Answers are grounded in this paper and cite their sources.
Machine-learning models of molecules usually encode nuclear geometry, while spatial electronic information is either omitted or requires costly molecule-specific quantum calculations. Here we present eMol, an electron-density-driven model that overcomes this trade-off by coupling atomic scaffolds with density-derived probes via a retrievable memory, allowing compatible electronic environments to be reused across molecules without repeated first-principles density calculations. To integrate the transferred electronic information with nuclear geometry, eMol employs an equivariant atom–electron architecture that captures atomic geometry, local electronic structure and higher-order many-body interactions within a unified representation. To avoid computing electron density for every query, eMol uses structure-only retrieval and geometric alignment to reuse probes from a memory of approximately two million previously characterized molecules. This design yields consistently strong performance across 24 quantum-chemical property prediction tasks and 24 molecular potential-learning systems, and reproduces reference structural distributions in biomolecular dynamics. Its learned attention patterns align with local electron density, and it prioritizes literature-supported compounds in drug-activity datasets. More broadly, decoupling expensive physical information from real-time computation provides a generalizable paradigm for molecular learning.
No takes yet. Share an insight, caveat, or question.
Zeng et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: