Title: From Holo Pockets to Electron Density: GPT-style Drug Design with Density

Abstract:

The landscape of structure-based drug design (SBDD) has been transformed by recent breakthroughs in generative modeling. However, current approaches generally rely on conditioning molecule generation on vacant binding pockets extracted from holo complexes, thereby neglecting critical contextual elements such as solvent molecules and filler ligands. To address this limitation, we propose utilizing low-resolution electron density (ED)—derived from these filler components—as a physically informed condition for de novo drug design. Our methodology accommodates two categories of ED: those computed via algorithms and those obtained through experimental techniques like cryo-EM and X-ray crystallography. This versatility facilitates unified pre-training strategies and seamless integration with experimental data. Unlike rigid pocket representations, experimental ED inherently accounts for conformational flexibility, offering a more accurate depiction of the binding microenvironment. Leveraging this insight, we present EDMolGPT, a decoder-only autoregressive architecture designed to generate molecular structures from low-resolution ED point clouds. By anchoring the generation process in physically significant density signals, EDMolGPT reduces structural bias and yields molecules with realistic 3D conformations. The efficacy of our approach is demonstrated through evaluations across 101 distinct biological targets. Further details can be found on our project page: https://jiahaochen1.github.io/EDMolGPT_Page/.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC