MSIGR-PLA: Integrating Multi-Scale Interaction and Global Representations for Protein-Ligand Affinity Prediction

Accurate prediction of protein-ligand affinity (PLA) is crucial for accelerating drug discovery. Current methods exhibit limitations in extracting local protein-ligand interaction features and global representations, thereby hindering predictive accuracy. To address these limitations, we propose MSIGR-PLA, an integrative framework that integrates local multi-scale interaction features with global protein-ligand representations to improve the accuracy of PLA prediction. MSIGR-PLA employs two feature encoders to obtain rich representations. The local feature encoder contains a multi-scale dynamic interaction (MSDI) module, which consists of a GCN module, a cross-attention mechanism, and a Graph Transformer module. The global feature encoder uses a pre-trained ESM-2 model to extract protein sequence features and employs a CNN-Transformer module to encode ligand sequence information and a pre-trained GIN module to encode ligand structural information. Experimental results demonstrate that MSIGR-PLA consistently outperforms existing methods on four benchmark datasets, achieving improvements of 3.7%-9.0% in the Pearson correlation coefficient (R). Ablation studies further validate the effectiveness of the key modules in improving overall performance. Additionally, a case study demonstrates that the MSDI module can adaptively model multi-scale interaction features to identify key binding residues around ligands. Our code is available at https://github.com/zhc-moushang/MSIGR-PLA.