TriMod-DTI

A Triple-Modal Contrastive Learning Framework with Sequence, Graph, and 3D Features for Drug–Target Interaction Prediction

• Overview
• Installation
• Data
• Using
• Acknowledgements

Overview

TriMod-DTI is an innovative drug-target interaction (DTI) prediction framework that integrates three modalities of data, including 1D sequences, 2D molecular graphs, and 3D structures, for feature representation of drugs and proteins. The framework employs a tri-modal contrastive learning strategy, constructing cross-modal positive and negative sample pairs to align feature representations of different modalities in the latent space, thereby achieving feature enhancement.

Installation

Create a virtual environment

conda create -n TriMod-DTI python=3.10
conda activate TriMod-DTI

Clone the repo and install requried python dependencies

git clone https://github.com/llleXu/TriMod-DTI.git
cd TriMod-DTI
pip install -r requirements.txt

Data

Datasets

The data directory includes all the experimental datasets utilized in TriMod-DTI, covering GPCR, Human, and DrugBank.

Download SDF and PDB Files

The drug SDF files and protein PDB files are too large to be hosted directly on GitHub. These files are available for download via Google Drive at the following links:

• GPCR Dataset : download
• Human Dataset : download
• DrugBnak Dataset : download

Using

Configuration

Update the file paths in main.py to match your local directory structure. Modify the Data_Encoder initialization as follows:

Data_set = Data_Encoder(
    txtpath="data/gpcr/train_gpcr.txt",                 # Path to the text file
    sdf_directory="data/sdf_files/",                    # Directory containing SDF files
    sdf_map_path="data/gpcr/sdf_train_id.txt",          # Path to the SDF mapping file
    pdb_directory="data/pdb_files/",                    # Directory containing PDB files
    pdb_map_path="data/pdb_train_id.txt"                # Path to the PDB mapping file
)

Training and evaluation

cd dataset
python main.py

dataset specifically refers to gpcr, human and drugbank.

Acknowledgements

This implementation takes inspiration from prior work [1],[2],[3] and introduces several improvements and extensions to the original approach.

[1]Lin X, Zhang X, Yu Z G, et al. CSCL-DTI: predicting drug-target interaction through cross-view and self-supervised contrastive learning[C]//2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2024: 707-712.
[2]Yazdani-Jahromi M, Yousefi N, Tayebi A, et al. AttentionSiteDTI: an interpretable graph-based model for drug-target interaction prediction using NLP sentence-level relation classification[J]. Briefings in Bioinformatics, 2022, 23(4): bbac272.
[3]Mu S, Liang S, Li X. Medication Recommendation via Dual Molecular Modalities and Multi-Substructure Enhancement[J]. arXiv preprint arXiv:2405.20358, 2024.