LGBind: A structure-guided deep learning framework for accurate prediction of protein-ligand binding sites

Unraveling the precise mechanisms of protein-ligand recognition remains a cornerstone of structure-based drug discovery. While recent deep learning advances have accelerated computational screening, a critical bottleneck persists: binding pocket characterization is often rigidly disconnected from the querying molecule's properties. Contemporary models frequently enforce a static, unidirectional mapping, which fails to capture the complex biomolecular micro-environment and limits generalization to unseen molecules. To overcome this, we introduce LGBind, a structure-guided deep learning framework for generalized, ligand-conditioned binding site identification. LGBind constructs a robust feature space by synergizing protein language models (ESM2 and Ankh) with 3D structural annotations, comprehensively extracting both evolutionary and structural information. Rather than using conventional network layers, we conceptualize the receptor topology as a 3D geometric graph to precisely track atomic-level structural dependencies. Crucially, LGBind employs a dynamic, mutually-guided interaction mechanism to evaluate the correlation between protein residues and molecular topologies. This architecture dynamically aligns the ligand's chemical properties with the receptor's geometric constraints to learn intrinsic physical binding motifs. Extensive evaluations across rigorous benchmarks confirm LGBind achieves new state-of-the-art performance, exhibiting profound robustness and generalization in predicting novel ligand-target interactions. Furthermore, LGBind demonstrates remarkable versatility, significantly enhancing downstream protocols including sequence-only predictions and precision-driven molecular docking.