odb9402 / MoleculeTransformer

The BERT-based embedding model for SMILES molecule representation

Molecule transformer using BERT based model

The BERT-based embedding model for SMILES molecule representation from the paper "Self-Attention Based Molecule Representation for Predicting Drug-Target Interaction" written by Shin et al 2019. These sources are pytorch-implemented codes.

Usage

awk '{OFS="\t"; FS="\t"; print $2}' CID-SMILES > CID-SMILES.txt
python preprocess.py -i CID-SMILES.txt -o CID-SMILES_train.txt -v vocab.voc
python main.py -i CID-SMILES_train.txt -e 5 --lossWeight none -v vocab.voc

usage: main.py [-h] [-i INPUT] [-e EPOCHS] [--lossWeight {none,log,sqrt,raw}]

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT, --input INPUT
                        Input training SMILES file.
  -e EPOCHS, --epochs EPOCHS
  --lossWeight {none,log,sqrt,raw}
                        The type of class weights for the cross-entropy loss.

Final accuracy

• Epochs 1, without loss weights: 0.9471
• Epochs 3, without loss weights: 0.9554