TransFoxMol

This is a Pytorch implementation of the following paper:

Jian Gao, Zheyuan Shen, Yufeng Xie, Jialiang Lu, Yang Lu, Sikang Chen, Qingyu Bian, Yue Guo, Liteng Shen, Jian Wu, Binbin Zhou, Tingjun Hou, Qiaojun He, Jinxin Che, Xiaowu Dong, TransFoxMol: predicting molecular property with focused attention, Briefings in Bioinformatics, Volume 24, Issue 5, September 2023, bbad306, https://doi.org/10.1093/bib/bbad306

The upgraded version of this model, KnoMol, is available at https://github.com/gaojianl/KnoMol

Installation

You can just execute following command to create the conda environment. ''' conda create --name TFM --file requirements.txt '''

Usage

1. Dataset preparation

Put your raw csvfile(DATASET_NAME.csv) in dataset/raw/.

python molnetdata.py --moldata DATASET_NAME --task clas --ncpu 10

This will save the processed dataset in dataset/processed/.

2. Hyper-parameter searching

python run.py search DATASET_NAME --task clas --device cuda:0

This will return the best hyper-params.

3. Training

python run.py train DATASET_NAME \
               --task reg \
               --device cuda:0 \
               --batch_size 32 \
               --train_epoch 50 \
               --lr 0.0005 \
               --valrate 0.1 \
               --testrate 0.1 \
               --seed 426 \
               --fold 3 \
               --dropout 0.05 \
               --attn_head 6 \
               --attn_layers 2 \
               --output_dim 256 \
               --D 4 \
               --metric rmse

This will save the resulting model in log/checkpoint/.

4. Testing

python run.py test DATASET_NAME \
               --task clas \
               --device cuda:0 \
               --batch_size 32 \
               --attn_head 6 \
               --attn_layers 2 \
               --output_dim 256 \
               --D 4 \
               --pretrain log/checkpoint/XXXX.pkl

This will load the model in log/checkpoint/ to make predictions and the results are saved in log/.