Unknown proteins in cryo-EM maps now get a ranked shortlist: intact maps searched directly, 11 of 18 radial spoke proteins in top 15

Identification of unknown proteins moves from high-resolution settings that rely on side-chain density to intermediate-resolution intact density maps. Deep-learning model-building tools generally need better than about 4 Å for side-chain density to be resolved, while intermediate-resolution maps lack reliable side-chain information and still offer thousands of candidates. On 14 held-out benchmark maps containing at least one manually confirmed true-positive candidate, the learning-to-rank model outperformed the original CoLoRes score and each individual engineered feature in map-wise average precision, Recall@P and NDCG.

The pipeline docks each candidate structure rigidly into the target density map, then a ranking model combines several agreement signals. CoLoRes reports the best pose per candidate by a correlation score, but that score depends on protein size, overlap volume and map resolution, so it cannot reliably compare thousands of candidates. Features include masked Pearson correlations over strict and locally dilated overlap regions, the largest connected overlap fraction, an overlap fragmentation penalty, explained map voxels and candidate atom count, integrated by an XGBoost ranker trained with map-level grouping.

On real unassigned data the pipeline returns verifiable candidates and narrows the field. These density maps previously had no molecular identity, and only the sample source, such as E. coli or human cells, was known. A 4.2 Å E. coli map ranked DnaK first, a 3.9 Å hexameric map ranked ArnA first with only ArnA covering the complete protomer density, and a 4.5 Å human extracellular vesicle map put five α2-macroglobulin superfamily members on top, with tetrameric symmetry narrowing the list to two.

The workflow targets single-particle cryo-EM reconstructions and cryo-ET subtomogram-averaged maps, takes intact density maps as input without prior segmentation, and draws candidates from AlphaFold2-predicted models or experimental PDB structures. Demonstrated cases include DnaK and ArnA from an E. coli recombinant sample, α2-macroglobulin-family candidates from human extracellular vesicles, the Chlamydomonas radial spoke complex, and several tomography averages. Computational cost grows with target map size, search resolution and candidate count; for example, searching the human AlphaFold2 database of 40,596 processed domains against an approximately 1-MDa density map at the default 15 Å resolution took about 48 hours on 288 CPU cores.

The ranking model depends on the quality of the initial CoLoRes rigid-body placement, so correct candidates can rank unexpectedly low when placements are suboptimal; in EMD-22475 and EMD-56117, four and two correct candidates respectively ranked well below their expected positions. Under controlled low-pass filtering, most true hits stayed highly ranked at cutoffs of about 4–12 Å, while rankings generally became less reliable beyond about 15 Å; however, EMD-65039 and EMD-12661 behaved in contrasting ways, showing that nominal resolution alone does not determine performance, and that local map quality, density continuity, sharpening and preservation of overall shape also matter. Proteins with highly similar folds remain intrinsically difficult to distinguish by rigid-body docking, as seen when actin and actin-like proteins or IMPDH and GMPR family members rank together, requiring biochemical or contextual information for discrimination.