EM3DFold: accurate de novo protein and nucleic acid model building for cryo-EM maps using language model-powered deep learning

Abstract

Cryo-electron microscopy (cryo-EM) has become one of the most powerful techniques for macromolecular structure determination. However, accurate model building from cryo-EM maps remains challenging, particularly for nucleic acids. Here, we present EM3DFold, a unified de novo model-building framework for accurate structure determination of proteins, nucleic acids, and protein-nucleic acid complexes from cryo-EM maps using a density-aware, large language model-powered three-track attention (TTA) network. The TTA network effectively integrates sequence, density, and structural information to enable accurate all-atom model building. EM3DFold was extensively evaluated on independent benchmarks of 298 experimental cryo-EM maps at < 4.0 Å resolutions, and achieves an unprecedentedly high median accuracy of 75% completeness (88% coverage and 94% sequence accuracy) for 178 nucleic acid maps, 90% completeness (95% coverage and 96% sequence accuracy) for 124 protein-nucleic acid complexes, and 95% completeness (97% coverage and 98% sequence accuracy) for 120 protein-only targets, substantially outperforming state-of-the-art methods including ModelAngelo, EM2NA, CryoREAD, and EMProt. In addition, EM3DFold also produces models with superior model-to-map fit and stereochemical quality, providing a robust and reliable solution for automated cryo-EM model building. The EM3DFold package is freely available at https://github.com/huang-laboratory/EM3DFold.