A Multimodal Deep Learning Framework for Predicting PPI-Modulator Interactions

Introduction

This repository contains the PyTorch implementation of the MultiPPIMI framework, a multimodal deep learning approach with a bilinear attention network that explicitly models interactions between PPI targets and modulators.

Framework

Acknowledgements

This implementation is inspired and partially based on earlier works [1], [2].

Environments

Install packages under conda env

conda create -n MultiPPIMI python=3.7
conda activate MultiPPIMI

conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
conda install pytorch-geometric -c rusty1s -c conda-forge
pip install ogb==1.3.5
pip install rdkit

Data

The data folder contains all experimental data used in MultiPPIMI. In data/folds, we have full data split under four settings (S1: random split, S2: modulator cold-start, S3: PPI cold-start, S4: cold pair).

Reproduce results

For the warm-start experiments with MultiPPIMI, for one fold as an example, you can directly run the following command.

python main.py --fold 1 --eval_setting S1

For the cold-start experiments on one fold as an example.

python main.py --fold 1 --eval_setting S2
python main.py --fold 1 --eval_setting S3
python main.py --fold 1 --eval_setting S4

Cite:

• Heqi Sun, Jianmin Wang, Hongyan Wu, Shenggeng Lin, Junwei Chen, Jinghua Wei, Shuai Lv, Yi Xiong, Dong-Qing Wei. A Multimodal Deep Learning Framework for Predicting PPI-Modulator Interactions. J. Chem. Inf. Model. 2023; doi: https://doi.org/10.1021/acs.jcim.3c01527

References

• [1] Bai, P., Miljković, F., John, B. et al. Interpretable bilinear attention network with domain adaptation improves drug–target prediction. Nat Mach Intell 5, 126–136 (2023). https://doi.org/10.1038/s42256-022-00605-1
• [2] Liu, S., Wang, H., Liu, W., Lasenby, J., Guo, H. and Tang, J., 2021. Pre-training molecular graph representation with 3d geometry. arXiv preprint arXiv:2110.07728.