GRAM-DTI: Adaptive Multimodal Representation Learning for Drug-Target Interaction Prediction

This repository contains the official PyTorch implementation of the ICLR 2026 paper: "GRAM-DTI: Adaptive Multimodal Representation Learning for Drug-Target Interaction Prediction".

📖 Overview

Drug target interaction (DTI) prediction is a cornerstone of computational drug discovery. While deep learning has advanced DTI modeling, existing approaches primarily rely on pairwise SMILES-protein interactions, failing to exploit the rich multimodal information available for small molecules.

GRAM-DTI is a novel pre-training framework that integrates four distinct modalities into unified representations:

1. SMILES Sequences (via MolFormer)
2. Text Descriptions / Molecular Functions (via MolT5)
3. Hierarchical Taxonomic Annotations (HTA) (via MolT5)
4. Protein Sequences (via ESM-2)

Key Contributions

• Gramian Volume-Based Multimodal Alignment: Extends contrastive learning to four modalities, capturing higher-order semantic alignments beyond conventional pairwise approaches.
• Gradient-Informed Adaptive Modality Dropout: Dynamically regulates each modality's contribution during pre-training based on its gradient informativeness, preventing dominant but less informative modalities from overwhelming complementary signals.
• Auxiliary Weak Supervision: Incorporates IC50 activity measurements (when available) to ground learned representations in biologically meaningful interaction strengths.

📂 Repository Structure

GRAM-DTI/
├── pretraining/
│   ├── __init__.py
│   ├── losses.py       # Implementation of Volume Loss, and Adaptive Dropout
│   ├── models.py       
│   ├── trainer.py      # Training loop, optimization, and logging logic
│   └── run.py          # Main entry point for distributed pre-training
├── README.md
└── requirements.txt    # Dependencies