Co-Diffuse: Generative Co-Design of Protein–Ligand Interactions via 3D Equivariant Diffusion Models with Induced-Fit Awareness

Mahule Roy, Subhas Roy

Abstract:

Protein flexibility and induced-fit effects are critical but often overlooked aspects of structure-based drug design. Most existing generative models treat proteins as rigid scaffolds, limiting their ability to predict realistic binding geometries. We present Co-Diffuse, a novel framework that jointly generates ligand molecules and their corresponding protein binding pocket conformations using SE(3)-equivariant diffusion models. By explicitly modeling the mutual adaptation between proteins and ligands, Co-Diffuse captures the dynamic nature
of molecular recognition. Comprehensive evaluations on PDBbind, CrossDocked2020, and cryptic site benchmarks demonstrate that Co-Diffuse achieves an average RMSD improvement of 1.8 over leading docking baselines and yields more accurate binding affinity predictions. Ablation studies highlight the importance of joint generation and physics informed constraints. Co-Diffuse represents a significant step
toward dynamic, physically grounded generative modeling for structure-based drug design.