; WASO: -0.42, 95% CI = -2.11, 1.27; SE: B = 0.09, 95% CI = -0.28, 0.46; SOL: B = 1.10, 95% CI = -0.81, 3.02; AHI: B = -0.25, 95% CI = -1.11, 0.62). Additionally, using commonly used cut points for subjective and objective sleep quality (i.e., PSQI>5, ESS≥10, actigraphic TST<5 hours, SE<80%, WASO>90 minutes, SOL>30 minutes, and AHI>15), we also observed no associations with hyperkyphosis as the predictor (PSQI: OR = 1.15, 95% CI = 0.80, 1.66; ESS: OR = 1.23, 95% CI = 0.82, 1.85; TST: OR = 0.94, 95% CI = 0.45, 1.99; WASO: OR = 1.26, 95% CI = 0.85, 1.85; SE: OR = 1.06, 95% CI = 0.72, 1.57; SOL: OR = 1.09, 95% CI = 0.76, 1.57; AHI: OR = 1.15, 95% CI = 0.80, 1.66). Given that long-duration sleep (in addition to short-duration sleep) may also be associated with poor health outcomes [ 20 ], we repeated the cut point analysis for TST but excluded those with >8 hours sleep, and results were still non-significant (OR = 0.99, 95% CI = 0.47, 2.10).
10.1371/journal.pone.0228638.t002
Table 2 Subjectively measured sleep quality by hyperkyphosis status.
Hyperkyphosis
Comparison
Normal Mean (SD)
Hyperkyphotic Mean (SD)
Bivariate β (95% CI) 1
Multivariate β (95% CI) 1 , 2
PSQI
Global score
5.4 (3.04)
5.7 (2.98)
0.26 (-0.29, 0.81)
0.10 (-0.48, 0.68)
Subjective quality
0.8 (0.69)
0.8 (0.64)
0.01 (-0.11, 0.13)
-0.01 (-0.15, 0.12)
Latency
0.8 (0.85)
0.9 (0.92)
0.13 (-0.02, 0.29)
0.07 (-0.09, 0.24)
Duration
0.7 (0.64)
0.6 (0.62)
-0.10 (-0.22, 0.01)
-0.03 (-0.15, 0.09)
Efficiency
0.6 (0.87)
0.6 (0.89)
0.05 (-0.11, 0.21)
0.05 (-0.12, 0.22)
Disturbance
1.3 (0.54)
1.2 (0.48)
-0.05 (-0.15, 0.04)
-0.08 (-0.19, 0.02)
Sleep medication
0.5 (1.03)
0.7 (1.17)
0.16 (-0.03, 0.35)
0.13 (-0.07, 0.34)
Daytime dysfunction
0.7 (0.68)
0.8 (0.63)
0.05 (-0.07, 0.18)
-0.04 (-0.16, 0.09)
ESS
6.8 (3.93)
7.3 (4.17)
0.50 (-0.22, 1.22)
-0.01 (-0.77, 0.75)
Notes: SD = Standard Deviation; β = Beta Coefficient; 95% CI = 95% Confidence Interval; PSQI = Pittsburgh Sleep Quality Index; ESS = Epworth Sleepiness Scale
1 Corresponds to beta coefficient in linear regression models for the difference in subjective sleep measures by kyphosis severity groups.
2 Models are adjusted for age, height, weight, marital status, PASE score, cardiovascular disease, and study site.
10.1371/journal.pone.0228638.t003
Table 3 Objectively measured sleep quality by hyperkyphosis status.
Hyperkyphosis
Comparison
Normal Mean (SD)
Hyperkyphotic Mean (SD)
Bivariate β (95% CI) 1
Multivariate β (95% CI) 1 , 2
Total Sleep Time, minutes
393.5 (65.95)
398.3 (67.60)
4.78 (-7.25, 16.80)
-3.53 (-16.15, 9.09)
WASO, minutes
78.0 (32.27)
78.9 (28.87)
0.90 (-4.84, 6.64)
-0.29 (-6.35, 5.78)
Sleep Efficiency, %
82.7 (7.00)
82.5 (6.77)
-0.19 (-1.45, 1.07)
-0.20 (-1.54, 1.15)
Sleep Latency, minutes
38.7 (33.73)
41.9 (44.05)
3.21 (-3.30, 9.73)
3.00 (-4.04, 10.04)
Apnea Hypopnea Index
18.5 (15.94)
19.4 (17.11)
0.98 (-2.00, 3.95)
0.95 (-2.17, 4.08)
Notes: SD = Standard Deviation; β = Beta Coefficient; 95% CI = 95% Confidence Interval; WASO = Wake After Sleep Onset
1 Corresponds to beta coefficient in linear regression models for the difference in objective sleep measures by kyphosis severity groups.
2 Models are adjusted for age, height, weight, marital status, PASE score, cardiovascular disease, and study site.
Finally, as participation in MrOS Sleep visit 2 required that a participant not use a respiratory device at night (e.g., CPAP) or other oxygen treatment, and may thus have excluded those with the most severe sleep disturbances, we also conducted a sensitivity analysis to determine whether hyperkyphosis was associated with subjective sleep quality (as measured by the PSQI) administered to the entire MrOS cohort at Visit 4 where this exclusion criteria was not applied. Among the individuals with complete data for hyperkyphosis at Visit 3 and PSQI data for Visit 4 an average of 7.4 (SD = 0.57) years later (N = 1,607), we found unadjusted associations between hyperkyphosis and the PSQI components of sleep duration (B = -0.08, 95% CI = -0.17, 0.00) and daytime dysfunction (B = 0.13, 95% CI = 0.05, 0.21). After adjusting for participant age at Visit 3, daytime dysfunction remained statistically significant (B = 0.12, 95% CI = 0.04, 0.20) ( S1 Table ).
Discussion
In this population-based sample of community-dwelling older men, we hypothesized that presence of hyperkyphosis would be associated with poorer subjective and objective sleep quality. Our hypothesis was not supported, finding no differences in sleep measurements between those with and without hyperkyphosis. In addition, severity of hyperkyphosis was not associated with future sleep quality measures. These results are in line with findings from our previous study [ 10 ], and suggests that the pulmonary and postural complications seen in hyperkyphosis may not contribute to future poor sleep quality in men. As our previous study showed a significant cross-sectional relationship between hyperkyphosis and subjective sleep quality in women, replication of our longitudinal analysis in a sample of older women is needed.
Why sleep quality might be adversely affected by hyperkyphosis in women, but not men is not clear. The prevalence of hyperkyphosis is higher in women as compared to men, with underlying conditions contributing to hyperkyphosis (including osteoporosis, degenerative disk disease, spinal weaknesses, etc.) that are all more prevalent in women [ 21 ]. It is possible that these underlying conditions may be contributing to poor sleep rather than hyperkyphosis per se. Another explanation might be gender differences in postural changes with aging. In older women, the hyperkyphosis changes predominantly occur in the thoracic region while in men the cervical changes that would be captured by the blocks measure of kyphosis are much more prominent [ 1 ]. It is conceivable that thoracic versus cervically prominent hyperkyphosis would more adversely influence pulmonary function and therefore, measures of sleep quality. In our previous paper, the kyphosis measure used was the flexicurve that is placed along the thoracic and not cervical spine. It will be important for future studies to directly make comparisons between men and women.
More broadly from a clinical context, understanding the association between hyperkyphosis and sleep quality is important for determining factors contributing to hyperkyphosis prognosis and other aging outcomes over time. Sleep quality is well acknowledged as a major contributor to health and well-being across the entire life course, and obtaining proper sleep may help in minimizing the deleterious effects of hyperkyphosis and other underlying health conditions that are contributing to impairments in functioning. Additionally, by maintaining quality sleep, one might be able to improve the health outcomes of hyperkyphosis treatment, which may improve quality of life among those affected.
This study’s strengths include a large population-based sample with well-validated subjective and objective sleep measures and a longitudinal design. However, our study should be interpreted in the context of its limitations. First, wrist actigraphy was not administered to those using a respiratory device at night and who were not willing to discontinue use during sleep assessment. This may have resulted in a sample of participants with better pulmonary function and by proxy may have excluded those with the most severe sleep apnea (e.g., obstructive sleep apnea) limiting sleep quality. Thus, our null findings may be explained by having a sample of subjects who are less disabled by hyperkyphosis. Our sensitivity analyses suggest there were correlations between hyperkyphosis and some subjective PSQI sleep components, but this may be due to the larger sample size increasing power. Second, our results are only generalizable to older men. Additional longitudinal investigations are needed in women and more diverse samples. Third, we did not have data on treatment for hyperkyphosis. Forth, while we controlled for numerous confounding variables, residual confounding due to unmeasured confounders may be possible.
In summary, while we did not find an association between hyperkyphosis and subsequent sleep quality, more work is needed to determine the ways in which hyperkyphosis may influence not only sleep but other biological, behavioral and social factors important for healthy aging and prevention of disability. Such findings may substantially contribute to improving the health of an aging population.
Supporting information
S1 Table
Sensitivity analysis of Pittsburgh Sleep Quality Index scores at Visit 4 by hyperkyphosis status (N = 1,607).
SD = Standard Deviation; b = Beta Coefficient; 95% CI = 95% Confidence Interval; PSQI = Pittsburgh Sleep Quality Index. 1 Corresponds to beta coefficient in linear regression models for the difference in subjective sleep measures by kyphosis severity groups. 2 Models are adjusted for age.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Development of new drugs for infectious diseases has taken new urgency in recent years given the emergence and spread of antimicrobial resistance [ 1 ] that threatens global health. Resource-poor countries, where the infectious disease burden is highest, are most at risk. Drug resistant veterinary pathogens seriously compromise global food security. Animal African trypanosomiasis (AAT or nagana) affects millions of domestic animals each year [ 2 ] causing billions of dollars’ worth of lost productivity in a part of the world where food scarcity impacts the population heavily. Given the rise of resistance to existing trypanocides [ 3 ], the Global Alliance for Livestock Veterinary Medicines has developed a programme to seek new drugs for AAT ( https://www.galvmed.org/livestock-and-diseases/livestock-diseases/animal-african-trypanosomosis/ ). The leading class are the benzoxaboroles [ 4 – 7 ], boron-containing compounds that display versatile therapeutic potential against various infectious diseases [ 8 ]. Acoziborole is undergoing Phase II/III clinical trials for human African trypanosomiasis (HAT) [ 9 , 10 ], a neglected tropical disease with unmet medical needs [ 11 ]. Acoziborole may play a key role in the HAT elimination programme [ 12 ], being active against both bloodstream and CNS involved stages of the disease after a single, oral dose [ 9 ].
More recently, another benzoxaborole, AN11736, was identified as a potential development candidate for AAT [ 13 ]. AN11736 cures cattle of both Trypanosoma congolense and Trypanosoma vivax infection as a single 10 mg/kg dose [ 13 ]. Compared to other benzoxaboroles, AN11736 is extremely potent against trypanosomes, killing at doses two to three orders of magnitude lower than that of the earlier AAT benzoxaborole candidate AN7973 [ 14 ] and of acoziborole [ 15 ], respectively.
As novel chemical entities, the benzoxaboroles are unlikely to display cross-resistance with current trypanocides. However, characterisation of mode of action and resistance mechanisms of these compounds are only now starting to emerge. Acoziborole resistance was initially associated with multiple genetic changes [ 16 ]. Subsequently, acoziborole, AN11736 and AN7973 were shown to target the Cleavage and Polyadenylation Specificity Factor 3 (CPSF3) which, when over-expressed, reduced drug sensitivity [ 15 ]. CPSF3 has also been identified as a target for benzoxaboroles in two apicomplexan parasites [ 17 , 18 ], although in highly divergent organisms, i.e. bacteria and fungi, other targets have been proposed, including tRNA synthases [ 19 , 20 ] and beta-lactamase [ 21 ]. In Trypanosoma brucei treated with acoziborole, metabolomics experiments revealed a profound change in methionine metabolism [ 22 ], that may relate to RNA processing defects, given multi-methylation of the spliced leader sequence used for trans -splicing by trypanosomatids [ 23 ]. Some processing mechanisms of particular benzoxaborole molecules by parasites have been identified. A trypanocidal benzoxaborole, of the amino-methyl subclass, was shown to be subject to two-step metabolic processing, involving a primary conversion by an amine oxidase in host serum to an aldehyde, that is further metabolised to a carboxylate via T . brucei aldehyde dehydrogenase [ 24 ]. More recently, the benzoxaborole AN13762 was found to be intracellularly hydrolysed in Plasmodium falciparum by a lysophospolipase homologue, whose loss of function was linked to resistance [ 25 ].
Here we report on the risk and mode of resistance to AN11736 in animal trypanosomes. Our results indicate that AN11736 acts as a prodrug that, once inside trypanosomes, is cleaved by specific serine carboxypeptidases, thus creating a concentration gradient resulting in more parent drug entering the trypanosome cell. Loss or reduction of this enzymatic activity renders trypanosomes highly resistant to AN11736 and to related benzoxaboroles containing a common peptidic linker between the boron head group and a secondary moiety.
Results
Selection of resistance to AN11736 in T . brucei and T . congolense
Resistance to AN11736 was selected in T . brucei and T . congolense by continuous culture in escalating doses of drug. T . brucei able to grow in the presence of 9–18 nM of AN11736 were obtained after ~30 days of culture ( Fig 1A ). Resistant clones (one each of two independent resistant lines) were around 200-fold (TbOX R _A) and >300-fold (TbOX R _C) less sensitive to AN11736 as compared to parent cells.
10.1371/journal.ppat.1008932.g001
Fig 1
AN11736 resistance selection in T . brucei and T . congolense and cross-resistance phenotype of the T . congolense AN11736-resistant clones, revealing a common chemical feature.
(A) In vitro selection of resistance to AN11736 in independent T . brucei lines TbOX R _A and TbOX R _C (left) and resistance levels of two resistant clones obtained from lines TbOX R _A (EC 50 58 nM) and line TbOX R _C (EC 50 97 nM) as compared to the wild type line (TbWT, 0.3 nM) (right). (B) Stepwise in vitro selection of resistance to AN11736 in independent T . congolense lines TcoOX R _B and TcoOX R _C (left) and resistance levels of two resistant clones obtained from lines TcoOX R _B (EC 50 15 nM) and line TcoOX R _C (EC 50 54 nM) as compared to the wild type parent line TcoWT (EC 50 0.3 nM) (right). (C) Cross-resistance of the T . congolense AN11736-resistant clones to other benzoxaboroles revealed the presence of a peptide-bond linker (highlighted in blue) in the highly cross-resistant compounds (>20-fold), whereas the same chemical feature was absent in non cross-resistant compounds (< 2-fold). See S2 Table for full data. Values in (A), (B) (right panels) represent means ± SEM of n ≥ 4 (A) or n = 3 (B) independent biological replicates, with data in (B) each generated from two technical replicates.
For T . congolense , in vitro cultivation for more than eight months, in the presence of increasing concentrations of the compound, was required to reach high-level resistance. One individual clone from each of two different resistant lines, able to grow in the presence of 24–50 nM of the benzoxaborole, was chosen for subsequent studies. The sensitivity of these parasites to AN11736 decreased by >50-fold (clone TcoOX R _B) to nearly 200-fold (clone TcoOX R _C) as compared to the parent line ( Fig 1B ).
The T . congolense resistant clones grew only slightly slower than the parent line, but the growth rate was further reduced when the trypanosomes were cultured in the presence of AN11736 ( S1A Fig ). The resistance phenotype of these clones was stable after three months of growth in the absence of AN11736 ( S1B Fig ). Parent but not AN11736-resistant T . congolense parasites were cleared in in vivo mouse infections upon treatment with 5 mg/kg AN11736 ( S1C Fig ).
The T . congolense AN11736-resistant parasites did not demonstrate cross-resistance to drugs currently licensed for AAT, nor to other trypanocides used for HAT ( S1 Table ). Notably, no cross-resistance to the clinical candidate acoziborole was found ( S1 Table ).
Further cross-resistance analysis in T . congolense using a diverse array of benzoxaboroles revealed that AN11736-resistant trypanosomes were cross-resistant to compounds with a peptide-bond linker containing a valinate-amide motif, whereas these parasites showed no cross-resistance to benzoxaboroles without the linker ( Fig 1C ; see S2 Table for all data). These results agree with the absence of cross-resistance with acoziborole, which lacks the linker. Similar findings were obtained for T . brucei AN11736-resistant parasites when tested against various trypanocides and a selection of the same benzoxaboroles array ( S3 Table ).
Resistant trypanosomes share genetic changes in a tandem array of serine carboxypeptidase (CBP) genes
Genome sequencing of the two T . brucei resistant clones TbOX R _A and TbOX R _C revealed a notable reduction of read depth in a region on chromosome 10 where a tandem repeat of the three serine peptidases TbCBP1A , TbCBP1B and TbCBP1C (Tb927.10.1030–1050 respectively) is present ( Fig 2A ). Evidently, a deletion of one or two TbCBPs alleles had occurred in TbOX R _A and TbOX R _C, which also exists in a region of apparent loss of heterozygosity.
10.1371/journal.ppat.1008932.g002
Fig 2
Serine carboxypeptidases are deleted in T . brucei and T . congolense resistant to AN11736.
(A) Coverage of whole genome sequencing data for wild type (TbWT) and AN11736-resistant (TbOX R _A, TbOX R _B) T . brucei at the genomic locus of CBP1 gene copies. The number of fragments mapping to 100 bp windows is shown as individual points, with a moving average of nine windows shown as a line plot. Coverage was normalised for depth of sequencing as described in Materials and Methods. The position of genes within this region is shown as blocks above the plot, with CBP1 genes in blue and other genes in black. (B) Coverage plot as in (A) but with T . congolense wild type (both parent, TcoWT, and high passage, TcoWT_HP) and AN11736-resistant lines (TcoOX R _A, TcoOX R _B).
Genome sequencing of the T . congolense resistant clones TcoOX R _B and TcoOX R _C and the parent lines (TcoWT, cultured for a limited number of passages, and TcoWT_HP, high passage, maintained in culture for the same time required for drug resistance selection), revealed reduced read coverage across the syntenic region of chromosome 10. This region comprises of nine annotated TbCBP1 paralogues in the T . congolense IL3000 reference genome (here referred to as TcoCBP1A-I ) ( Fig 2B ), although the read counts across the T . congolense CBP locus suggest a different arrangement of paralogues in our experimental strain relative to the reference genome assembly. We observed a pronounced drop in read depth at the 5’ end of TcoCBP1A and rise across the gene TcoCBP1I , indicating a loss of several CBPs in resistant cells. The high sequence conservation between the genes complicates analysis. TcoCBP1A and TcoCBP1I are the most divergent from the other seven TcoCBPs ( S2 Fig ). Five genes ( TcoCBP1C , TcoCBP1D , TcoCBP1F-H ) appeared to have the least mappable reads and hence were most likely deleted in both lines. TcoCBP1A , TcoCBP1B and TcoCBP1I did not appear to be affected. We were not able to identify homozygous SNPs in TbCPSF3 or TcoCPSF3 that could explain the resistance phenotype.
Knockdown of serine carboxypeptidases causes AN11736 resistance in T . brucei
RNA interference (RNAi) target sequencing (RIT-seq) provides a means to identify genes whose knockdown promotes drug resistance [ 26 ]. Selecting a library of T . brucei cells containing RNAi-inducing constructs covering the whole genome in the presence of AN11736 also identified CBPs knockdown as the dominant ‘hit’ conferring resistance ( Fig 3A ). Moreover, silencing the expression of the TbCBP1 genes by targeted RNAi confirmed their importance for sensitivity to AN11736, as tetracycline induction of RNAi in these trypanosomes increased the EC 50 ~25-fold ( Fig 3B ).
10.1371/journal.ppat.1008932.g003
Fig 3
Serine carboxypeptidase disruption confers resistance to AN11736.
(A) Genome-wide RIT-seq screen in T . brucei identified the CBP1 genes on chromosome 10 as ‘hits’ for AN11736 resistance. In the lower panel, blue and red peaks are forward and reverse barcoded reads from individual RNAi target fragments; grey, all other reads. (B) Targeted CBP1 RNAi knockdown of the three serine carboxypeptidases in T . brucei conferred resistance to AN11736 under tetracycline induction (Tet+ EC 50 7.7 nM, Tet- EC 50 0.3 nM). (C) Cumulative cell growth in the presence of 10 nM AN11736 (arrows indicate addition of fresh drug) for uninduced T . brucei and two independent induced Cas9/sgRNA CBP1A-C clones, following 24 h of Cas9 induced editing in the latter case. (D) A PCR assay of the Cas9/sgRNA CBP1A-C clones revealed TbCBP1 editing in two independent drug-resistant clones (upper panel, duplicate samples are shown); the locus from the uninduced line yielded a product of approximately 5.7 kbp, while both clones yielded a product of approximately 2 kbp. B (blank), no genomic DNA; control, uninduced line genomic DNA; predicted wild type CBP1 fragment size, 5,696 bp. The maps in the lower panel indicate the edited loci, as determined by sequencing. The small red bars indicate the gRNA target-sites; the red arrowheads indicate the primers used for the PCR-assay; blue indicates >99% identical regions among multiple paralogues; grey, unique to Tb927.10.1050; green, unique to Tb927.10.1030. (E) Dose-response curves for AN11736 of the two CRISPR-Cas9 edited clones analysed, both displaying a drug-resistant phenotype: when CBP1 function was disrupted by Cas9 editing, T . brucei became, on average, 250-fold more resistant to AN11736. Data in (B), (E) represent means ± SD of n = 3 independent biological replicates.
Disruption of TbCBP1A-C (Tb927.10.1030–1050) function by CRISPR-Cas9 gene editing [ 27 ] corroborated these results. Cas9 programmed to target TbCBP1A-C was induced for 24 h and then cells were selected with AN11736 using two independent Cas9/sgRNA CBP1 clones. The growth profiles indicated robust drug-resistance with induction of TbCBP1A-C editing that was not observed in wild type control cells ( Fig 3C ). A PCR-based assay confirmed that the CBP1 locus was disrupted in both independent edited clones ( Fig 3D , upper panel). Consistent with repair by single-strand annealing [ 28 ], sequencing of the products revealed recombination within blocks of identity in the 1050 ( TbCBP1C ) and 1030 ( TbCBP1A ) genes ( Fig 3D , lower panel; see S3 Fig for clone 1 sequence). Notably, these cells retained a chimeric copy of CBP1A and CBP1C , suggesting that the chimeric protein fails to sensitise parasites to the drug. This may also be the case for the paralogues retained by the resistant strains described above that emerged following drug selection ( Fig 2 ). Assessment of clones’ sensitivity to AN11736 revealed a near 200-fold increase in EC 50 for clone 1 and a 300-fold increase in EC 50 for clone 2 ( Fig 3E ). Thus, CRISPR-Cas9 editing confirmed the role of these serine carboxypeptidases in sensitivity to AN11736.
Re-expression of serine carboxypeptidases re-sensitises AN11736-resistant trypanosomes to the drug
Re-expression of a functional copy of TbCBP1B re-sensitised T . brucei TbOX R _A to AN11736 ( Fig 4A ). Trypanosomes retain the catalytic triad Ser-Asp-His of carboxypeptidases, identified by alignment with other serine carboxypeptidases belonging to the S10 family, well characterised in yeast [ 29 , 30 ] and the other trypanosomatid T . cruzi [ 31 ] ( S4 Fig ). Disruption of the T . brucei catalytic triad in the active site of TbCBP1B, by substituting the nucleophilic S179 with a hydrophobic alanine, failed to re-sensitise the cells to the drug using the same approach ( Fig 4A ). Re-expression of TcoCBP1A and TcoCBP1H in TcoOX R _C also partially restored sensitivity to AN11736 ( Fig 4B ). Heterologous re-expression of TbCBP1B in TcoOX R _C ( Fig 4C ) partially re-sensitised the parasites to AN11736. A similar effect was obtained when re-expressing the only predicted CBP1 serine carboxypeptidase annotated in the T . vivax genome in TbOX R _A ( Fig 4D ). Heterologous expression of TcoCBP1H but not TcoCBP1A in TbOX R _A restored sensitivity ( Fig 4E ).
10.1371/journal.ppat.1008932.g004
Fig 4
Re-expression and heterologous expression of serine carboxypeptidases re-sensitise AN11736-resistant trypanosomes.
(A) Re-expression of TbCBP1B (Tb927.10.1040) in resistant line TbOX R _A partially re-established sensitivity to AN11736 (EC 50 1.6 nM), except when serine 179 was replaced with alanine in the catalytic triad (TbOX R _A + TbCBP1B (S179A), EC 50 65 nM). (B) Add-back of TcoCBP1H and, to a lesser extent, TcoCBP1A partially restored sensitivity to AN11736 in the resistant line TcoOX R _C (EC 50 5.7 nM and EC 50 11.7 nM respectively, TcoOX R _C EC 50 16.3 nM). (C) Heterologous expression of TbCBP1B (EC 50 6.5 nM) partially restored sensitivity to AN11736 in resistant TcoOX R _C line (EC 50 16.3 nM) compared to TcoWT (EC 50 1.2 nM). (D) Expression of TvCBP1 (EC 50 1.4 nM) partially restored sensitivity to AN11736 in resistant TbOX R _A line (EC 50 70.7 nM) compared to TbWT (0.46 nM). (E) Expression of TcoCPB1H (EC 50 0.65 nM) but not TcoCBP1A (EC 50 75.1 nM) restored sensitivity to AN11736 in resistant TbOX R _A line (EC 50 70.7 nM) compared to TbWT (0.46 nM). Data represent means ± SD of n = 3 independent biological replicates.
These results prove that serine carboxypeptidases sensitise different Trypanosoma species to the benzoxaborole AN11736 and that loss of these genes renders the parasites less sensitive to the compound.
Trypanosome serine carboxypeptidases cleave AN11736 to a carboxylate derivative
Mass spectrometry analysis of T . brucei wild type and resistant parasites treated for 6 h with a high dose of AN11736 (0.9 μM, > 1,000-fold EC 50 ) showed the compound was present in the parent and both resistant lines, indicating there was no defect in uptake ( Fig 5A ). Further analysis revealed a compound fragment that was present in the wild type, but not the resistant lines (m/z 292.1347, retention time 10 minutes) ( Fig 5B ). This fragment had a boron isotope distribution ( S5 Fig ) and a predicted formula of C 14 H 19 O 5 NB. A simulation of the isotopic distribution of C 14 H 19 O 5 NB matched the pattern for m/z 292.1347 ( S5 Fig ), further corroborating the identification of this metabolite as a fragment of AN11736, generated by cleavage within the linker at the level of the ester bond.
10.1371/journal.ppat.1008932.g005
Fig 5
AN11736 is cleaved at the ester bond to a carboxylate derivative, which accumulates at high levels in wild type, but not resistant T . brucei or T . congolense .
(A) LC-MS analysis revealed presence of AN11736 in TbWT and resistant lines TbOX R _A and TbOX R _C after 6 h of incubation. (B) A peak of m/z 292.1347 was detected in positive mode in TbWT cells after 6 h of incubation but was barely detectable in resistant cell lines TbOX R _A and TbOX R _C. The peak was identified as the AN11736 carboxylate derivative AN14667, whose chemical structure is shown above the graph. (C) AN11736 was detected in TcoWT, the resistant lines TcoOX R _C, TcoOX R _B and the CBP1H-complemented TcoOX R _C line after 6 h of incubation. (D) The peak of m/z 292.1347 was detected in positive mode in TcoWT cells after 6 h of incubation and at around half the intensity for re-sensitised CBP1H add-back line TcoOX R _C + CPB1H, while the peak was detected at very low levels in resistant cell lines TcoOX R _C and TcoOX R _B, with intensities even lower than that measured for TcoWT at 0 h. (E) Mass spectrometry quantification of AN11736 (full bars) and the metabolite AN14667 (empty bars) in T . brucei wild type (black) and AN11736 resistant line TbOX R _A (blue) over a period of 6 h. (F) as in (E) but for T . congolense wild type and AN11736 resistant line TcoOX R _C. Data represent means ± SD of n = 3 (A-D) or n = 2 (E, F) independent biological replicates.
In T . congolense treated for 6 h with 2 μM AN11736 we could detect AN11736 in parent wild type, both resistant lines and the CBP1H-complemented line ( Fig 5C ). In T . congolense the C 14 H 19 O 5 NB fragment was identified as a large peak in the wild type line but at substantially lower levels in the resistant lines TcoOX R _C and TcoOX R _B ( Fig 5D ). The CBP1H-complemented resistant line TcoOX R _C regained higher levels of the product, proportional to regaining sensitivity to the drug.
Taken together, these data reveal that AN11736 acts as a prodrug that, once inside trypanosomes, is cleaved by serine carboxypeptidases at the ester bond to give a carboxylate derivative (m/z 292.1347, later synthesized under the code name AN14667). In resistant trypanosomes, where genes encoding for serine carboxypeptidases have been deleted or disrupted, this activation does not occur, or does so with substantially reduced efficiency.
When tested against trypanosomes, the carboxylate derivative AN14667 showed much reduced activity compared to AN11736 (~15,000-fold less active against T . brucei wild type and ~800-fold less active against T . congolense wild type), most likely explained by the charged carboxylate derivative poorly traversing the parasite membrane ( S6 Fig ).
AN11736 metabolism causes accumulation of AN14667 and sustains further internalization of the parent compound in sensitive cells
Absolute quantification of AN11736 and its carboxylate metabolite AN14667 by UPLC-MS/MS in wild type and resistant parasites substantiated these findings ( Fig 5E and 5F ). Over a period of 6 h, relatively unchanged intracellular amounts of AN11736 (1 ng ml -1 at 0 h–essentially cells centrifuged as soon as possible after addition of drug, and 0.6 ng ml -1 at 6 h) were measured in TbWT, while in these cells levels of AN14667 were already 1,600-fold higher at time 0 h and further increased at 6 h (1,560 ng ml -1 at 0 h and 11,436 ng ml -1 at 6 h), indicating a very fast processing of the parent compound. In the resistant line TbOX R _A the opposite was observed: levels of AN11736 were higher at both timepoints (6.4 ng ml -1 at 0 h and 3.85 ng ml -1 at 6 h), while AN14667 levels remained much lower (34.5 ng ml -1 at 0 h and 150.5 ng ml -1 at 6 h) than those found in TbWT ( Fig 5E ). Quantification of these metabolites in T . congolense corroborated the T . brucei data. Levels of AN11736 remained low and slightly decreased over time in TcoWT (1.6 ng ml -1 at 0 h and 0.6 ng ml -1 at 6 h), while during the same time levels of AN14667 were markedly higher (17,272 ng ml -1 at 0 h and 12,814 ng ml -1 at 6 h). In TcoOX R _C, AN11736 was present at higher concentration (7.5 ng ml -1 at 0 h and 9.2 ng ml -1 at 6 h) than in the wild type line, while its metabolite levels remained lower than in the wild type (767 ng ml -1 at 0 h and 1,757 ng ml -1 at 6 h) ( Fig 5F ).
T . brucei over-expressing CPSF3 were less sensitive to both acoziborole and AN11736 [ 15 ] indicating that both drugs act by inhibiting this target. The differential activity between acoziborole and AN11736 would, therefore, appear to be related to the metabolism of the parent drug by CBP1, creating a more potent, charged derivative that is retained in the cell, as previously shown for another benzoxaborole by a distinct metabolic route [ 24 ]. Hence, AN11736 would enter cells down a concentration gradient that is perpetuated by drug metabolism with the cleaved derivative accumulating to concentrations much higher than the parent drug. As absolute quantification shows, the metabolite AN14667 reaches vastly higher levels than the parent compound, supporting the superior activity of AN11736 as compared to other benzoxaboroles that do not undergo enzymatic activation, relating to a far greater intracellular accumulation of active compound inside parasites.
Discussion
The benzoxaborole class of compounds has produced multiple clinical development candidates against a range of conditions, including infectious disease [ 4 – 7 ]. Acoziborole, for example, is in clinical trials for human African trypanosomiasis [ 9 , 10 ] and AN11736 is a member of a highly potent benzoxaborole subclass currently under consideration for treatment of AAT [ 13 ]. Recently, the conserved splicing factor CPSF3 was proposed as the major cellular target for benzoxaboroles in trypanosomes [ 14 , 15 ], supported by an earlier study that revealed, among other changes, amplification of the CPSF3, selected during induction of resistance in T . brucei [ 16 ]. AN11736 and a series of related compounds have potency against trypanosomes that exceed that of acoziborole by two to three orders of magnitude. Given requirements of high potency (to keep costs as low as possible for use in cattle), these compounds have received particular interest.
As part of the development process, understanding the risk and mechanisms of resistance is crucial. Here we reveal that the high potency of AN11736 is related to prodrug processing: once the compound has entered trypanosomes it is cleaved by serine carboxypeptidase(s) to a carboxylate product trapped within the cell. This enables accumulation of the benzoxaborole to greatly exceed that where no prodrug conversion and entrapment occurs. The same process, however, leads to a less desirable situation where selection of resistance becomes possible due to loss of enzyme(s) involved in prodrug processing.
Resistance to AN11736 occurs by disruption of expression of serine carboxypeptidase (CBP) genes, which results in diminished AN11736 cleavage. This mechanism of prodrug activation appears analogous to one recently observed in P . falciparum , where benzoxaborole AN13762 is cleaved by esterase activity, whose loss confers resistance to the compound [ 25 ].
The CBPs have been characterised in the trypanosomatid T . cruzi [ 31 ]. In this parasite, the C group serine peptidases in the S10 serine peptidase family proteolytically cleave at C-termini at acidic pHs. This cleavage happens in lysosomes [ 31 ], where the enzymes also have esterase and deamidase activities [ 32 , 33 ]. In T . cruzi , activation of serine peptidases may be achieved through cleavage of a pro-domain by cruzipain [ 34 ]. T . brucei serine carboxypeptidases also have a pro-domain, but it is not known whether brucipain, the cruzipain homologue, is required for T . brucei serine peptidase activation. It is possible that mutations in brucipain would result in a secondary mechanism of resistance, although this was not observed in our analysis. It is probable that the T . brucei and T . congolense CBP serine carboxypeptidases play similar roles to that in T . cruzi where its lysosomal localisation and multiple hydrolytic capabilities [ 32 , 33 ] likely play a generic role in macromolecule turnover. We have not ascertained whether the genes are essential in procyclic form parasites which are resident in the tsetse fly. It will be important to understand this in the future, since a fitness cost in this lifecycle stage would hinder the transmission of parasites that develop resistance via this route in the mammalian bloodstream.
Gene deletions in the serine carboxypeptidase arrays of both T . congolense and T . brucei could clearly be linked to resistance to AN11736. Due to the high degree of sequence homology of the CBPs in both T . brucei and T . congolense we were unable to identify the precise CBP gene deletion(s). However, resistance to AN11736 occurred relatively quickly in T . brucei in vitro , while for T . congolense , which possesses a larger array of CBP genes, resistance took longer to emerge. Importantly, T . congolense AN11736-resistant parasites retained infectivity and resistance phenotype in mice.
The ~200-fold level of resistance obtained for both T . brucei and T . congolense indicates that potency of the otherwise hyper-potent AN11736 would be similar to that of many other benzoxaboroles, including acoziborole (500 nM against T . congolense and 270 nM against T . brucei ) [ 22 ] in absence of drug processing. This suggests that maintaining a concentration in animals that would still kill, even if drug activation were lost, could be possible, albeit using much higher doses of AN11736, which might compromise economic development.
Conversely, a similar cleavage of AN11736 could occur through peptidases present in the blood of treated animals, hence affecting pharmacokinetics. This possibility could reduce the amount of parent compound in circulation, an occurrence particularly important in view of potential prophylactic applications. Our data suggest that the pre-processed compound is of much reduced activity, presumably as the charged derivative is membrane impermeant, consistent with the sink effect resulting from intracellular generation of a carboxylate product.
Experiments with heterologous expression of serine carboxypeptidases suggest that benzoxaborole activation by ester cleavage identified for T . congolense and T . brucei is most likely shared with other trypanosomes, or at least with the major veterinary species T . vivax . In T . vivax the CBP locus in the Y486 strain reference genome consists of a single gene. Whether this would make resistance easier to acquire, or conversely more difficult, given the lack of redundancy, should be investigated once an in vitro culturing system for T . vivax has been developed.
As well as elucidating the resistance mechanism to a class of potent benzoxaboroles, the discovery of a particular moiety that is specifically cleaved by trypanosomal carboxypeptidases offers the potential to exploit that linker to create novel prodrugs with targeted activity against trypanosomatids, i.e. drugs that may not have the ability to traverse membranes could be linked, via the CBP cleavable bridge, to hydrophobic moieties enabling diffusion into cells where they would be cleaved to release the specific inhibitor. A risk of resistance to such compounds emerging through mutation to the CBP genes, though, could limit such a use.
Materials and methods
Ethics statement
The mouse experiment was carried out in accordance with the Animals (Scientific Procedures) Act 1986 and the University of Glasgow care and maintenance guidelines. All animal protocols and procedures were approved by the Home Office of the UK government and the University of Glasgow Ethics Committee. Work was covered by Home Office Project Licence 60/4442.
Trypanosome culture
Bloodstream form (BSF) T . b . brucei strain Lister 427 was cultured at 37°C in a humidified, 5% CO 2 environment in HMI-11 (Gibco), supplemented with 10% FBS (Gibco). BSF T . congolense strain IL3000 was cultured at 34°C in a humidified, 5% CO 2 environment in TcBSF-3 [ 35 ], containing 20% commercial goat serum (Gibco) and 5% Serum Plus II (SAFC Biosciences). Red blood cell lysate was not present.
Generation of oxaborole-resistant lines
T . congolense and T . brucei parasites were selected for resistance to AN11736 by subculturing cells in vitro in the continuous presence of increasing concentrations of the compound. Multiple, independent cell lines were selected in parallel. Resistant lines were cloned by limiting dilution.
Trypanocidal activity
In vitro trypanocidal activity was measured using the Alamar Blue method as previously described [ 36 ]. T . brucei were seeded at 2×10 4 cells ml -1 and T . congolense at 2.5×10 5 cell ml -1 and EC 50 values determined after a total drug incubation time of 72 h. All experiments were carried out in duplicate and on at least three independent occasions unless stated otherwise. Although the assay measures metabolic conversion of resazurin to resorufin reagent and this can also be hindered by trypanostatic compounds we believe in this case it is a true surrogate for cell death hence we labelled the y-axis as percentage parasite survival in Fig 1 .
In vivo virulence of benzoxaborole-resistant trypanosomes
200 μl of 4.5×10 7 T . congolense AN11736 resistant cells and 2.8×10 7 wild type in fresh TcBSF-3 were injected intravenously into immunocompromised NIH female mice (5 per group). Once high parasitaemia had developed (day 16), AN11736 (prepared as a suspension in 10% DMSO) was administered i.p.at a dose of 5 mg/Kg. Parasitaemia was monitored daily by tail blood examination and mice humanely culled when parasitaemia reached 10 8 cells ml -1 .
Whole genome analysis
DNA from T . brucei and T . congolense wild type and resistant clones was extracted using the NucleoSpin Tissue Kit (Macherey-Nagel). Sequencing of paired 75 bp reads was performed using the NextSeq 500 platform (Illumina). Libraries were prepared with 500 ng input gDNA using QIAseq FX DNA library kit (Qiagen) and fragments of 300 bp, including adaptors, were selected with Agencourt AMPure XP (BeckmanCoulter), according to manufacturer instructions. Reads were trimmed for quality and adaptor contamination using Trim Galore! v0.6.2 (Babraham Bioinformatics) and reads were aligned to either the T . b . brucei TREU 927 reference genome, release 43 (available from TriTrypDB, https://tritrypdb.org/tritrypdb/ ), or the T . congolense TcIL3000 2019 reference genome assembled from Pacific Biosciences sequencing data by N. Hall group (also available from TriTrypDB), using Bowtie2 v2.3.5 [ 37 ]. Alignment rates for all samples were 85% for T . brucei and 98% for T . congolense . Reads were sorted and duplicates marked using SAMtools v1.9 [ 38 ] and Picard Tools v2.20.2 ( http://broadinstitute.github.io/picard/ ). To determine depth of coverage, the number of fragments mapping to 100 bp windows along the genome was quantified using featureCounts v1.6.3 [ 39 ] with fragments mapping to multiple windows counted as 1/n (where n is the total number of windows to which a given fragment maps). For comparison of depth of coverage, the number of fragments mapping to each window was normalised for sequencing depth by dividing it by the ratio of number of reads aligned to the parent chromosome (chromosome 10) for that sample to the mean aligned reads for that chromosome across all samples.
Genetic manipulation of T . congolense and T . brucei
For the re-expression of CBPs in T . brucei the amplified open reading frame sequences were cloned into the pRM481 vector [ 40 ] using Xba I and Bam HI (for primers list see S4 Table ). Due to high sequence homology between the TbCBP genes the Tb427.10.1040 ORF was synthesized (BaseClear) and used as template for the PCR. The S179-encoding codon was modified in the above-described pRM481-derived vector containing wild type TbCBP1B by using Q5 Site-Directed Mutagenesis (SDM) Kit (NEB). Primers designed for this purpose were TGTTGGGGAAgcCTACGGTGGC and ACAAAGAAGTCGTTTTCAC.
A plasmid was generated that targets the tubulin locus of T . congolense and transcribes blasticidin S deaminase (BSD) and the C-terminal 6×HA tagged trypanosomal CBP described herein. T . congolense 5’ and 3’-tubulin and actin intergenic sequences were amplified by PCR with Q5-Polymerase (NEB) from genomic DNA. Blasticidin S deaminase ( BSD ) was amplified from a plasmid pGL2271 (for primers list see S4 Table ). Vector pRM481 was digested with Asc I and the plasmid backbone was purified and used for Gibson assembly (NEB) with PCR products. In the resulting vector the CBP genes were flanked upstream by actin intergenic sequence and BSD . Upstream and downstream of the BSD gene were the 5’ and 3’ tubulin intergenic sequences that allowed integration into the tubulin locus upon Asc I digest. This vector was further modified to accept the CBP genes: at 3’ of BamH I site a Hpa I site was introduced and the Cla I site, separating the gene ( GFP ) and the actin intergenic region, was exchanged with a Sal I site by SDM (for primers list see S4 Table ).
Plasmid pRM481 encoding for TcoCBP1A was modified by SDM. A Hpa I site was introduced 3’ of the 6×HA tag and the Xma I site. Then a PCR was performed to amplify the open reading frame and introduce a Sal I site 5’ of the TcoCBP1A ATG and a primer that bound in the T . brucei 3’ tubulin intergenic region (AAACCTACACATGGTGCGACG). TcoCBP1A was inserted into the pRM481 derivative with Sal I and Hpa I. Further gene exchanges were done by amplifying the ORF and inserting into the BamH I and Sal I sites.
10 μg of Asc I-digested plasmid DNA were transfected into 3×10 7 T . congolense (TcoOX R _C) and T . brucei (TbOX R _A) cells as described [ 41 ]. T . congolense and T . brucei clones were selected with 0.4 μg ml -1 blasticidin (InvivoGen) in TcBSF-3 and 5 μg ml -1 blasticidin in HMI-11, respectively. Expression was verified by detection of 6×HA tagged protein and EF1α as loading control on the LiCor’s Odyssey Imaging system.
Putative catalytic serine of TbCBP1B (S179) was identified by aligning its ORF sequence with a series of S10 serine peptidase ORF sequences downloaded from MEROPS Peptidase Database [ 42 ] including Carboxypeptidase Y from yeast, whose catalytic triad is well characterized [ 29 , 30 ]. Alignment was conducted in CLC Genomics Workbench.
RNAi screen
Determinants of AN11736 resistance were identified using an RNAi library screen as previously described [ 26 ]. Briefly, cultures from the screen were split and supplemented with fresh AN11736 as required and DNA was extracted from drug-resistant cells. RNAi target fragments were amplified by PCR using the LIB2f and LIB2r primers. The products were then subjected to high-throughput RIT-seq. Sequencing was carried out on an Illumina HiSeq platform at BGI (Beijing Genomics Institute). Reads were mapped to the T . brucei 927 reference genome (v9.0, tritrypdb.org ) with Bowtie2 using the parameter: very- sensitive-local-phred33. The generated alignment files were manipulated with SAMtools and a custom script was used to identify reads with barcodes (GCCTCGCGA) [ 43 ]. Total and bar-coded reads were then quantified using the Artemis genome browser [ 44 ].
Targeted RNAi of CBP locus
A single RNAi construct targeting all three TbCBPs in a common region of their DNA sequence was produced using plasmid pGL2084 as a backbone, and the resulting vector was transfected into T . brucei 2T1 BSF as previously described [ 45 ]. Primers to amplify the RNAi target sequence included AttB gateway flanks: Fw: GGGGACAAGTTTGTACAAAAAAGCAGGCT CGTTAATCAATGGAGCGGAT, Rev: GGGGACCACTTTGTACAAGAAAGCTGGGT GCTTTCCCCAACAACAAAGA. Genetically modified parasites were selected in HMI-11 complemented with 0.5 μg ml -1 phleomycin (InvivoGen), and 2.5 μg ml -1 hygromycin B (Calbiochem). RNAi induction was obtained with tetracycline (Sigma-Aldrich) at 1 μg ml -1 24 h before experiments.
CRISPR-Cas9 gene editing
T . brucei Lister 427 parasites were grown and manipulated as described previously [ 27 ]. The Cas9 gRNA oligonucleotide pair comprised CBP1.G1 ( AGGG CCTCTTGCAGGATTGGCTGT) and CBP1.G2 ( AAAC ACAGCCAATCCTGCAAGAGG); the overhanging ends that facilitate cloning are in bold. These were annealed, ligated to Bbs I-digested pT7 sgRNA , confirmed by sequencing and introduced into 2T1 T7-Cas9 cells as described [ 27 ]. DNA analysis and drug-sensitivity analysis were carried out as described [ 27 ]. The primer pair used for the Cas9 PCR assay comprised FwCBP1PCRout (GTTACAACATAACCACCGCGG) and RvCBP1PCRout (GGTGGAGTGGGCACAACCAC).
Metabolomics analysis
T . congolense and T . brucei metabolites were extracted for untargeted metabolomics analysis following treatment with test compounds at 10×EC 50 or with the DMSO vehicle control (below 1% v/v) for 6 hours. For each sample 1×10 8 cells were collected and their metabolism was quenched by rapidly cooling to 4°C using a dry ice/ethanol bath. The cells were kept at 4°C from hereon. After a wash in ice cold PBS, cells were resuspended in 200 μl of extraction solvent (Chloroform:Methanol:Water 1:3:1) and shaken at 4°C for 1 h. Extracts were centrifuged at 17,000×g, 10 min, 4°C and the supernatants collected and stored under argon at -80°C until analysis by LC-MS. Four replicates of each sample were prepared. Samples were analysed on an Orbitrap Fusion mass spectrometer (Thermo Fisher Scientific) in both positive and negative modes (switching mode). Hydrophilic interaction liquid chromatography (HILIC) was carried out on a Dionex UltiMate 3000 RSLC system (Thermo Fisher Scientific) using a ZIC-pHILIC column (150 mm Å~ 4.6 mm, 5 μm column, Merck Sequant). HPLC mobile phase A was 20 mM ammonium carbonate in water and mobile phase B was 100% acetonitrile. The column was maintained at 30°C and samples were eluted with a linear gradient from 80% B to 20% B over 24 minutes, followed by 8 minutes wash with 5% B and 8 minutes re-equilibration with 80% B, at the flow rate of 300 μl/minute. Orbitrap data were acquired as previously described [ 46 ]. Untargeted peak-picking and peak matching from raw LC-MS data were obtained using XCMS and mzMatch respectively. Metabolite identification and relative quantitation was performed using IDEOM interface [ 46 ] and PIMP [ 47 ], by matching accurate masses and retention times of authentic standards or, when standards were not available, by using predicted retention times. p-values were adjusted for multiple testing using the Benjamini-Hochberg method. Identifications were supported by fragmentation pattern match to MzCloud database ( https://www.mzcloud.org/home.aspx ) and isotope distribution. The Xcalibur software package from Thermo Fisher Scientific was used for targeted peak picking and fragmentation analysis.
Intracellular quantification of AN11736 and AN14667
Metabolism studies were performed at 1 μM AN11736 with TbWT, TbOX R _A, TcoWT and TcoOX R _C BSF trypanosomes. At 0 h, 1 h and 6 h timepoints 5×10 8 parasites for T . brucei and 7×10 8 parasites for T . congolense were collected and cell pellets resuspended in 100 μl and 50 μl 1×PBS, respectively, precipitated by addition of a 2-fold volume of acetonitrile and centrifuged at 1,700×g, 10 min at room temperature. The supernatant was diluted with water to maintain a final solvent concentration of 50% and stored at -80°C prior to UPLC-MS/MS analysis, following a similar protocol as the one described in Wyllie and colleagues [ 48 ].
Supporting information
S1 Fig
Phenotype of the T . congolense AN11736 resistant clones.
(A) Growth curves of T . congolense AN11736 resistant clones B (TcoOX R _B, left) and C (TcoOX R _C, right) in presence (open symbols) or absence (full symbols) of AN11736 at selection concentration (12 nM for clone B and 24 nM for clone C) and of the parental wild type either cultured for a limited number of passages (TcoWT), or maintained in culture for the same time required for drug resistance selection (TcoWT_HP, high passage); doubling times for each line are indicated within brackets. (B) Stability of the resistant phenotype of clones TcoOX R _B and TcoOX R _C as measured by Alamar Blue assay following 90 days of culture in absence (-OX) or presence (+OX) of AN11736. (C) The resistant clone TcoOX R _C retained virulence and AN11736 resistance in NIH female mice (5/5) reaching high levels of parasitaemia not cleared by treatment with 5 mg/kg of AN11736 at day 16 post infection (arrow), a dose sufficient to clear all TcoWT trypanosomes (5/5). Data in (A, B) represent means ± SD of n = 3 independent biological replicates.
(TIF)
S2 Fig
Nucleotide alignment of the nine annotated CBP1 genes of T . congolense .
The alignment was made with CLC genomics workbench using the TcIL3000 reference genome available from TriTrypDB ( https://tritrypdb.org/tritrypdb/ ).
(PDF)
S3 Fig
Sequencing of the Cas9/sgRNA CBP1 clone 1, a Tb427.10.1050/103 chimera.
(A) Nucleotide sequence. (B) Translated amino acid sequence. In bold, identical to only Tb427.10.1050; bold and italicised, identical to only Tb427.10.1030; normal, no difference to Tb427.10.1040; italicised and underlined are differences that are not present in 1050 or 1030. Blue and underlined, identical to Tb427.10.1050 and Tb427.10.1040. Red, underlined and bold, unique to the chimera. Underlined sequence, Cas9 targeting sequence.
(PDF)
S4 Fig
Amino acid alignment of serine carboxypeptidases from T . b . brucei , T . congolense , T . vivax and T . cruzi .
Extract of alignment of the annotated serine carboxypeptidases from T . congolense (TcoCPB1A-I), T . b . brucei (TbCBP1A-C, Tbb427.10.1030–50), T . vivax (TvCBP1, TvY486_1000990) and T . cruzi (TcCBP1, TcCLB.508671.20). The sequences were blasted against the entire collection of S10 carboxypeptidases stored at MEROPS Peptidase Database [ 42 ]. Family S10 has residues of the catalytic triad in the order Ser, Asp and His [ 29 , 30 ] and carboxypeptidase Y (MER0002010) from Saccharomyces cerevisiae is the most representative gene of the family. Indicated with asterisks is the polar catalytic serine (S179) of the triad. This Ser was targeted in Tb927.10.1040 for site directed mutagenesis, substituting with a hydrophobic alanine (S179A). The alignment was made with CLC genomics workbench.
(PDF)
S5 Fig
Isotopic distribution for fragment m/z 292.1347 (metabolite AN14667).
The fragment had an isotopic distribution that matched the simulated isotopic distribution of C 14 H 19 O 5 NB (example obtained for a TbWT replicate treated for 6 h with AN11736).
(TIFF)
S6 Fig
Sensitivity of T . brucei and T . congolense to the AN11736 metabolite (or fragment) AN14667.
(A) No difference in susceptibility to AN14667 was found for TbWT and the resistant line TbOX R _A (EC 50 9.5 μM and 10.55 μM respectively). (B) The AN11736 resistant T . congolense line TcoOX R _C was more than 2.5-fold more resistant to the metabolite than TcoWT (EC 50 2.49 μM and 0.92 μM respectively); Data represent means ± SD of n = 3 independent biological replicates.
(TIFF)
S1 Table
Cross-resistance of T . congolense AN11736-resistant clones B (TcoOX R _B) and C (TcoOX R _C) to the main veterinary trypanocides and other trypanocidal compounds.
(PDF)
S2 Table
In vitro cross-resistance of T . congolense AN11736-resistant clones TcoOX R _B and TcoOX R _C to a diverse array of other benzoxaboroles.
(PDF)
S3 Table
Cross-resistance of T . brucei AN11736-resistant clones A (TbOX R _A) and to a selection of trypanocides and other benzoxaboroles.
(PDF)
S4 Table
Primer sequences for integration into pRM481, generation of T . congolense tubulin locus integration plasmid and for insertion into and modification of T . congolense tubulin locus integration plasmid.
(PDF)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
The historical invisibility of lesbian, gay, bisexual, and transgender (LGBT) lives is something of a pattern in LGBT health research, driven by invisibility in public health surveillance systems [ 1 ]. Nonetheless, a growing number of high-quality data sources have documented health inequalities in chronic disease, infectious disease, mental health, and violent victimization [ 2 ]. A growing research agenda seeks to examine the origins of these inequalities and evaluate interventions to address them [ 2 ]. As the body of LGBT health research grows, evidence synthesis through reproducible systematic review and meta-analysis methodologies becomes increasingly important [ 3 ].
Systematic reviews provide a rigorous approach to identifying existing literature thereby limiting bias through the selection of studies [ 3 ]. Additionally, systematic reviews and meta-analyses can show trends across multiple smaller studies that are individually difficult to interpret given their small size [ 3 ]. Searches of the grey literature (i.e., unpublished in academic journals) can help counteract the effect of publication bias [ 4 ]. Systematic reviews and meta-analyses can inform evidence-based interventions and identify practice-based evidence from community organizations [ 5 ]. Systematic reviews are particularly important when study results are spread across multiple disciplines and academic as well as non-academic journals.
To achieve these important goals, however, systematic reviews and meta-analyses must be conducted in a high-quality manner [ 6 ]. In the initial stages of identifying the existing literature through a systematic search process, bias can be introduced by failing to identify relevant studies. Implementing high-quality searches and reporting them according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines remains a challenge for systematic reviews in general [ 7 , 8 ].
Although the importance of systematic review and meta-analysis methodologies and reporting are of general concern to health researchers, the incredible diversity of terminology used to describe and define LGBT communities by researchers, advocates, and community members provides an additional challenge to systematically reviewing LGBT health literature [ 9 , 10 ]. To assist LGBT health researchers in the literature search process, we sought to examine the keyword searches used, report on the searches, and propose additional terminology for use in LGBT health searches. We operationalized this in two aims: (1) to describe characteristics of search strategies used in LGBT health systematic reviews and (2) to identify a comprehensive set of LGBT search terms that can be used to increase the sensitivity of LGBT health systematic review searches.
Methods
Search
Using PubMed/MEDLINE, we developed keywords and MeSH terms in two domains (systematic reviews and homosexuality). We based our starting search keywords on resources from the University of Texas Libraries [ 11 ] for systematic review terminology and our previous work reviewing the literature on tobacco interventions for LGBT populations [ 12 ]. After iteratively testing and improving our search strings, we then translated our search into the controlled vocabulary of other databases. We excluded certain unrelated terms because their abbreviations are used in LGBT health, for example: “markov state model” and “men who have sex with men” are both abbreviated MSM. A matrix of controlled vocabulary and individual database search strings are reported in S1 File . We implemented our search on May 28–29, 2015, in three health databases: Embase, PsycINFO, and PubMed/MEDLINE. We hand-searched the Web site of the journal LGBT Health on September 11, 2015. We set no date, geographic, or language limits in our search or in our inclusion process. Our final PubMed/MEDLINE search string was:
((bisexual[tiab] OR bisexuality[MeSH Terms] OR bisexuality[tiab] OR bisexuals[tiab] OR gay[tiab] OR gays[tiab] OR GLB[tiab] OR GLBT[tiab] OR homosexual[tiab] OR homosexualities[tiab] OR homosexuality[MeSH Terms] OR homosexuality[tiab] OR homosexuals[tiab] OR intersex[tiab] OR lesbian[tiab] OR lesbianism[tiab] OR lesbians[tiab] OR LGB[tiab] OR LGBT[tiab] OR "men who have sex with men"[tiab] OR msm[tiab] OR queer[tiab] OR "sexual minorities"[tiab] OR "sexual minority"[tiab] OR "sexual orientation"[tiab] OR transgender[tiab] OR transgendered[tiab] OR transgenders[tiab] OR transsexual[tiab] OR transsexualism[MeSH Terms] OR transsexualism[tiab] OR transsexuality[tiab] OR transsexuals[tiab] OR "women loving women"[tiab] OR "women who have sex with women"[tiab] OR WSW[tiab]) NOT (gay[au] OR "laparoscopic gastric bypass"[tiab] OR "markov state model" OR "multiple source method"[tiab]))
AND
(systematic*[tiab] AND (bibliographic*[tiab] OR literature[tiab] OR review[tiab] OR reviewed[tiab] OR reviews[tiab])) OR (comprehensive*[tiab] AND (bibliographic*[tiab] OR literature[tiab])) OR "integrative literature review"[tiab] OR "integrative research review"[tiab] OR "integrative review"[tiab] OR “research synthesis”[tiab] OR “research integration”[tiab] OR meta-analys*[tiab] OR meta-analyz*[tiab] OR meta-analyt*[tiab] OR metaanalys*[tiab] OR metaanalyz*[tiab] OR metaanalyt*[tiab] OR “meta-analysis as topic”[MeSH:noexp] OR Meta-Analysis[ptyp] OR ((review[tiab] AND (rationale[tiab] OR evidence[tiab])) AND review[pt])
Inclusion
We set our criteria for inclusion as being a systematic review related to LGBT health. We defined systematic review as (a) using a set of keywords in (b) two or more databases with (c) independent coders assessing all identified records for inclusion or exclusion. Guidelines from Agency for Healthcare Research and Quality (AHRQ) (recommendation 7.6.6) [ 13 ], Cochrane (recommendation 7.2.4) [ 14 ], and the U.S. Institute of Medicine (IOM; recommendation 3.3.3) [ 15 ] all recommend dual independent coding for inclusion to reduce error and increase confidence in the findings.
In defining LGBT health, we sought to include studies that addressed domains such as injury prevention, chronic disease, mental health, violence, and sexual health and well-being. We a priori excluded: (a) HIV/AIDS-specific studies (because they often focus exclusively on same-sex behavior and we wished to focus this search on a broader definition of LGBT health), (b) studies about same-sex contact and resulting risk for HIV and sexually transmitted infections, (c) studies about the impact of LGBT parents on children (because the children may not be LGBT), (d) studies about treatment of homosexuality or gender dysphoria (including hormone therapy), and (e) studies about the origins of homosexuality.
After de-duplication, two authors independently screened the title and abstract of 1,226 records for potential inclusion or exclusion, removing studies clearly not related to the research question. Two authors then independently screened each of the 134 full text records identified for possible inclusion. At each stage, differences in coding were reconciled through discussion and consensus of at least two authors. We did not calculate reliability because we viewed the goal of independent coders being one of enhancing sensitivity to eligible records rather than one of establishing uniformity. We used Covidence (covidence.org) to manage the screening and coding process. Fig 1 shows the inclusion process.
10.1371/journal.pone.0156210.g001
Fig 1
PRISMA Flow Diagram, May 28–29, 2015.
Abstraction
Two authors independently abstracted the following information from the included records: (a) if the review reported a search string in keeping with the PRISMA guideline #8 (“Present full electronic search strategy for at least one database, including any limits used, such that it could be repeated”) [ 16 ], (b) the keywords used to define the LGBT population of interest, (c) the databases searched, (d) any hand-searched journals, (e) inclusion of grey literature, (f) if the authors reviewed reference lists of included studies, (g) involvement of a librarian (because inclusion of a librarian has been shown to improve search quality [ 17 ]), (h) if the study assessed publication bias, (i) whether the study included meta-analysis, and (j) the area of LGBT health covered. We discussed any discrepancies in extraction coding and obtained consensus among authors, then exported the data into an evidence table.
In the interest of assembling maximum data on search terms used for LGBT health, we e-mailed the corresponding author to request the full search string if it was not reported in the manuscript. We then cross-referenced the abstracted search strategies with our own search strategy to create the most comprehensive search string for LGBT health systematic reviews.
Results
We identified 19 studies meeting our inclusion criteria. These studies examined aging [ 18 , 19 ], alcohol use [ 20 ], breast cancer [ 21 ], cardiovascular outcomes for transgender users of sex steroids [ 22 ], health information–seeking behaviors [ 23 ], intimate partner violence [ 24 – 26 ], mental health [ 27 ], stressful childhood experiences [ 28 ], substance abuse [ 29 , 30 ], suicide [ 31 ], tobacco use [ 12 , 32 , 33 ], and victimization/abuse [ 34 ].
Review terminology and databases are reported in Table 1 . In accordance with PRISMA reporting guidelines for searches, 13 presented a final search string from a specific database, including any limits. The number of LGBT-related keywords ranged from 1 [ 24 , 31 ] to 31 [ 35 ]. One study reported conflicting information about what databases were searched [ 31 ]. For the remaining 18 studies, the number of databases ranged from 2 [ 24 , 30 , 34 ], the minimum required for study inclusion, to 15 [ 18 ]. The most commonly used databases in the 19 identified studies were PubMed/MEDLINE (17 studies), PsycINFO (15 studies), CINAHL (8 studies), Web of Science/Knowledge (7 studies), and Embase (6 studies). Nine studies searched the grey literature, 17 reported searching reference lists, and 6 reported involvement of a librarian as an author in the methods or in the acknowledgements. Six of the 19 included studies assessed publication bias. Nine studies conducted a meta-analysis. Table 2 reports the search characteristics.
10.1371/journal.pone.0156210.t001
Table 1 Study databases and search keywords (or final search strategy, if reported), N = 19, May 28–29, 2015.
Study
N: Databases
Page: Search or Keywords
Badenes-Ribera et al ., 2015 [ 24 ]
2: PubMed and PsycINFO
p. 47: Lesbian
Batejan et al ., 2015 [ 35 ]
4: PsycINFO, Medline, SocINDEX, and ERIC
Appendix: bicurious OR bisexual(s) OR bisexuality OR gay(s) OR GLB OR GLBQ OR GLBs OR GLBT OR GLBTQ OR heteroflexible OR homosexual(s) OR homosexuality OR lesbian(s) OR LGB OR LGBQ OR LGBS OR LGBT OR LGBT OR lesbigay OR men who have sex with men OR MSM OR queer(s) OR same sex attracted OR same sex attracted youth OR SSA OR SSAY OR same-sex relations OR sexual minority OR sexual orientation OR women who have sex with women OR WSW
Blosnich et al ., 2013 [ 32 ]
10: Academic Search Elite, Alt HealthWatch, CAB Abstracts 1990-Present, CINAHL with Full Text, Health Source Consumer Edition, Health Source: Nursing/Academic Edition, MEDLINE, PsycARTICLES, PsycINFO and Social Work Abstracts
p. 67: ((homosexual* OR gay OR ‘sexual minority’ OR ‘sexual minorities’ OR lesbian* OR bisexual* OR queer OR ‘sexual orientation’ OR ‘men who have sex with men’ OR MSM OR ‘women who have sex with women’ OR WSW)
Buller et al ., 2014 [ 25 ]
13: MEDLINE, EMBASE, Global Health, PsycINFO, the Health Management Information Consortium database [HMIC], Social Policy and Practice, the Cumulative Index to Nursing and Allied Health Literature [CINAHL], the International Bibliography of the Social Sciences [IBSS], Web of Science, Africa Web, Index Medicus for South-East Asia Region [IMSEAR], Index Medicus for the Eastern Mediterranean Region [IMEMR], and Latin American and Caribbean Health Sciences Literature [LILACS])
In S1 File (PubMed via OVID): Homosexuality/ OR Homosexuality, Male/ OR Transsexualism/ OR Bisexuality/ OR Homosexual*.mp. OR Transexual*.mp. OR Bisexual*.mp. OR Transgender.mp. OR (MSM or men who have sex with men or ((man or men or male*) adj3 (gay or homosexual or queer or bisexual* or transsexual* or transgender)) or LGBT).mp.
Elamin et al ., 2010 [ 22 ]
5: Ovid MEDLINE, Ovid Embase, Ovid PsycInfo, Thomson Scientific Web of Science and Elsevier Scopus
Upon Request: 1. (trans adj (sexual$ or gender$ or male or men or women or female or people or person$)).mp. [mp = title, original title, abstract, name of substance word, subject heading word] 2. gender identity/ and su.fs. 3. sex reversal, gonadal/ 4. ((sex$ or gender) adj (transition$ or transform$ or reassign$ or chang$)).mp. [mp = title, original title, abstract, name of substance word, subject heading word] 5. transsexualism/ or (trans adj sexual$).mp. or transexual$.mp. or transsexual$.mp. [mp = title, original title, abstract, name of substance word, subject heading word] 6. ((gender or sexual$) adj2 (dysphor$ or identity)).mp. [mp = title, original title, abstract, name of substance word, subject heading word] 7. (crossgender or (cross adj (sex$ or gender$))).mp. [mp = title, original title, abstract, name of substance word, subject heading word] 8. (transgender$ or (trans adj gender$)).mp. [mp = title, original title, abstract, name of substance word, subject heading word] 9. (m2f or f2m or "male-to-female" or "female-to-male").mp. and (1 or 2 or 3 or 4 or 5 or 6 or 7 or 8) [mp = title, original title, abstract, name of substance word, subject heading word] 10. or/1-9 11. limit 10 to humans
Finkenauer et al ., 2012 [ 18 ]
15: CINAHL (1942-), Medline (1942-), Health Services/Technology Assessment Texts, Web of Science, EMBASE (1947-), Sociological Abstracts (1952), Social Services Abstracts (1806-), Gender Studies Database (1972-), LGBT Life with Full Text, Ageline (1978-), PsycINFO (1806-), Scopus, ERIC, The New York Academy of Medicine Grey Literature Report, and Dissertations & Theses: Full Text.
p. 313: transgender or transsexual or transexual or transman or transwoman or genderqueer or “gender queer” or LGBT or GLBT or transvestite or crossdress or “cross dress ” or “cross-dress ” or “drag queen” or “drag queens” or “drag king” or “drag kings” or “gender identity disorder” or “gender dysphori ”
Friedman et al ., 2011 [ 34 ]
2: MEDLINE and PsycINFO
P. 1482: NR but example keywords include gay, lesbian, bisexual, sexual orientation, homosexual, and homosexuality
Goldbach et al ., 2014 [ 29 ]
3: PsychINFO, PubMED, and EBSCO.
p. 351: lesbian OR gay OR bisexual OR sexual minority
Harding et al ., 2012 [ 19 ]
4: Medline (1950-present), PsycINFO (1806–2010), Cinahl (1982–2010), and ASSIA (1987–2010).
p. 603: homosexual OR lesbian OR gay OR transgender OR bisexual
King et al ., 2008 [ 27 ]
12: Medline, Embase, PsycINFO, Cinahl, the Cochrane Library Database, the Web of Knowledge, the Applied Social Sciences Index and Abstracts, the International Bibliography of the Social Sciences, Sociological Abstracts, the Campbell Collaboration and grey literature databases for Google and Google Scholar
NR
Langhinrichsen-Rohling et al ., 2012 [ 26 ]
7: Academic Search Premier, Education Resources Information Center, Medical Literature Analysis and Retrieval System Online, PsycINFO, CINAHL, Biomedical Reference Collection, and SocINDEX
p. 205: N/A (authors manually selected LGBT-related studies from a search on intimate partner violence)
Lee et al ., 2009 [ 33 ]
7: Seven databases were searched for peer-reviewed research articles (Cumulative Index to Nursing and Allied Health Literature (CINAHL), Cochrane Library via Wiley InterScience, Education Resources Information Center (ERIC), Health Source: Nursing/Academic, Institute for Scientific Information (ISI) Web of Science, PsycINFO via EBSCO Host and PubMed)
p. 282: homosexuality OR homosexual OR gay OR "sexual minority" OR "female homosexuality" OR "homosexuality, female" OR lesbian OR bisexuality OR bisexual OR transgender OR transsexual OR transsexualism OR transsexuality OR MSM OR queer OR "sexual orientation" OR "men who have sex with men” OR WSW OR “women loving women” OR “women who have sex with women” OR lesbianism
Lee et al ., 2014 [ 12 ]
8: Cochrane Central Register of Controlled Trials via Wiley Online Library; Cumulative Index to Nursing and Allied Health Literature (CINAHL), Global Health, PsycINFO, and Social Work Abstracts via EBSCO; Embase; PubMed; and Scopus.
p. 824: (homosexuality[MeSH Terms] OR homosexuality[tiab] OR homosexual[tiab] OR gay[tiab] OR LGBT[tiab] OR GLBT[tiab] OR LGB[tiab] OR “sexual minority”[tiab] OR “sexual minorities”[tiab] OR lesbian[tiab] OR bisexuality[MeSH Terms] OR bisexuality[tiab] OR bisexual[tiab] OR transsexualism[MeSH Terms] OR transsexualism[tiab] OR transgender[tiab] OR transsexual[tiab] OR trans- sexuality[tiab] OR msm[tiab] OR queer[tiab] OR “sexual orientation”[tiab] OR “men who have sex with men”[tiab] OR WSW[tiab] OR “women loving women”[tiab] OR “women who have sex with women”[tiab] OR lesbianism[tiab])
Liu et al ., 2014 [ 20 ]
3: PubMed, WanFang Data, Google Scholar
p. 2: “MSM” OR “men who have sex with men” OR “gay” OR “homosexual”
Marshal et al ., 2008 [ 30 ]
2: PsycINFO and Medline
p. 548: NR but example keywords include: gay, lesbian, bisexual, LGB
Meads et al ., 2013 [ 21 ]
8: Cochrane library (CDSR, CENTRAL, HTA, DARE, NHSEED), MEDLINE, EMBASE, PsycINFO, CAB abstracts, Web of Science (SCI, SSCI), SIGLE, Social Care Online
p. 2: lesbian, gay women, queer, bisexual, sexual preference, sexual orientation
Pompili et al ., 2014 [ 31 ]
Unknown: MedLine, Excerpta Medica, PsycLit and PsycINFO, and Index Medicus reported in methods section; in results 3 are reported as being searched: PubMed, Scopus, and Web of Knowledge.
p. 1904: Bisexuality
Rose et al ., 2013 [ 23 ]
4: MEDLINE, Applied Social Services Index and Abstracts, Sociological Abstracts and Social Service Abstracts.
p. 419: lesbian OR gay OR bisexual
Schneeberger et al ., 2014 [ 28 ]
5: MEDLINE (Ovid), PubMed, Web of Science, Google Scholar, and PsycNet (includes PsycINFO, PsycBOOKS, PsycARTICLES, PsycTESTS),
p. 2–3: lesbian, gay, bisexual, transgender, transsexual, homosexual, men who have sex with men
Note: NR = Not Reported
10.1371/journal.pone.0156210.t002
Table 2 Study Characteristics, N = 19, May 28–29, 2015.
Study
Replicable String Reported
Reported Librarian Involvement
Searched Citations
Grey Literature Search
Hand Search
Assessed Publication Bias
Meta-analysis
Research Area
Badenes-Ribera et al ., 2015 [ 24 ]
Yes
No
Yes
No
Yes *
No
Yes
Intimate Partner Violence
Batejan et al ., 2015 [ 35 ]
No
No
Yes
Yes
No
Yes
Yes
Self-Harm, Non-Suicidal
Blosnich et al ., 2013 [ 32 ]
Yes
No
Yes
No
No
No
No
Tobacco
Buller et al ., 2014 [ 25 ]
Yes
No
Yes
No
Yes **
Yes
Yes
Intimate Partner Violence
Elamin et al ., 2010 [ 22 ]
No
Yes
Yes
Yes
No
No
Yes
Cardiovascular Disease, Sex Steroids
Finkenauer et al ., 2012 [ 18 ]
Yes
Yes
Yes
Yes
No ***
No
No
Aging
Friedman et al ., 2011 [ 34 ]
No
No
Yes
Yes
No
Yes
Yes
Victimization and Abuse
Goldbach et al ., 2014 [ 29 ]
Yes
No
Yes
No
No
Yes
Yes
Substance Abuse
Harding et al ., 2012 [ 19 ]
Yes
No
Yes
No
No
No
No
Aging
King et al ., 2008 [ 27 ]
No
Yes
Yes
Yes
No
No
Yes
Mental Health
Langhinrichsen-Rohling et al ., 2012 [ 26 ]
No
No
Yes
No
No
No
No
Intimate Partner Violence
Lee et al ., 2009 [ 33 ]
Yes
Yes
Yes
No
No
No
No
Tobacco
Lee et al ., 2014 [ 12 ]
Yes
Yes
No
Yes
No
No
No
Tobacco
Liu et al ., 2014 [ 20 ]
Yes
Yes
Yes
No
No
Yes
Yes
Alcohol
Marshal et al ., 2008 [ 30 ]
No
No
Yes
Yes
No
Yes
Yes
Substance Abuse
Meads et al ., 2013 [ 21 ]
Yes
No
Yes
Yes
No
No
No
Breast Cancer
Pompili et al ., 2014 [ 31 ]
Yes
No
Yes
No
No
No
No
Suicide
Rose et al ., 2013 [ 23 ]
Yes
No
No
No
No
No
No
Health Information Seeking Behaviors
Schneeberger et al ., 2014 [ 28 ]
Yes
No
Yes
Yes
No
No
No
Stressful Childhood Experiences
* Hand search: Journal of Homosexuality; Journal of Lesbian Studies; Journal of Gay & Lesbian Social Service; Journal of GLBT Family Studies; Journal of LGBT Health Research; Journal of LGBT Issues in Counseling
** Hand search: Journal of Homosexuality, Journal of Gay & Lesbian Social Services, and Journal of LGBT Issues in Counseling
*** But, completed a "manual search of published reference material"
Our initial PubMed/MEDLINE search contained 36 search terms. Cross-referencing these with the identified search terms, we identified an additional 46 LGBT-related terms. We excluded three of these—“cross dress,” “drag king(s),” and “drag queen(s)”—used in a review on transgender aging [ 18 ] because these terms do not necessarily indicate a sexual orientation or gender identity [ 10 ]. We excluded two additional terms, the abbreviations “SSA” and “SSAY” (for same-sex attracted [youth]), because these picked up thousands of un-related articles, leaving 41 new terms. We added five additional terms that were not used in any search. These are terms we have seen used in LGBT health research: “same gender loving” [ 36 ], “same sex couple” [ 37 ], “same sex couples” [ 38 ], “sexual and gender minority” [ 39 ], and its plural version, “sexual and gender minorities.” This full list of 82 terms is presented below with bolded terms coming from the identified reviews and italicized terms added based on their use in the field.
(bicurious[tiab] OR bisexual[tiab] OR bisexuality[MeSH Terms] OR bisexuality[tiab] OR bisexuals[tiab] OR “cross sex”[tiab] OR crossgender[tiab] OR F2M[tiab] OR “female-to-male”[tiab] OR gay[tiab] OR gays[tiab] OR “gender change”[tiab] OR “gender dysphoria”[tiab] OR “gender identity”[tiab] OR “gender queer”[tiab] OR “gender reassign”[tiab] OR “gender transform”[tiab] OR “gender transition”[tiab] OR genderqueer[tiab] OR GLB[tiab] OR GLBQ[tiab] OR GLBs[tiab] OR GLBT[tiab] OR GLBTQ [tiab] OR heteroflexible [tiab] OR homosexual[tiab] OR homosexualities[tiab] OR homosexuality[MeSH Terms] OR homosexuality[tiab] OR homosexuals[tiab] OR intersex[tiab] OR lesbian[tiab] OR lesbianism[tiab] OR lesbians[tiab] OR lesbigay[tiab] OR LGB[tiab] OR LGBQ[tiab] OR LGBS[tiab] OR LGBT[tiab] OR M2F[tiab] OR “male-to-female”[tiab] OR “men who have sex with men”[tiab] OR msm[tiab] OR queer[tiab] OR “same gender loving”[tiab] OR “same sex attracted”[tiab] OR “same sex couple”[tiab] OR “same sex couples”[tiab] OR “same sex relations”[tiab] OR “sex change”[tiab] OR “sex reassign”[tiab] OR “sex reversal”[tiab] OR “sex transform”[tiab] OR “sex transition”[tiab] OR “sexual and gender minorities”[tiab] OR “sexual and gender minority”[tiab] OR “sexual identity”[tiab] OR “sexual minorities”[tiab] OR “sexual minority”[tiab] OR “sexual orientation”[tiab] OR “sexual preference”[tiab] OR “trans female”[tiab] OR “trans male”[tiab] OR “trans man”[tiab] OR “trans men”[tiab] OR “trans people”[tiab] OR “trans person”[tiab] OR “trans woman”[tiab] OR “trans-sexuality”[tiab] OR transexual[tiab] OR transgender[tiab] OR transgendered[tiab] OR transgenders[tiab] OR transsexual[tiab] OR transsexualism[MeSH Terms] OR transsexualism[tiab] OR transsexuality[tiab] OR transsexuals[tiab] OR transvestite[tiab] OR “women loving women”[tiab] OR “women who have sex with women”[tiab] OR WSW[tiab] NOT ("laparoscopic gastric bypass"[tiab] OR gay[au] OR "markov state model" OR "multiple source method"[tiab]))
The body of evidence identified from the LGBT search domain in PubMed/MEDLINE on November 4, 2015, is 40,759 and 53,451 for our original and the expanded search, respectively.
Discussion
There is room for improvement in the implementation and reporting of literature searches in LGBT health systematic reviews and meta-analyses. Strong evidence synthesis is essential to address a multitude of health concerns for LGBT populations. Authors have an ethical obligation to the field to reduce bias from study identification to ensure limited available resources are used effectively.
A strong evidence base for documenting, understanding, and intervening on LGBT health inequalities requires high-quality systematic reviews and meta-analyses. Comprehensive guidelines are available from AHRQ [ 13 ], Cochrane [ 14 ], and IOM [ 15 ]. Based on this assessment of the state of LGBT health systematic reviews, we recommend that authors of systematic reviews in LGBT health use and report (and peer reviewers hold to account): (a) including a librarian or information specialist as collaborator to improve the search quality [ 17 ], (b) using more than one academic database, (c) using the controlled vocabulary of databases, (d) conducting searches of the reference lists of included studies, (e) reporting a complete specific search string so that the review can be updated as new literature emerges, (f) using dual coders for inclusion to improve data quality, and (g) using dual coders for abstraction or, at minimum, a reviewer to confirm and validate evidence tables [ 40 ]. The work presented in this paper contributes to the development of better searches given the complex terminology used in LGBT health [ 9 ], but each of these recommendations on its own would contribute to stronger evidence synthesis in the field of LGBT health.
There are important limitations to this study. First, we used a somewhat restrictive definition of systematic review requiring dual, independent coding of titles and abstracts. Although AHRQ [ 13 ], Cochrane [ 14 ], and IOM [ 15 ] recommend dual independent coding for inclusion, many systematic reviews—some with strong search strategies—were ineligible due to not reporting the number of coders or having a single author decide which papers to include. Second, we did not empirically test our comprehensive search against other strings used by each of the studies identified in our search, thus we cannot be certain to what extent our search would improve the identification of relevant studies. We viewed this as being an unfair comparison because the identification of studies is a multi-step process that is unique to the aims of a given study. Third, searches must balance sensitivity and specificity; our work represents a preliminary effort to address search coverage by increasing sensitivity to LGBT health-related articles. Further work is needed to ensure a balance between sensitive searches and more specific searches. Fourth, changes in terminology to define and describe LGBT populations are likely already happening [ 9 ]; although our work provides a thorough list of keywords for searching, future reviewers should consider the ever-shifting landscape of LGBT terminology. Fifth, we conducted our original search in three academic databases; searching a larger number of databases could have resulted in inclusion of additional reviews. Sixth, we did not assess the role of publication bias in our identification of search terminology; results could be influenced by unpublished reviews that may have poorly designed search strings.
The lives of LGBT individuals have historically been invisible in health data [ 1 ] and in popular culture [ 41 ]. With growing research to address health inequalities, it is imperative that rigorous methods to identify and synthesize existing research be employed. With diverse and shifting terminology being used, researchers should carefully consider the terminology used to identify as much of the relevant literature as possible. Efforts to combat health inequalities are only as strong as the evidence available to know what inequalities exist, how they come into being, and how to intervene against them.
Supporting Information
S1 File
Systematic review protocol.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Background
Unkept hospital appointments, also known as “did not attends” (DNAs), are a dilemma facing multiple healthcare systems worldwide. In 2017–2018, 8 million National Health Service (NHS) hospital appointments, almost 1 in 10, were unkept in England. Each outpatient hospital appointment is estimated to cost the NHS £120, yielding an overall cost for the system of approximately £1 billion in unkept appointments [ 1 – 3 ]. It is estimated that unkept appointments cost the healthcare system in the US $150 billion a year [ 4 ]. The financial and public health impacts of unkept appointments are therefore vast. It is wasteful of resources and may also increase patient morbidity and lengthen waiting lists from 1 week to up to 6 months [ 5 , 6 ]. A nationwide study in Scotland reported that those who missed more than 2 appointments had a 3-fold increase in hazards of mortality compared to those who did not miss appointments [ 6 ]. The NHS is a service with limited resources and is ever under the stress of financial limitations; hence, unkept appointments need to be addressed to ensure resources are allocated appropriately.
Reasons for unkept appointments often include forgetfulness, being unaware of the appointment, feeling too unwell to attend, hospital administrative errors, work commitments, transport difficulties, and resolution of symptoms [ 7 – 20 ].
In the absence of a more predictive and proactive approach to mitigate unkept appointments, outpatient clinics often overbook appointments. This puts constraints on those delivering clinic care. Other proactive interventions such as appointment reminders by phone call, letter, or text message have been implemented to try to reduce the number of unkept appointments. Other strategies have included giving patients the responsibility of booking their appointment, either through Freephone service or online [ 5 ]. Interventions may fail as they are often targeted using a blanket approach, without knowing which cohort of patients or clinical factors are most effective to target. Interventions need to be targeted effectively to increase patient engagement and therefore aid the ongoing issue of tackling health inequalities.
In order to understand how best to target interventions to reduce the number of unkept outpatient clinic appointments, it is first important to understand the predictors and characteristics of unkept appointments. Previous studies have employed machine learning and statistical modelling to predict unkept appointments [ 21 – 24 ]. However, such modelling has not been carried out in a UK setting across all medical specialties, but rather has been limited to individual specialties or restricted to primary care or community settings.
Our aim was to demonstrate the rates of unkept outpatient clinic appointments across hospital trusts in England, with an added breakdown by specialty. To appraise the potential utility of this approach at a local level, we examined the predictors of unkept outpatient clinic appointments across specialties at a single trust within England, Imperial College Healthcare NHS Trust (ICHT). Finally, we trained machine learning models to determine the effectiveness of a potential intervention in reducing unkept appointments.
Methods
Data capture
We present a data flow diagram in Fig 1 . We used Hospital Episode Statistics (HES) outpatient data spanning England from April 2016 to March 2018 to generate and train the models used for this study. Data codings and definitions can be found on the NHS Digital website [ 2 ]. HES is a database detailing all admissions and emergency and outpatient appointments at NHS hospitals. HES provides a number of patient characteristics including age, sex, ethnicity, and geographical information. Index of Multiple Deprivation (IMD) demographic data were also used—a dataset that is openly available [ 25 ].
10.1371/journal.pmed.1003783.g001
Fig 1
Data flow diagram.
Numbers of appointments and unkept appointments were obtained for the top 50 trusts by appointment volume. Unkept appointments within Imperial College Healthcare NHS Trust were further broken down by specialty. DNA, did not attend; HES, Hospital Episode Statistics; IMD, Index of Multiple Deprivation.
Data cleaning
The submission of kept outpatient appointment data to NHS Digital is mandatory in England. However, the submission of unkept appointments is not mandatory. Some sites have a reported 0% unkept appointment rate in the data they submit. We excluded these sites from this analysis, along with the bottom and top 10% of outliers.
Our goal was to predict outpatient non-attendance without warning. Hence, appointments that were cancelled in advance, either by patients or consultants, were also excluded from this analysis.
Appointments from April 2016 to March 2017 were used to train the models. The recency (in days) and frequency (over the previous 12 months) of appointments and unkept appointments were calculated and included as predictors (see S1 Text for details).
We tested the models using ICHT-specific data from April 2017 to March 2018. This again included the number of appointments, unkept appointments, and cancellations, calculated by specialty, from the previous 12 months.
Statistical analysis
The recorded outcome variable was binary, indicating whether the patient attended their appointment or did not attend their appointment without providing advance warning. Note however that the output of the models was a probability between 0 and 1, indicating the likelihood of a given patient not showing up to an appointment prior to the fact. Advance cancellations were not included in the data.
Variables used in the model were defined as shown in Table 1 .
10.1371/journal.pmed.1003783.t001
Table 1 Predictor variable definition.
Predictor variable
Description
Unkept appointments last 12 months
Count of number of unkept appointments in last 12 months
Appointments last 12 months
Count of number of outpatient appointments in last 12 months
Cancellations last 12 months
Count of number of cancellations in last 12 months
Days since last unkept appointment
Days since the previous unkept appointment
Days since last appointment
Days since the previous outpatient appointment (kept or unkept)
Days since last cancellation
Days since the previous cancellation
Lead care professional
Patient to be seen by the lead care professional versus another member of the professional team
APPTAGE CALC
Age at appointment—babies under 1 year decimalised
REFSOURC
Source of referral
Health deprivation score
Health Deprivation and Disability subscale from Index of Multiple Deprivation (IMD)
IDAOPI score
Income Deprivation Affecting Older People Index subscale from IMD
IDACI score
Income Deprivation Affecting Children Index subscale from IMD
IMD score
Overall Index of Multiple Deprivation Scale
Sex
Sex of patient
Weekday
Day of week (Monday, Tuesday, Wednesday, etc.)
Consultation
Service type requested
Appointment type
First attendance, follow-up, or telephone
Using the R statistical programming language, we trained machine learning models for each of the top 100 treatment specialties by national volume in 2016–2017 to determine predictors of unkept appointments and their relative importance ( S1 Text ). These models were gradient boosting machines (GBMs). HES data from 2016–2017 were used to train the models.
The models were then tested using 2017–2018 HES outpatient data. We conducted an analysis based on a hypothetical intervention targeting the top 10% of outpatient appointments by risk. The intervention here could be a phone call reminder or a virtual consultation. This was a post hoc analysis, based on observational data, and did not have a prespecified study plan.
Test metrics
Model sensitivity is the proportion of unkept appointments captured by the intervention. For example, a sensitivity of 0.33 implies that 33% of unkept appointments are captured; that is, the intervention could at best reduce unkept appointments by 33%. The positive predictive value (PPV) is the proportion of the time the model is right. For example, suppose that 100 people are chosen for the intervention. If the PPV is 0.5, then this tells us that 50 out of 100 would have missed their appointment if there was no intervention. The likelihood ratio (LR) tells us how much more likely those who are selected for an intervention are to have an unkept appointment, in comparison to those who are not selected. The area under the receiver operating characteristic curve (AUROC) tells us how good the model is at distinguishing classes, in this instance between patients who will attend their appointment versus those who will not.
Prediction metrics
The prediction metrics included the percentage importance of the different predictors for a given speciality (defined in terms of reduction in predictive error), the average importance for a predictor across all outpatient specialties, and the variation in importance for a predictor across all outpatient specialties.
See S1 Table (RECORD Checklist) for details on our reporting in this study.
As per the Health and Social Care Act 261 and the Data Protection Act 2018, as a national institution, NHS Digital is directed to store and analyse secondary care data. Internal approval for this project was granted by the information asset owner.
Results
Across the whole NHS
In both 2016–2017 and 2017–2018, there were approximately 97–98 million outpatient appointments after removing advance cancellations, and data cleaning. The rate of unkept appointments in these datasets was 8.0% and 8.1%, respectively.
Across trusts in England, focusing on the top 50 trusts by appointment volume, the rate of unkept appointments ranged from 3.9% to 14.8%. Fig 2 shows the top 20 trusts with the highest unkept appointment rates. The full table can be found in S2 Table . Eight of the 10 trusts with the highest unkept appointment rates were within the London region.
10.1371/journal.pmed.1003783.g002
Fig 2
Unkept appointment rates by UK hospital trust (excluding advance cancellations).
Within Imperial College Healthcare NHS Trust
Focusing on ICHT only, there were approximately 1.1 million outpatient appointments, excluding advance cancellations and sites with very low volumes. Approximately 910,000 remained after data cleaning. The rate of unkept appointments in this data set was 11.2%. The rates of the highest and lowest 5 specialties are presented in Table 2 . Hepatology had the highest rate of unkept appointments (17%), and medical oncology had the lowest (6%). See S3 Table for all specialties.
10.1371/journal.pmed.1003783.t002
Table 2 Highest and lowest unkept appointment rates by speciality.
Speciality name
Number of appointments
Number of unkept appointments
Unkept appointment rate (%)
Highest rates of unkept appointments
Hepatology
14,731
2,494
17%
Diabetic medicine
12,182
1,929
16%
Ophthalmology
74,442
11,188
15%
Ear, nose, and throat
34,820
5,109
15%
Vascular surgery
12,666
1,772
14%
Lowest rates of unkept appointments
Gynaecology
62,526
5,508
9%
Breast surgery
16,838
1,428
8%
Anaesthetics
15,690
1,180
8%
Audiological medicine
20,577
1,647
8%
Medical oncology
36,466
2,052
6%
Predictors of outpatient non-attendance overall
Predictor importance, sorted by average importance, for our composite prediction model across all specialties is shown in Table 3 . Days since the previous unkept appointment (recency) was responsible on average for 25%, and the number of unkept appointments a patient had in the previous 12 months (frequency) was responsible on average for 13% of the predictive value of these models. Age at time of appointment accounted for around 10% of the predictive value. A larger number of previous unkept appointments was associated with an increased risk of failing to keep a future appointment. A larger number of previous cancellations was associated with an increased risk of failing to keep a future appointment. Older patients were least likely to miss an appointment. More deprived areas (lower IMD decile) were associated with an increased risk of an unkept appointment. Seeing a lead care professional was associated with a decreased risk of an unkept appointment.
10.1371/journal.pmed.1003783.t003
Table 3 Overall predictor importance.
Predictor
Mean
Rank average
Variation
Days since last unkept appointment
25%
1.2
0.2
Unkept appointments last 12 months
13%
3.0
1.7
Age at appointment
10%
4.6
8.8
Lead care professional
9%
5.1
6.5
Appointments last 12 months
7%
5.3
4.7
Days since last appointment
6%
6.0
4.2
Referral source
5%
7.1
8.2
Health deprivation score
5%
7.1
4.0
IDAOPl score
5%
7.7
7.6
IDACI score
3%
10.8
3.8
IMD score
2%
12.0
4.4
Weekday
2%
11.7
3.6
Appointment type
2%
12.8
5.9
Days since last cancellation
2%
12.7
3.1
Consultation
2%
13.9
5.4
Cancellations last 12 months
1%
15.6
2.5
Sex
0%
16.5
0.8
IDACI, Income Deprivation Affecting Children Index; IDAOPI, Income Deprivation Affecting Older People Index; IMD, Index of Multiple Deprivation.
Predictors of outpatient non-attendance by specialty
Fairly consistently across the specialties with the highest and lowest unkept appointment rates, days since the previous unkept appointment and the number of previous unkept appointments in the last 12 months were among the most important predictors (Tables 4 and 5 ). The full table for all specialties can be seen in S4 Table . Age at time of appointment and number of appointments in the last 12 months were also important. Age at appointment was the most variable in terms of its importance for a given specialty. For instance, age at appointment is relatively important as a predictor of attendance for audiological medicine; ear, nose, and throat; and ophthalmology, but relatively unimportant for hepatology and vascular surgery. Referral source also varied in importance by specialty. Number of days since the last unkept appointment was consistently among the top 2 predictors, while sex was consistently among the bottom 4.
10.1371/journal.pmed.1003783.t004
Table 4 Predictor rank for each specialty with the highest and lowest unkept appointment rates: Mean percentage.
Speciality name
Appointments last 12 months
Unkept appointments last 12 months
Cancellations last 12 months
Days since last appointment
Days since last unkept appointment
Days since last cancellation
Age at appointment
IDAOPI score
IDACI score
IMD score
Health deprivation score
Consultation
Lead care professional
Appointment type
Sex
Referral source
Weekday
Highest rates of unkept appointments
Hepatology
9.3%
16.4%
0.6%
7.8%
29.5%
1.7%
5.9%
2.4%
2.3%
4.3%
4.8%
3.7%
3.7%
0.9%
0.4%
3.3%
3.1%
Diabetic medicine
12.3%
15.1%
0.6%
9.5%
19.3%
1.4%
6.7%
1.5%
1.3%
0.9%
5.2%
0.8%
14.8%
4.5%
0.9%
4.0%
1.3%
Ophthalmology
5.9%
11.4%
0.5%
4.3%
28.3%
1.1%
13.6%
6.4%
3.9%
3.1%
4.8%
1.1%
6.3%
0.7%
0.3%
6.9%
1.4%
Ear, nose, and throat
3.8%
9.5%
0.4%
4.4%
25.4%
1.6%
15.2%
5.0%
2.6%
1.6%
7.2%
0.8%
16.0%
1.4%
0.4%
3.3%
1.6%
Vascular surgery
5.6%
14.3%
1.2%
4.9%
25.9%
1.1%
5.4%
3.5%
1.2%
1.3%
5.7%
2.6%
20.1%
1.6%
0.4%
3.4%
1.7%
Lowest rates of unkept appointments
Gynaecology
3.4%
7.8%
0.6%
8.5%
26.9%
2.4%
6.1%
4.4%
2.7%
2.2%
7.6%
1.5%
11.4%
3.7%
0.0%
8.2%
2.5%
Breast surgery
5.7%
7.5%
0.3%
7.0%
37.1%
1.4%
10.6%
4.3%
3.0%
1.8%
4.9%
1.4%
6.5%
4.0%
0.0%
2.8%
1.6%
Anaesthetics
4.5%
7.6%
0.7%
5.3%
34.8%
2.3%
14.0%
6.1%
2.5%
1.8%
3.2%
0.2%
3.8%
2.8%
0.8%
6.5%
3.3%
Audiological medicine
2.9%
11.1%
0.6%
2.3%
13.2%
1.4%
29.0%
11.9%
3.5%
2.5%
4.3%
0.3%
7.3%
0.5%
0.3%
7.7%
1.3%
Medical oncology
7.2%
8.2%
1.6%
8.2%
13.7%
3.6%
6.3%
10.5%
6.9%
6.4%
8.3%
5.3%
5.3%
1.8%
0.4%
3.3%
3.2%
IDACI, Income Deprivation Affecting Children Index; IDAOPI, Income Deprivation Affecting Older People Index; IMD, Index of Multiple Deprivation.
10.1371/journal.pmed.1003783.t005
Table 5 Predictor rank for each specialty with highest and lowest unkept appointment rates: Rank average.
Speciality name
Appointments last 12 months
Unkept appointments last 12 months
Cancellations last 12 months
Days since last appointment
Days since last unkept appointment
Days since last cancellation
Age at appointment
IDAOPI score
IDACI score
IMD score
Health deprivation score
Consultation
Lead care professional
Appointment type
Sex
Referral source
Weekday
Highest rates of unkept appointments
Hepatology
3
2
16
4
1
14
5
12
13
7
6
8
9
15
17
10
11
Diabetic medicine
4
2
17
5
1
11
6
10
12
15
7
16
3
8
14
9
13
Ophthalmology
7
3
16
9
1
13
2
5
10
11
8
14
6
15
17
4
12
Ear, nose, and throat
8
4
17
7
1
12
3
6
10
13
5
15
2
14
16
9
11
Vascular surgery
5
3
15
7
1
16
6
8
14
13
4
10
2
12
17
9
11
Lowest rates of unkept appointments
Gynaecology
10
5
16
3
1
13
7
8
11
14
6
15
2
9
17
4
12
Breast surgery
6
3
16
4
1
14
2
8
10
12
7
15
5
9
17
11
13
Anaesthetics
7
3
16
6
1
13
2
5
12
14
10
17
8
11
15
4
9
Audiological medicine
9
4
14
11
2
12
1
3
8
10
7
17
6
15
16
5
13
Medical oncology
6
4
16
5
1
12
9
2
7
8
3
11
10
15
17
13
14
Green corresponds to the most important predictors, and red to the least important. IDACI, Income Deprivation Affecting Children Index; IDAOPI, Income Deprivation Affecting Older People Index; IMD, Index of Multiple Deprivation.
GBM model
In order to predict rates of unkept appointments, GBM models were trained on the top 100 specialties by appointment volume. Data from 2016–2017 were used to train the models, and data from 2017–2018 were used to test them. The test metrics for the specialties with the highest and lowest unkept appointment rates are shown in Table 6 . From these models, we calculated the sensitivity, LR, and PPV for generating interventions in selected proportions of non-attenders to assess potential clinical improvements in attendance.
10.1371/journal.pmed.1003783.t006
Table 6 Gradient boosting machine validation metrics of specialties with the highest and lowest unkept appointment rates.
Speciality name
Number of appointments
Number of unkept appointments
Unkept appointment percent
Sensitivity
PPV
LR
AUROC
Highest rates of unkept appointments
Hepatology
14,731
2,494
17%
0.28
0.46
4.10
0.74
Diabetic medicine
12,182
1,929
16%
0.29
0.47
4.62
0.74
Ophthalmology
74,442
11,188
15%
0.26
0.39
3.63
0.71
Ear, nose, and throat
34,820
5,109
15%
0.26
0.38
3.61
0.69
Trauma and orthopaedics
44,966
6,460
14%
0.25
0.37
3.45
0.69
Lowest rates of unkept appointments
Gynaecology
62,526
5,508
9%
0.31
0.27
3.82
0.72
Breast surgery
16,838
1,428
8%
0.35
0.29
4.48
0.72
Audiological medicine
20,577
1,647
8%
0.24
0.19
2.67
0.67
Anaesthetics
15,690
1,180
8%
0.26
0.20
2.99
0.67
Medical oncology
36,466
2,052
6%
0.29
0.16
3.30
0.69
AUROC, area under the receiver operating characteristic curve; LR, likelihood ratio; PPV, positive predictive value.
Sensitivity at 10% cutoff
A sensitivity of 0.28 for hepatology suggests that 28% of patients who do miss their appointment would be successfully targeted if the top 10% least likely to attend received an intervention. As a result, an intervention targeting the top 10% of likely non-attenders, in the full population of patients, would be able to capture 28% of unkept appointments if successful.
Likelihood ratio
The LR for the top 5 specialties was greater than 3, meaning those patients selected by the models for an intervention were at least 3 times as likely to miss their appointment than those who were not selected for an intervention.
Positive predictive value
The PPV for the top 5 specialties was between 37% and 47%, meaning that of those selected for the targeted intervention, 37%–47% would be expected to miss their appointment prior to the intervention. This is in comparison to a 14%–17% unkept appointment rate across all appointments for the top 5 specialties.
Among the bottom 5 specialties, of those selected for an intervention roughly 16%–29% would be expected to miss their appointment, in comparison to a 6%–8% unkept appointment rate across all appointments in the bottom 5 specialties.
Area under the curve
As a metric of model performance, the AUROC was fairly consistent across both the top 5 and bottom 5 specialties, ranging from 0.67 to 0.74.
So long as the cost of an intervention is less than one-third of the average cost of the potential reduction in unkept appointments, then using these models for targeted interventions would theoretically be cost-effective.
Discussion
Unkept appointments are a worldwide issue, causing inequalities in health and inefficient use of resources. Using a national data-driven approach, we determined national and local unkept outpatient appointment rates across secondary care in the United Kingdom, and across multiple specialties at a single hospital trust. Previous nationwide studies have looked primarily at general practice data [ 6 ]. In our study, rate of unkept appointments varied across NHS hospital trusts from 4.3% to 15.1%. The higher rates seen in London may be explained by the more heterogenous population within London, with language barriers, transport failures, and administrative failures.
Predictors of unkept appointments can be divided into clinical, behavioural, and sociodemographic. At ICHT, there was great heterogeneity in unkept appointment rates across all specialties. The highest rates were seen in hepatology; diabetes; ophthalmology; ear, nose, and throat; and vascular surgery patients. Interestingly, these patients’ co-morbidities may overlap. For example, a patient with diabetes may require ophthalmology review of diabetic retinopathy or vascular review for diabetic foot ulcers. It is not clear why hepatology had the highest rate of unkept appointments; however, studies looking at gastroenterology patients found that causes of unkept appointments included forgetting their appointment or clerical errors [ 10 ]. A cohort study of 521 unkept ophthalmology appointments found that the top reasons for non-attendance were not feeling well enough to attend, forgetting the appointment, administrative errors, and that their condition had improved [ 26 ].
When stratifying by the predictors of unkept appointments, there were similarities and differences across all specialties. Consistently, having a prior unkept appointment was the greatest predictor across all specialties except medical oncology. This suggests that behaviour is the most important predictor, and hence behaviour-related interventions targeting those with recurrent unkept appointments is necessary. Therefore, adopting a targeted approach to reducing unkept appointments maybe more effective than a blanket approach.
Sex was the least important predictor for the majority of specialties. This contradicts other findings in the literature. Previous studies suggested that sex was a predictor of non-attendance, with males having a higher risk [ 6 , 12 , 17 , 27 ]. Similarly, deprivation did not rank very highly, in contrast to existing literature [ 11 , 13 , 28 – 30 ]. A possible reason for this is that the models generated here had greater granularity and included predictors of greater importance that could not always be captured in other studies.
Gynaecology, breast surgery, anaesthetics, audiological medicine, and medical oncology had the lowest rates of unkept appointments. Oncology patients may have better adherence to treatment due to the mortality associated with their disease and hence may be more likely to attend their appointments. Whilst not all, a significant proportion of gynaecological and breast patients may also fall under oncology.
Targeted interventions could be implemented at multiple levels: organisational, psychosocial, or through information dissemination. For example, virtual clinics could be a practical solution, and have been trialled across multiple specialties [ 31 – 33 ]. Other interventions include stating the cost of the appointment when sending SMS reminders to patients, which has been shown in a trial to reduce unkept appointment rates compared to SMS reminders not stating appointment costs [ 34 ]. Shared appointments, where patients receive consultations with their doctor in the presence of other patients with similar conditions, may provide another means of reducing unkept appointment rates [ 35 ]. Such interventions would have to be trialled, and an assessment of utility, safety, cost-effectiveness, and patient satisfaction would have to be undertaken.
Whilst text message reminders have been shown to reduce unkept appointment rates, patients are still missing their appointments, and hence the present findings will allow us to go a step further by introducing targeted interventions. In addition, vulnerable, elderly, and deprived patients may not have access to a mobile phone, and therefore would not benefit from a blanket approach using SMS reminders. They may also be most at harm, should they miss their appointment, which again calls for a more targeted approach to ensure they receive the appropriate care.
Aside from the cost implications of unkept appointments, there is increased mortality associated with missing appointments, as seen in general practice. A nationwide study in Scotland reported that those who missed more than 2 appointments had a 3-fold increase in hazards of mortality compared to those who did not miss appointments [ 6 ]. Such data are lacking in the secondary care setting.
The GBM models output unkept appointment propensity scores, helping us rank patients in order of which patients are most likely to miss their appointments. The idea is to target a certain proportion of patients, implement an intervention, and decrease unkept appointment rates with minimal effort, rather than targeting all patients. In this way, interventions can be introduced more cost-effectively. Using the 5 specialties with the highest rates of unkept appointments, the models suggest that so long as the cost of an intervention is less than one-third of the average cost of the potential reduction in unkept appointments, using these models for targeted interventions would theoretically be cost-effective. In the context of analysing many specialties, we used a uniform sensitivity cutoff of 10%. In a live service, we could have different cutoffs for different specialties, based on their resources and respective rates of unkept appointments.
Overall, the top 2 predictors—namely, recency and frequency of previous unkept appointments—accounted for 38% of the average predictive value. This highlighted how a simple intervention based on these 2 predictors might have some utility. But it also highlights one of the advantages of applying machine learning models to predictive problems such as this—namely, that the contribution of many factors, along with the complexities of their interactions, can be accounted for in a way that focusing on just a few key factors does not allow.
It is likely that in order to run and implement these models and then apply targeted interventions across the population, technology, and possibly artificial intelligence, will be utilised. In February 2019, the Topol review was published [ 36 ]. This review, commissioned by the UK secretary of health, was designed to elucidate how the NHS can make the most of technology to improve services and help ensure their sustainability [ 37 ]. Digital medicine and artificial intelligence can aid in decision-making processes such as booking systems and targeting interventions, as well as utilising the vast volume of data available to generate the models.
As with any study relying on the use of routinely collected data, there are a number of limitations. The submission of kept outpatient appointment data to NHS Digital is mandatory in England. However, the submission of unkept outpatient appointment data is not. Not all trusts report unkept appointments consistently, so the data here may not reflect the true unkept appointment rate. This too may be the case within specialties at a single trust. Hospitals reporting a 0% rate of unkept appointments were excluded from this analysis, along with the bottom and top 10% of outliers. Data with the most missingness would have been among the top 10%. In addition, the results by specialty here are based on a single trust in England; hence, the generalisability of the results across the whole country or other countries may be questioned. Furthermore, our data did not include mental health services, as mental health data are found in the Mental Health Services Data Set (MHSDS) rather than HES. In addition, ICHT does not have a dedicated psychiatry department and utilises liaison psychiatry services from partner London trusts.
We excluded the 10% of trusts with the highest rates of unkept appointments. Including them would have potentially resulted in the model underpredicting unkept appointments. However, we also excluded the 10% of outliers with the lowest rates of unkept appointments, and so there is likely to be some offsetting. Additionally, excluding records that were likely to have lower data quality would have improved the accuracy of the predictions. However, as a limitation, trusts thus excluded would be less represented in the data, so the model would again be less generalisable. Furthermore, in our analysis, we excluded cancellations as we did not have cancellation dates. This may have affected the applicability of the model. Appointments can be cancelled by patients or by the care provider, and cancellations may occur shortly before an appointment, or well in advance. In an ideal scenario, we would know when an intervention for reducing unkept appointments was applied, and filter cancellations accordingly. For example, if patients are called 3 days before an appointment, then we could include cancellations within 3 days of appointments as unkept appointments, while excluding all cancellations that had happened before the phone call reminder.
Nonetheless, the study highlights the importance of repeating such an exercise across other datasets. We would recommend digital health policy makers mandate trusts to record and submit unkept appointments, in addition to those attended, to avoid this issue for future related research.
This study has identified the prevalence of unkept appointments nationally, by trust and broken down by specialty within a single UK trust. The clinical implications are that those locations and specialties with the highest rates may require intervention. The granularity of the predictors allows us to identify which patients are best targeted to implement such interventions, to ensure a reduction in unkept appointments, to ultimately reduce the morbidity associated with them and the waste of resources. Further study is needed in other UK trusts and in other countries to better understand this issue globally, so as to tackle healthcare inequalities. Understanding the complications of unkept appointments in a secondary care setting would also be pertinent as it may aid in clinical decision making and follow-up planning.
The new methods of modelling unkept appointments introduced in this study allow us to have a deeper understanding of the root causes of unkept appointments at the national and local level and may offer a path to offer novel interventions in order to address these causes. Future work should break down these unkept appointments and their causes into clinical, behavioural, and psychosocial domains so that specific targets in these areas can be generated to minimise unkept appointment loads. These approaches will need validation with other datasets and in formalised clinical trial settings to address the global issue of unkept appointments at the national and local level. The lessons derived from these approaches to unkept appointments may therefore in turn be a route to increase efficacy and efficiency in an era of healthcare rationing and financial constraint.
Supporting information
S1 Table
RECORD checklist.
(DOCX)
S2 Table
Unkept appointment rates by UK hospital trust.
(DOCX)
S3 Table
Unkept appointment rates and model metrics for ICHT in 2017–2018 by specialty.
(DOCX)
S4 Table
Predictor importance for each specialty.
Filtered for specialties with at least 10,000 appointments in 2016–2017 (after data cleaning).
(DOCX)
S1 Text
Model specification.
Fig A: Risk ratio of non-attendance for patients with an unkept appointment in the past 12 months.
(DOCX)
|
Vytvor v slovenčine zhrnutie totho článku. Laicky vysvetli všetky základné pojmy v článku. Ak používaš úvodzovky, nezabudni, že v slovenčine sa prvé úvodzovky píšu dole. Využi 8000 až 8100 tokenov. Tu je text článku: |
Introduction
Under-five mortality remains a major public health issue in sub-Saharan Africa, despite a remarkable decline during the Millennium Development Goal (MDG) era from 2000 to 2015 [ 1 ]. The under-5 mortality rate (U5MR) estimates in 2016 suggest that one in twelve children of sub-Saharan Africa did not reach their fifth birthday [ 2 ]. Pneumonia, preterm birth complications, intrapartum events, and diarrhoea constitute the main causes of under-5 deaths in sub-Saharan Africa [ 3 ]. Indirect factors related to child, maternal, family, community, and the environment are also strongly associated with under-5 mortality, and hence underlie theses direct causes [ 3 – 5 ]. Most of the direct and indirect causes are preventable. During the MDG era, facilitated by the commitment of donors, local governments, and other stakeholders cost-effective interventions were scaled up. The effects of the interventions show considerable spatial heterogeneity.
Since 2006, Burkina Faso scaled up child health interventions consisting of subsidy of deliveries, artemisinin-based combination therapies (ACTs), rapid diagnostic tests (RDTs) for malaria at health facility and community levels, and universal distribution of long lasting insecticidal nets (LLINs). Furthermore, new vaccines have been included in the Expanded Programme of Immunization (EPI). The U5MR declined from 146.9 per 1000 to 88.6 per 1000 between 1998 and 2015 [ 6 ]. Likewise in other Sub-Saharan Africa countries the reduction was heterogeneous between and within the different administrative regions of the country [ 7 – 11 ]. For example, the DHS 2010 data showed that the U5MR in the region of Sahel was four times higher than that of Centre-Est [ 6 ]. In the rural district of Nouna, the U5MR declined by more than 50% during the last 25 years, however strong disparities in mortality have been observed within the district, which are likely to be related to disparities in risk factors and coverage of health interventions [ 10 ]. Even though the coverage of child and maternal interventions has increased globally at national level, there are large regional disparities within the country. For example, household ownership of at least one LLIN was more than 90% in the regions of Plateau Central and Nord, while it was only 41% in the Centre-Nord region [ 6 ]. The breastfeeding rate within one hour after birth varied from 27% to 67%. Regional variation also exists in the skilled antenatal and birth attendance across the country. Results of the Malaria Indicator Survey (MIS) in 2014 showed that in the region of Centre and Centre-Ouest, around 2% of children received an ACT on the same or the next day after the onset of fever, while in the region of Cascades, ACT coverage was much higher (42%) [ 12 ]. Moreover, data from routine health management information system confirm the same heterogeneity observed in the MIS data [ 12 ]. The variability in the coverage of health interventions is related to external factors including deficiencies in the health system, which affect their effectiveness. In Burkina Faso, very few studies have assessed the association between health interventions and U5MR at subnational level. The few available studies assessed a subset of interventions [ 13 ].
Our aim was to assess the magnitude of association between child, maternal and household health interventions and under-five mortality at national and sub-national levels. We hypothesize that there is a geographical variation in the effects of health interventions and we aim to identify the interventions and the regions where there is a statistically important association between intervention coverage and child mortality. Our data will support decision making for delivering the most effective interventions in those regions where the highest rate were predicted.
Methods
Study area
Burkina Faso is a country in West Africa, with an estimated population of 18.5 millions inhabitants in 2015. Around 40.1% of the population lives below the poverty threshold and the Human Development Index (HDI) is 0.402. The population is relatively young with 21% and 54% below 5 and 18 years, respectively. The country is part of the Sudanian zone with a dry tropical climate and two seasons: a dry season from November to June characterised by a peak of respiratory diseases and a wet season from July to October with malaria as the most important communicable disease.
Data source
U5MR and health intervention data were extracted from the Burkina Faso Demographic and Health Survey (DHS) carried out in 2010. The data were collected from a two-stage cluster design and are representative at the national level, for urban and rural areas and for the 13 administrative regions. The survey was carried out in 574 georeferenced clusters and included 17 087 women of reproductive age who provided information for 15 375 lives births in the five years before the survey.
We extracted information of selected key health interventions of the countdown to 2015 initiative with less than 15% of missing values [ 14 ]. In particular, we included the following child-related health interventions: all antigens, measles, and DPT3 immunization, vitamin A supplementation, use of long lasting insecticidal nets (LLINs), malaria treatment by any anti-malarial, exclusive breastfeeding, immediate breastfeeding after birth, and baby post-natal check. We considered maternal interventions, such as skilled birth and antenatal care, post-natal check, family planning and intermittent preventive treatment of malaria during pregnancy (IPTp). The household specific interventions included in the study were improved drinking water source and sanitation, wealth index, and ownership of LLIN. We include malaria interventions (sleeping under ITN, household ownership of ITNs and malaria treatment) in the list of child health intervention because children under-5 years old are at higher risk of morbidity and mortality in malaria endemic countries such as Burkina Faso. Data measuring coverage of health interventions were aggregated at regional level. A description of the health intervention coverage indicators used in this study is given in S1 Table . Furthermore, we extracted information on socio-demographic characteristics of mothers and children under-5years of age such as birth order, sex, and place of delivery, mother’s age at first birth, her literacy level, as well as the number of live births.
Environmental and climatic factors, such as land surface temperature (LST), vegetation indices (Enhanced Vegetation Index (EVI) and Normalized Difference Vegetation Index (NDVI)), distance to water bodies, and type of land cover were compiled from satellite sources. Day and night LST and rainfall data were averaged for year 2010. Permanent water bodies were obtained from the land cover category. Details on the source of climatic data and their spatio-temporal resolution are given in S2 Table . The shapefile of Burkina Faso was extracted from the “The Humanitarian Data Exchange” database at: https://data.humdata.org .
Statistical methods
Bayesian geostatistical proportional hazard models with a Weibull baseline hazard were fitted on child mortality data to assess the association between child, maternal, and household health interventions and U5MR. The models were adjusted for child, maternal, socio-demographic characteristics, climatic and environmental factors. Spatial correlation was introduced by a Gaussian process, adopted on cluster-specific random effects with an exponential correlation function of the distance between survey clusters. Spatially varying regression coefficients for the interventions were used to capture the geographical variation of the association at sub-national level and they were modelled by regional random effects with a conditional autoregressive prior distribution. That is, let s = { s 1 , s 2 , …, s m }, s i ∈ D ⊂ R 2 be the set of locations at which mortality data are observed, t j ( s i ) be the time to death or the censoring time (in months) for child j at location s i , X j ( s i ) be the vector of child, maternal, socio-demographic and climatic factors and Z ( s i ) be the coverage of a given intervention at location s i . We modeled the mortality hazard as, h ( t j ( s i )) = h 0 ( t j ( s i ))exp( β T X j ( s i ) + ( a + w q ( i ) ) Z ( s i ) + φ ( s i )) and assumed a Weibull baseline hazard i.e. h 0 ( t j ( s i )) = δ ( t j ( s i )) δ −1 where δ is the shape parameter, β T = ( β 1 , …, β p ) is the vector of regression coefficients with exp( β l ), l = 1, … p , corresponding to the hazard ratio (HR). φ ( s i ) is a cluster-specific random frailty which captures spatial correlation in mortality i.e. clusters in closer proximity are expected to have similar mortality hazard due to common exposures. We modeled φ ( s ) = ( φ ( s 1 ), φ ( s 2 ), …, φ ( s m )) T by a Gaussian process, i.e. φ ( s )~ N (0, σ 2 R ), with an exponential correlation function of the distance d kl between locations s k and s l , that is R kl = exp(− d kl ρ ). The parameter σ 2 corresponds to the variance of the spatial process and ρ controls the rate of correlation decay with distance. For the exponential correlation function, - l o g ( 0 . 05 ) ρ determines the distance at which the correlation drops to 0.05 (i.e. effective range of spatial process). The geographical variation in the association between interventions and U5MR was modelled by the spatially varying coefficients, a + w q ( i ) where a quantifies the magnitude of the association at global (national) level and w = ( w 1 , …, w Q ) T are the varying effects at regional (sub-national) levels q = 1, … Q with q ( i ) indicating the region q corresponding to the location s i . We introduced spatial dependence among the regions via a conditional autoregressive (CAR) prior for w , that is w ~ N ( 0 , σ q 2 R q ) with R q = ( I - γ C ) - 1 D . σ q 2 is the variance of the spatially varying intervention coefficients, D is a diagonal matrix with entries D kk = g k −1 where g k is the number of neighbors of region k , γ measures overall spatial dependence and C is a proximity matrix with normalized entries that is C kl = ω kl / g k , ω kl is 1 if region k neighbors l and 0 otherwise [ 15 ]. To complete Bayesian model formulation, we assumed inverse gamma priors for all spatial variances with known parameters, i.e. σ 2 , σ q 2 ~ I G ( 2 . 01 , 1 . 01 ) , a uniform prior distribution for ρ ~ U ( a , b ), where a and b chosen such as the effective range is within the maximum and minimum distances of the observed locations and a uniform prior for γ ~ U ( λ 1 - 1 , λ 2 - 1 ) , where λ 1 , λ 2 are the smallest and largest eigenvalue of D - 1 2 C D 1 2 The shape parameter was assigned an exponential prior δ ~ exp (0.01). Non-informative normal priors were adopted for the regression coefficients β l , a ~ N (0, 10 3 ) for l = 1, … p .
Model parameters were estimated using Markov Chain Monte Carlo (MCMC) simulation. We run a two chain algorithm for 300 000 iterations with an initial burn-in of 15,000 iterations. Convergence was assessed by the Gelman and Rubin diagnostic [ 16 ].
Prior to Bayesian spatial analysis, bivariate, non-spatial, Weibull proportional hazards models were fitted to identify potential child, maternal and socio-economic confounders. Variables with p-value less than 0.15 were included in the geostatistical model.
The statistical analyses were carried out in STATA version 14 (StataCorp.; College Station, TX, USA) and OpenBUGS version 3.2.3 (Imperial College and Medical Research Council; London, UK). Maps were produced in ArcGIS version 10.2.1 (Esri Inc.; Redlands, CA) and graphs in R (R Core Team; Vienna, Austria).
Ethical approval
We used secondary data that was made available by the MEASURE Demographic Health Survey (DHS) Program based in the United States of America. According to the survey report [ 6 ], ethical approval was obtained by the institutional review board of ICF of Calverton, Maryland, USA and the national ethics committee for health research of Burkina Faso under deliberation N°2014-7-072. The survey was anonymous. Blood samples were taken from all eligible children for whom parents or responsible adults had given their informed consent.
Results
Our sample included 541 (94.3%) clusters and 13 505 (87.8%) children under the age of five years, after removing clusters with missing coordinates. In total, 1209 (9%) children died before their fifth birthday owing to an estimate for U5MR of 128 per 1000. The geographical distribution of U5MR is shown in Fig 1 . The highest U5MRs were observed in the regions of Est, Sahel, and Sud-Ouest with respective U5MR of 172, 197, and 223 per 1000. The Centre-Est had the lowest rate of 81 per 1000.
10.1371/journal.pone.0218163.g001
Fig 1
Regional distribution U5MR in Burkina Faso based on DHS 2010.
Approximately two-third of the children were born in health facilities and 86% lived in rural areas. About 5% of mothers were younger than 19 years of age, two-third gave their first birth before age 19 years and 84% were not educated. Around one quarter of the respondents have more than five children and about 44% of the households were relatively poor (i.e. household asset in the first two quintiles). The socio-demographic characteristics of the sample are summarised in Table 1 .
10.1371/journal.pone.0218163.t001
Table 1 Child, maternal and household characteristics and hazard rate ratios estimated by bivariate Weibull proportional hazards models.
Covariate
Percentage (%) N = 13 505
Number of death (%)
Hazard rate ratio (95%CI)
P value
Children characteristics
Sex
Female
49.1
566 (8.5)
1.00
Male
50.9
643 (9.4)
1.12 (0.99–1.25)
0.063
Place of residence
Urban
13.6
104 (5.7)
1.00
Rural
86.4
1 105 (9.5)
1.71 (1.40–2.10) *
<0.001
Place of delivery
Health facility
66.3
650 (7.3)
1.00
Home
32.7
559 (12.3)
1.57 (1.40–1.76) *
<0.001
Birth order
1–5
75.3
863 (8.5)
1.00
>5
24.7
346 (10.4)
1.26 (1.11–1.43) *
<0.001
Mothers characteristics
Age group (years)
< 19
4.6
76 (12.3)
1.00
20–35
75.3
870 (8.6)
0.48 (0.38–0.61) *
<0.001
>35
20.1
262 (9.7)
0.51 (0.40–0.69) *
<0.001
Age at first birth
≤19
66.01
850 (9.5)
1.00
>19
34.0
359 (7.8)
1.22 (1.07–1.39) *
0.001
Number of live birth
1–5
71.0
750 (7.8)
1.00
>5
29.0
459 (11.7)
1.47 (1.31–1.65) *
<0.001
Mother education level
Primary and above
14.0
104 (5.5)
1.00
No education
86.0
1 11 (9.5)
1.72 (1.41–2.11) *
<0.001
Households characteristics
Asset index
Richest
33.7
301 (6.6)
1.00
Middle
22.5
280 (9.2)
1.41 (1.20–1.67) *
<0.001
Poorer
43.8
628 (10.6)
1.62 (1.42–1.86) *
<0.001
*: Statistically significant association (i.e.P-value<5%)
Coverage estimates of the child, maternal, and household health interventions used in the study, stratified by region are given in Table 2 . The corresponding intervention coverage maps are shown in S1 and S2 Figs.
10.1371/journal.pone.0218163.t002
Table 2 U5MR and coverage of child, maternal and household health interventions stratified by region, as assessed by the Burkina Faso DHS 2010.
Health Interventions (%)
Administrative regions
Boucle du Mouhoun
Cascades
Centre
Centre-Est
Centre-Nord
Centre-Ouest
Centre-Sud
Est
Hauts Basins
Nord
Plateau Central
Sahel
Sud-Ouest
National level
Child interventions
Use of ITNs by under-5 years old
42.6
49.8
33.1
35.9
32.0
46.6
38.2
46.0
36.2
65.7
73.5
37.3
44.7
43.4
Malaria treatment
22.5
19.2
36.1
47.9
23.1
36.8
45.7
42.2
42.0
42.0
40.5
20.1
31.7
35.7
Exclusive breastfeeding
6.7
10.1
2.4
8.8
8.6
6.8
7.3
9.0
7.3
6.0
7.3
5.2
5.9
7.0
Breastfeeding after birth within 24 h
36.4
45.6
55.3
26.6
36.1
29.7
45.4
55.4
38.6
46.8
66.9
35.4
37.1
41.4
Baby post-natal check within 24 hours
15.0
18.1
15.2
31.8
20.7
19.5
27.7
30.8
4.5
21.5
19.7
5.4
9.0
18.3
Measles immunization
90.3
90.5
96.8
95.2
95.6
85.9
94.9
75.3
88.5
91.7
95.3
79.2
85.4
88.5
DPT3 immunization
97.0
79.2
93.7
98.1
97.6
91.6
96.2
83.0
87.8
93.6
96.1
82.0
92.0
91.3
All-antigen * immunization
86.9
71.7
88.0
93.0
94.9
82.8
92.4
69.6
81.1
88.8
90.1
74.0
81.0
83.7
Vitamin A
75.7
49.2
59.8
80.6
83.4
41.2
89.2
46.0
64.1
72.4
70.8
36.5
65.4
63.6
Maternal interventions
Skilled birth attendance
64.5
77.5
96.4
84.6
73.5
60.7
86.4
55.4
73.1
61.4
80.8
40.1
42.9
67.1
Antenatal visits
93.4
95.5
98.7
99.6
97.1
95.2
99.3
92.4
96.0
94.8
98.7
88.3
92.5
95.1
Family planning
12.9
16.5
32.2
8.4
8.7
8.8
16.8
12.0
28.3
9.0
13.9
7.8
10.0
14.0
Intermittent preventive treatment in pregnancy
36.4
43.8
32.3
52.2
46.9
47.2
51.3
32.3
29.6
40.0
53.2
19.9
58.1
39.2
Household’s interventions
Improved sanitation
24.3
33.2
76.9
15.0
22.4
16.8
11.1
5.1
27.5
25.7
39.7
9.2
8.2
22.5
Improved drinking water
62.5
89.6
94.6
87.3
86.9
65.0
83.7
65.2
76.4
64.3
93.4
61.6
47.6
73.2
Household ownership of ITNs
73.3
81.5
72.7
62.7
52.8
74.0
62.8
83.2
63.3
98.1
95.4
81.5
71.2
74.3
U5MR (per 1000)
117
106
85
81
93
101
110
172
101
153
151
197
223
128
*: include BCG, Polio3, DPT3, and measles immunization
Among the child health interventions, those related to immunization had coverage above 80%, the required level for the universal health coverage [ 17 ]. Baby post-natal check 24 hours after birth and exclusive breastfeeding had the weakest coverage; 18% and 7%, respectively.
The coverage of maternal health interventions related to family planning, intermittent preventive treatment of malaria during pregnancy, and antenatal care visit were 14%, 39% and 95%, respectively. Household-based interventions for safe drinking water and ownership of at least one insecticide-treated nets (ITNs) were covering each around 73%. The proportion of households in the country with access to improved sanitation was low; around 20%. The distribution of the health interventions within the 13 regions showed strong heterogeneity. In general, the Sahel, Sud-Ouest, Est and the Centre-Est were the regions with the lowest coverage of most interventions. Geographical disparities were observed in socio-economic proxies, such as the household asset index and the mother’s education level. Fig 2 indicates that the wealthiest and the most educated tended to have high coverage of child health interventions.
10.1371/journal.pone.0218163.g002
Fig 2
Frequency distribution of the coverage of child health intervention by household asset index (A) and by mother’s education level (B) in Burkina Faso based on DHS 2010.
Results of the bivariate survival analysis in Table 1 indicate that all child, maternal, and household socio-demographic covariates (except sex) were associated with child survival. Tables 3 and 4 show the hazard rate ratios (HRR) of child, mother, and household-specific interventions, estimated by Bayesian geostatistical models, adjusted for socio-economic and climatic covariates. HRR are also provided graphically in Fig 3 and as maps in S3 and S4 Figs.
10.1371/journal.pone.0218163.t003
Table 3 Estimates (posterior median and 95% Bayesian credible intervals) of the association between child health interventions at national and sub-national levels and U5MR obtained by Bayesian geostatistical Weibull proportional hazards models with spatially varying regression coefficients for the intervention coverage covariates.
All antigen Immunization
DPT3 immunization
Measles immunization
ITN use by under five
Malaria treatment
Breastfeeding within 24 hours
Exclusive breastfeeding
Baby post-natal check within 24 hours
Vitamin A
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Geographical scale
National
0·92 (0·87–0·96) *
0·89 (0·86–0·98) *
0·91 (0·89–0·95) *
0·95 (0·90–0·97) *
1·02 (0·93–1·05)
1·07 (0·98–1·16)
0·90 (0·81–0·93) *
0·89 (0·86–0·92) *
0·94 (0·89–0·98) *
Regions
Boucle du Mouhoun
0·99 (0·81–1·10)
0·97 (0·83–1·22)
0·87 (0·82–1·00)
0·99 (0·89–1·17)
1·04 (0·83–1·34)
1·14 (0·85–1·48)
0·99 (0·81–1·16)
0·79 (0·65–1·03)
1·01 (0·81–1·03)
Cascades
0·97 (0·78–1·40)
1·11 (0·87–1·28)
0·91 (0·73–0·92) *
0·87 (0·61–1·00)
1·17 (0·91–1·49)
1·34 (0·93–1·93)
0·85 (0·63–1·00)
0·82 (0·71–0·87) *
0·92 (0·82–0·96) *
Centre
1·09 (0·84–1·50)
0·93 (0·78–1·11)
1·09 (0·80–1·29)
0·99 (0·89–1·03)
1·04 (0·83–1·22)
1·17 (0·88–1·60)
1·05 (0·79–1·34)
0·98 (0·71–1·18)
1·00 (0·69–1·06)
Centre-Est
0·74 (0·64–0·90) *
0·72 (0·61–0·82) *
0·81 (0·72–0·89) *
0·84 (0·75–0·99) *
1·04(0·86–1·32)
1·25 (0·89–1·76)
0·83 (0·68–0·93) *
0·86 (0·73–0·99) *
0·94 (0·72–0·96) *
Centre-Nord
0·85 (0·69–1·04)
0·79 (0·61–0·95) *
1·09 (0·76–1·11)
1·17 (0·84–1·31)
1·15 (0·85–1·35)
1·11 (0·81–1·55)
0·81 (0·62–0·88) *
0·87 (0·76–1·06)
0·96 (0·67–1·06)
Centre-Ouest
0·82 (0·68–0·97) *
0·82 (0·69–0·95) *
1·14 (0·79–1·15)
1·13 (0·91–1·14)
1·11 (0·72–1·20)
0·93 (0·66–1·27)
0·99 (0·74–1·20)
1·21 (0·99–1·36)
1·11 (0·96–1·14)
Centre-Sud
1·10 (0·80–1·60)
1·00 (0·73–1·33)
1·11 (0·86–1·17)
1·30 (0·89–1·40)
0·96 (0·70–1·28)
1·02 (0·71–1·46)
0·73 (0·55–0·90) *
1·03 (0·82–1·19)
1·21 (0·95–1·27)
Est
0·97 (0·84–1·05)
0·83 (0·71–1·01)
0·95 (0·81–1·05)
0·88 (0·73–0·93) *
1·16 (0·84–1·34)
0·93 (0·72–1·21)
0·95 (0·73–1·11)
0·81 (0·69–1·38)
1·16 (0·97–1·19)
Hauts Bassins
1·10 (0·78–1·21)
1·05 (0·95–1·54)
0·98 (0·80–1·10)
1·36 (0·94–1·37)
1·11 (0·86–1·26
1·10 (0·75–1·62)
1·25 (0·96–1·63)
1·14 (0·84–1·21)
1·06 (0·82–1·30)
Nord
0·96 (0·74–1·19)
0·99 (0·81–1·24)
1·06 (0·80–1·12)
0·86 (0·68–0·89)
0·89 (0·72–0·96) *
1·05 (0·82–1·38)
0·85 (0·73–1·12)
0·97 (0·85–1·31)
0·85 (0·74–0·93) *
Plateau Central
0·91 (0·60–0·97) *
0·83 (0·65–1·00)
0·92 (0·80–1·09)
1·08 (0·77–1·16)
1·04 (0·78–1·12)
0·98 (0·73–1·30)
0·82 (0·63–1·01)
0·96 (0·70–1·04)
1·11 (0·87–1·17)
Sahel
0·96 (0·80–1·15)
0·94 (0·76–1·12)
0·93 (0·85–1·05)
0·78 (0·62–0·81) *
1·30 (1·04–1·36)
1·14 (0·81–1·50)
0·89 (0·64–0·99) *
0·70 (0·47–0·86) *
0·82 (0·67–0·85) *
Sud-Ouest
0·80 (0·66–0·90) *
0·96 (0·82–1·14)
0·74 (0·68–0·76) *
1·00 (0·80–1·26)
0·72 (0·48–0·91) *
0·93 (0·70–1·24)
1·04 (0·70–1·20)
0·61 (0·55–0·95) *
1·00 (0·76–1·14)
Spatial parameters
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Range (km)
17·7 (15·0–59·6)
34·5 (14·3–39·3)
27·6 (19·7–37·0)
37·6 (29·0–47·4)
49·3 (22·6–71·4)
56·0 (8·2–69·5)
39·0 (20·1–64·7)
29·9 (9v6-66·0)
36·4 (8·2–42·5)
Spatial variance
0·37 (0·32–0·41)
0·16 (0·14–0·23)
0·16 (0·12–0·19)
0·15 (0·13–0·23)
0·16 (0·12–0·23)
0·19 (0v11-0·33)
0·13 (0·10–0·16)
0·15 (0·13–0·17)
0·16 (0·12–0·34)
Variance of spatially varying effect
0·56 (0·36–0·63)
0·17 (0·11–0·20)
0·27 (0·15–0·47)
0·34 (0·20–0·48)
0·33 (0·25–0·57)
0·23 (0·11–0·56)
0·19 (0·12–0·41)
0·31 (0·15–0·37)
0·14 (0·11–0·29)
* Statistically important association.
10.1371/journal.pone.0218163.t004
Table 4 Estimates (posterior median and 95% Bayesian credible intervals) of the association between maternal and household health interventions at national and sub-national levels and U5MR obtained by Bayesian geostatistical Weibull proportional hazards models with spatially varying regression coefficients for the intervention coverage covariates.
Skill birth attendance
Antenatal visit
Family planning
IPT
Improved drink water
Improved sanitation
Household ownership of nets
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Hazard rate ratio (95% BCI)
Geographical scale
National
0·93 (0·88–0·96) *
0·95 (0·92–0·98) *
0·91 (0·85–0·94) *
1·01 (0·94–1·08)
1·01 (0·90–1·03)
1·00 (0·96–1·04)
1·09 (0·95–1·13)
Regions
Boucle du Mouhoun
0·95 (0·72–1·02)
1·02 (0·82–1·09)
0·94 (0·71–1·06)
1·11 (0·88–1·48)
0·92 (0·82–1·18)
0·90 (0·69–0·97) *
1·33 (0·87–1·75)
Cascades
1·03 (0·81–1·29)
1·11 (0·96–1·42))
1·01 (0·85–1·22)
0·98 (0·79–1·24)
1·31 (0·85–1·40)
1·18 (0·95–1·21)
0·87 (0·67–1·00)
Centre
1·03 (0·85–1·33)
0·86 (0·58–1·19)
0·89 (0·71–0·96) *
1·01 (0·81–1·51)
0·96 (0·72–1·05)
1·11 (0·88–1·22)
1·44 (0·73–1·77)
Centre-Est
0·81 (0·60–0·88) *
0·68 (0·62–0·90) *
1·17 (0·84–1·55)
0·80 (0·69–0·96) *
0·73 (0·62–0·99) *
1·08 (0·77–1·11)
1·15 (0·84–1·24)
Centre-Nord
1·08 (0·90–1·14)
1·01 (0·88–1·13)
0·90 (0·75–1·19)
0·96 (0·66–1·20)
0·96 (0·71–0·99) *
1·17 (0·86–1·27)
1·27 (0·97–1·47)
Centre-Ouest
1·11 (0·88–1·25)
0·99 (0·82–1·29)
0·92 (0·75–0·96)
1·04 (0·79–1·23)
0·86 (0·59–0·91) *
1·19 (0·87–1·25)
0·98 (0·80–1·16)
Centre-Sud
1·12 (0·91–1·37)
1·15 (0·70–1·47)
0·88 (0·60–1·01)
1·09 (0·89–1·57)
1·23 (0·95–1·41)
1·16 (0·89–1·31)
1·43 (0·96–1·75
Est
0·79 (0·66–0·91) *
0·76 (0·71–0·83) *
1·09 (0·88–1·19)
0·93 (0·69–1·06)
1·05 (0·84–1·17)
1·12 (0·91–1·32)
1·05 (0·75–1·22)
Hauts Bassins
1·07 (0·83–1·18)
1·19 (0·92–1·73)
0·89 (0·68–1·23)
1·17 (0·89–1·56)
1·12 (0·90–1·15)
1·04 (0·88–1·11)
1·11 (0·78–1·32)
Nord
0·85 (0·74–0·94) *
1·04 (0·82–1·09)
1·10 (0·83–1·33)
1·23 (0·91–1·52)
0·98 (0·79–1·03)
0·83 (0·71–0·98) *
0·96 (0·79–1·10)
Plateau Central
1·02 (0·79–1·17)
1·01 (0·79–1·39)
0·73 (0·58–0·87) *
1·02 (0·71–1·35)
0·92 (0·71–0·97) *
1·29 (0·96–1·41)
1·12 (0·82–1·21)
Sahel
0·80 (0·73–0·89) *
0·89 (0·79–0·97) *
1·12 (0·72–1·35)
0·75 (0·57–0·97) *
0·91 (0·74–0·97) *
0·89 (0·80–0·98) *
1·13 (0·79–1·28)
Sud-Ouest
0·73 (0·58–0·85) *
0·90 (0·82 0·95) *
0·70 (0·50–0·75) *
1·16 (0·83–1·71)
1·73 (1·35–3·20)
0·88 (0·75–1·15)
0·98 (0·66–1·04)
Spatial parameters
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Median (95% BCI)
Range (km)
21·0 (38·1–53·5)
31·2 (23·9–50·3)
18·3 (30·0–68·4)
53·3 (42·2–71·1)
58·9 (33·9–68·9)
37·4 (15·4–70·1)
37·7 (12·4–69·6)
Spatial variance
0·13 (0·13–0·16)
0·15 (0·13–0·20)
0·17 (0·12–0·23)
0·16 (0·11–0·19)
0·15 (0·11–0·18)
0·14 (0·10–0·17)
0·16 (0·11–0·22)
Variance of spatially varying effect
0·18 (0·11–0·43)
0·23 (0·16–0·37)
0·21 (0·18–0·98)
0·22 (0·11–0·45)
0·33 (0·22–0·53)
0·19 (0·15–0·40)
0·28 (0·13–0·39)
* Statistically important association.
10.1371/journal.pone.0218163.g003
Fig 3
Hazard rate ratios (posterior median and 95% BCI) of child (A) maternal and household (B) health interventions estimated by Bayesian geostatistical Weibull proportional hazards models with spatially varying regression coefficients for the intervention coverage covariates.
DHS 2010, Burkina Faso. The horizontal line corresponds to a HRR equal to one.
Breastfeeding within 24 hours after birth and household ownership of ITNs did not have an important association neither at national nor at regional level. At national level, DPT3 immunization and baby post-natal check within 24 hours were associated with a reduction of U5MR (HRR = 0.89, 95% BCI: 0.86–0.98 and HRR = 0.89, 95% BCI: 0.86–0.92, respectively).
There was a considerable variation in the magnitude of the association between the interventions and U5MR within the country. The number of interventions by region associated with a reduction of U5MR ranged from zero to 11 with a median number of three. The region, with the highest number of statistically important interventions were Centre-Est, Sahel, and Sud-Ouest having 11, eight, and seven interventions, respectively. A second group of regions is composed by Cascades, Plateau Central, Nord, Centre-Ouest, Centre-Nord, and Est with three to four interventions associated with child survival. The remaining regions had at most one intervention with a statistically important regression coefficient. No important intervention was identified in the region of Hauts Bassins. The combination of interventions with statistically important coefficients varied by region.
Four out of 13 regions had child-specific interventions with protective effects: all-antigen immunization (Centre-Est, Centre-Ouest, Plateau Central, and Sud-Ouest), exclusive breastfeeding (Centre-Est, Centre-Nord, Centre-Sud, and Sahel), vitamin A supplementation (Cascades, Centre-Est, Nord, and Sahel) and baby post-natal check (Cascades, Centre-Est, Sahel, and Sud-Ouest). Use of ITNs (Centre-Est, Est, and Sahel), measles (Cascades, Centre-Est and Sud-Ouest) and DPT3 immunization (Centre-Est, Centre-Nord, and Centre-Ouest) were associated with U5MR in three regions. Intermittent preventive treatment of malaria in pregnant women was not associated with mortality hazard at national level; however, it was associated with mortality hazard in the regions of Centre-Est and Sahel. Among all child interventions at regional level, baby post-natal check showed the strongest negative association with U5MR in the Sud-Ouest region (HRR = 0.61, 95% BCI: 0.55–0.95). Skilled birth delivery, antenatal care attendance, and family planning are maternal interventions with statistically important coefficients at both national and regional levels. Antenatal care had the highest association with U5MR in Centre-Est (HRR = 0.68, 95% BCI: 0.62–0.90). At national level, none of the interventions related to household was associated with child survival. However, improved drinking water is associated with a reduction in U5MR in five regions (Centre-Est, Centre-Nord, Centre-Ouest, Plateau Central, and Sahel) while improved sanitation in three (Boucle du Mouhoun, Nord, and Sahel) ( Fig 3B ).
Malaria is one of the main causes of high U5MR in Burkina Faso. Use of ITNs and treatment with any antimalarial showed a protective effect on U5M in three (i.e. Centre-Est, Est, and Sahel) and two regions (i.e. Nord and, Sud-Ouest), respectively.
Association between U5M with socio-demographic characteristics remained the same across the health interventions. That is, male, born at home, birth order higher than five, first birth younger than 19 years, lack of education, mother’s age less than 19 or above 35 years, and poorer household were associated with increased mortality ( S3 and S4 Tables).
The models with spatially varying coefficients for IPT and improved drinking water have the highest estimates (i.e. posterior median) of their range parameters suggesting that the spatial correlation in their residuals extend over longer area compared to that in the rest of the models. This implies a weaker spatial correlation structure for the corresponding interventions. However, the credible intervals of parameter estimates across the different health interventions are overlapping; indicating that there are no statistically important differences in the spatial correlation of the intervention coverage indicators. The same result is also observed in the variance parameter of the Gaussian process estimated by the models.
Discussion
This is the first study to assess geographical variation in the association of child, maternal and household health interventions with child survival in Burkina Faso at regional level, taking into account socio-demographics characteristics and climatic disparities. The geographical distribution of the coverage of the health interventions showed considerable heterogeneity. Interventions with the highest coverage (> 80%) are those related to child immunization and antenatal care visits. Skilled birth attendance, improved drinking water, and vitamin A supplementation had coverage of 60–80% at country level. The promotion of the above mentioned health interventions have history of several decades. Conversely, interventions with national coverage of less than 40% are those whose implementation was strengthened only the year 2000. These include treatment of malaria with ACTs, exclusive breastfeeding, and early breastfeeding after childbirth.
In our analysis, the child and maternal socio-demographic factors associated with child survival (place of residency, place of delivery, number of live birth, mother’s education level, age at first birth, and age group) are similar to those reported by several authors [ 18 , 19 ]. Our findings showed that boys under-five years old have higher mortality hazard than girls as some studies have found in developing countries [ 20 ]. The mother’s characteristics (number of live birth, education level, age at first birth, and age group) are interrelated with the household socio-economic status. In African settings, the poorest exhibit highest child mortality because poverty influences their health seeking behaviour. Women from poorer households are most often less educated, less autonomous, and make less use of maternal and child care services [ 21 ]. This is highlighted in Fig 2 . Socio-economic status is associated with failure to complete immunization, which is an effective child intervention [ 22 ]. The Ministry of Health has addressed the financial barriers by subsidizing childbirth and new-born care in 2006 and removed completely the users’ fees for children under-5 years old in 2016. As a result, the use of health services by the mothers has been increased [ 23 ]. The high proportion of home delivery in Burkina Faso can be explained by the high proportions of poorest households and of women not educated or living in remote rural areas. As known, home delivery exposes to asphyxia and neonatal sepsis. Several studies reported that women delivering at young age are at higher risk of preterm birth and complicated delivery [ 24 , 25 ]. These are major causes of neonatal mortality which accounts for 44% of under-five mortality [ 3 , 25 ]. Multiple parity (more than five live births) is another risk factor and must be taken into account during antenatal care visit. Multiple parity leads to weak reconstitution of the mother’s nutritional stock, and hence children with higher birth order are prone to low birth weights which impact negatively their survival. Climatic and environmental factors are linked to child malnutrition, family’s income and the development of water-borne diseases in sub-Saharan Africa [ 26 , 27 ].
Burkina Faso has two main seasons; wet and dry. Malaria and water-borne diseases are prominent during the wet season. Acute respiratory infections and meningitis are more prevalent during the dry season. Low precipitation may be protective in the Sud-Ouest region of the country, which receives most rain, but might be associated with increased mortality in the dry north [ 26 ] Rainfall is protective in our analyses. We also found a positive association between distance to water bodies and mortality hazard. Similar findings have been reported with regard the spatial distribution of malaria risk in the country [ 28 ] A possible explanation could be that people living in close proximity to open surface water bodies are more aware about the risks and then protect themselves.
The interventions assessed in the current study are almost those whose implementation have been regularly monitored at country level to evaluate progress towards the attainment of MDGs [ 14 ] They are effective at national and subnational level except the household ownership of net and breastfeeding within 24 hours after birth which did not show any impact on child mortality hazard at subnational level.
The variation of the association between the health interventions and U5MR in Burkina Faso may be explained by factors related to health system performance (such as health workers density, quality of care, accessibility of health facilities, availability of drugs and supplies …), variations in the coverage of interventions and of the climatic and environmental differences across the country. Centre-Est, Sahel and Sud-Ouest are the regions with more than seven interventions with protective effect on child mortality. U5MR is low in Centre-Est but high in the Sahel and Sud-Ouest. Geographically, the three regions belong to the three climatic areas of the country. The Sahel is the driest region and is part of the Sahelian zone; the Centre-Est belongs to the intermediate Sudano-sahelian zone, while the Sud-Ouest belongs to the Sudanian part with the highest rainfall. Furthermore, in 2010, the Sahel and the Sud-Ouest were the poorest regions, while the Centre-Est was one of the richest. In Sahel and Sud-Ouest, even if the interventions are effective, their coverage is below the national average. The financial barrier might play a crucial role in the uptake of health services and care by the population. On the contrary in the wealthiest region of Centre-Est the population makes better use of the health programme.
Cascades, Plateau Central, Nord, Centre-Ouest, Centre-Nord, and Est have three to four interventions with protective effect and the U5MR in these regions is above 100 per 1000 (with exception of the region of Centre-Nord). These regions present weak coverage of various interventions: malaria treatment in Cascades, and Centre-Nord, immunization, skilled birth, and antenatal attendance, improved drinking water and sanitation in Est, family planning in Centre-Nord, and Nord. They belong to the Sudano-sahelian and Sudanian climatic areas. Furthermore, Est, Nord, Centre-Ouest, and Centre-Nord have high proportion of poor households. The health interventions with protective effects are mostly immunization, malaria-related interventions, vitamin A, baby post-natal check, skilled birth, and antenatal attendance. The finding in Plateau Central region is rather surprising. All interventions showed higher coverage than at national level; however the U5MR is among the highest in the country at around 151 per 1000. The interventions with important effects are family planning, improved drinking water and all antigen immunization. Plausible explanations for those results may be a long delay in health seeking behaviour and a low quality of health care in this region.
Centre, Boucle du Mouhoun, and Centre-Sud have only one intervention with protective effect, while there was no intervention associated with U5MR in Hauts-Bassins. The U5MRs in these regions are below the national rate of 128 per 1000. Centre and Hauts Bassins are the wealthiest regions of Burkina Faso with the highest availability of health infrastructures. In Centre region, family planning is the only intervention associated with a reduction in U5MRn. The mortality rate in Hauts-Bassins is 20 per 1000 higher than Centre, and the average coverage of health interventions is higher, however, no health intervention showed a statistically important association with U5MR. Emphasis should be put on ITNs, baby post-natal check within 24 hours, and IPT, which coverage are below the national average. Thus, the reinforcement of these interventions can impact positively child survival. Boucle du Mouhoun and Centre-Sud belong to regions with high proportion of households in the lower wealth index category. Only improved sanitation in Boucle du Mouhoun and exclusive breastfeeding in Centre-Sud are associated with the reduction of U5Mr. These regions have low coverage of improved drinking water (in Boucle du Mouhoun), use of ITNs and improved sanitation (in Centre-Sud).
Child mortality is also influenced by health system related factors, such as the density of health professional, availability of medical products, and quality of health care. It is interesting to note that Centre and Hauts Bassins that have only one or no intervention associated with child survival, respectively, have the highest density of health professionals in the country, although they cover about a quarter of total population [ 29 ]. Our results are in the same direction with previous studies that proved association of ACTs with malaria parasitemia at regional level in Centre-Est, Nord and Sahel [ 28 ]. Several authors have highlighted the protective effect of net ownership on child mortality, which, however could not be confirmed in the current study [ 30 ]. A possible explanation is that the DHS was carried out after a mass distribution of nets, but the mortality estimates are covering the 5-year period prior to the survey and therefore they are not influenced by the household ownership of nets. Breastfeeding within 24 hours after birth was a second health intervention not associated with U5MR. In the literature, early initiation of breastfeeding (less than one hour) is related to reduced neonatal mortality [ 31 ]. Our indicator included a delay of 24 hours which may explain the above finding.
A limitation of our study is that the coverage of health interventions is based on self-reporting information, therefore, the measurement may be prone to recall bias. Our analysis assumed that there is no systematic bias recall neither for a given health intervention nor for a given specific region. There are two more limitations which may have led to under-estimation of the association between the intervention coverage and U5MR. In particular, the mortality data are covering the 5-year period prior to the survey. During the latest years the coverage of interventions has increased. The analysis could not take into account the study period because there were no year specific data available. Data were also aggregated at regional level. The few regions in the analysis most likely contributed to a reduced geographical variation in the coverage of interventions. A better approach would have been to further disaggregate the regions according to rural/urban type.
Concluding, the most effective interventions related to U5MR at national level is the DPT3 immunization, followed by the baby post-natal check within 24 hours, exclusive breastfeeding, measles immunization, all antigen immunization, and vitamin A. Low coverage of DPT3 immunization was observed in Cascades, Plateau Central, and Est, of baby post-natal check in Cascades, Centre Est, and Sud-Ouest and of exclusive breastfeeding in Centre, Sud-Ouest, Boucle du Mouhoun, and Hauts Bassins. Skilled birth and antenatal care attendance are the most effective maternal interventions and should be reinforced in the regions of Sahel, Sud-Ouest, Nord, and Est. Child survival can be enhanced by increasing the coverage of improved drinking water in Sahel, Sud-Ouest, and Boucle du Mouhoun and the coverage of improved sanitation in Est, Centre-Sud, Sahel, and Sud-Ouest.
Supporting information
S1 Table
Description of the intervention coverage indicators used in the study.
(DOCX)
S2 Table
Climatic covariates, sources and spatial and temporal resolution.
Data were extracted during the year of 2010.
(DOCX)
S3 Table
Hazard rates ratio (posterior median and 95% Bayesian credible intervals) of child, maternal, household socio-demographic and climatic factors used to adjust the association between child health interventions and U5MR.
Estimates are obtained by Bayesian geostatistical Weibull proportional hazards model with spatially varying regression coefficients for the intervention coverage covariates.
(DOCX)
S4 Table
Hazard rates ratio (posterior median and 95% Bayesian credible intervals) of child, maternal, household socio-demographic and climatic factors used to adjust the association between maternal and household health interventions and U5MR.
Estimates are obtained by Bayesian geostatistical Weibull proportional hazards models with spatially varying regression coefficients for the intervention coverage covariates.
(DOCX)
S1 Fig
Geographical distribution of the coverage of child health interventions.
The coverage are based on quartile cut-offs: (A) all antigen immunization, (B) DPT3 immunization, (C) measles immunization, (D) malaria treatment, (E) ITN use, (F) baby post-natal check, (G) exclusive breastfeeding, (H) breastfeeding within 24 hours, (I) vitamin A supplementation.
(TIF)
S2 Fig
Geographical distribution of the coverage of maternal and household health interventions.
The coverage are based on quartile cut-offs: (A) skilled birth attendance, (B) skilled antenatal care, (C) family planning, (D) intermittent preventive treatment of malaria in pregnancy, (E) household ownership of bed nets, (F) improved sanitation, (G) improved drinking water.
(TIF)
S3 Fig
Spatially varying coefficients of child health interventions on U5MR.
Hazard rates ratio estimates (posterior median) obtained by Bayesian geostatistical Weibull proportional hazards model with spatially varying regression coefficients for the intervention coverage covariates. The distribution of the hazard rates ratio are based on quartile cut-offs: (A) all antigen immunization, (B) DPT3 immunization, (C) measles immunization, (D) malaria treatment, (E) ITN use, (F) baby post-natal check, (G) exclusive breastfeeding, (H) breastfeeding within 24 hours, (I) vitamin A supplementation.
(TIF)
S4 Fig
Spatially varying coefficients of maternal and household health interventions on U5MR.
Hazard rates ratio estimates (posterior median) obtained by Bayesian geostatistical Weibull proportional hazards model with spatially varying regression coefficients for the intervention coverage covariates. The distribution of the hazard rates ratio are based on quartile cut-offs: (A) skilled birth attendance, (B) skilled antenatal care, (C) family planning, (D) intermittent preventive treatment of malaria in pregnancy, (E) household ownership of bed nets, (F) improved sanitation, (G) improved drinking water.
(TIF)
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.