HCGT-PL: a heterogeneous contrastive graph transformer unifying protein-ligand affinity prediction and...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in pharmacology

Background:

• Structure-based virtual screening and binding affinity prediction face challenges from solvation/entropy effects, protein flexibility, and induced fit.
• Accurate prediction of protein-ligand interactions is crucial for efficient drug discovery and development.

Purpose of the Study:

• To introduce a novel framework, Heterogeneous Contrastive Graph Transformer for Protein-Ligand (HCGT-PL), to address limitations in current protein-ligand modeling.
• To enhance the accuracy and generalizability of binding affinity prediction and virtual screening.

Main Methods:

• Representing protein-ligand complexes as directed heterogeneous graphs with diverse node and relation types.
• Employing relation-specific multi-head attention for effective message passing and aggregation within the graph structure.
• Utilizing unsupervised augmentations for transferable interaction representations, followed by fine-tuning for affinity regression and virtual screening tasks.

Main Results:

• Achieving robust accuracy, strong ranking capability, and pronounced early-enrichment across diverse benchmarks and hold-out evaluations.
• Demonstrating consistent generalization across various protein families and binding pocket conditions.
• Visualizations indicating the model's focus on key ligand functional groups and contacting receptor side chains within the binding pocket.

Conclusions:

• The HCGT-PL framework offers a unified, transferable, and interpretable solution for protein-ligand modeling.
• This approach advances computational methods for drug discovery by effectively handling complex molecular interactions.
• The study highlights the potential of integrating heterogeneous graph modeling, graph transformers, and contrastive learning for biological applications.