EDMolGPT: A Decoder-Only Framework for 3D Drug Design via Electron Density

Jiahao Chen, Letian Gao, Yanhaozhu, Wenbiao zhou, Bing Su, Zhi John Lu, Bo Huang

Keywords: Drug design, Electronic density

TL;DR: We propose EDMolGPT, a decoder-only generative framework designed to produce 3D ligand molecules in an autoregressive manner via electron density.

Abstract: Electron density-guided drug design is a promising structure-based drug discovery (SBDD) frontier, crucial for delineating dynamic molecular features and intermolecular interactions. Existing methods leveraging electron density for \textit{de novo} molecule generation employ a two-stage process: generating hypothetical binder electron densities within a pocket, then interpreting them into molecules. While mitigating bias from binders pre-existing in the pocket, these approaches' two-stage nature can lead to error accumulation. Furthermore, these methods are limited by rigid pocket assumptions, which may compromise the diversity of the generated electron density. These limitations often result in drug-like molecules lacking favorable three-dimensional (3D) conformations or conversely, 3D conformations without assured drug-likeness. We introduce EDMolGPT, a novel decoder-only framework that directly synthesizes molecules from the low-resolution electron density point cloud derived from an existing binder. By leveraging this existing binder’s low-resolution electron density and avoiding explicit pocket structures, our strategy effectively mitigates bias, circumvents two-stage error, and negates rigid pocket limitations. EDMolGPT's autoregressive decoder-only architecture, guided by robust low-resolution electron density, efficiently generates binding molecules with high drug-likeness and favorable 3D conformations. Rigorous validation across 101 biological targets underscores its potential to accelerate novel therapeutic agent discovery.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 4236