Graph-Based Deep Learning for Protein–Ligand Interaction Modelling

Overview

Predicting how a small molecule binds a protein target is central to structure-based drug discovery, and most existing models are trained on broad, largely wild-type complexes — which generalise poorly to the closely related mutant pockets that matter most for drug-resistance biology. This project develops a mutation-sensitive graph neural network architecture aimed at that gap: a model that can pick up on how a single point mutation reshapes a binding pocket and, with it, ligand affinity.

The architecture fuses several complementary views of a protein–ligand complex — protein-language-model sequence embeddings, a graph encoding of pocket structure, and a graph representation of the ligand — combined through a learned cross-attention fusion mechanism, and is benchmarked against standard structure-based binding-affinity datasets before being applied to case studies drawn from my own drug-resistance research (including antifolate binding in PfDHFR).

Key Highlights

• Mutation-sensitive architecture, aimed at generalising across near-identical mutant binding pockets
• Multimodal fusion of sequence, pocket-structure, and ligand-graph representations via cross-attention
• Benchmarked against standard structure-based protein–ligand affinity datasets
• Applied to case studies in antifolate resistance drawn from my broader research programme
• Developed in collaboration with the Upadhyay Lab (Dept. of CSIS)

Researcher

Led by Danny Muzata, a biomedical data scientist at BITS Pilani, Hyderabad Campus, in collaboration with Dr. Prajna D. Upadhyay's group in the Dept. of CSIS.

Google Scholar  •  ORCID