CryoEM maps no longer just show a shape: cryoID2 pairs sequence and structure to pick PfA-M17 out of 166 proteins
Synopsis
CryoEM can reconstruct a density map yet often cannot say which molecule the map contains, and matching hundreds or thousands of mass-spectrometry or genome-derived candidates by hand is a bottleneck. CryoID2 uses three 3D U-Nets to predict structure classes, key-atom positions and reduced residue or nucleotide types from the map, turns them into short sequence queries that rank candidates, then rigidly fits the candidates' PDB or AlphaFold models back into the density and ranks them by agreement. In a 166-protein mass-spectrometry pool it placed PF3D7_1446200/PfA-M17 first and selected PDB 4ZX8 as the best structural model; on the RNA side it ranked the Ca-raiA-containing interval fourth in a genome-derived non-coding pool, and structural evaluation then put PDB 9E73 first.
Interpretation
Once a cryoEM reconstruction yields a three-dimensional density map, which protein or RNA it represents still has to be determined separately, and cryoID2 makes that step a single route from density map to candidate sequences to structural models. The earlier cryoID was aimed mainly at proteins and relied on algorithmic feature extraction and sequence matching, without deep-learning prediction or structural comparison in the identification step. On the P. falciparum E1_D3 experimental density map, PF3D7_1446200/PfA-M17 ranked first among 166 mass-spectrometry-derived candidates, and its corresponding experimental structure PDB 4ZX8 received the highest structural score among retrieved models.
The pipeline first translates the density map voxel by voxel into structure classes, key-atom positions and reduced residue or nucleotide types, then reads short sequence fragments along traced backbones. The 20 amino acids are grouped into six classes and nucleotides into two, a reduced alphabet used to absorb local prediction uncertainty. Three U-Nets handle structure classes, key atoms and reduced types, with voxel-level recall of 0.97, 0.99 and 0.92 and precision of 0.24, 0.16 and 0.22, which clustering and tracing then consolidate.
After candidate ranking, the pipeline retrieves the candidates' experimental or AlphaFold models, fits them to the map by deterministic rigid-body optimization using predicted features and experimental density, and re-ranks them by map agreement. Sequence ranking only narrows the candidate pool, and structural evaluation then discriminates among the narrowed candidates, including structurally similar RNA models. Across 52 published protein cryoEM cases, cryoID2-selected models after fitting had a mean Q-score of 0.4120 versus 0.4466 for deposited models, a median difference of −0.0001; an alignment-only control averaged 0.4138.
Perspective
The demonstrated scope is module-level benchmarks, a structural retrieval and fitting benchmark over 52 published protein cryoEM cases, one protein-identification case from a 166-sequence mass-spectrometry pool, and one RNA proof-of-concept case from a genome-derived non-coding pool whose candidates were constrained to within ±50% of the known 204-nucleotide reference length.
The RNA analysis is explicitly a proof of concept rather than a general benchmark of genome-scale RNA identification, evaluating a single case with a deposited structure, and performance may depend on genomic-interval definition, candidate-length constraints, candidate-pool composition and the availability of suitable structural models; rigid-body fitting does not accommodate internal conformational changes; the nucleic-acid branch can accept thymine-containing input at the software level, but that does not show RNA-trained predictions generalize to DNA.