3D Geometric Equivariant Graph Neural Networks with Contrastive Learning for Protein-Ligand Binding Affinity Prediction.

Protein-Ligand Binding Affinity (PLBA) prediction is a key task in computer aided drug design, underpinning virtual screening, lead optimization, and drug repurposing. Despite recent advances in deep learning, many existing methods fail to fully exploit three-dimensional (3D) geometric information and remain sensitive to conformational variability. To address these limitations, we propose 3D-GEGCL, a framework that integrates 3D geometrically equivariant graph neural networks with self-supervised contrastive learning for robust and accurate PLBA prediction. Protein-ligand complexes are represented as atomic-level 3D geometric graphs and encoded using an E(3)-equivariant network to ensure physical consistency under rotations and translations, while contrastive learning enforces representation consistency across geometrically perturbed views. Comprehensive evaluations conducted on the PDBbind v2019 and CSAR-HiQ benchmarks reveal that 3D-GEGCL achieves superior performance over existing leading methods across a range of assessment criteria. Additional ablation analyses validate the synergistic roles of geometric equivariance and contrastive learning, underscoring the robustness and broad applicability of the introduced framework. Implementation scripts for the various model components are publicly accessible via https://github.com/ligaili01/3D-GEGCL .