acceptodds
Under review as a conference paper at ICLR 2027

Protein Localization & Identification Directly from CryoEM Maps

Abstract

Cryo-electron microscopy (cryoEM) produces three-dimensional maps of the electrostatic potential of protein complexes. Interpreting a map requires building an atomic model, which in turn requires knowing which protein occupies each region. Automated methods answer this by tracing the protein backbone and identifying side chains, and both steps fail below 5Å resolution. In this work, we propose a method that assigns proteins to regions of the map without having to build an atomic model. We align a frozen foundation model of cryoEM maps with a frozen protein language model in a shared embedding space, training only two lightweight heads on cached features. Two components are essential: a soft-target contrastive objective that treats solvent as an explicit class and distributes target mass over every chain of a complex, and a rigid-motion-invariant attention module over the sampled voxels of a map. Given a map and the sequences of the proteins it contains, our model assigns regions of the map to their corresponding proteins, outperforming controls. This performance is retained at resolutions where backbone tracing is no longer reliable. When the protein composition is unknown, the same embedding can instead retrieve candidate sequences from a large catalogue, placing the correct protein among the top-ranked candidates in more than 84% of cases. Together, these results show that protein identity can be inferred directly from cryoEM density without first constructing an atomic model.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
cryo-electron microscopy, cryo-EM, protein identification, protein localization, protein sequence, representation learning, multimodal learning, contrastive learning, protein language models, foundation models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VQ6bOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a method for identifying and localizing proteins to cryo-EM density maps without the need of atomic model. The approach is inspired by CLIP and uses two frozen models: a cryo-EM map encoder and protein language model to create a shared representation. The main contributions are distilled map encoder; soft-target, background-aware, cross-chain contrastive objective with independently controlled negatives; VoxMix, distance-based spatial attention module. The shared representation is used for voxel classification and sequence retrieval. Evaluation on 225 test maps show 65% assignment accuracy and retrieval of at least one correct sequence among the top 20 for 84% of maps. Validation includes feature extractor representation probing and spatial-mixing ablations but lacks direct comparisons with established protein identification and structure building methods. All original maps have resolution between 1.5-3.9A, with lower resolutions simulated by filtering. Furthermore, reference atomic models determine input voxels and assignment metrics score only foreground voxels. The experiments do not establish performance in the intended atomic model free, low resolution application.

Critical strengths

none

Critical weaknesses
  1. The intended low resolution application is not validated on real low resolution data. The main argumentation for this paper is that the standard approach for identifying and localizing proteins in maps requires building atomic models, which, according to authors, becomes unreliable at resolutions worse than 5A. That said, all tested maps are 1.5-3.9A, with lower resolutions only simulated by filtering the high resolution maps, which does not reproduce the noise and processing characteristics of real data.
  2. The claim of atomic model free inference is not demonstrated end-to-end. Although the network does not explicitly build an atomic model, the evaluation pipeline uses the reference atomic structure to select its input voxels. The reported metrics additionally score only selected foreground voxels, leaving false protein assignments outside that foreground unmeasured.
Questions for authors
  1. Could you evaluate the method on real low resolution experiment maps, without using reference atomic models for voxel selection or inference, only for ground truth labels?
2 comments

We thank the reviewer for their feedback. We have addressed both experimental-validation concerns through new experiments detailed in General Response:

  • Section 1: End-to-end inference without atomic coordinates, with modest reductions in voxel accuracy and retrieval mAP.
  • Section 2: Validation on genuine experimental maps at 4.3–5.8 Å, including direct comparisons demonstrating successful protein identification in cases where ModelAngelo (SOTA) identifies no proteins.

Regarding critical strengths and novelty, we believe the assessment does not fully recognize the following conceptual and methodological contributions:

  • Direct 3D density-to-sequence alignment: To our knowledge, the first ML approach to directly match protein sequences to spatial regions of experimental 3D cryo-EM density without atomic reconstruction.
  • Two-tower cross-modal learning: A framework aligning frozen cryo-EM and protein language models to support both spatial protein localization and sequence retrieval.
  • VoxMix spatial attention: A rigid-motion-invariant architecture that improves voxel accuracy from 39.7% to 65.0% and macro IoU from 24.4% to 48.1%, outperforming post-hoc Gaussian smoothing.
  • CleanDIFT for CryoFM2: Adaptation of diffusion-feature distillation to produce an efficient clean-input cryo-EM encoder, enabling lightweight alignment of frozen foundation models.

We do not claim that two-tower learning, attention, or distillation are individually new. Rather, our contribution lies in their formulation for spatially resolved density-to-sequence correspondence, supported by controlled ablations and now by experimental low-resolution validation.

In light of these contributions and the new evidence addressing both stated concerns, we respectfully ask the reviewer to reconsider their assessment and raise their score accordingly.

Thank you for the detailed rebuttal and additional experiments.

The new foreground segmentation experiment substantially addresses my concern regarding atomic-model-free inference. The evaluation on genuine low-resolution maps and comparison with ModelAngelo are also valuable additions.

My main remaining concern is the limited scope of the low-resolution validation. All nine experimental maps are ribosome-related, and only three have resolutions worse than 5 Å. Existing benchmarks such as Emap2sec and Emap2sec+ contain experimental maps spanning 5–10 Å with associated atomic structures.

Regarding novelty, I recognize the value of direct density-to-sequence alignment and the improvements introduced by VoxMix, even if the individual components are largely based on established techniques. My original concerns were primarily about experimental validation rather than methodological novelty.

Overall, the additional experiments substantially address one of my main concerns and provide encouraging evidence for the proposed approach. While the low-resolution validation remains limited, I acknowledge the novelty of the task and the difficulty of evaluating it against existing methods. I will therefore increase my rating from 1 to 2.

Reviewer_fkb1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies protein localization and sequence retrieval from cryo-EM maps by aligning a CryoFM2-derived map encoder with ESM-C sequence representations. CryoFM2 is first distilled into a clean-input encoder; the alignment stage uses frozen towers with trainable projection heads, a background-aware soft-target objective, and distance-biased voxel attention (VoxMix). The model assigns sampled voxels to candidate protein sequences and ranks sequences for a map. Evaluation uses 225 held-out maps with deposited atomic models, a catalogue of 1,966 test sequences, and artificially low-pass-filtered maps to explore lower resolutions.

Critical strengths
  1. The cross-modal formulation is potentially useful. The results suggest that cryo-EM map representations contain information that can be aligned with sequence representations without explicitly tracing a backbone in the prediction network. The reported voxel assignment accuracy of about 0.65, compared with about 0.31 for the longest-sequence prior, provides evidence of a useful signal under the evaluated conditions.

  2. The paper includes thoughtful controls and candidly reports limitations. Length/composition controls, shuffled pairing, a random-encoder control, and comparison with spatial smoothing help distinguish learned correspondence from simpler shortcuts. The authors also explicitly acknowledge dependence on atomic-model-based voxel selection, synthetic resolution degradation, and limited retrieval precision.

Critical weaknesses
  1. The evaluated pipeline does not yet establish the central practical claim of localization and identification without prior atomic modeling. Both training and evaluation voxels are selected using distances to deposited protein atoms. Appendix F states that density-threshold-based foreground selection failed and that model-free foreground selection remains future work. This does not directly give the network the sequence labels, but it supplies privileged spatial information that would be unavailable in the intended unknown-structure workflow. Moreover, all native maps have resolutions of 1.5–3.9 Å; the motivating lower-resolution regime is evaluated only by low-pass filtering these maps. Such filtering does not reproduce the noise and disorder of genuinely low-resolution data. Together, these limitations leave a substantial gap between the controlled benchmark and the setting in which the method is intended to be most useful.

  2. The methodological advance and demonstrated downstream benefit are currently limited. The approach combines established feature distillation, two-tower alignment, soft targets, and distance-based attention. Such an adaptation can be valuable, but its significance needs convincing evidence on the target problem. The full-pool retrieval results currently support shortlist generation rather than reliable identification: the best reported hit@20 is 83.6%, meaning that at least one correct sequence appears among twenty candidates, with corresponding recall@20 of 33.4%; Table 3 separately reports a top-ranked sequence present in only about 20.4% of maps. I did not find a real unknown-identity case independently validated after prediction, or a demonstrated downstream workflow benefit. Thus the current evidence is closer to a controlled proof of concept than a validated solution to the stated practical bottleneck.

Overall, I appreciate the controls and transparency, but currently lean toward weak rejection because the methodological contribution is modest and the most consequential application claims remain insufficiently validated.

Questions for authors
  1. Can the authors evaluate with foreground voxels selected without any atomic-model coordinates, even on a representative subset, and quantify the change in both localization and retrieval performance? This would directly affect my assessment of the claimed model-free workflow.

  2. Is there evidence on genuinely low-resolution experimental maps, preferably with identity or localization independently established, rather than only low-pass-filtered high-resolution maps? If not, can the scope of the conclusions be narrowed accordingly?

  3. Can the authors demonstrate that the retrieved shortlist improves an actual identification or fitting workflow relative to a practical existing approach, reporting both accuracy and computational cost? In particular, how useful are the false positives and missed components when a complex contains several proteins?

  4. Which result most clearly isolates the benefit of the proposed alignment design beyond standard two-tower alignment with spatial smoothing, and establishes that this gain matters in the intended low-resolution setting?

1 comment

We thank the reviewer for their constructive questions. Full experimental details are provided in General Response.

1. Can foreground selection avoid atomic coordinates?

Yes. Section 1 in General Response replaces structure-derived selection with a U-Net. Voxel accuracy changes from 0.647 to 0.617, macro IoU from 0.435 to 0.324, and retrieval mAP from 0.227 to 0.217. This removes the requirement for atomic coordinates at inference while retaining most localization accuracy and retrieval performance.

2. Is there evidence on genuinely low-resolution experimental maps?

Section 2 in General Response evaluates the end-to-end pipeline on nine experimental ribosomal-complex maps at 4.3–5.8 Å, with known protein identities. These experiments directly test real low-resolution data rather than low-pass-filtered high-resolution maps.

3. Does retrieval offer practical value relative to existing methods, including computational cost?

On EMD-8621 (5.8 Å), our method localizes 10/20 proteins and ranks four true sequences above all 2,000 decoys, while neither ModelAngelo workflow correctly identifies a protein. We interpret this as useful candidate prioritization, not complete recovery of complex composition.

On one H200 and four CPU cores, our pipeline typically takes approximately one minute per map with cached sequence embeddings. Embedding the candidate list adds 1.5–6 minutes once and is reusable across maps. ModelAngelo’s sequence-free building averages 10 minutes, followed by less than one minute for HMM search.

4. What is gained beyond standard two-tower alignment and spatial smoothing?

In the original ablations, VoxMix improves voxel accuracy from 39.7% to 65.0% and macro IoU from 24.4% to 48.1%, while also outperforming Gaussian smoothing. This supports learned chain-discriminative context beyond spatial smoothing alone. These ablations isolate the architectural benefit in the original benchmark; the new experiments demonstrate the complete pipeline’s low-resolution utility.

In light of these clarifications and new experimental evidence, we respectfully ask the reviewer to reconsider their assessment and raise their score.

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

The paper aligns cryo-EM map representations and protein sequence representations to localize known candidate proteins and retrieve candidate sequences without constructing an atomic model during network inference. It distills CryoFM2 into a clean-input encoder, freezes that encoder and ESM-C, and trains projection heads with a background-aware soft contrastive objective. Distance-biased voxel attention, VoxMix, adds spatial context. On 225 held-out maps, the depth-16 model achieves 0.6495 foreground voxel accuracy and 0.4813 macro IoU. Retrieval against 1,966 test sequences provides useful shortlists, but its top-ranked sequence is present in only 20.4% of maps under the reported presence score. Low-resolution behavior is studied through synthetic low-pass filtering, including mixed-resolution training.

Critical strengths
  1. The paper demonstrates a useful correspondence task beyond residue identification. The gain from the pointwise model to VoxMix is large: top-1 accuracy improves from 0.3968 to 0.6495 and macro IoU from 0.2440 to 0.4813 (Table 2). The two-block model also exceeds post-hoc Gaussian smoothing, so the improvement cannot be explained entirely by spatial coherence (Section 3.3; Table 5). This supports learning chain-discriminative context from frozen representations with relatively little trainable capacity.
  2. The evaluation takes shortcuts and statistical dependence seriously. Within-map assignment, volume/composition/length controls, shuffled pairing, and component-level bootstrap intervals address several plausible alternative explanations. Appendix B.4 explicitly accounts for the fact that 225 maps form only 89 independent sequence-cluster components. The paper also reports weak top-1 retrieval and poor cross-resolution transfer rather than hiding them behind the shortlist hit rate.
Critical weaknesses
  1. The practical model-free pipeline is not yet evaluated. Foreground sampling uses distance to deposited heavy atoms, excludes the intermediate shell, and removes voxels near map faces (Sections 2.2 and 4; Appendix B.4). Retrieval likewise uses these foreground voxels (Section 3.4). Consequently, the evaluation provides the method with a structure-derived selection of where to look, even when protein composition is supposed to be unknown. The network does not consume the atomic model, but the demonstrated pipeline still does. An inference-time selection based only on density, together with end-to-end assignment/background metrics, is necessary to establish the advertised application.
  2. The strongest application regime and practical retrieval comparison are insufficiently established. All real maps are 1.5–3.9 Å; the intended lower-resolution regime is simulated by filtering. The native model falls approximately to the volume prior after filtering to 8 Å, and retraining or mixed-resolution training is required (Section 3.5; Table 6). These are informative results, but they do not establish performance on noisy, disordered experimental maps beyond 5 Å. Furthermore, retrieval is evaluated against 1,966 sequences drawn from the test corpus, and the method is compared mainly with diagnostic controls rather than an operational identification workflow. The evidence currently supports a promising correspondence proof of concept more strongly than a ready identification method.
Questions for authors
  1. Model-free input selection: Can you replace structure-derived foreground sampling with a density-only mask and report assignment, background rejection, and full-pool retrieval on the same 225 maps? How much performance is lost when boundary/interface voxels are retained?
  2. Target application: Do you have results on genuinely experimental maps worse than 5 Å, using a model selected without those test maps? A small well-documented set would materially strengthen the low-resolution claim.
  3. Retrieval interpretation: Which exact checkpoint, residue reduction, and metric support “more than 84%” in the abstract? Please distinguish the probability of retrieving at least one true sequence within 50 candidates from per-sequence recall@50, and provide a comparison with a practical identification method where it is applicable, including time cost and failure cases.
2 comments

We thank the reviewer for recognizing the value of the spatial architecture and experimental controls. We address your questions below, with full results in General Response.

1. Can foreground selection be independent of atomic coordinates?

Yes. Section 1 in General Response replaces structure-derived foreground selection with a U-Net. Voxel accuracy changes from 0.647 to 0.617 and retrieval mAP from 0.227 to 0.217. Macro IoU decreases more substantially, from 0.435 to 0.324, identifying foreground segmentation as a remaining limitation rather than a prerequisite requiring atomic coordinates.

2. Does the method work on genuinely low-resolution experimental maps?

Section 2 in General Response evaluates the end-to-end pipeline on nine experimental maps at 4.3–5.8 Å, rather than artificially filtered maps. On EMD-8621 at 5.8 Å, our method correctly localizes 10/20 proteins from the known candidate list, compared with 0/20 for ModelAngelo (SOTA). This directly demonstrates useful correspondence in the intended low-resolution regime.

3. What does the retrieval percentage mean, and how useful is retrieval in practice?

We will clarify the abstract to report, explicitly distinguishing the fraction of maps with at least one correct candidate from per-sequence recall.

The new experiments additionally test retrieval against 2,000 decoys. On EMD-8621, four true proteins rank above every decoy, whereas ModelAngelo’s sequence-free building and HMM-search workflow returns no significant hits. Performance remains variable: on EMD-47303, no true protein outranks every decoy. We therefore describe retrieval as candidate prioritization, not complete composition recovery.

Our pipeline typically takes approximately one minute per map with cached embeddings; embedding adds 1.5–6 minutes once per candidate list. ModelAngelo’s sequence-free building averages 10 minutes, followed by less than one minute for HMM search on the same hardware.

In light of these clarifications and additional evidence, we respectfully ask the reviewer to reconsider their assessment and raise their score.

Thanks for the clarification and the additional experiments. The evaluation on nine real experimental low-resolution maps, the decoy-based retrieval analysis, and the runtime comparison cover a good part of what I originally raised.

That said, some important limitations remain. The low-resolution evaluation rests on only nine maps and a known candidate list, so it doesn't yet show robust blind localization or identification on a broader set of noisy, heterogeneous experimental maps. The result on EMD-8621 is encouraging, but the substantial variation across maps—including the failure on EMD-47303—should come through more prominently in the main conclusions.

The retrieval analysis is now described more appropriately as candidate prioritization rather than complete composition recovery. Even so, the revised evidence doesn't yet establish reliable protein identification in a realistic operational setting, especially with larger or independently constructed candidate pools. The exact checkpoint, candidate-pool construction, residue reduction procedure, and retrieval metric should also be stated clearly in the main paper.

Overall, the new results strengthen the low-resolution and retrieval analyses, but the method is still better supported as a correspondence and candidate-prioritization proof of concept than as a broadly validated protein identification pipeline. My recommendation remains unchanged.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Addressing Shared Reviewer Concerns

All three reviewers raised two common concerns, which we address collectively in the following sections.

1. End-to-end validation set

We agree that the lack of an atomic model-free inference was an important omission. Since our initial submission we have trained a simple U-Net to classify foreground from solvent using our CryoFM2 CleanDIFT features and re-evaluated our pipeline in a manner faithful to how it would be used in practice. We note only minor decreases in performance, with segmentation reducing from 0.647 using the atomic model to 0.617 with the faithful end-to-end approach, and retrieval reducing from 0.227 to 0.217 mAP. We have updated the manuscript to reflect this necessary change.

MetricAtomic modelUNet
Voxel-level accuracy0.6470.617
Macro IoU0.4350.324
Presence scoring (mAP)0.2270.217

2. Protein identification in experimental 4.3–5.8 Å cryoEM maps: our pipeline vs ModelAngelo

We agree with the reviewers that simulated low-resolution maps obtained through low-pass filtering do not fully capture the noise and artifacts present in experimental data. To address this limitation, we evaluated our end-to-end pipeline on genuine low-resolution held-out cryo-EM maps and compared its performance against ModelAngelo, a state-of-the-art method for automated atomic model building and protein identification.

mapsampleresolution (Å)proteins in mapours: foundours: above all 2k decoysours: median rankMA: foundMA HMM: found
EMD-8621E. coli 30S + YjeQ5.8201043800
EMD-47303E. coli iSAT 50S intermediate5.3135051600
EMD-3661E. coli 30S + RsgA5.218937770
EMD-45666E. coli 50S intermediate + RluB5.011104740
EMD-16502E. coli 50S intermediate4.9143151420
EMD-19807K. lactis 48S preinitiation complex4.73922532120
EMD-61625E. coli 50S precursor + B. subtilis Erm4.717151091715
EMD-37604Sulfolobus 30S + tRNA fragment4.626357920
EMD-54130human mitoribosome SSU intermediate4.332191332016
  • ours: found — with the known list only, proteins whose density is mostly assigned to the right protein.
  • ours: above all 2k decoys — with 2000 decoys added, known proteins ranked above every decoy.
  • ours: median rank — with decoys added, median position of the known proteins among ~2010 candidates (1 = best; ~1000 = no signal).
  • MA: found — model_angelo build with the known list; proteins whose built residues are mostly assigned to the right sequence.
  • MA HMM: found — ModelAngelo's route for unknown proteins: build_no_seq, then its HMM profiles searched against the known + decoy sequences (E < 1e-3); known proteins with a hit. Zero decoys were hit on every map below.

These results demonstrate that our method can identify and localize proteins in genuine experimental cryo-EM maps at 4.3–5.8 Å, extending beyond the 1.5–3.9 Å training range. Our method consistently outperforms ModelAngelo in the known-sequence evaluation on all tested maps with resolutions poorer than 4.9 Å. For example, at 5.0 Å (EMD-45666), we correctly localize 10/11 proteins, compared with 4/11 identified by ModelAngelo, while ranking four true proteins above all 2,000 decoys. Although performance varies across maps, these results demonstrate the practical value of direct density-to-sequence alignment at resolutions where reliable atomic model building and subsequent sequence assignment become challenging.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.