Hierarchical Contrastive Learning for Multi-Domain Protein-Ligand Binding

Researchers introduced HCLBind, a self-supervised binding model for protein-ligand affinity prediction.

It uses hierarchical contrastive pre-training with local coordinate perturbations and inter-domain rotations to better capture protein flexibility.

The model also adds domain-aware attention and foundation-model adaptation with LoRA for multi-domain proteins.

Results suggest improved binding prediction reliability, which could strengthen drug-discovery screening.