Multi-modal Protein-Drug Interaction Prediction Via Attention-based Network

A novel attention mechanism-driven deep model for estimating the binding affinity between protein and drug by exploiting multi-modal information including protein sequence, protein structure, and drug fingerprint.

• ✔ Our pre-trained encoder is at ./files/pretrain/BertModel.pth.
• ✔ Code of fine-tuning phase is released.

Table of Contents

• Multi-modal Protein-Drug Interaction Prediction Via Attention-based Network
  • Table of Contents
  • Environment & Dependencies
  • Usage
  • Addition

Environment & Dependencies

• Python==3.8.13

• torch==1.8.0

# create new environ
conda create -n pdi_fusionner python=3.8.13

# build dependencies
pip install -r requirements.txt
pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html

Usage

1. Data preparation

• Place the processed data file, namely pafnucy_total_rdkit-smiles-v1.csv, to ./files/data. Notice that this file combines both the protein datasets, i.e., PDBbind, CASF-2013 and Astex datasets, and the drug SMILES dataset. An entry of this file should look like:

  	PDB-ID	seq	SMILES	rdkit_smiles	Affinity-Value	set
  0	11gs	PYTVVYF...	CCC(=O)...	CC[C@@H]...	5.82	train

• Place the protein contact map that processed from 3D crystal structure data, namely pdbbind_2016, to ./files/3d_pdb

• Update the configurations config.yaml to make sure everything is matched

2. Fine-tuning

python finetune_main.py --model_idx=0 --epochs=150 --gpu_start=0 --batch_size=64 \
                        --patience=10 --factor=0.5 --min_lr=1e-5  --tb

Specify model_idx to choose our proposed model or its ablation variants. More options please refer to config.yaml and the argument parser in finetune_main.py.

Addition

More detailed information will be released and updated soon!