CryoACE Automates Accurate Atomic Model Building in Cryo-EM.

Reconstructing protein models from cryo-electron microscopy (cryo-EM) density maps is a complex task, particularly due to the need to ensure physicochemical validity and manage the inherent conformational heterogeneity of biological molecules. Existing solutions often provide only static predictions or rely on computationally intensive heuristic searches, limiting their practical application. Researchers have developed CryoACE, an innovative end-to-end framework designed to accurately and automatically build atomic models for both homogeneous and heterogeneous protein structures. CryoACE introduces two key innovations. First, it employs an atom-centric reconstruction paradigm where density features are directly sampled at atomic coordinates and iteratively refined, replacing less efficient voxel convolutions for multimodal fusion. Second, CryoACE incorporates a training-free guidance mechanism that utilizes predicted local resolution priors. This mechanism is crucial for resolving dynamic ambiguity within the density maps. Validated on a new high-quality dataset, CryoACE not only significantly outperforms current baselines on static benchmarks but also, for the first time, reveals atomic-level dynamic conformations on complex real-world datasets, such as EMPIAR-10345, without requiring pre-existing static structural information.