Awesome

An effective self-supervised framework for learning expressive molecular global representations to drug discovery

This repository is the official implementation of Our Paper Published on BIB.

📋 We provide the code of MPG implementation for pre-training, and fine-tuning on downstream tasks including molecular properties, DDI, and DTI predictions.

Requirements

To install requirements:

pip install -r requirements.txt

Pre-training

To pre-train the model from scrach, please run the following command to preprocess the data by PSD and AttrMasking:

python pretraining/loader.py 

Then, run:

bash pretraining/run_pretraining.sh

📋 Here, we provide an input example of unlabelled molecular data at data/pretraining/raw/little_set_smi. To train the model on 11 M molecules thoroughly, please download the complete pre-training dataset first and put it in the data/pretraining/raw/ folder.

Fine-tune

📋 To finetune the model, please download the pretrained models first and put the model in the pretrained_model/ folder.

Results

To fine-tune on molecular propertied prediction, please run:

Classification tasks

python property/finetune.py --dataset toxcast --lr 0.0001 --input_model_file pretrained_model/MolGNet.pt

Note: If you want to run the fine-tune on CPU, please add an argument --cpu to the command above. There are more hyper-parameters which can be tuned during finetuning. Please refer to the arguments in finetune.py.

Regression tasks

python property/finetune_regress.py --dataset esol --lr 0.0001 --input_model_file pretrained_model/MolGNet.pt

To fine-tune on DDI prediction, please run:

BIOSNAP

python DDI/finetune_snap.py --input_model_file pretrained_model/MolGNet.pt

TWOSIDES

python DDI/finetune_twosides.py --input_model_file pretrained_model/MolGNet.pt

To fine-tune on DDI prediction, please run:

python CPI/cross_validate.py --dataset human --input_model_file pretrained_model/MolGNet.pt

Evaluation for reproductivity

Due to the non-deterministic behavior of the function index_select_nd(See link) and the randomless of dataset split, it is hard to exactly reproduce the training process of finetuning. Therefore, we provide the finetuned model and the splitted datasets for thirteen datasets to guarantee the reproducibility of the experiments. Note: these results are fine-tuned in different hardware environments, resulting in slightly difference from reported statistics in the paper.

Molecular property prediction

• BBBP
• SIDER
• ClinTox
• BACE
• Tox21
• ToxCast
• FreeSolv
• ESOL
• Lipo

DDI prediction

• TWOSIDES
• BIOSNAP

DTI prediction

• human
• elegans

We provide the eval.py function in property, DDI and DTI folders to reproduce the experiments. For example, to evaluate the performance on BBBP dataset, suppose the finetuned model is placed in finetuned_model/, please run:

python property/eval.py --dataset bbbp --model_dir finetuned_model/property

Citation

Please cite our paper if you find this code useful for your research:

@article{li2021effective,
  title={An effective self-supervised framework for learning expressive molecular global representations to drug discovery},
  author={Li, Pengyong and
          Wang, Jun and
          Qiao, Yixuan and
          Chen, Hao and
          Yu, Yihuan and
          Yao, Xiaojun and
          Gao, Peng and
          Xie, Guotong and
          Song, Sen},
  journal={Briefings in Bioinformatics},
  volume={22},
  number={6},
  pages={bbab109},
  year={2021},
  publisher={Oxford University Press}
}