CoaDTI

CoaDTI predicts and annotates drug–target interactions by jointly modeling chemical molecules (SMILES-derived graphs) and protein sequences to identify interacting molecules and protein residues within binding pockets.

Key Features:

• Co-Attention Mechanism: Employs a co-attention mechanism to model interaction information across drug and protein modalities simultaneously.
• Transformer and GraphSage Integration: Uses transformers to learn protein representations from raw amino acid sequences and GraphSage to extract molecule graph features from SMILES.
• Transfer Learning Strategy: Incorporates transfer learning by pre-training transformers on protein sequences to enhance protein feature encoding with limited labeled data.
• Competitive Performance and Novelty Detection: Demonstrates competitive results across three public datasets relative to state-of-the-art models and enables identification of novel DTIs, including candidate drug interactions with SARS-CoV-2 proteins.
• Interpretability through Visualization: Visualizes co-attention scores to illustrate which parts of the drug and protein contribute to predicted interactions.

Scientific Applications:

• High-throughput DTI prediction: Enables large-scale prediction of drug–target interactions for in silico drug discovery workflows.
• Residue-level binding annotation: Provides residue-level annotation of protein binding pockets by identifying interacting protein residues.
• Novel interaction discovery: Supports discovery of potential therapeutic interactions, including applications to emerging pathogens such as SARS-CoV-2.

Methodology:

Integrates a co-attention framework that combines transformers trained on raw amino acid sequences and GraphSage on SMILES-derived molecule graphs, employs transfer learning via pre-training of transformers on protein sequences, and produces co-attention score visualizations.