Macromolecular structure modeling for cryo-EM using deep learning

Structure modeling of proteins, nucleic acids, and ligands is an essential step in interpreting cryo-EM maps, but it remains challenging, particularly when map resolution is limited. We have been developing a series of structure modeling tools for cryo-EM that leverage deep learning to identify atom positions and local structures in the maps. Here, we report our recent progress. In our latest developments, we integrate a protein structure prediction architecture with cryo-EM map information to enable macromolecular complex structure modeling directly within cryo-EM maps. When the map resolution is low, structure fitting, rather than de novo modeling, often provides a more practical solution. To address this, we previously developed DiffModeler, a deep learning–assisted structure fitting tool. We now present an updated version, DMcloud, which supports local structure fitting, making it particularly effective when the input structure models contain local conformational errors. All of our modeling tools, including DiffModeler and DMcloud, are freely available through our web server: https://em.kiharalab.org.