Update README.md
Browse files
README.md
CHANGED
@@ -30,4 +30,82 @@ configs:
|
|
30 |
path: data/valid-*
|
31 |
- split: test
|
32 |
path: data/test-*
|
|
|
|
|
33 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
path: data/valid-*
|
31 |
- split: test
|
32 |
path: data/test-*
|
33 |
+
size_categories:
|
34 |
+
- 1B<n<10B
|
35 |
---
|
36 |
+
# ZINC_22 Pretraining Dataset
|
37 |
+
|
38 |
+
## Dataset Description
|
39 |
+
This dataset is derived from the **ZINC-22** database (~70B synthesizable compounds as of Sept 2024) and was prepared for large-scale pretraining of molecular language models. We randomly sampled **1.5 billion molecules** using a **stratified heavy-atom count split** (4–49 atoms) to ensure coverage of diverse chemical sizes.
|
40 |
+
All molecules were **deduplicated** to remove repeats, **canonicalized** in SMILES format, and **converted** into multiple string representations: SMILES, SELFIES, SAFE, DeepSMILES.
|
41 |
+
|
42 |
+
---
|
43 |
+
|
44 |
+
## Precomputed Statistics
|
45 |
+
This repository includes precomputed reference statistics (`*_stats.pkl`) for evaluating generated molecules against validation and test sets.
|
46 |
+
These statistics are used to compute the following metrics:
|
47 |
+
|
48 |
+
- **FCD** – Fréchet ChemNet Distance
|
49 |
+
- **SNN** – Similarity to Nearest Neighbor
|
50 |
+
- **Frag** – Fragment similarity (BRICS decomposition)
|
51 |
+
- **Scaf** – Scaffold similarity (Bemis–Murcko scaffolds)
|
52 |
+
|
53 |
+
### File Naming Convention
|
54 |
+
Files are provided for multiple reference set sizes:
|
55 |
+
- `_175k` → 175,000 molecules
|
56 |
+
- `_500k` → 500,000 molecules
|
57 |
+
- `_1M` → 1 million molecules
|
58 |
+
- `_3M` → 3 million molecules
|
59 |
+
- *(no suffix)* → full set
|
60 |
+
|
61 |
+
By convention:
|
62 |
+
- `valid_stats_*` → computed from the **random validation split**
|
63 |
+
- `test_stats_*` → computed from the **scaffold-based split**
|
64 |
+
|
65 |
+
These statistics enable **consistent and reproducible** evaluation across experiments.
|
66 |
+
|
67 |
+
---
|
68 |
+
|
69 |
+
## How to Use
|
70 |
+
|
71 |
+
Before running the example below, make sure you have these packages installed:
|
72 |
+
```bash
|
73 |
+
pip install rdkit fcd-torch
|
74 |
+
```
|
75 |
+
### Example: Download stats from the Hub and compute FCD
|
76 |
+
|
77 |
+
```python
|
78 |
+
from huggingface_hub import hf_hub_download
|
79 |
+
import pickle
|
80 |
+
from fcd_torch import FCD as FCDMetric
|
81 |
+
|
82 |
+
# 1. Download the precomputed stats file from Hugging Face Hub
|
83 |
+
stats_path = hf_hub_download(
|
84 |
+
repo_id="chandar-lab/ZINC_22",
|
85 |
+
repo_type="dataset",
|
86 |
+
filename="valid_stats_175k.pkl" # change to desired file
|
87 |
+
)
|
88 |
+
|
89 |
+
# 2. Load the reference stats
|
90 |
+
with open(stats_path, "rb") as f:
|
91 |
+
reference_stats = pickle.load(f)
|
92 |
+
|
93 |
+
# 3. Compute FCD for your generated molecules
|
94 |
+
generated_smiles = ["CCO", "CCN", "CCCN", "CCCN"] # replace with your generated set
|
95 |
+
fcd_calculator = FCDMetric(batch_size=4)
|
96 |
+
|
97 |
+
fcd_value = fcd_calculator(gen=generated_smiles, pref=reference_stats["FCD"])
|
98 |
+
print(f"FCD score: {fcd_value:.4f}")
|
99 |
+
```
|
100 |
+
|
101 |
+
## Citation
|
102 |
+
|
103 |
+
```bibtex
|
104 |
+
@article{chitsaz2024novomolgen,
|
105 |
+
title={NovoMolGen: Rethinking Molecular Language Model Pretraining},
|
106 |
+
author={Chitsaz, Kamran and Balaji, Roshan and Fournier, Quentin and
|
107 |
+
Bhatt, Nirav Pravinbhai and Chandar, Sarath},
|
108 |
+
journal={arXiv preprint},
|
109 |
+
year={2025},
|
110 |
+
}
|
111 |
+
```
|