An environment-conditioned QM/MM resource for protein–ligand electron densities and quantum-informed pose rescoring
Structure-based drug design relies on geometric and empirical features, whereas electron density encodes polarization and weak interactions but is costly to obtain at biomolecular scale. We present a multilevel QM/MM resource from MISATO trajectories and PDBbind complexes, comprising 104,636 electronic-structure records across 96,467 geometries. Physics-embedded equivariant surrogates achieve 0.326–0.384% density NMAE on PDB-disjoint tests with explicit TIP3P water; predicted receptor electrostatic potentials agree with reference values (Pearson r = 0.99529). For pose rescoring, we evaluated quantum electrostatic and non-covalent interaction descriptors on 405 four-filter generalization complexes, each with 20 poses from Vina, smina and AutoDock4, including at least one with heavy-atom RMSD <2.0 Å. Matched ablations improve Top-1 recovery by 7.41 percentage points over Vina-derived inputs and by 3.70–4.69 points over geometry-only arms. Full models achieve 69.14–70.12% Top-1 success.
Authors
- Chengbin Hou (ORCID: https://orcid.org/0000-0001-6648-793X)
- Qian Xie (ORCID: https://orcid.org/0000-0001-9901-0396)
- Gang Logan Liu (ORCID: https://orcid.org/0000-0001-8841-1514)
- Changzhi Jiang (ORCID: https://orcid.org/0000-0003-0765-6320)
- Tianying Zhang
- Xiangrong Liu
- Bingli Jiao
- Junrong Huang
- Kun Hou
- Juan Liu
Institutions
- Xiamen University (CN)
- Peking University (CN)
- Fuyao University of Science and Technology (CN)
- Huazhong University of Science and Technology (CN)
- Fujian University of Technology (CN)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-16
- DOI
- https://doi.org/10.26434/chemrxiv.15012250/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint