Datasets:
metadata
tags:
- chemistry
size_categories:
- 1M<n<10M
configs:
- config_name: default
data_files: allmolgen.tar.xz
Downloaded using PyTDC (https://tdcommons.ai/).
Contains the unique canonicalized SMILES molecules from MOSES, ZINC-250K, and ChEMBL-29, done with RDKit.
Distribution of tokenized SMILES sequence lengths below. The following regex string was used to split the SMILES molecule into tokens: ([[^]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|(|)|.|=|#|-|+|\|/|:|~|@|?|>>?|*|$|%[0-9]{2}|[0-9])
Included in the .csv (after extracting the .tar.xz file) is a column "smi_len". If using the same SMILES tokenization regex string as above, you can simply filter using the values in this column ("smi_len"). I'd recommend post-processing since clearly a majority of the sequences are of a much shorter length than the highest, which is above 1400 (using my regex string).