Learning from Hetero Density for Cryo-EM Protein Reconstruction

1 INTRODUCTION

• Hetero components are widespread in analyzed maps, but reconstruction supervision mainly focuses on protein annotations despite maps also containing nucleic acids, glycans, cofactors, ligands, and lipids or detergents.Reliable reference structures are time- and expertise-intensive to build, limiting training data.Source: §1 · p. 1
• Hetero density can resemble protein density and create ambiguity, yet nearby hetero components can also provide structural clues about neighboring protein sites.Examples include N-linked glycans attached to asparagine, protein–RNA interfaces, ADP binding pockets, and lipids adjacent to helices.Source: §1 · pp. 1–2
• Uniformly filtering hetero neighborhoods can improve some chain constructions but degrade others, even within the same hetero class, because neighborhoods may contain false predictions and reference protein sites.The analysis motivates selecting hetero cues according to component type and local protein context.Source: §1 · p. 2
• CRYOCUE uses anchor-supervised multiscale hetero representations for backbone localization and conditions structure refinement on predicted hetero candidates’ class, confidence, and relative geometry.The framework adaptively exploits hetero information as the protein structure evolves.Source: §1 · p. 2
• Experiments report improved backbone localization and final protein reconstruction over protein-only baselines, including in regions associated with hetero components.The contribution combines a unified hetero representation with adaptive conditioning during localization and refinement.Source: §1 · p. 3

PRELIMINARY ANALYSIS

• The analysis uses Cryo2Struct’s explicit candidate stage to inspect predicted protein sites before chain construction, then tests filtering nearby candidates.The figure summarizes candidate patterns and filtering outcomes as improved, degraded, or unchanged.Source: §2 · p. 3, Figure 3 · p. 4

Hetero effects on backbone localization

• Nucleic acids, glycans, and lipids or detergents are often near false protein candidates without protein support, while nucleotides and cofactors often have false candidates alongside protein support.Other ligands show a mixture of these patterns, so hetero density is not a uniform non-protein signal.Source: §2 · p. 3

Hetero effects on chain construction

• Filtering candidates near annotated hetero components often improves chain construction, particularly around nucleic acids, but degradations occur across all hetero classes.Filtering can remove misleading predictions or discard useful protein information; remaining inputs were kept unchanged.Source: §2 · p. 3

3 METHODOLOGY

• CRYOCUE’s two-stage pipeline uses multiscale hetero features to localize initial protein sites, then uses frame-relative hetero conditioning to refine them before final sequence-guided chain assembly.Protein sequences enter only during final chain assembly.Source: §3 · p. 4, §3.3 · p. 6

3.1 ANCHOR-SUPERVISED HETERO REPRESENTATION

• Representative reference atoms act as anchors for localized hetero targets, with positive voxels requiring a matching class and other annotated hetero regions ignored.The detector predicts background, protein backbone, protein side chain, and five hetero classes.Source: §3.1 · p. 4

3.2 BACKBONE LOCALIZATION WITH HETERO FEATURES

• Parallel protein and fixed hetero streams use coarse cross-attention and fine-scale gated fusion so hetero context can inform backbone localization at different resolutions.At coarse scales, protein features query hetero features across the crop or local windows; at finer scales, a gate varies the hetero contribution by location.Source: §3.2 · p. 5

3.3 STRUCTURE REFINEMENT WITH HETERO CANDIDATES

• During structure refinement, nearby hetero candidates contribute class, confidence, and geometry relative to each evolving protein frame.Candidate neighborhoods remain fixed, while distance and local direction are recomputed as the frames move; a learned gate updates protein features.Source: §3.3 · pp. 5–6
• Refined frames and residue features predict backbone geometry and torsions, while residue-type profiles are aligned to input sequences for fragment assembly and all-atom construction.The structure network is trained with backbone-coordinate and torsion supervision.Source: §3.3 · p. 6

4.1 EXPERIMENTAL SETUP

• The evaluation uses family-aware training, validation, and test splits and compares global structure, local geometry, density support, backbone localization, and hetero-associated recovery.Source: §4.1 · p. 6

4.2 PROTEIN STRUCTURE RECONSTRUCTION

• CRYOCUE improves TM-score, lDDT, and coverage, with lower break rate; gains center on recovery and continuity rather than matched-coordinate accuracy or Q-score.The test-set table reports map-level mean values.Source: Table 1 · p. 7, §4.2 · p. 7

4.3 BACKBONE LOCALIZATION NEAR HETERO COMPONENTS

• CRYOCUE achieves the highest overall backbone precision and recall and highest hetero-neighborhood recall, while maintaining high precision and improving protein–hetero contact recall.Source: Table 2 · p. 7, §4.3 · p. 7

4.5 HYPERPARAMETER SENSITIVITY AND EFFICIENCY

• TM-score rises, then saturates and slightly declines as candidate caps and neighborhood radii increase; selected settings are stable, and runtime is close to EMProt but faster than ModelAngelo and EModelX.Source: Figure 4 · p. 8, §4.5 · p. 9

4.4 ABLATION OF HETERO REPRESENTATION AND CONDITIONING

• Localization-only hetero features improve reconstruction and chain continuity over protein-only, but removing anchor supervision drops structural metrics below the protein-only variant.Source: §4.4 · p. 8
• Dynamic frame-relative geometry further improves reconstruction; backbone-fusion-only leads Q-score, while the full model leads structure reconstruction and connectivity.Source: §4.4 · p. 8

4.6 CASE STUDY

• In 8IH5, CryoCue preserves the local protein trace near a hetero component where the compared methods miss reference sites, alongside greater global recovery.Source: Figure 5 · p. 9, §4.6 · p. 9

• CryoCue uses hetero information during backbone localization and structure refinement, improving structure recovery and chain continuity.Ablations show the importance of anchor supervision and dynamic hetero conditioning.Source: §5 · p. 9
• The approach depends on hetero-prediction quality, and its gains are clearer in structural recovery than in atom-level density fit.Source: §5 · p. 9
• The authors identify improving hetero prediction and introducing sequence information earlier than final chain construction as future directions.Source: §5 · p. 9

• Generative AI assisted with manuscript organization, language editing, drafting, literature retrieval, research execution, and software development.Source: AI USE STATEMENT · p. 10
• The authors reviewed and revised AI-assisted text and tested AI-assisted scripts and code.Source: AI USE STATEMENT · p. 10
• The authors take full responsibility for the final manuscript, code, results, and scientific claims.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The study uses publicly available cryo-EM maps and macromolecular structures, without human participants, private data, or personally identifiable information.Source: ETHICS STATEMENT · p. 10

A RELATED WORK

• Prior protein-reconstruction methods primarily use protein-derived information, whereas this work studies hetero density as evidence for backbone localization and refinement.Existing non-protein methods model specific component types; CryoCue instead uses predicted hetero information to guide protein reconstruction.Source: A RELATED WORK · p. 13

B.1 DATA COLLECTION AND QUALITY CONTROL

• The dataset pairs released primary maps with fitted structures through official EMD–PDB associations, retaining eligible pairs and excluding unresolved duplicates and ambiguously classified components.Source: §B.1 · p. 13, Table 4 · p. 14

B.2 DATASET SPLITS

• The dataset contains 18,420 pairs—17,820 training, 300 validation, and 300 testing—with protein-family-based filtering to reduce overlap across splits.Training candidates are excluded when a chain alignment meets the stated identity and coverage thresholds or files duplicate validation or test data.Source: §B.2 · pp. 13–14, Table 5 · p. 14

B.3 MAP PREPROCESSING

• Maps and fitted coordinates are placed in a common frame, and each full map is normalized once so overlapping crops retain consistent density values.Source: §B.3 · p. 14, Table 6 · p. 14

B.4 LABEL CONSTRUCTION

• Labels distinguish background, protein backbone, protein side chain, and five hetero classes; anchor supervision targets representative hetero sites while structural support assigns voxel labels.Hetero positives are restricted to voxels near accepted anchors whose class agrees with structural support; unsupported positive density is ignored.Source: §B.4 · p. 15

B.5 DATASET STATISTICS

• Pair-level counts report the presence of each of the five hetero classes, and classes can co-occur within a pair.Consequently, class counts are not additive.Source: §B.5 · p. 16, Table 8 · p. 16
• The masked structural-support census contains 484,570,312,002 training voxels, with protein backbone and side-chain support occupying 1.8063% and five hetero classes occupying 0.3451%.These are pre-anchor-restriction structural-support counts, not the distribution of sampled training inputs.Source: §B.5 · p. 16, Table 9 · p. 16
• The reported occupancy percentages describe structural support before anchor restriction, not sampled training-input frequencies.Source: §B.5 · p. 16, Table 9 · p. 16

C.1 ANALYSIS SUBSET AND HETERO STATISTICS

• Hetero voxels make up 14.06% of protein- and hetero-associated voxels, while 78.05% of the 1,868 analyzed map–structure pairs contain nonzero hetero-labeled voxels.Candidate-neighborhood analysis uses 1,362 pairs with successful atom and amino acid predictions.Source: §C.1 · p. 16, Table 10 · p. 17

C.2 PROTEIN CANDIDATES NEAR HETERO COMPONENTS

• FP-only neighborhoods dominate around nucleic acids, glycans, and lipids or detergents; nucleotide/cofactor neighborhoods most often combine false positives with protein support, while other ligands are more balanced.Protein support includes a true-positive candidate or a reference protein site, including sites missed by the predictor.Source: §C.2 · p. 17, Table 11 · p. 18

C.3 CANDIDATE FILTERING AND CHAIN CONSTRUCTION

• Filtering near hetero components produces both improvements and degradations in every group; most nucleic-acid cases improve, but uniform filtering can also remove useful candidates.The effect is most pronounced for nucleic acids, where the total number of matched protein sites increases substantially.Source: §C.3 · p. 18, Table 12 · p. 18

D.1 HETERO DETECTOR

• The hetero detector combines a multiscale local 3D SCUNet backbone with a global Transformer that updates the bottleneck and modulates decoder skip features before an eight-class prediction head.Its feature pyramid is also exposed to the backbone predictor.Source: §D.1 · p. 19

D.2 BACKBONE PREDICTOR AND MULTISCALE FUSION

• The backbone predictor fuses hetero features at scales 8, 4, 2, and 1 through coarse-scale cross-attention and fine-scale gated fusion, while keeping the detector fixed during backbone training.The final fused protein feature produces continuous N, Cα, and C response maps.Source: §D.2 · p. 20

D.3 STRUCTURE REFINEMENT WITH HETERO CANDIDATES

• Structure refinement encodes hetero candidates by class, confidence, and uncertainty, then conditions protein features using candidate geometry recomputed relative to the current protein frames.Each protein site uses up to 32 confidence-ranked candidates within 12 Å; hetero conditioning is applied at blocks 4, 8, 12, and 16.Source: §D.3 · pp. 21–22

D.4 END-TO-END INFERENCE

• End-to-end inference processes the map with the hetero detector, backbone predictor, and amino acid predictor, then refines protein frames in three rounds while keeping hetero candidates fixed in the map.Candidate geometry relative to protein frames is recomputed during each round.Source: §D.4 · p. 23
• Inference initializes protein frames, encodes hetero candidates, selects neighbors, and applies three refinement rounds before tracing fragments, aligning residue profiles to sequence, pruning, and building atoms.Refinement applies hetero conditioning at layers 4, 8, 12, and 16, then predicts torsions.Source: §D.4 · p. 23

E.1 TRAINING PROCEDURE

• Training proceeds from anchor-supervised hetero detection to backbone localization, independent density-only amino-acid prediction, and structure refinement using predicted hetero candidates.Models are trained from scratch with separate parameters, optimizers, and objectives; structure training uses density, initial protein frames, predicted hetero candidates, and reference structure targets.Source: §E.1 · p. 24

E.2 VOXEL-STAGE TRAINING

• Training requests mix natural crops, hetero-positive crops, and protein-containing crops at a 94/4/2 ratio, with the latter streams increasing exposure to those regions.The hetero-positive stream also maintains coverage across five hetero classes.Source: §E.2 · pp. 24–25
• Density and supervision share one of 24 orientation-preserving cube rotations, while backbone and amino acid supervision use the valid central 16^3 region.Detector supervision covers all valid positions in the 48^3 crop.Source: §E.2 · p. 25
• The three voxel models train independently; detector parameters are fixed before optimizing the backbone predictor and multiscale fusion modules.The amino acid predictor is trained separately from density and is used later for sequence assignment and chain construction.Source: §E.2 · p. 25

E.3 STRUCTURE-STAGE TRAINING

• At inference, structure refinement starts from protein sites predicted by backbone localization, not the reference-based initialization used during training.Training perturbs reference coordinates with 1 Å Gaussian noise and sometimes randomizes frame rotations.Source: §E.3 · p. 25
• Structure training selects 128 nearby residues around a sampled center and jointly optimizes the structure backbone, torsion head, candidate encoder, and conditioning modules.The structure objective uses backbone-coordinate and torsion losses, with one refinement forward pass and no recycling.Source: §E.3 · pp. 25–26
• For each initial protein frame, up to 32 predicted hetero candidates within 12 Å are selected; their positions and neighborhood membership stay fixed during a forward pass.Candidate geometry is recomputed as protein frames change.Source: §E.3 · p. 25

F EVALUATION DETAILS

• TM-score uses protein-only US-align and is normalized by reference protein length, while lDDT is Cα-only and includes mapped intra-chain and inter-chain reference pairs within 15 Å.Coverage is the fraction of reference residues in the US-align correspondence and differs from backbone recall because it has no 3 Å positional criterion.Source: F EVALUATION DETAILS · p. 26
• Backbone precision and recall use greedy one-to-one matching of predicted and reference Cα sites within 3 Å in the original map coordinate frame, without structural superposition.Precision corresponds to backbone accuracy and recall to backbone completeness.Source: F EVALUATION DETAILS · p. 27
• Backbone RMSD measures coordinate error only among matched sites, without additional rigid-body fitting, so it reflects localization conditional on recovery.Unmatched sites are excluded.Source: F EVALUATION DETAILS · p. 27
• Backbone break rate evaluates missing connections only when both reference edge endpoints have been recovered; edges with a missing endpoint are excluded.Reference and predicted edges are formed from adjacent residues with consecutive identifiers in the same chain and Cα distances from 2.5 to 4.5 Å.Source: F EVALUATION DETAILS · p. 27
• Q-score compares sampled map density around recognized protein atoms with an ideal atom-centered Gaussian profile and reports an atom-level mean.Sampling spans radii from 0 to 2.0 Å in 0.1 Å increments, with eight positions per radius and Gaussian width σ = 0.6 Å.Source: F EVALUATION DETAILS · p. 27
• Hetero-neighborhood metrics assess backbone localization within 4 Å of fixed reference hetero atoms, and contact metrics count protein-residue–hetero-object pairs within 4 Å.The contact measures evaluate the reconstructed protein environment around reference hetero components, not hetero detection or hetero structure prediction.Source: F EVALUATION DETAILS · p. 28

G.1 HETERO DETECTION RESULTS

• The detector is designed to provide hetero context, not reconstruct complete hetero structures.The paper contrasts this setting with component-specific approaches that may assume additional information such as sequences, ligand identities, binding sites, or fitted protein structures.Source: §G.1 · p. 28
• Stage 1 uses multiscale dense detector features, whereas Stage 2 uses sparse candidates whose contributions are controlled by confidence and learned gates.The paper states that partial but reliable hetero predictions can still provide useful information for protein reconstruction.Source: §G.1 · p. 28
• The class-wise detection table reports precision, recall, and F1 for each of the five hetero classes.Source: Table 16 · p. 29