acceptodds
Under review as a conference paper at ICLR 2027

Learning from Hetero Density for Cryo-EM Protein Reconstruction

Abstract

Reconstructing protein structures from cryo-electron microscopy (cryo-EM) maps is essential for understanding macromolecular assemblies. Although learning-based methods have improved protein reconstruction, information from hetero components remains underused. Our analysis finds both false predictions and reference protein sites near hetero components; filtering nearby candidates can improve or impair chain construction. We introduce CryoCue, a framework that uses hetero information to guide protein reconstruction. An anchor-supervised detector learns hetero representations across five component classes. Multiscale hetero features guide backbone localization, while predicted hetero candidates condition structure refinement through their class, confidence, and frame-relative geometry. Experiments show that CryoCue improves backbone localization near hetero components and achieves more accurate protein structure reconstruction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cryo-electron microscopy, protein structure reconstruction, non-protein density

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_cNhyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CRYOCUE uses predicted non-protein density to improve protein model building from cryo-EM maps. An anchor-supervised detector represents five hetero classes and supplies multiscale features to a backbone-localization network. Sparse predicted hetero candidates subsequently condition structure refinement through class information, confidence, and geometry relative to evolving protein frames. The method improves structural agreement, coverage, and connectivity over the reported baselines,

Critical strengths
  1. The scientific motivation is specific and well supported. The preliminary analysis shows why uniformly suppressing candidates near hetero components can help some cases and damage others.
  2. The design connects the representation to both stages of reconstruction.** Anchor targets provide a consistent localization task across disparate component shapes; coarse attention and fine-scale gates expose hetero context during backbone localization; frame-relative conditioning makes that context responsive to structure updates.
  3. The performance is good on benchmarks.
  4. The data protocol may be instructive to other cryo-em methods.
Critical weaknesses
  1. The most important issue is that the ablations do not fully isolate hetero semantics from capacity and auxiliary supervision. The additional detector is supervised on protein backbone and side-chain classes as well as hetero classes. It is hard to tell whether the structure prediction network use the protein feature or use the hetero feature. Maybe the detector provide additional protein signals and do not use hetro features at all. A matched auxiliary stream trained only on protein/background labels, or a collapsed-hetero-class control, would test whether the benefit requires the proposed semantic representation.

  2. External-baseline comparability is not sufficiently documented. The paper specifies its own large training inventory and staged optimization, but does not provide comparably detailed baseline configurations, training data, checkpoint provenance, tuning budgets, or overlap checks.

  3. As Table 16 reports, the different classes of hetero have largely different accuracies. F1 values are 76.04% for nucleic acid, 22.77% for glycan, 26.85% for nucleotide/cofactor, and 14.07% for lipid/detergent. Partial detections may still provide useful context, as the authors explain, but aggregate reconstruction gains do not show that all component classes benefit. A per-class dropout experiment may be valuable.

  4. Anchor supervision brings large improvement, from 74.7 to 84.6. But 74.4 is even lower than protein-only baseline, maybe because anchor change how we positive region number or ratio. So this need be claimed whether the improvement is brought from the anchor data curation or real anchor. Maybe author should clarify the supervision region data curation in details.

Questions for authors

Is it necessary to distinguish between different classes of hetero as the final task is for protein structure. Will merging all classes of hetero into one imparing the final performance?

7 comments
AuthorsOpenReview ↗

Response to W1: Hetero Semantics and Model Capacity

Response to Reviewer cNhy

We thank the reviewer for the thoughtful and constructive review. We are pleased that the reviewer recognizes the well-supported scientific motivation, hetero-aware conditioning across both reconstruction stages, strong experimental performance, and the potential value of our data processing protocol. Below, we address each concern and provide further clarification of our experimental setup and methodological design.

Responses to Critical Weaknesses

W1: The most important issue is that the ablations do not fully isolate hetero semantics from capacity and auxiliary supervision.

W1.1: The additional detector is supervised on protein backbone and side-chain classes as well as hetero classes. It is hard to tell whether the structure prediction network use the protein feature or use the hetero feature. Maybe the detector provide additional protein signals and do not use hetro features at all.

A: We understand the reviewer's concern that CryoCue's improvements may arise from additional model capacity or auxiliary protein supervision rather than hetero-specific representations. Our architecture-matched Protein-only and Collapsed-hetero ablations provide evidence that both hetero-label supervision and class-specific representations contribute to the improvement beyond matched model capacity and protein supervision. Specifically:

(1) Matched model capacity and protein supervision do not explain the full performance gain. The Protein-only variant retains the same detector and reconstruction architecture, including auxiliary protein supervision, but removes hetero-class supervision. CryoCue improves TM-score from 0.783 to 0.846, lDDT from 0.712 to 0.789, and Coverage from 77.73% to 85.02%. Since the architecture and auxiliary protein supervision are matched, these gains cannot be attributed to additional model capacity or protein supervision alone.

(2) Preserving hetero-class distinctions improves reconstruction beyond identifying hetero regions. We further introduce a Collapsed-hetero control that merges all five hetero classes into a single category while retaining the same architecture and protein supervision. As shown in Table A, Collapsed-hetero already achieves a TM-score of 0.833, outperforming the Protein-only variant (0.783). Importantly, distinguishing the five hetero classes further improves TM-score from 0.833 to 0.846, lDDT from 0.773 to 0.789, and Coverage from 83.82% to 85.02%. This indicates that CryoCue benefits not only from identifying hetero regions but also from their class-specific representations.

W1.2: A matched auxiliary stream trained only on protein/background labels, or a collapsed-hetero-class control, would test whether the benefit requires the proposed semantic representation.

A: Both suggested controls are included: the architecture-matched Protein-only control isolates the additional benefit of hetero-label supervision, while the Collapsed-hetero control demonstrates the benefit of retaining separate hetero classes. As shown in Table A and detailed in W1.1, the controls achieve TM-scores of 0.783 and 0.833, respectively, compared with 0.846 for full CryoCue.

Table A. Ablation of hetero supervision and semantic granularity.

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
EMProt0.8050.7481.070.5760.6041.2
Protein-only (matched architecture)0.7830.71277.730.6180.5961.55
Collapsed-hetero0.8330.77383.820.5790.5880.86
CryoCue (five hetero classes)0.8460.78985.020.5710.5930.69
AuthorsOpenReview ↗

Response to W2: Baseline Comparability and Reproducibility

W2: External-baseline comparability is not sufficiently documented.

W2.1: The paper specifies its own large training inventory and staged optimization, but does not provide comparably detailed baseline configurations, training data, checkpoint provenance, tuning budgets, or overlap checks.

A: We understand the reviewer's concern regarding baseline configurations, checkpoint provenance, tuning budgets, and potential training–test overlap. All external baselines were evaluated using their official pretrained checkpoints and recommended inference procedures, without retraining or additional hyperparameter tuning; we also examined potential direct training–test overlap using available documentation. Specifically:

(1) Baselines use official pretrained checkpoints without additional training or tuning. We evaluated EMProt, E3-CryoFold, ModelAngelo, EModelX, and CryoBoltz using their publicly released pretrained checkpoints and the inference procedures provided in their official GitHub repositories and README instructions. All methods were evaluated on the same 300-map test set using their respective supported inputs and a consistent evaluation protocol. No baseline was retrained, fine-tuned, or subjected to additional hyperparameter searches. Table B summarizes the experimental configurations.

(2) Training–test overlap checks are based on available inventories and documentation. We examined the available training-data information and checkpoint documentation for each baseline. For EMProt, E3-CryoFold, and EModelX, comparisons against their available training inventories identified no direct PDB/EMDB entry overlap with our test set. For ModelAngelo, the documented training-data cutoff (April 2022) precedes the release of our test structures. For CryoBoltz, we additionally checked for direct PDB entry overlap with the Boltz-1 structural training data, which uses a September 30, 2021 cutoff.

Table B. External baseline configurations and evaluation protocol.

MethodCheckpoint sourceInference configurationAdditional training or tuning
EMProtOfficial pretrained weightsOfficial pipeline; three de novo recycling rounds and sequence-guided assemblyNone
E3-CryoFoldOfficial pretrained weightsOfficial Chroma-based backbone generation and sequence-alignment pipelineNone
ModelAngeloOfficial pretrained weights (v1.0.18)Official Cα/GNN inference and sequence-guided model buildingNone
EModelXOfficial pretrained weightsOfficial backbone prediction, sequence alignment, and structure completionNone
CryoBoltzOfficial pretrained weightsOfficial GitHub/README inference procedureNone
AuthorsOpenReview ↗

Response to W3: Class-wise Hetero Contributions

W3: As Table 16 reports, the different classes of hetero have largely different accuracies.

W3.1: F1 values are 76.04% for nucleic acid, 22.77% for glycan, 26.85% for nucleotide/cofactor, and 14.07% for lipid/detergent. Partial detections may still provide useful context, as the authors explain, but aggregate reconstruction gains do not show that all component classes benefit.

A: We understand the reviewer's concern that the large variation in hetero detection accuracy makes it unclear whether all five component classes contribute to protein reconstruction. Our per-class evaluation and newly added dropout experiments provide evidence that all five hetero classes offer useful reconstruction cues, including classes with relatively low detection F1 scores. Specifically:

Class-wise reconstruction improvements extend to all five hetero classes. Although detector F1 varies considerably across hetero classes, CryoCue consistently improves local Cα localization and protein–hetero contact recovery over EMProt for all five classes, as shown in Table C. This indicates that imperfect hetero detection can still provide useful structural context.

Table C. Class-specific detection and reconstruction performance.

Hetero classDetector F1 (%)EMProt Cα F1 (%)CryoCue Cα F1 (%)EMProt Contact F1 (%)CryoCue Contact F1 (%)
Nucleic acid76.0485.93789.90275.83678.621
Glycan22.7780.4985.42775.24378.426
Nucleotide/cofactor26.8586.47590.33579.25281.734
Other ligand49.195.79998.48583.01785.074
Lipid/detergent14.0793.20996.35579.68682.55

W3.2: A per-class dropout experiment may be valuable.

A: Removing the conditioning cues of any individual hetero class degrades reconstruction performance across all six reported metrics. We further remove the conditioning cues from each hetero class individually while keeping the trained model unchanged. As shown in Table D, removing any class consistently degrades all six reconstruction metrics. For example, removing nucleic-acid cues reduces TM-score from 0.846 to 0.835, while removing lipid/detergent cues reduces it to 0.843. These results provide additional evidence that conditioning cues from each hetero class contribute to reconstruction, rather than the improvements being driven solely by classes with high detection accuracy.

Table D. Effect of removing individual hetero-class cues during inference.

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
w/o nucleic-acid cues0.8350.77683.910.5840.5880.84
w/o glycan cues0.8420.78484.610.5760.5910.75
w/o nucleotide/cofactor cues0.840.78284.430.5780.590.78
w/o other-ligand cues0.8390.78184.280.580.5890.8
w/o lipid/detergent cues0.8430.78584.730.5750.5920.74
CryoCue (Full)0.8460.78985.020.5710.5930.69
Reviewer_otLzOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a novel atomic model building method from cryo-EM density maps, CryoCue. The method is motivated by the observation that in numerous cases, density maps consist not only of the protein structures, but also non-protein molecules such as nucleic acids, glycans, cofactors, and ligands. These molecules are named “hetero” components, and the CryoCue method serves as a specialized deep learning method focusing on maps that consist of both protein and hetero components. The hetero components provide “cues” for the protein structure model to guide the generation. This has been achieved by a novel 3D convolutional neural network architecture adapted from SCUNet (Zhang et al., 2023). The model pipeline consists of two stages. The first stage consists of two parallel streams: in the first stream, the voxels which are candidates for being labeled a hetero component are identified; whereas in the second stream, the protein backbone locations (Ca) are predicted. In the second stage, the backbone structure is refined on the map through the conditioning provided by the hetero candidate embeddings. Using the refined geometry and torsions, sequence information is used to construct the all-atom model. The merit of the method is measured on a curated dataset of deposited map-model pairs, achieving state-of-the-art results in structure reconstruction. The paper provides ablation and hyperparameter studies as well, substantiating the design rationale behind the proposed architecture.

Critical strengths

The method addresses an important problem in a detailed study. Utilizing the knowledge of hetero components in a structure can, in theory, contribute to better automated model building into density maps. This paper demonstrates a successful application of this hypothesis. The proposed adaptation of the 3D neural network architecture to the domain is well-motivated and its pipelines correlate clearly to the dual concerns of identifying hetero components within a map, and guiding the model generation accordingly. Results reported highlight the strength of the proposed method, and the appendices of the work are well-prepared with the dataset filtering steps as well as algorithmic method details.

Critical weaknesses

The quantitative success reported in the results do not directly imply practical applicability. Numerous common bottlenecks and design trade-offs explored in atomic model building literature have been left unexplored. The number of chains and residues that the model building methods could handle is one of them. Another one is map resolution for which the atomic model is being built on. A few directional questions in this regard are provided beow.

Questions for authors

(Q1) Does this method work only on a single chain or multiple chains? What are some model/resource constraints on the number of residues that the method could build? (Q2) It is my understanding that the model currently takes in maps on a D=48 box sized voxel grid. For larger complexes, this box size might be quite small so that the Nyquist resolution might not allow necessary precision levels for building an atomic model. Could the authors provide an explanation for what sorts of systems were studied, e.g. with respect to molecular weight and map pixel sizes (A/pix)? (Q3) The decision to model the hetero structures only through anchor points is supported with ablation experiments that show decreased performance when all atoms in a hetero structure are considered. It could be the case that the undersampling on the map is clumping many atoms into a single voxel, hence lending itself for more for an anchor approach. Is this observed? (Q4) Can the method perform if there are no hetero components in a map, or does it provide spurious predictions? (Q5) Can the authors provide a few more case studies similar to Figure 5? (Q6) How does the method against more recent baselines, such as CryoBoltz*?

  • Raghu et al., Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data, NeurIPS 2025.
8 comments
AuthorsOpenReview ↗

Response to W1 and Q1: Practical Applicability and Large Complexes

Response to Reviewer otLz

We thank the reviewer for the thoughtful and constructive review. We are pleased that the reviewer recognizes the importance of the research problem, well-motivated architecture and reconstruction pipeline, strong experimental results, and comprehensive supplementary materials. Below, we address each concern and provide further experimental evidence and clarification.

Response to Critical Weakness

W1: The quantitative success reported in the results do not directly imply practical applicability. Numerous common bottlenecks and design trade-offs explored in atomic model building literature have been left unexplored. The number of chains and residues that the model building methods could handle is one of them. Another one is map resolution for which the atomic model is being built on.

A: We understand the reviewer's concern regarding practical applicability, particularly complex size, computational resources, and map sampling. Our test-set analysis shows that CryoCue handles multi-chain and large protein complexes through local sliding-window processing, with computational costs that depend on map volume and the number of predicted candidates, rather than a fixed input-size constraint imposed by the local crops. We address chain and residue scale in Q1, and voxel sampling and map resolution in Q2, with detailed results in Tables A and B.

Responses to Questions for Authors

Q1: Does this method work only on a single chain or multiple chains? What are some model/resource constraints on the number of residues that the method could build?

A: We clarify how CryoCue handles multi-chain complexes and the computational constraints associated with large protein structures. CryoCue supports both single-chain and multi-chain systems, with 81% of our test maps containing multiple protein chains. Our evaluation covers structures with up to 180 chains and 37,860 resolved protein residues. Specifically:

(1) CryoCue models multi-chain complexes without a fixed chain or residue limit imposed by local input crops. CryoCue performs backbone localization over the entire density map without requiring predefined chain identities. The predicted protein sites are subsequently refined and assembled into chains using sequence information. Its local density crops and structure refinement neighborhoods therefore do not impose a fixed limit on the total number of chains or residues.

(2) Inference remains feasible for large complexes, although runtime increases with structure size. The main computational costs depend on map volume, the number of predicted candidates, and subsequent structure assembly. As shown in Table A, inference time increases with complex size, averaging 5.15 minutes per map across the test set. The maximum observed peak GPU memory allocation is approximately 8.1 GiB on an A100 40GB GPU. For example, CryoCue processes 9YH1, a complex containing 33 protein chains and 16,929 resolved residues, in 48.62 minutes.

Table A. Test-set scale and inference runtime.

Protein residuesMapsMean runtime (min/map)Median runtime (min/map)
≤1,0001133.272.84
1,001–5,0001594.634.08
>5,0002815.6912.76
Overall3005.153.91

Runtime measurements use one A100-SXM4-40GB GPU per map and include Stage 1, amino-acid prediction, Stage 2 refinement, and chain assembly, starting from prepared inputs.

AuthorsOpenReview ↗

Response to Q2: Sliding-window Processing and Spatial Resolution

Q2: It is my understanding that the model currently takes in maps on a D=48 box sized voxel grid. For larger complexes, this box size might be quite small so that the Nyquist resolution might not allow necessary precision levels for building an atomic model. Could the authors provide an explanation for what sorts of systems were studied, e.g. with respect to molecular weight and map pixel sizes (A/pix)?

A: We understand the reviewer's concern that a 48³ voxel input may restrict the spatial coverage and resolution required for large-complex model building. The 48³ grid is a local sliding window rather than a downsampled representation of the entire map. CryoCue processes large complexes through overlapping windows at a fixed spatial sampling interval, without restricting the overall map size. Specifically:

(1) Sliding-window processing preserves the sampling interval when processing larger maps. CryoCue resamples maps to 1 Å/voxel and processes overlapping 48³ crops with a stride of 16 voxels. Predictions are merged into the full-map coordinate frame for subsequent structure refinement and chain assembly. Thus, larger complexes require more windows rather than coarser spatial sampling. The 48³ crop determines the local field of view, not the Nyquist resolution; the 1 Å sampling grid corresponds to a nominal Nyquist resolution of 2 Å. Resampling does not improve the intrinsic resolution of the original density map.

This patch-based strategy is also used in existing cryo-EM model-building methods. For example, EMProt uses 48³ density crops, while Cryo2Struct adopts 32³ crops. Within each crop, CryoCue further employs global self-attention and multiscale feature fusion to capture structural context at different spatial scales.

(2) The benchmark includes large complexes and diverse map sampling characteristics. As shown in Table B, our 300-map benchmark covers protein complexes ranging from 14.6 to 4,212.7 kDa, with original pixel spacings of 0.445–1.486 Å/pixel and reported resolutions of 1.45–4.00 Å. The largest complex, 9K3N, contains 180 protein chains and 37,860 resolved residues, demonstrating the applicability of our sliding-window approach to large macromolecular assemblies.

Table B. Molecular size and map sampling characteristics of the 300-map benchmark.

PropertyMinMedianMax
Protein polymer molecular weight (kDa)14.625176.4374,212.74
Total polymer molecular weight, including nucleic acids (kDa)14.625181.8534,376.47
Resolved protein residues1101,21637,860
Protein chains14180
Original pixel spacing (Å/pixel)0.4450.8651.4863
Reported map resolution (Å)1.453.034
AuthorsOpenReview ↗

Response to Q3: Atomic Overlap and Anchor Supervision

Q3: The decision to model the hetero structures only through anchor points is supported with ablation experiments that show decreased performance when all atoms in a hetero structure are considered. It could be the case that the undersampling on the map is clumping many atoms into a single voxel, hence lending itself for more for an anchor approach. Is this observed?

A: We understand the reviewer's suggestion that multiple hetero atoms may fall into the same voxel due to limited spatial sampling, potentially explaining the benefit of anchor supervision. Our analysis shows that direct atom-to-voxel collisions are rare. A key benefit of anchor supervision is to convert heterogeneous all-atom regions into more consistent representative-point localization targets. Specifically:

(1) Direct voxel collisions are rare, while anchor supervision reduces redundant and uneven targets. We examine atom-to-voxel collisions within 439 sampled density crops. Among 143,939 voxels occupied by 144,033 hetero heavy atoms, only 94 voxels (0.0653%) contain multiple atoms, indicating that direct voxel collisions are rare. However, supervising regions around all hetero atoms still produces redundant and uneven targets due to differences in component size and geometry. Anchor supervision simplifies this learning task by providing localized and consistent supervision. As shown in Table C, it reduces the positive hetero voxel ratio from 6.46% to 1.47%, cross-class target overlap from 0.0333% to 0.0002%, and class-wise variation from 1.29 to 0.97.

(2) Random-anchor ablation supports benefits from both localized supervision and representative anchor selection. To distinguish the benefit of localized supervision from chemically representative anchor selection, we introduce a Random-anchor control that selects one fixed random heavy atom per component under the same supervision procedure. As shown in Table D, Random-anchor improves TM-score from 0.747 to 0.817 over No-anchor, while chemically defined anchors further improve it to 0.846. These results support the benefits of both localized supervision and representative anchor selection.

Table C. Comparison of hetero supervision characteristics.

Supervision propertyNo-anchorAnchor
Positive hetero voxels (%)6.4641.468
Hetero support changed to ignore (%)077.289
Cross-class target overlap (%)0.033340.00018
Positive voxels per component-window instance305.11469.295
Class-wise variation (CV) ↓1.2870.967

Statistics are computed from 439 hetero-enriched 48³ crops, comprising 48,545,280 valid voxels and 10,285 component-window instances. Percentages refer to valid window voxels unless otherwise specified.

Table D. Ablation of anchor supervision.

MethodTM-score ↑lDDT ↑Coverage (%) ↑
Protein-only0.780.71177.44
No-anchor0.7470.69275.8
Random-anchor0.8170.75882.31
CryoCue (Anchor)0.8460.78985.02
Reviewer_i8ytOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

CryoCue uses non-protein (hetero) density to guide cryo-EM protein model building. A preliminary analysis with Cryo2Struct shows that protein candidates near hetero components can be either false or useful, so uniform filtering has mixed effects. CryoCue therefore trains an anchor-supervised detector for five hetero classes. Its multiscale features guide backbone localization, and its sparse candidates condition IPA-based refinement through frame-relative geometry. On 300 test maps, CryoCue reports the best TM-score, lDDT, coverage, and break rate, along with higher recall near hetero components. However, its Q-score and hetero-neighborhood precision remain below ModelAngelo.

Critical strengths
  1. Most model-building methods treat non-protein density as a nuisance or ignore it. This paper asks whether that density can help, which is a useful shift for a field that has focused on protein-derived information. Its preliminary analysis also gives the design a concrete motivation by showing that hetero neighborhoods can contain both false predictions and real protein support.

  2. The method outperforms strong baselines on most structural metrics at a practical runtime.

  3. The paper is open about its limits, including the lack of Q-score improvement.

Critical weaknesses
  1. The main claim is confounded. The detector also predicts protein backbone and side-chain classes, so its features are not purely hetero features. The fusion stream also adds capacity, and “Protein-only” is not capacity-matched. Part of the +0.066 TM-score gain may therefore come from an auxiliary protein-aware network rather than hetero information. A capacity-matched control is needed, such as the same stream trained with protein labels only. Results should also be reported separately for maps with no hetero components.

  2. The results do not show that the gains are specific to hetero regions. Recall improves by a similar amount near hetero components and overall. Table 2 also pools all classes, so the paper cannot show the class-dependent behavior that its premise requires.

  3. Detector F1 scores are very low for lipids/detergents (14.07%), glycans (22.77%), and cofactors (26.85%), limiting reliable cues for non-nucleic classes.

  4. The tables report only means, and the drop of the no-anchor variant below Protein-only needs a better explanation than the one given.

Questions for authors
  1. How the metrics in Table 2 look when classes are considered separately?
  2. Stage 2 is trained only on perturbed reference residues, so it never sees false protein sites near hetero density. It is unclear how hetero conditioning handles the problem that motivates the method. Could you explain more in this regard?
9 comments
AuthorsOpenReview ↗

Response to W1.1: Hetero Information and Model Capacity

We thank the reviewer for the thoughtful and constructive review. We are pleased that the reviewer recognizes the effectiveness of hetero-aware reconstruction, strong overall structural performance, and improvements in backbone recovery and chain continuity. Below, we address each concern and provide additional experimental evidence and methodological clarifications.

Responses to Critical Weaknesses

W1: The main claim is confounded.

W1.1: The detector also predicts protein backbone and side-chain classes, so its features are not purely hetero features. The fusion stream also adds capacity, and “Protein-only” is not capacity-matched. Part of the +0.066 TM-score gain may therefore come from an auxiliary protein-aware network rather than hetero information.

A: We understand the reviewer's concern that CryoCue's improvements may arise from additional model capacity, global context, or auxiliary protein supervision rather than hetero information. The architecture-matched Protein-only control suggests that hetero-label supervision contributes additional benefits beyond model capacity and protein supervision.

(1) The architecture-matched control shows substantial additional benefits from hetero-label supervision. We introduce a Protein-only control with exactly the same architecture as CryoCue, differing only in the removal of hetero-label supervision. This preserves model capacity, global context, and auxiliary protein supervision. As shown in Table A, the control achieves a TM-score of 0.783, only slightly above the original Protein-only baseline (0.780), while CryoCue achieves 0.846. These results indicate that additional model capacity and protein supervision alone cannot explain the full improvement.

(2) Protein-density supervision helps distinguish protein and hetero signals. The hetero detector also predicts protein backbone and side-chain density to better distinguish protein from hetero signals. Since hetero components are generally smaller and sparser than protein regions, treating protein density as background can increase class confusion. Unlike precise Cα localization, the detector learns the spatial support of protein density, providing discriminative context for identifying different hetero classes.

Table A. Comparison with the architecture-matched Protein-only control.

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
Protein-only0.780.71177.440.6090.6021.54
Architecture-matched Protein-only0.7830.71277.7340.6180.5961.546
CryoCue0.8460.78985.020.5710.5930.69
AuthorsOpenReview ↗

Response to W1.2: Capacity Control and Hetero-stratified Evaluation

W1.2: A capacity-matched control is needed, such as the same stream trained with protein labels only. Results should also be reported separately for maps with no hetero components.

A: We provide both the capacity-matched control in Table A and a hetero-stratified evaluation. The gains are larger on hetero-containing maps than on maps without hetero annotations.

The hetero-containing subset shows larger improvements than the subset without hetero annotations. We divide the 300 test maps into 273 with hetero annotations and 27 without. As shown in Tables B and C, CryoCue performs well on both subsets, with larger gains over EMProt on hetero-containing maps. TM-score improves by 0.043 versus 0.023, lDDT by 0.051 versus 0.029, and Coverage by 4.20 versus 1.43 percentage points, respectively.

This behavior is consistent with our two-stage design. Joint protein–hetero supervision encourages discriminative density representations that generalize to maps without annotated hetero components. When hetero information is present, multiscale cross-attention and gated fusion further guide backbone localization, while hetero-conditioned refinement exploits predicted class, confidence, and frame-relative geometry to improve protein model building.

Table B. Structural model-building performance on maps with and without hetero annotations.

With hetero annotations (n = 273)

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
EMProt0.8080.74481.380.5720.6031.19
E3-CryoFold0.7250.52668.311.0780.3758.24
ModelAngelo0.7270.65773.140.590.6122.65
EModelX0.8120.69583.070.8380.4496.79
CryoCue (Ours)0.8510.79585.580.5680.5930.66

Without hetero annotations (n = 27)

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
EMProt0.7760.70177.970.6160.6161.26
E3-CryoFold0.5650.3754.311.1580.31113.62
ModelAngelo0.7370.66874.030.6160.6242.18
EModelX0.7390.6175.940.8870.4328.64
CryoCue (Ours)0.7990.7379.40.6060.5970.97

Table C. Backbone localization and protein–hetero contact performance on maps with and without hetero annotations (%).

With hetero annotations (n = 273)

MethodCα P ↑Cα R ↑Hetero Cα P ↑Hetero Cα R ↑Contact P ↑Contact R ↑
EMProt97.1783.10793.03586.18185.64673.993
E3-CryoFold86.73680.44571.63876.60667.76259.699
ModelAngelo94.28481.38397.51184.93787.76573.145
EModelX93.54984.0875.84686.83271.41964.636
CryoCue (Ours)97.55986.16796.27189.88386.39677.877

Without hetero annotations (n = 27)

MethodCα Precision ↑Cα Recall ↑
EMProt90.10478.307
E3-CryoFold78.70373.478
ModelAngelo87.66279.627
EModelX86.84978.269
CryoCue (Ours)90.20380.9

Hetero-neighborhood and protein–hetero contact metrics are undefined for maps without reference hetero annotations.

AuthorsOpenReview ↗

Response to W2.1: Overall and Hetero-neighborhood Improvements

W2: The results do not show that the gains are specific to hetero regions.

W2.1: Recall improves by a similar amount near hetero components and overall.

A: The reviewer notes that the improvements in overall and hetero-neighborhood recall are similar. The gains extend beyond immediate hetero neighborhoods, consistent with multiscale hetero-context fusion.

Compared with EMProt, CryoCue improves overall Cα recall by 3.018 percentage points and hetero-neighborhood Cα recall by 3.702 percentage points. These observations suggest that CryoCue's benefits are not restricted to immediate hetero neighborhoods, which is consistent with the broader contextual modeling enabled by multiscale feature fusion. Coarse-scale cross-attention incorporates broader hetero context, while fine-scale gated fusion adaptively modulates local features, allowing hetero information to benefit protein localization beyond nearby residues.

Reviewer_HivsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper asks if non-protein ("hetero") density from cryo-EM maps can help automated atomic structure building, which includes nucleic acids, glycans, lipids, nucleotides and cofactors, ligands, etc. Most existing tools for automated model building focus on protein and DNA/RNA structures only.

The preliminary analysis presented in the paper shows that (1) the regions/neighborhoods around nucleic acid, glycan, and lipid residues can contain false-positive protein backbone atoms (C_alpha), whereas nucleotides/cofactors are typically found around the protein backbone, which agrees with intuition; and (2) simply removing candidate protein backbone atoms around hetero residues can improve model building on some maps but hurts for most hetero classes.

Based on this motivation, the authors design a multistage architecture to (1) detect and derive representations for hetero components, which adapts SCUNet for 3D with a global transformer as an anchor-supervised voxel detector; (2) predict protein backbone based on the (frozen) identified hetero component features using both coarse-scale features (via cross-attention) and fine-scale features (via gated convolution); and (3) "refine" the backbone into a full atomic structure using invariant point attention, similar to AlphaFold. While these individual components are standard, their combination for this particular task is unique and a new contribution.

On a homology-filtered test set, their method CryoCue reports improved performance against other common automated structure building baselines, including fewer chain breaks and better C_alpha recall near hetero residues. Ablations attribute these gains to both stages in the architecture.

Critical strengths
  1. The task is motivated well, particularly given growing interest in cryo-EM for resolving complexes that contain non-protein components. The authors note that 78% of the analyzed maps in the preliminary study contain non-protein density. The preliminary analysis section clearly shows that non-protein density affects model building with significant class dependence. This shows that masking, removing, or ignoring such density is suboptimal for model building, even though many model builders treat it as a nuisance rather than a source of information. The broader idea of using predictions of auxiliary ‘distractor’ classes as conditioning context (rather than masking them), may also be transferrable to other "dense" prediction problems, though this is not discussed in the paper.

  2. CryoCue improves upon recent baselines (i.e. EMProt, ModelAngelo, EModelX) on a 300-map held-out set, with gains in TM-score (0.846 vs. 0.805), coverage, and chain continuity (break rate of 0.69% vs. 1.20%), at an inference cost comparable to EMProt. The authors also transparently include expected metrics even where CryoCue does not improve (i.e. Q-score, near-hetero precision), which adds credibility.

  3. The documentation of details of the algorithm, data preparation and preprocessing, training, etc. in the appendices is particularly comprehensive and appreciated, and is above average for many conference papers. This includes definitions for all evaluation criteria and design choices like anchor supervision, particularly given the heavy class-imbalance of the training data (since nucleic acids are overrepresented hetero residues). The train/test split excludes training pairs with any chain at ≥25% identity and ≥80% coverage to a test chain, which I believe is a fairly stringent homology filter.

Critical weaknesses
  1. The central causal claim of the paper – that the performance gains come primarily from better handling of hetero density information – is not directly supported and is confounded with other factors. The hetero density detector is also trained to recognize protein backbone and side chains (it also includes a global transformer that the protein-only backbone predictor appears to lack). Thus, adding it contributes extra capacity, global context, and protein supervision, not just non-protein information. The protein-only pipeline appears to be weaker than the existing baselines (TM 0.780 in Table 3); it is unclear if the hetero modules actually add information or compensate for weaknesses in the protein pipeline. As one way to address this, the results should be split by maps with vs. without non-protein components. If the gains are also observed on maps without hetero content, they are not driven by hetero information. Additional ablations could also address this.

  2. The selective or class-dependent use of hetero density information is claimed but not demonstrated from what I can tell. The papers motivation is that hetero density should be treated differently by class, but there is no analysis to suggest that the model does this. Given the heavy class imbalance of the different hetero classes (primarily nucleic acids with 88% hetero voxels) and relatively small size of the test set, it would be helpful to include an analysis of the learned gates or attention weights by hetero class. Table 16 shows that nucleic acids are the only reliably detected class (F1 score of 76% vs. 14-49% for the others). In particular, a per-class breakdown of the results would be particularly insightful in Table 2; without that reporting, the hetero-neighborhood gains could be mostly nucleic acid-driven. I acknowledge that the per-class test subsets are small (e.g., 21 maps with lipids, 32 with glycans), which makes the variance reporting in point 3 more important.

  3. All of the reported results are map-level means from a single run. Given the limited size of the test set, running multiple seeds and comparing pairs of maps would be helpful and demonstrate which settings CryoCue improves upon the baselines. The case study is single example; additional examples would be nice to see (even if in the appendix). In addition, the baseline protocol is not specified. It is unclear if the baselines were run with released weights or retrained from the sample splits, for example. This affects how the comparison should be interpreted. F1 scores should also be reported for the hetero neighborhood and protein-hetero contact columns to clarify if the trade-off between recall and precision performance is a consistent improvement. DeepTracer 2.0, which segments protein from nucleic-acid density before modeling, is essentially the uniform 'segment then model' strategy that the paper argues against. DeepTracer 2.0 should be discussed and would be a natural baseline if feasible.

  4. Minor points but important for clarity: The manuscript uses some non-standard terminology that conflicts with established cryo-EM usage. "Reconstruction" for example conventionally refers to producing a density map from a set of particle images; "atomic model building" is typically used for the task in this paper. "Non-protein density" may be a clearer term than "hetero" particularly in the abstract, since cryo-EM involves many forms of "heterogeneity" and thus “hetero” may be confusing. Other key terms like "protein support" and "chain construction" should also be defined in the main text. Overall, much of the text is very jargon-heavy and dense; consider more proofreading to improve clarity. In addition, Figure 2 should not be a line plot since the x-axis is categorical; a stacked bar plot would be more reasonable. Figure 3 is very dense; a simplified pipeline overview may be better suited and save space, and detailed figures could be moved to the appendix.

Questions for authors
  1. Why were only 300 maps each (<2% of the data) used for validation and testing? The per-class test subsets are small as previously mentioned, which limits any per-class conclusions.
  2. Please describe the procedure for final chain assembly (after refinement) more clearly – for example, how were fragments identified, were residues filtered or removed, etc.?
  3. Report results separated by maps with vs. without non-protein components, as previously mentioned. This will help clarify if gains can be attributed to non-protein information rather than added capacity and global context.
  4. Report a breakdown of hetero-neighborhood and contact metrics in Table 2 by hetero class, and report the learned gate or attention statistics by class if feasible. Given that nucleic acids dominate the non-protein voxels and the detection results, are the gains present for other classes?
  5. Provide more detail in how the baselines were run (released weights or retrained, same input possible test overlap with their training data, etc.), as previously described. Also a discussion of performance at different resolution ranges would be helpful.
  6. The structure network is trained on only perturbed positive/true-labeled resides (no false positives). I'm a bit confused as to how the Stage 2 network learns to handle the spurious predictions near non-protein density that motivate the method. Please clarify.
  7. Consider making some of the rewording and figure changes suggested in point 4 above.
16 comments
AuthorsOpenReview ↗

Response to W1.1–W1.2: Hetero Information and Model Capacity

We sincerely thank the reviewer for the careful and constructive assessment. We appreciate the reviewer's recognition of our research motivation, which turns non-protein density from a traditionally overlooked source of interference into useful information for atomic model building. We are also grateful for the positive assessment of CryoCue's novelty, performance improvements over existing methods, and comprehensive experimental and implementation details. We address the reviewer's concerns point by point below.

Responses to Critical Weaknesses

W1: The central causal claim of the paper – that the performance gains come primarily from better handling of hetero density information – is not directly supported and is confounded with other factors.

W1.1: The hetero density detector is also trained to recognize protein backbone and side chains (it also includes a global transformer that the protein-only backbone predictor appears to lack). Thus, adding it contributes extra capacity, global context, and protein supervision, not just non-protein information.

A: We understand the reviewer's concern that CryoCue's performance gains may arise from additional model capacity, global context, or auxiliary protein supervision rather than hetero information. Our architecture-matched Protein-only control shows that additional architecture and protein supervision alone yield only marginal improvements, while hetero-label supervision provides substantial additional benefits.

(1) The architecture-matched control confirms that hetero supervision provides substantial additional benefits. We introduce a Protein-only control with exactly the same architecture as CryoCue, differing only in the removal of hetero-label supervision. This ensures that model capacity, global context, and auxiliary protein supervision are preserved, allowing us to better isolate the contribution of hetero-label supervision. The results show only marginal improvements over the original Protein-only baseline, whereas the full CryoCue achieves substantially better performance.

(2) Protein-density supervision helps distinguish protein and hetero signals. The hetero detector also predicts protein backbone and side-chain density to better distinguish protein from hetero signals. Since hetero components are generally much smaller than protein regions, treating protein density as background can exacerbate class confusion. Unlike precise Cα voxel localization, this module learns the spatial support of protein density, providing discriminative context for identifying different hetero classes.

Table A. Comparison with the architecture-matched Protein-only control.

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
Protein-only0.780.71177.440.6090.6021.54
Architecture-matched control0.7830.71277.7340.6180.5961.546
CryoCue0.8460.78985.020.5710.5930.69

W1.2: The protein-only pipeline appears to be weaker than the existing baselines (TM 0.780 in Table 3); it is unclear if the hetero modules actually add information or compensate for weaknesses in the protein pipeline.

A: Protein-only is an internal ablation designed to assess the contribution of CryoCue's hetero-aware components, rather than an independently optimized protein model-building method. EMProt and EModelX employ different structural modeling, sequence-guided chain completion, and postprocessing strategies. Therefore, their performance differences relative to Protein-only cannot be directly attributed to the presence or absence of hetero information.

AuthorsOpenReview ↗

Response to W1.3: Performance With and Without Hetero Components

W1.3: If the gains are also observed on maps without hetero content, they are not driven by hetero information. Additional ablations could also address this.

A: We understand the reviewer's concern about whether CryoCue's improvements depend on the presence of hetero components. CryoCue performs well on both subsets, with larger gains on maps containing hetero components, highlighting the benefits of hetero-aware protein model building. This can be explained by two aspects of our design:

(1) Joint supervision improves generalization even without annotated hetero components. Joint protein–hetero density supervision encourages the detector to learn discriminative representations that distinguish protein density from non-protein signals. Together with global contextual modeling, these representations can improve protein backbone localization even when no hetero components are annotated.

(2) Explicit hetero conditioning provides additional benefits when hetero components are present. When hetero density is present, CryoCue explicitly exploits hetero information at both stages. In Stage 1, multiscale cross-attention and gated fusion incorporate hetero features into backbone localization. In Stage 2, hetero-conditioned IPA refinement further leverages predicted hetero classes, confidence scores, and frame-relative geometry to optimize protein structures. These mechanisms enable more effective use of hetero context when it is available.

Experimental results. We divide the 300 test maps into 273 with hetero annotations and 27 without. Compared with EMProt, CryoCue achieves gains of 0.043 in TM-score, 0.051 in lDDT, and 4.20 percentage points in Coverage on maps with hetero annotations, compared with 0.023, 0.029, and 1.43 percentage points, respectively, on maps without them. Detailed results are provided below.

Table B. Structural model-building performance on maps with and without hetero annotations.

With hetero annotations (n = 273)

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
EMProt0.8080.74481.380.5720.6031.19
E3-CryoFold0.7250.52668.311.0780.3758.24
ModelAngelo0.7270.65773.140.590.6122.65
EModelX0.8120.69583.070.8380.4496.79
CryoCue (Ours)0.8510.79585.580.5680.5930.66

Without hetero annotations (n = 27)

MethodTM-score ↑lDDT ↑Coverage (%) ↑RMSD (Å) ↓Q-score ↑Break rate (%) ↓
EMProt0.7760.70177.970.6160.6161.26
E3-CryoFold0.5650.3754.311.1580.31113.62
ModelAngelo0.7370.66874.030.6160.6242.18
EModelX0.7390.6175.940.8870.4328.64
CryoCue (Ours)0.7990.7379.40.6060.5970.97

Table C. Backbone localization and protein–hetero contact performance.

With hetero annotations (n = 273)

MethodCα P ↑Cα R ↑Hetero Cα P ↑Hetero Cα R ↑Contact P ↑Contact R ↑
EMProt97.1783.10793.03586.18185.64673.993
E3-CryoFold86.73680.44571.63876.60667.76259.699
ModelAngelo94.28481.38397.51184.93787.76573.145
EModelX93.54984.0875.84686.83271.41964.636
CryoCue (Ours)97.55986.16796.27189.88386.39677.877

Without hetero annotations (n = 27)

MethodCα Precision ↑Cα Recall ↑
EMProt90.10478.307
E3-CryoFold78.70373.478
ModelAngelo87.66279.627
EModelX86.84978.269
CryoCue (Ours)90.20380.9
AuthorsOpenReview ↗

Response to W2.1: Class-dependent Gating and Attention

W2: The selective or class-dependent use of hetero density information is claimed but not demonstrated from what I can tell.

W2.1: Given the heavy class imbalance of the different hetero classes (primarily nucleic acids with 88% hetero voxels) and relatively small size of the test set, it would be helpful to include an analysis of the learned gates or attention weights by hetero class.

A: The reviewer asks whether CryoCue learns to utilize different hetero classes differently during protein model building. CryoCue adaptively incorporates hetero context through learnable attention and gating mechanisms rather than explicitly prescribing how each class should affect Cα predictions. Our class-wise analyses reveal distinct responses across hetero categories. Specifically:

(1) Learned gating responses differ across hetero classes. We analyze attention-weighted gate values across predicted classes. As shown in Table D, nucleotide/cofactor (0.1127) and nucleic acid (0.0893) exhibit stronger responses than protein backbone (0.0382), while glycan (0.0718), other ligand (0.0556), and lipid/detergent (0.0437) show different levels of modulation. These results indicate that CryoCue does not apply uniform gating across hetero classes.

(2) Attention allocation is not simply proportional to candidate frequency. We compare each class's candidate proportion with its attention share. As shown in Table E, nucleotide/cofactor, glycan, and nucleic acid receive 2.83×, 2.37×, and 1.80× their candidate proportions, respectively, whereas lipid/detergent receives 0.73×. This indicates that attention allocation is not simply proportional to candidate frequency.

Table D. Attention-weighted gate responses across predicted classes.

Predicted classMean weighted gate95% CI
Protein backbone0.0382[0.0367, 0.0398]
Nucleic acid0.0893[0.0821, 0.0976]
Glycan0.0718[0.0634, 0.0807]
Nucleotide/cofactor0.1127[0.1009, 0.1263]
Other ligand0.0556[0.0493, 0.0621]
Lipid/detergent0.0437[0.0369, 0.0518]

Gate statistics are grouped by predicted class. The 95% confidence intervals are estimated using map-level bootstrap resampling across the test set.

Table E. Candidate proportions and attention allocation across predicted hetero classes.

Predicted classCandidate shareAttention shareAttention / Candidate
Nucleic acid1.77%3.19%1.799×
Glycan0.06%0.14%2.367×
Nucleotide/cofactor0.04%0.12%2.830×
Other ligand1.77%2.07%1.172×
Lipid/detergent0.27%0.20%0.726×

Both shares are normalized over all predicted classes, including protein and background. Only hetero classes are displayed; therefore, the listed percentages do not sum to 100%.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.