Reducing Atomic Clashes in Geometric Diffusion Models for 3D Structure-Based Drug Design
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Structure Based Drug Design, Geometric Molecular Generation, Diffusion Models
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In the domain of Three-dimensional Structure-Based Drug Design (3D SBDD), the 3D spatial structures of target pockets serve as inputs for the generation of molecular geometric graphs. The Geometric Diffusion Model (GDM) has been recognized as the state-of-the-art (SOTA) method in 3D SBDD, attributed to its exceptional generation capabilities on geometric graphs. However, the inherent data-driven nature of GDM occasionally neglects critical inter-molecular interactions, such as Van der Waals force and Hydrogen Bonding. Such omissions could produce molecules that violate established physical principles. Particular evidence is that GDMs exhibit atomic clashes during generation due to the overly close proximity of generated molecules to protein structures. To address this, our paper introduces a novel constrained sampling process designed to obviate such undesirable collisions. By integrating a non-convex constraint within the current Langevin Dynamics (LD) of GDM and utilizing the proximal regularization techniques, we force molecular coordinates to obey the imposed physical constraints. Notably, the proposed method requires no modifications to the training process of GDMs. Empirical evaluations show a significant reduction in atomic clashes via the proposed method compared to the original LD process of GDMs.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3193
Loading