"MSCoD: A Hierarchically-Adaptive Bayesian Flow Networks for Structure-Based Drug Design"

"Structure-Based Drug Design (SBDD) is apowerful strategy in computational drug discovery, lever-aging three-dimensional protein structures to designmolecules with high binding affinity. However, capturingcomplex protein-ligand interactions across multiple scalesremains challenging, since current methods often ignorethe hierarchical organization and asymmetry of these inter-actions. To tackle these problems, we developed a novelhierarchically-adaptive framework that integrates a multi-scale feature processing (MSFP) module and a multi-headcooperative attention (MHCA) module, called MSCoD. TheMSFP module employs parallel pathways with distinct com-pression ratios, enabling hierarchical feature extraction andallowing the model to effectively capture both global con-text and fine-grained molecular details. The MHCA mod-ule applies asymmetric cooperative attention to effectivelymodel protein-ligand interactions and adapt ligand fea-tures to the protein environment. This combination enablesMSCoD to generate molecules with enhanced binding affin-ity, stability, and drug-like properties, while maintaininginformative feature representations. Experimental resultsshow that MSCoD outperforms existing methods in gener-ating molecules with more stable 3D structures and higherbinding affinities. MSCoD demonstrates strong capability inmodeling protein-ligand interactions and further extends todrug\u2013target affinity interaction prediction tasks, highlight-ing its versatility and broad applicability in structure-baseddrug discovery. Our code is available at https:\/\/github.com\/xulong0826\/MSCoD."