Allo-DiffHopping: A Computational Approach for Allosteric Scaffold Hopping Using Conditional E(3) Equivariant Diffusion Models

Fragment-based drug design presents significant computational challenges, necessitating advanced methods for efficient molecular generation. Diffusion generative models have shown promise in this domain; however, prior approaches have overlooked the critical role of pocket-ligand interactions. In response, we introduce "Allo-DiffHopping," an innovative computational model employing conditional E(3) equivariant diffusion for allosteric scaffold hopping. This method processes an arbitrary number of fragments and anchor points to construct new molecules via scaffold hopping. The generative process is conditioned on detailed pocket-ligand interactions, including binding affinity and 3D structural data. These interactions are essential in fragment-based drug design. Our evaluation demonstrates that Allo-DiffHopping produces linkers with better binding affinity to proteins, outperforming baseline models in compound validity, synthetic accessibility, Vina score, and SCRDKit metrics. Additionally, in allosteric batch tests, Allo-DiffHopping exhibits potential for generating high-performance inhibitors, highlighting its applicability and effectiveness in computational drug design.