This is a repo containing some work I've done to investigate how deep learning models encode chemical structures. Pretraining dataset and SMILES string tokenizer vocabulary are taken from Aksamit et al. (2024).
Data preprocessing is done using the misc_scripts.py file. You'll need to provide a csv of SMILES strings. Aksamit et al. (2024) use a 140MB file of drug-like molecules for this.
Training is done using the training.py script. This script runs thte training loop for a set number of epochs.
Each SMILES string is converted into a
Aksamit, N., Tchagang, A., Li, Y. et al. Hybrid fragment-SMILES tokenization for ADMET prediction in drug discovery. BMC Bioinformatics 25, 255 (2024). https://doi.org/10.1186/s12859-024-05861-z