Modeling Protein-Ligand Binding Affinity Using Graph Neural Networks: Integrating Molecular Interactions and Physics-Based Properties

Abstract

Predicting protein-ligand binding affinity has always been one of the primary challenges in drug discovery. Though many machine learning approaches have been applied and reached substantial progress, despite advancements, the accurate prediction by integrating molecular interactions and physics-based properties of ligands and proteins remains challenging. In this project, we present a graph neural networks (GNNs)-based framework for predicting protein–ligand binding affinity, using publicly accessible CrossDocked2020 dataset. Our project compares three message-passing architectures—Linear, Set Transformer Aggregation (STA), and Graph Attention Network (GAT). Our best model achieves performance of Pearson’s R ≈ 0.79 and Kendall’s τ ≈ 0.58. We present a practical GNNs-based framework with plausible binding affinity prediction capabilities, designed to effectively differentiate correct poses from incorrect ones.