acceptodds
Under review as a conference paper at ICLR 2027

CryoProt: A Protein Pretraining Framework with Cross-Box Interactions on Cryo-EM Density Maps

Abstract

Despite the growing availability of cryo-electron microscopy (cryo-EM) density maps, effectively leveraging them for protein representation remains challenging. First, current methods lack a general-purpose protein pretraining framework tailored for cryo-EM density maps, designed for protein-related property prediction. Second, existing approaches typically partition density maps into local box regions and model them independently, overlooking interactions across boxes which are essential for capturing global structural context in cryo-EM density map. To address these challenges, we propose CryoProt, a protein pretraining framework designed for cryo-EM density maps. CryoProt introduces a Map Encoder based on multi-head latent attention (MLA), where box-level representations interact through a shared latent space, enabling explicit modeling of cross-box dependencies within the density map. Furthermore, we adopt a multi-task pretraining strategy to learn generalizable representations that can be effectively transferred to diverse downstream tasks, such as protein flexibility prediction, where cryo-EM density maps are not required and can be inferred implicitly by the pretrained model. Experimental results demonstrate that CryoProt consistently outperforms existing state-of-the-art methods across multiple benchmarks, achieving up to 12% improvement over the best-performing baselines, highlighting the importance of modeling cross-box interactions in cryo-EM data. The source code is publicly available at https://anonymous.4open.science/r/CryoProt.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cryo-EM Density Map, Protein, Pretraining

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WoNNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Leveraging cryo-EM maps for protein representations is challenging. In this paper they suggest CryoProt - a pre training framework with multi head latent attention between boxes and multi-task training strategy. Almost all the results are within std levels of other pre-trained baselines.

Critical strengths

The paper is clear, well written, has beautiful and informative figures and presents a complete set of experiments.

Critical weaknesses

Almost all the results are within std levels of a different baseline.

Questions for authors
  1. How many parameters is CryoProt model compared to the other baselines? And what is the difference in training time? Is it more efficient?
  2. Is it more efficient to fine tune CryoProt for each task rather than training the per domain model?
  3. Based on the experiments it is unclear to me what is the advantage of the method as the results are within std level of other baselines in both Table 2 and ablations (except for ∆∆G).
2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Thanks for your thoughtful and encouraging review. We will incorporate this discussion into the revised manuscript.

W1: Almost all the results are within std levels of a different baseline.

Response: Although some improvements are within the standard deviation of competing baselines, CryoProt achieves higher mean performance than the baselines across multiple downstream tasks. More importantly, our primary contribution is not merely achieving substantial numerical improvements over existing methods, but rather introducing a protein pretraining framework that leverages experimental cryo-EM density maps for downstream property prediction. In particular, CryoProt explicitly models cross-box interactions, addressing the limitation of existing approaches that process local density boxes independently and overlook their spatial dependencies. Our experimental results demonstrate the effectiveness of this framework, with consistently competitive performance and improvements in mean metrics across downstream tasks. We believe these contributions extend beyond numerical gains and provide a new perspective on learning protein representations from experimental cryo-EM density maps.

Q1: How many parameters is CryoProt model compared to the other baselines? And what is the difference in training time? Is it more efficient?

Response: Since CryoProt is a protein pretraining framework, we compare its model size with representative pretrained models rather than domain-specific models. As shown in Table R1, CryoProt contains 178M parameters, approximately 3.7× fewer than ESM-1b, and has a model size comparable to ESM2-150M and OmegaFold. CryoProt completes pretraining in 26 hours on two RTX 5090 GPUs, and the resulting pretrained representations can be reused across multiple downstream tasks without retraining from scratch.

For the pretrained baselines, we directly use their officially released weights, as retraining these large-scale models from scratch would require substantial computational resources. Moreover, their official implementations do not provide validated pretraining configurations for RTX 5090 GPUs, making a reliable comparison under identical hardware conditions difficult. Therefore, we avoid estimating their pretraining times or claiming superior pretraining efficiency without direct evidence. Nevertheless, CryoProt's moderate model size and reusable pretrained representations support efficient adaptation to downstream tasks. As further demonstrated in Q2, CryoProt requires only 5 hours of fine-tuning for protein flexibility prediction, compared with 7–20 hours for domain-specific models, highlighting its advantage in task-specific training efficiency.

Table R1. Model Size and Training Time Comparison.

ModelParameters (M)Training Time (h)
ESM-1b650—
ESM2-150M150—
OmegaFold124—
CryoProt (Ours)17826

Q2: Is it more efficient to fine tune CryoProt for each task rather than training the per domain model?

Response: CryoProt is more efficient for downstream task adaptation than training domain-specific models from scratch. Taking protein flexibility prediction as an example, CryoProt requires only 5 hours of fine-tuning, compared with 7–20 hours for competing models, achieving up to 4× faster training, as shown in Table R2.

Table R2. Training Time Comparison for Protein Flexibility Prediction.

ModelTraining / Fine-tuning Time (h)Speedup (vs. CryoProt)
DEFMap204.0×
RMSF-Net102.0×
ResaPred71.4×
Flexpert-3D122.4×
CryoProt (Ours)51.0×

Q3: Based on the experiments it is unclear to me what is the advantage of the method as the results are within std level of other baselines in both Table 2 and ablations (except for ∆∆G).

Response: We have addressed this question in Response W1.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer WoNN,

Thank you again for your insightful comments. As the discussion period is approaching its conclusion, we would appreciate any further feedback on whether our response has addressed your concerns. If you have any remaining questions or concerns, please feel free to raise them, and we would be happy to provide further clarification. Thank you for your time and consideration.

Best regards, Authors

Reviewer_LUnCOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CryoProt, a pretraining framework that uses experimental cryo-EM density maps to learn protein representations for downstream property prediction. It argues that existing map-based methods are task-specific and process residue-centered regions independently, ignoring interactions between them. To address this, CryoProt partitions each map into residue-centered boxes, encodes them with an MLP plus 3D sinusoidal positional encodings, and models cross-box dependencies with a Transformer equipped with multi-head latent attention (MLA), alongside a frozen ESM2-150M sequence encoder. The model is pretrained on about 20k EMDB maps with three objectives: sequence-to-map prediction, masked box reconstruction, and modality classification. A key design property is that downstream inference requires only sequence: map representations are inferred from sequence via the cross-modal prediction module and concatenated with sequence embeddings. The authors evaluate CryoProt on four tasks (flexibility, active-site identification, binding affinity, and ΔΔG\Delta\Delta G prediction), and report consistent improvements over task-specific and pretrained baselines, supported by ablations and a Top-L contact-coverage analysis.

Critical strengths
  • A practical way to use experimental cryo-EM data for general protein representation. The paper treats the large, mostly unlabeled body of EMDB maps as a pretraining resource rather than a task-specific input. Through cross-modal distillation, downstream tasks need only sequence, which removes the main barrier to using map-informed models: most proteins of interest have no density map.
  • Broad evaluation with a fairly extensive set of controls. CryoProt is tested on four tasks covering residue-level and protein-level prediction. The paper includes ablations that address natural alternative explanations: synthetic versus experimental maps, PDB- or sequence-derived inputs in place of maps, sequence-encoder choice, and a sequence-similarity leakage check.
Critical weaknesses
  • Mischaracterized prior work undermines the problem formulation. The central motivation (Challenge 2) is that existing cryo-EM methods partition maps into residue-centered boxes and model them independently, but the cited works do not support this. In Sec. 2, CryoDRGN and CryoNeFEN are described as density-map methods that "often partition maps into local box regions", yet both reconstruct density from 2D particle images with coordinate-based decoders; they take no density map as input and use no box partitioning. EModelX is an atomic model-building method, not a reconstruction or enhancement method. In Sec. 1, the residue-based partitioning claim cites EModelX, CryoTEN, and CryoFM, but CryoTEN splits maps into uniform overlapping 48348^3 blocks to fit GPU memory, and CryoFM tokenizes maps into uniform 434^3 patches; neither uses residue positions. Moreover, CryoFM is itself a pretrained foundation model over cryo-EM densities, yet the "first general-purpose" claim is never delineated against it. Among the cited works, only the task-specific DEFMap and RMSF-net match the described setting, so the "overlooked cross-box interactions" gap is not established as a general limitation.

  • Insufficient positioning against existing protein foundation models. The discussion of protein pretraining (Sec. 2) is limited to models from 2021–2022 and omits recent multimodal models that jointly learn from sequence, structure, and/or functional annotations and transfer to property prediction (e.g., SaProt, ESM-3, ProstT5, ProtST). Given how mature these models are, the key question is what experimental density contributes beyond them, and the experiments do not settle it: the pretrained baselines are limited to ESM-1b, ESM2-150M, ProtBert, and OmegaFold, and GearNet is discussed but not benchmarked.

  • The central claim "MLA as the key to cross-box interaction" is technically misdescribed and not supported by the ablations. First, the paper contrasts MLA with "conventional attention mechanisms that compute pairwise interactions over all boxes, leading to quadratic complexity" (Sec. 3.1) and claims that MLA "improves scalability for long sequences" (App. A). Both claims are incorrect. Eq. 10 is standard full softmax attention over all LL boxes with low-rank key/value projections, so the cost remains O(L2)O(L^2), with only a smaller per-pair constant (dc=128d_c = 128 vs. d=240d = 240). The claimed "shared latent space" (Abstract; Sec. 3.1) is also not what is implemented: each latent cj=WDhj\mathbf{c}_j = \mathbf{W}_D \mathbf{h}_j is computed per box, and nothing is shared or aggregated across boxes. Second, "w/o MLA" removes cross-box interaction entirely (Sec. 4.4), so it tests interaction versus no interaction rather than MLA itself. The informative comparison is MLA versus standard self-attention, and there the results are close. On flexibility (Table 3), all three metrics differ by less than one fold-level standard deviation (e.g., RMSE 0.484±0.0210.484 \pm 0.021 vs. 0.497±0.0170.497 \pm 0.017). On the other tasks (Tables 8–10), MLA is consistently ahead, but by margins comparable to the fold variance, and no seed replicates or significance tests are reported.

  • The performance gains are not attributable to cross-box modeling or to experimental density. At inference, map features are predicted from frozen ESM2-150M embeddings by an MLP (Eq. 14, App. D.2) and concatenated with the sequence features, so the cross-box Map Encoder never sees a downstream protein. The gains could therefore come from added capacity or from the pretraining targets acting as a regularizer rather than from cross-box modeling. A matched baseline (ESM2-150M plus a head of equal capacity, trained with the same protocol) is needed to separate these explanations and is missing. Relatedly, the local-window baseline uses a window of 30 to match the average number of interacting boxes (Sec. 4.4). Since boxes are ordered by residue index (Sec. 3.1), the window is presumably defined in sequence order, whereas the motivating statistic (Fig. 1a) concerns 3D proximity (<10 Å), and spatial neighbors are generally not sequence neighbors. Unless the window is defined spatially, which is not described, the "local vs. global" comparison does not test the paper's own premise.

Questions for authors
  • Can you compare against stronger or more recent pretrained models, e.g., larger ESM2 variants and structure-aware or multimodal models such as SaProt? Ideally, also include a structure-pretrained baseline trained on the same proteins' deposited atomic models, to isolate what experimental density adds beyond the atomic structure.

  • For SKEMPI v2 and HER2, are folds split by complex/antibody family or by individual mutant entry? With mutant-level splits, variants of the same complex appear in both training and test folds. In addition, how are the mutation, the wild-type reference, and the binding partner represented when only sequence is used downstream, given that mean pooling over a single-residue change yields nearly identical embeddings?

  • Why do several ablated variants (w/o MLA, w/ sequence, w/ PDB structure, w/o tasks 2+3) perform worse than plain ESM2-150M (Tables 2, 3), even though they contain the same encoder plus additional components? Were their hyperparameters tuned with the same budget as the full model?

  • Could you add two baselines: (a) ESM2-150M plus a head of equal capacity to CryoProt's downstream path (cross-modal MLP + concatenation + task head), trained with the same protocol and without map pretraining; (b) a spatial (radius ≤ 10 Å or kNN) attention variant? These two controls would directly show whether the gains come from map pretraining and from spatial cross-box interaction.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1)

We appreciate your detailed review and constructive suggestions. We will expand the new discussion and results in the revised manuscript.

W1: Mischaracterized prior work undermines the problem formulation.

Response: We acknowledge that our descriptions of CryoDRGN and CryoNeFEN were inaccurate and will correct them in the revised manuscript. We will also revise Challenge 2 by removing the claim that existing methods generally use residue-centered boxes. Nevertheless, local volumetric or patch-based processing remains common in cryo-EM density map modeling, as exemplified by CryoTEN and CryoFM. This motivates further investigation of spatial dependencies across local density regions. Importantly, CryoProt aims to be the first protein pretraining framework leveraging experimental cryo-EM density maps for downstream protein property prediction. While CryoFM learns generative density representations primarily for map-related tasks, such as reconstruction and restoration, CryoProt learns transferable protein representations for diverse property prediction tasks. Furthermore, CryoProt explicitly models cross-box interactions, integrating local density information with global spatial dependencies. We will further clarify these distinctions and refine our novelty claims in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2)

W2: Insufficient positioning against existing protein foundation models.

Response: We have added four recent protein foundation models, including ESM3 [1], SaProt [2], ProstT5 [3], and ProtST [4], covering both sequence-only and multimodal approaches, and evaluated them across all four downstream tasks under the same experimental protocol. As shown in Table R1, CryoProt achieves competitive performance across all four tasks, outperforming these models on most evaluation metrics. Although CryoProt does not achieve the best results on a few individual metrics, its consistently strong overall performance supports the effectiveness of incorporating experimental cryo-EM density information and explicitly modeling cross-box interactions for transferable protein representation learning.

Table R1. Performance Comparison with Recent Protein Foundation Models. The best results are in bold.

A. Protein Flexibility Prediction

ModelSeqStrMapRMSE ↓Pearson ↑Spearman ↑
ESM3-Open-1.4B✓✗✗0.619 ± 0.0520.248 ± 0.0090.260 ± 0.014
SaProt-650M-PDB✓✓✗0.498 ± 0.0360.314 ± 0.0330.327 ± 0.025
ProstT5✓✓✗0.515 ± 0.0330.327 ± 0.0420.314 ± 0.030
ProtST-ESM-1b✓✗✗0.503 ± 0.0690.307 ± 0.0150.307 ± 0.022
CryoProt✓✗✓0.484 ± 0.0210.310 ± 0.0410.315 ± 0.046

B. Active Site Identification

ModelSeqStrMapAUPRC ↑MCC ↑FPR ↓
ESM3-Open-1.4B✓✗✗0.653 ± 0.0170.612 ± 0.0220.027 ± 0.002
SaProt-650M-PDB✓✓✗0.645 ± 0.0140.601 ± 0.0060.025 ± 0.002
ProstT5✓✓✗0.744 ± 0.0100.658 ± 0.0110.016 ± 0.001
ProtST-ESM-1b✓✗✗0.742 ± 0.0120.659 ± 0.0190.022 ± 0.004
CryoProt✓✗✓0.746 ± 0.0230.669 ± 0.0190.021 ± 0.001

C. Binding Affinity Prediction

ModelSeqStrMapRMSE ↓Pearson ↑Spearman ↑
ESM3-Open-1.4B✓✗✗0.629 ± 0.0250.230 ± 0.0130.221 ± 0.011
SaProt-650M-PDB✓✓✗0.493 ± 0.0230.456 ± 0.0430.480 ± 0.047
ProstT5✓✓✗0.512 ± 0.0450.413 ± 0.0750.446 ± 0.079
ProtST-ESM-1b✓✗✗0.492 ± 0.0300.450 ± 0.0710.459 ± 0.074
CryoProt✓✗✓0.486 ± 0.0390.524 ± 0.0820.527 ± 0.087

D. ΔΔG Prediction

ModelSeqStrMapRMSE ↓Pearson ↑Spearman ↑
ESM3-Open-1.4B✓✗✗1.161 ± 0.0600.491 ± 0.0370.415 ± 0.026
SaProt-650M-PDB✓✓✗1.121 ± 0.0360.513 ± 0.0290.429 ± 0.018
ProstT5✓✓✗1.126 ± 0.0350.622 ± 0.0230.439 ± 0.014
ProtST-ESM-1b✓✗✗1.126 ± 0.0360.506 ± 0.0270.432 ± 0.016
CryoProt✓✗✓1.036 ± 0.0330.611 ± 0.0250.488 ± 0.007

[1] T. Hayes, R. Rao, H. Akin, et al., “Simulating 500 million years of evolution with a language model,” Science, vol. 387, no. 6736, pp. 850–858, 2025.

[2] J. Su, C. Han, Y. Zhou, et al., “SaProt: Protein language modeling with structure-aware vocabulary,” in Proceedings of the International Conference on Learning Representations (ICLR), 2024.

[3] M. Heinzinger, K. Weissenow, J. G. Sanchez, et al., “Bilingual language model for protein sequence and structure,” NAR Genomics and Bioinformatics, vol. 6, no. 4, Art. no. lqae150, 2024.

[4] M. Xu, X. Yuan, S. Miret, et al., “ProtST: Multi-modality learning of protein sequences and biomedical texts,” in Proceedings of the International Conference on Machine Learning (ICML), 2023, pp. 38749–38767.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 3)

W3: The central claim "MLA as the key to cross-box interaction" is technically misdescribed and not supported by the ablations.

Response: We acknowledge that our MLA implementation employs low-rank query and key-value projections followed by full attention, retaining quadratic complexity with respect to the number of boxes. We will correct the complexity analysis and clarify that latent representations are computed independently for each box within a common low-dimensional feature space. Importantly, our primary contribution lies in explicitly modeling cross-box interactions for cryo-EM-based protein representation learning, rather than reducing the asymptotic complexity of attention. In our ablation study, “w/o MLA” specifically represents the removal of cross-box interactions, allowing us to evaluate their contribution. Furthermore, we compare MLA with two alternative interaction mechanisms, local-window attention and standard self-attention, to assess the effectiveness of our design. As shown in Table 3, MLA achieves the lowest RMSE (0.484), compared with 0.546 without interactions, 0.519 with local-window attention, and 0.497 with self-attention. Although some improvements are within the fold-level standard deviations, MLA consistently achieves better mean performance across the evaluated downstream tasks, supporting the effectiveness of our design. We will clarify these distinctions and refine the corresponding claims in the revised manuscript.

W4: The performance gains are not attributable to cross-box modeling or to experimental density.

Response: We clarify the two points as follows:

  • Capacity-Matched Baseline and Experimental Density Information: We introduce a capacity-matched baseline, consisting of ESM2-150M and an additional prediction module with the same downstream model capacity as CryoProt, trained and evaluated under the same protocol. As shown in Table R2, CryoProt outperforms this baseline across four downstream tasks, indicating that the improvements cannot be attributed solely to increased model capacity. Furthermore, as shown in Table 3 of the manuscript, replacing experimental cryo-EM maps with ChimeraX-synthesized maps, PDB structures, or sequence representations while retaining the MLA architecture and pretraining objectives consistently degrades performance, further supporting the contribution of experimental cryo-EM density information.

    Table R2. Performance Comparison with the Capacity-Matched Baseline.

    A. Protein Flexibility Prediction (RMSF)

    ModelRMSE ↓Pearson ↑Spearman ↑
    ESM2-150M + Capacity-Matched Head0.612 ± 0.0660.264 ± 0.0140.227 ± 0.029
    CryoProt0.484 ± 0.0210.310 ± 0.0410.315 ± 0.046

    B. Active Site Identification

    ModelAUPRC ↑MCC ↑FPR ↓
    ESM2-150M + Capacity-Matched Head0.695 ± 0.0090.628 ± 0.0030.038 ± 0.002
    CryoProt0.746 ± 0.0230.669 ± 0.0190.021 ± 0.001

    C. HER2 Binding Affinity Prediction

    ModelRMSE ↓Pearson ↑Spearman ↑
    ESM2-150M + Capacity-Matched Head0.528 ± 0.0520.434 ± 0.0950.503 ± 0.101
    CryoProt0.486 ± 0.0390.524 ± 0.0820.527 ± 0.087

    D. ΔΔG Prediction

    ModelRMSE ↓Pearson ↑Spearman ↑
    ESM2-150M + Capacity-Matched Head1.289 ± 0.0390.554 ± 0.0320.411 ± 0.009
    CryoProt1.036 ± 0.0330.611 ± 0.0250.488 ± 0.007
  • Clarification of Local-Window Attention: We clarify that local-window attention is defined by three-dimensional spatial proximity, not sequence order. Specifically, for each box, we calculate the Euclidean distances between box centers and select the 30 spatially nearest neighboring boxes for attention computation. Thus, the local-window baseline captures local spatial interactions regardless of sequence positions, enabling a comparison with global cross-box interactions. We will clarify this implementation in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 4)

Q1: Can you compare against stronger or more recent pretrained models?

Response: We have addressed this question in Response W2.

Q2: How are data leakage and mutation-related information handled in SKEMPI v2 and HER2?

Response: We clarify the two points as follows:

  • Data Splitting: As described in the manuscript, we adopt sequence similarity-based protein-level splitting for five-fold cross-validation rather than residue-level or random mutant-level splitting. Highly similar protein sequences are grouped into the same fold, ensuring that different mutant variants of the same protein are not distributed across training and test folds. This strategy effectively prevents potential data leakage arising from closely related mutant entries.

  • Mutation and Binding Partner Representation: For SKEMPI v2, we use the wild-type and mutant sequences of the mutated chain, which are processed through separate backbone branches. Each branch combines sequence representations with sequence-inferred map representations. The resulting features are concatenated and aggregated through masked mean pooling for ΔΔG prediction. For HER2, we use antibody sequence representations to predict binding affinity against the fixed HER2 target. Importantly, CryoProt adopts ESM2 as its sequence encoder, which produces contextualized residue-level representations. Therefore, even a single-residue mutation can influence representations beyond the mutated position, enabling the model to capture mutation-related differences after pooling. Although the binding partner is not explicitly encoded, the learned representations provide useful information for predicting mutation-induced binding affinity changes.

Q3: Why do several ablated variants (w/o MLA, w/ sequence, w/ PDB structure, w/o tasks 2+3) perform worse than plain ESM2-150M (Tables 2, 3), even though they contain the same encoder plus additional components? Were their hyperparameters tuned with the same budget as the full model?

Response: These ablated variants are not simply ESM2-150M with additional components; they retain the cross-modal learning framework while removing or replacing key mechanisms. Specifically, removing MLA eliminates cross-box interactions, replacing experimental density maps with sequence or PDB-derived representations changes the information available during pretraining, and removing Tasks 2 and 3 reduces the constraints on representation learning. Consequently, the remaining components may introduce redundant or less informative features, and increased model capacity alone does not guarantee better performance than ESM2-150M. These results highlight the importance of combining experimental density information, cross-box interactions, and multi-task pretraining in the complete CryoProt framework. Regarding hyperparameters, all ablation variants were evaluated using the same data splits, training schedule, optimization settings, and evaluation protocol as the full model, without additional variant-specific hyperparameter tuning. This ensures that the observed differences primarily reflect the effects of the ablated components rather than differences in experimental settings.

Q4: Could you add two baselines: (a) ESM2-150M plus a head of equal capacity to CryoProt's downstream path (cross-modal MLP + concatenation + task head), trained with the same protocol and without map pretraining; (b) a spatial (radius ≤ 10 Å or kNN) attention variant? These two controls would directly show whether the gains come from map pretraining and from spatial cross-box interaction.

Response: We have addressed this question in Response W4.

Reviewer_13VEOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CryoProt, a pretraining framework intended to learn protein representations from cryo-EM density maps and transfer them to protein-property prediction. The method uses residue-centered local map boxes, an MLA-based map encoder intended to model cross-box context, and a multi-task pretraining objective combining sequence-to-map prediction, masked map reconstruction, and modality classification. The downstream setting uses sequence input only. The paper reports pretraining on 20,530 curated EMDB entries and five-fold evaluation on protein flexibility, active-site identification, binding-affinity prediction, and Delta-Delta-G prediction. It reports improvements over task-specific and pretrained baselines.

Critical strengths

1.The paper identifies a meaningful and specific problem. Cryo-EM density maps are large and have different spatial sizes, so many map-based protein models partition them into local boxes. For tasks that benefit from global context and a large receptive field, such as flexibility and active-site prediction, independently processing local boxes can discard useful long-range information. The motivation for cross-box interaction is therefore practically relevant. 2.The overall paradigm is useful. Pretraining on large unlabeled cryo-EM data, transferring to multiple property-prediction tasks, and performing inference from sequence alone is a practical deployment direction. It could make cryo-EM-derived structural information available when maps or structures are absent at test time. 3.The study covers residue-level and protein-level tasks and includes released pretrained and fine-tuned checkpoints. This provides a useful starting point for investigating whether cryo-EM-aware pretraining can transfer beyond a single benchmark.

Critical weaknesses
  1. The primary residue-level metric is pooled across proteins According to Appendix D, for residue-level tasks S is the total number of residues across all proteins. The released fine-tuning code likewise concatenates predictions and targets from all validation proteins and computes Pearson and Spearman once per fold. This pooled estimand gives long proteins disproportionate influence and mixes between-protein and within-protein variability. It also does not directly measure whether the predicted flexibility pattern is correct within each protein. In contrast, RMSF-net computes one correlation per protein and reports the mean over 335 proteins. The paper does not make this protocol distinction explicit, even though it compares CryoProt with RMSF-net in the same table.
  2. Per-protein evaluation shows a large performance gap and unreproduced results I evaluated the public RMSF annotations using the released CryoProt fine-tuned checkpoints and compared them with the per-protein RMSF-net results reported in the authors' supporting workbook. Of the 335 published proteins, 292 had sufficient finite annotations to compute correlations. Using an ensemble of the five released CryoProt folds, the mean per-protein Pearson was 0.354 and the mean per-protein Spearman was 0.412. The workbook reports a mean per-protein RMSF-net Pearson of 0.746 over all 335 proteins. On the 291 proteins represented in both evaluations, RMSF-net's mean per-protein Pearson was 0.760 versus 0.354 for CryoProt, a mean gap of 0.407. The two methods' per-protein correlation values had almost no association (Pearson correlation approximately 0.007). PDB Metric CryoProt RMSF-net 3J9D RMSE 0.963 0.376 Pearson 0.419 0.840 Spearman 0.481 0.815 3J9E RMSE 2.088 0.710 Pearson 0.088 0.892 Spearman 0.238 0.787 5I68 RMSE 1.214 0.752 Pearson 0.512 0.837 Spearman 0.723 0.768

CryoProt's Appendix I reports RMSE values of 0.527, 0.519, and 0.583 for these same three examples, whereas the released-checkpoint pipeline produces 0.963, 2.088, and 1.214. None of the five CryoProt fold checkpoints brings 3J9E close to 0.519. Because the official CryoProt inference entry point is missing, I treated the numerical reproduction as a best-effort reconstruction using the released model definition, preprocessing, and checkpoints. As a sanity check, I rendered the three structures with the same PDB, residue set, view, and color scale. The reproduced CryoProt flexibility patterns qualitatively resembled those in Appendix I, reducing concern that the reimplementation is grossly incorrect. Nevertheless, both the original Appendix I visualization and the reproduced CryoProt visualization appeared visibly less consistent with the ground-truth high- and low-flexibility regions than the RMSF-net visualization. The qualitative comparison corroborates the quantitative gap. These results suggest that the current flexibility model has not been adequately optimized or calibrated, and that the flexibility task has not been evaluated in a way that supports the claimed advantage over RMSF-net. 3. The documented inference entry point is missing The README instructs users to run python -m Case.predict, but no Case directory or corresponding module is released, so the documented command fails. Because this is a required inference entry point rather than omitted training data, the repository does not support the paper's reproducibility claim. The missing prediction code should be provided.

Questions for authors

Please provide the prediction code and exact entry point used for protein flexibility prediction. If you believe that my flexibility-prediction inference code or the scripts used to calculate the three per-protein metrics have any problems, please also let me know.

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank you for your detailed reviews. We address your concerns point by point.

W1: CryoProt uses pooled correlations across all residues, whereas RMSF-Net computes correlations per protein and averages them, potentially leading to unfair comparisons and disproportionate influence from longer proteins.

Response: As described in the paper, all baselines, including RMSF-Net, were retrained under the same experimental conditions as CryoProt, including the computational environment, hardware, data splits, training settings, and evaluation procedures. The reported RMSF-Net results were obtained through our own reimplementation rather than directly taken from the original paper. Thus, all methods were compared under a consistent evaluation protocol.

We emphasize that pooled correlation is a meaningful metric for residue-level RMSF prediction, as it measures the overall agreement between predicted and ground-truth values across all valid residues. In contrast, per-protein correlation evaluates flexibility variations within individual proteins. Both capture complementary aspects of prediction performance. Although pooled correlation may be influenced by protein length and between-protein variability, this does not invalidate its usefulness or the fairness of our comparisons. We will further clarify these details in the revised manuscript.

W2: Per-protein evaluation shows a large performance gap and unreproduced results I evaluated the public RMSF annotations using the released CryoProt fine-tuned checkpoints and compared them with the per-protein RMSF-net results reported in the authors' supporting workbook.

Response: We would like to clarify that the RMSF-Net results reported in our paper were obtained by retraining the model using the same training setup as CryoProt. Therefore, the per-protein RMSF-Net results (Pearson = 0.746) are not directly comparable to our reported results, as they were obtained under a different evaluation setting.

Regarding the three case studies, we have verified that the RMSE values reported in paper can be reproduced using our official prediction code. The discrepancies reported by the reviewer may arise from differences in the independently reconstructed inference pipeline, including preprocessing and prediction procedures. To address this concern, we have updated our anonymous repository with the complete prediction code and provided screenshots of the results obtained during our reproduction tests. These materials confirm the reproducibility of our reported case-study results and facilitate independent verification.

W3: The missing prediction code should be provided.

Response: We have updated our anonymous repository to include the complete inference code and the previously missing Case module. The documented command python -m Case.predict is now supported, allowing reviewers to directly reproduce the case-study predictions reported in the manuscript. In addition, we have provided detailed instructions and screenshots of our local reproduction tests, which confirm that the obtained RMSE values are consistent with those reported in the paper. These updates address the missing implementation and facilitate independent verification of our results.

Q1: Please provide the prediction code and exact entry point used for protein flexibility prediction.

Response: We have addressed this question in Response W3.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer 13VE,

Thank you again for your time and consideration. As the discussion period is about to conclude, we would appreciate any further feedback on whether our responses have addressed your concerns. Please let us know if there are any remaining questions or clarifications needed. Thank you very much for your attention.

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely thank you for your careful evaluations and constructive feedback. In response to your concerns, we have conducted additional experiments and revised the manuscript, with all revisions highlighted in blue. The key updates are summarized below.

1. Additional baselines and performance attribution.

We added four recent protein foundation models (ESM3, SaProt, ProstT5, and ProtST) and a capacity-matched ESM2 baseline. CryoProt achieves competitive performance across four downstream tasks, outperforming the capacity-matched baseline on all 12 metrics (Section 4.3 Comparison Results).

2. Methodological clarifications and ablation studies.

We corrected the descriptions of related work and MLA complexity, clarified the spatial definition of local-window attention, and expanded the ablation analysis to support the contributions of experimental cryo-EM density information and cross-box interactions (Related Work; Section 4.4 Ablation Study).

3. Reproducibility.

We released the complete inference code, reproduction instructions, and verification results to facilitate independent evaluation (Anonymous Repository).

We believe these revisions strengthen the experimental evidence and clarify the contributions of CryoProt. We sincerely appreciate the reviewers' valuable feedback and consideration.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.