Protein-Ligand Binding Affinity Prediction Using Graph Neural Networks

Abstract

Protein-ligand binding affinity prediction is a central challenge in computational drug discovery, as it enables the rapid in silico screening of medical candidates before costly experimental validation. While classical physics-based methods and empirical scoring functions have long been applied to this problem, they suffer from either prohibitive computational cost or limited generalization across diverse molecular systems. Graph Neural Networks (GNNs) have emerged as a powerful alternative by representing molecules in their natural graph form and learning atomic interaction patterns directly from experimental data. However, the field of GNN architectures is vast, and it remains unclear which designs are most effective for structure-based binding affinity prediction. This master’s dissertation presents a systematic comparative evaluation of multiple GNN architectures on the protein-ligand binding affinity prediction task using the PDBbind dataset. We design a unified heterogeneous molecular graph representation that encodes the ligand and the protein pocket jointly, incorporating both covalent and non-covalent interaction information. All models are evaluated within an identical base architecture to ensure a fair and rigorous comparison, with the goal of identifying the most effective GNN design choices for structure-based drug discovery