HCGT-PL: A heterogeneous contrastive graph transformer unifying protein–ligand affinity prediction and structure-based virtual screening

Abstract

Structure-based virtual screening and binding affinity prediction remain challenging due to solvation/entropy effects, protein flexibility, and induced fit. We present a Heterogeneous Contrastive Graph Transformer for Protein-Ligand (HCGT-PL) framework. Each complex is represented as a directed heterogeneous graph with multiple node and relation types; relation-specific multi-head attention enables message passing and aggregation. Unsupervised augmentations yield transferable interaction representations that are fine-tuned for affinity regression and virtual screening. Across diverse benchmarks and hold-out evaluations, the approach delivers robust accuracy, strong ranking capability, and pronounced early-enrichment, with consistent generalization over varied protein families and pocket conditions. Interpretability visualizations indicate that the model prioritizes ligand functional groups and contacting receptor side chains within the binding pocket. By unifying heterogeneous graph modeling, graph Transformers, and contrastive learning, this framework provides a general, transferable, and interpretable solution for protein–ligand modeling.