MSCA-PLA: multi-scale cross-attention with differentiable pooling for protein-ligand binding affinity prediction

Accurately predicting protein-ligand binding affinity is essential for drug discovery. However, protein-ligand binding involves not only pairwise atomic contacts but also synergistic effects between atomic clusters; meanwhile, adaptively identifying which atoms can form effective atomic clusters to drive protein-ligand interactions remains a key challenge. To address this challenge, we propose MSCA-PLA, a multi-scale cross-attention framework with differentiable pooling for protein-ligand binding affinity prediction. Specifically, MSCA-PLA adaptively aggregates atom-level embeddings into binding-relevant atom clusters through differentiable pooling, and performs cluster-level cross-attention between protein and ligand clusters to model cooperative interactions among atom groups. Meanwhile, an atom-level bidirectional cross-attention module is employed to preserve fine-grained pairwise atomic contacts at the binding interface. Furthermore, to effectively integrate interaction information from different scales, we design a staged gated fusion strategy. Experiments on multiple benchmark datasets, diverse-protein settings, virtual-screening tasks, and a small external structural subset support the effectiveness of MSCA-PLA across the evaluated settings.