Improved model building for cryo-EM maps using local attention and 3D rotary position embedding

Abstract

Constructing atomic models from cryogenic electron microscopy (cryo-EM) density maps is essential for interpreting molecular mechanisms. In this study, we present CryFold, an approach for de novo model building for cryo-EM maps, leveraging recent advancements in AlphaFold2 (1) to improve the state-of-the-art method ModelAngelo (2). To accommodate the cryo-EM map information, CryFold replaces the global attention mechanism in AlphaFold2 with local attention, which is further enhanced by a novel 3D rotary position embedding. CryFold produces more complete models, reduces the resolution requirement, and accelerates the modeling. The application of CryFold to three new maps with unknown structure demonstrates its ability to accurately distinguish between paralog sequences in noisy regions, detect previously uncharacterized proteins with unknown functions, precisely compartmentalize the map to isolate non-protein components, and improve the modeling of conformational changes. A particular case includes a 104-protein complex that has been modeled within a few hours, and a minor conformational change of a single protein domain has been detected at the periphery when models from two different maps were compared. CryFold stands as an accurate method currently available for model building of proteins in cryo-EM structure determination. The source code and model parameters are available at https://github.com/SBQ-1999/CryFold.