Protein model building for intermediate-resolution cryo-EM maps by integrating evolutionary and experimental information.

Accurate model building in intermediate-resolution cryo-EM maps normally requires flexible fitting of reliable initial structures. However, while deep learning-based methods such as AlphaFold2 can predict highly accurate structures, the predicted structures often differ from experimental EM maps on both global and local scales, which poses a great challenge to accurate model building in intermediate-resolution EM maps with such initial structures. Addressing the challenge, we propose CryoEvoBuild, an automated method for improved protein model building from intermediate-resolution EM maps through the effective integration of evolutionary and experimental information. CryoEvoBuild implements a novel domain-wise fitting, refinement, assembly, and rebuilding pipeline with a recycling framework guided by AlphaFold2. Extensive benchmarking on a diverse test set of 117 maps at 4.0-10.0 Å resolutions demonstrates that CryoEvoBuild significantly improves the accuracy of AF2-predicted structures and outperforms state-of-the-art approaches, including EMBuild and phenix.dock_and_rebuild.