Physics-informed generative model for drug-like molecule conformers

David C Williams

Keywords: deep learning, graphical models, Probabilistic Models and Methods, Bioinformatics and Systems Biology

TL;DR: Diffusion-based generation of molecule conformers using a model informed by physics and employing denoising score matching

Abstract: Diffusion-based methods have been successfully applied to molecule conformer generation using implicit physical modeling. In contrast, conventional, rules-based approaches employ an explicit physical model such as a classical force field parameterization. In order to combine the advantages of both approaches, we present a diffusion-based, physics-informed denoising model (PIDM) for conformer generation that is constructed from molecule subgraph patterns borrowed from classical force fields. The result is a model that is resistant to overfitting and explainable. Using recent advances in denoising score matching, we naturally separate the task of training and generation while providing a smooth transition between deterministic and stochastic generative schemes that adapt to any number of denoising steps. We demonstrate conformer generation quality that outperforms the current state-of-the-art while employing a fraction of parameters.

Submission Number: 14608