CryoID2: deep learning-enhanced identification of protein and RNA from cryoEM maps

Abstract

Identifying the molecular components of cryoEM/cryoET maps of endogenous or poorly characterized complexes remains a major bottleneck. CryoID2 addresses this by integrating deep learning-based feature prediction with sequence prioritization and structural evaluation. Three 3D U-Nets extract structure classes, key-atom positions, and reduced residue/nucleotide types from the input map. These predictions support backbone tracing and the generation of short sequence queries that rank candidates from a user-supplied pool-using BLASTp for proteins and a maximin strategy for RNA. Corresponding PDB or AlphaFold models are then retrieved, rigidly fitted using predicted features and experimental density, and ranked by map agreement. Module-level tests show high-recall feature prediction and resolution-dependent atom clustering, while a 52-case benchmark recovers near-reference fitting quality. In practice, cryoID2 correctly identified PfA-M17 from a 166-protein mass-spectrometry pool and the unannotated Ca-raiA RNA from a genome-derived non-coding sequence set, demonstrating a unified sequence-plus-structure route to molecular identification in cryoEM maps.