acceptodds
Under review as a conference paper at ICLR 2027

From Physics to Experiment: Learning Resolution-Controllable Cryo-EM Density with Structured Diffusion

Abstract

Cryo-electron microscopy (cryo-EM) density maps are central to macromolecular structure determination, yet generating synthetic maps that faithfully approximate experimental reconstructions from an atomic model remains challenging. Analytical renderers provide explicit resolution control but rely on simplified physical approximations, whereas learned methods can capture experimental characteristics but often provide limited controllability. We introduce CryoNet.Density, a 3D Brownian Bridge Diffusion Model that formulates cryo-EM density-map generation as structured physical-to-experimental translation. Given an atomic model, a differentiable, resolution-matched renderer first produces a physically rendered map; CryoNet.Density then learns the correction toward the paired experimental reference map rather than generating density from noise. Fourier-feature conditioning enables continuous inference from 1.5–10 Å, while a one-transition stochastic sampler requires only two denoiser evaluations. On a held-out 1,923-case test set, CryoNet.Density outperforms analytical renderers and Struc2mapGAN across seven spatial- and Fourier-domain metrics. The generated density maps further improve secondary-structure recognition and residue identification, supporting their structural fidelity. We additionally construct the CryoNet.Density Database, comprising 3,640,616 traceable generated density maps from 260,044 eligible PDB entries at 14 controlled resolutions. Together, these results establish structured physical-to-experimental diffusion as a scalable framework for controllable synthetic cryo-EM density-map generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cryo-electron microscopy, Density map generation, Brownian Bridge Diffusion Model, Continuous resolution control, Structure determination

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SxotOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents CryoNet.Density, a resolution-conditioned 3D Brownian Bridge Diffusion Model that generates realistic, experiment-like cryo-EM density maps from atomic structures. A differentiable Gaussian renderer first produces a physically simulated map from the atomic model, and the diffusion model then learns a structured correction that maps it toward the corresponding experimental map. Target resolution is controlled continuously through FiLM conditioning on Fourier features. Sampling uses a single stochastic transition that needs only two denoiser evaluations.

The authors build a benchmark of paired atomic models and experimental maps, with 1,923 held-out test cases. They compare CryoNet.Density with analytical rendering methods and Struc2mapGAN on seven metrics covering the spatial and Fourier domains, and report large gains in map fidelity. Further experiments test resolution control and downstream usefulness, the latter by running frozen Emap2sec+ and ModelAngelo models on the generated maps. The paper also releases the CryoNet.Density Database, about 3.64 million generated maps at controlled resolutions.

Critical strengths
  1. A well-motivated formulation with thorough evaluation.

The paper frames density generation as translating a physically rendered map into an experimental one. Analytical rendering preserves the structural content, and the model learns only the remaining correction. This is a sensible way to combine explicit physics with the flexibility of a generative model.

The evaluation is fairly thorough, covering 1,923 held-out cases and seven spatial- and Fourier-domain metrics. CryoNet.Density outperforms all reported baselines, with the largest gains in Fourier-domain agreement. The ablations, including comparisons with direct regression and alternative sampling trajectories, support the main design choices. Taken together, the empirical case is convincing.

  1. Downstream utility and a potentially valuable large-scale dataset.

The authors go beyond map similarity and test whether the generated maps preserve features needed for structural interpretation. With frozen Emap2sec+ and ModelAngelo models, CryoNet.Density maps support downstream analysis better than the other synthetic maps, which is good evidence of practical value.

The accompanying database of about 3.64 million resolution-controlled maps could also be a useful resource for training and benchmarking data-driven cryo-EM methods. This extends the potential impact of the work beyond map generation itself.

Critical weaknesses
  1. Limited validation of experimental variability, and an unclear physical meaning for resolution control.

CryoNet.Density agrees well with the paired experimental maps, but the evaluation does not show whether its outputs reproduce the variability of real experimental reconstructions.

Real cryo-EM maps vary with acquisition conditions, particle count, reconstruction pipeline, local resolution, and structural heterogeneity. The model, by contrast, is conditioned only on the atomic structure and a nominal target resolution. It is therefore unclear whether the stochastic variation in its samples reflects realistic experimental variability or is simply an artifact of the learned sampling process.

The supplementary analysis of conditional diversity and residual structure is a useful first step. A more direct test would compare generated samples against multiple independent experimental reconstructions of the same structure, which exist for many entries in EMDB.

Resolution control raises a similar concern. It is evaluated mainly through spectral bandwidth and how well bandwidth tracks the requested resolution. This shows that the conditioning works, but bandwidth is not the same as experimentally measured resolution, and it says nothing about spatially varying local resolution. Validating local structural fidelity and other resolution-dependent properties would clarify what the resolution control actually achieves.

  1. Gaps in the downstream evaluation, and no direct evidence that the database is useful.

The downstream experiments suggest that the generated maps retain structural information, but some choices in the protocol limit how far the conclusions extend.

The ModelAngelo results are reported on 711 of 1,524 eligible protein test cases, keeping only cases where all four inputs yield nonzero sequence matches. This ensures that every method is evaluated on the same cases, but it hides how often each method fails. It may also bias the comparison: requiring success on every input, including the weakest baseline, tends to keep the easier cases. Reporting per-method success rates and aggregate results over all 1,524 cases would give a fuller picture.

The value of the CryoNet.Density Database is asserted but not demonstrated. No experiment trains or augments a downstream model with the generated maps. Showing improved performance on held-out experimental data after such training would be much stronger evidence of the resource’s impact.

These issues do not undermine the reported improvements, but they leave the broader practical benefit of the generated maps less certain.

Questions for authors

Q1. Does CryoNet.Density capture the variability of real reconstructions, beyond matching a single paired map? For structures with several independent experimental reconstructions, how does the spread of generated samples compare with the spread across those reconstructions?

Q2. Does resolution conditioning reproduce the properties of real maps at a given resolution? Beyond radial spectral bandwidth, how do local resolution and structural fidelity change with the target resolution? Does this behavior hold at intermediate resolutions not seen during training?

Q3. What are the ModelAngelo results over all 1,524 eligible cases? The reported results cover only the 711 cases where every input yields a nonzero sequence match. Could the authors report each method’s success rate, and aggregate performance over the full set with failures included?

Q4. How much of the gain comes specifically from the Brownian bridge formulation? The same-backbone comparison with direct regression is useful. How does CryoNet.Density compare with a standard conditional diffusion model, or another bridge formulation, at matched compute? This would separate the benefit of the bridge from that of conditional generative modeling in general.

6 comments
AuthorsOpenReview ↗

Overview and experimental variability

OVERALL RESPONSE:

Your suggestions led us to compare repeated reconstructions, evaluate local map agreement, and add a conditional-diffusion baseline. The revised manuscript includes four main results:

1. Improved downstream performance with stronger evaluation. On 60 real maps, generated-map fine-tuning improves ModelAngelo Cα F1 from 50.10% to 53.71% and sequence recovery from 40.32% to 43.83%. The corrected frozen-model comparison now includes all 1,524 cases, preserving the leading ranking (Weakness 2 , Question 3; Tables 3–4, lines 415–427 and 447–462).

2. CryoNet.Density generated maps agree better with local experimental density. We compared generated and physically rendered maps with six experimental references and tested resolution values absent from training. CryoNet.Density achieved higher local PCC at five of six reference points*and the best performance in six of seven metrics on the 17-case unseen-resolution test set (Question 2; Tables 16–17, lines 1806–1816 and 1958–1978).

3. CryoNet.Density outperforms standard conditional diffusion. We trained a conditional DDPM with the same network architecture and compared several sampling step counts. CryoNet.Density achieved Map CC 0.774 versus 0.642 for the best DDPM setting, using two rather than ten network evaluations (Question 4; Table 32, lines 2896–2912).

4. Repeated reconstructions quantify the variation in generated maps. We compared five experimental reconstructions with 20 generated samples at fixed structure and resolution. The mean maps had PCC 0.801, while the generated samples showed less variation than the experimental reconstructions (Question 1; Table 29, lines 2594–2623).

We first compare generated samples with repeated experimental reconstructions, then discuss resolution control, downstream performance, and the conditional DDPM baseline.

Comment 1: For structures with several independent experimental reconstructions, how does the spread of generated samples compare with the spread across those reconstructions?

Response: We compared all five PDB-6knf/EMD-0730 particle-subsampled reconstructions at 4.57 Å with 20 generated samples. Coordinates, resolution, physical input map, grid, and molecular mask were fixed. All PDB-6knf entries belong to validation. Voxel variance uses the n−1 denominator; map pairs that share a reconstruction are not independent.

EnsembleMapsMap pairsMean pairwise PCCMean normalized RMSDMean voxel variance (×10⁻³)
Experimental reconstructions5100.96780.25411.8137
Generated samples201900.99740.07150.0698

CryoNet.Density generated maps have 3.85% of the experimental voxel variance, and the two variance maps have PCC 0.090. Standardizing each map separately gives a variance ratio of 7.92% and variance-map PCC 0.091. The mean maps have PCC 0.801. Thus, generated maps agree with the average experimental density but vary less than the repeated reconstructions. These are reconstructions from one parent dataset, and particle-subset overlap is unrecorded.

The other sampling analyses address where these differences occur. Across 1,923 cases, the experimental-minus-rendered residual has median neighboring-voxel correlation 0.635 and fraction of residual squared density inside the molecular mask 0.845. On 96 structures, phase randomization preserves the power spectrum but reduces gradient alignment 0.4611 → −0.0002 and curvature alignment 0.5512 → −0.0001 (Tables 30–31, lines 2722–2759; Figure 23).

For five samples of each of three structures, molecular-region PCCs are 0.9939–0.9947. The region contains 78.9%–92.5% of sample variation, with variance 26.3–42.8 times the background variance. Full-volume PCCs to each structure's experimental reference are 0.828191 ± 0.000170, 0.828770 ± 0.000263, and 0.793743 ± 0.000156 (mean ± standard deviation across seeds). These results show consistent experimental-map agreement across samples, with local density variation (Table 28, lines 2533–2544; Figure 21).

CryoNet.Density generated maps also retain recognizable structural features: frozen ModelAngelo residue accuracy improves 0.496 → 0.629 and sequence match 84.1% → 92.8% over 1,524 cases; Emap2sec+ helix F1 improves 0.595 → 0.733 over 136 cases (Tables 2–3, lines 408–427).

Revised in the manuscript: the direct reconstruction comparison is in Appendix J.4.1, Table 29 and Figure 22 (lines 2594–2623); complementary analyses are cited above.

AuthorsOpenReview ↗

Resolution control and local structural fidelity

Comment 2 : Beyond radial spectral bandwidth, how do local resolution and structural fidelity change with the target resolution? Does this behavior hold at intermediate resolutions not seen during training?

Response: We added local comparisons with experimental maps and evaluations at resolution values absent from training. Section 4.3 now introduces these analyses (lines 380–383).

Local experimental-map agreement. Using fixed PDB-7kod coordinates, we compare generated maps and physical renders against six experimental references spanning 1.655–10.18 Å. EMD-22972 is the map deposited with PDB-7kod; the other five are Glacios reconstructions from the ice-thickness series in EMPIAR-11397. Before inspecting predictions, we select 24 Cα-centered spheres of 32 Å diameter. Correlations are averaged across three seeds per sphere, then summarized by the median and IQR across spheres.

Experimental referenceNominal resolution (Å)Physical-render PCC, median [IQR]CryoNet.Density PCC, median [IQR]Median paired Δ
EMD-229721.6550.636 [0.634, 0.638]0.844 [0.840, 0.846]+0.208
EMD-295663.400.755 [0.748, 0.762]0.884 [0.881, 0.889]+0.130
EMD-295674.640.759 [0.746, 0.767]0.844 [0.823, 0.853]+0.088
EMD-295699.630.716 [0.691, 0.723]0.712 [0.682, 0.723]−0.006
EMD-295709.840.576 [0.552, 0.588]0.584 [0.559, 0.596]+0.009
EMD-2956810.180.860 [0.853, 0.876]0.887 [0.866, 0.896]+0.019

Generated maps have higher local PCC at five of six reference resolutions, with the largest improvements at 1.655–4.64 Å. The value at 9.63 Å is slightly lower. Δ is the median of paired regional differences, not the difference between the column medians. This measures local map–map correlation at each reported global resolution (Appendix I.3.1, Table 16, lines 1784–1816).

Intermediate resolutions. We generated maps at 20 resolutions from 1.655 to 10.18 Å, keeping coordinates, grid, regions, and random draws fixed. Local PCC between adjacent outputs ranges from 0.931 to 0.997; the four 0.25 Å steps from 5.5 to 6.5 Å give 0.997, 0.991, 0.977, and 0.96. Five targets—1.655, 6.25, 9.63, 9.84, and 10.18 Å—are absent from both training and validation. Local density remains correlated between adjacent target resolutions (lines 1836–1849; Figures 16–17).

Experimental maps at resolution values absent from training. We checked all 13,659 training pairs and 1,923 test cases and identified 17 qualifying test cases. Their resolutions are within the 1.52–15.0 Å training range and 0.004–0.100 Å from the nearest training value. All methods use the same 17 cases and the predictions from Table 1; results are averaged across cases.

MethodnMap–map CC ↑Masked FSC AUC ↑FSC@target ↑SSIM ↑Correlation ↑log-RAPS MAE ↓PCC ↑
ChimeraX molmap170.7590.2040.2620.8800.7691.3460.564
phenix.model_map170.7710.2280.2920.6630.6081.9490.629
EMAN2170.7380.2090.2610.8780.7471.8220.532
TEMPy170.7710.2130.2480.8710.7853.6890.672
CryoNet molmap170.7380.2060.2610.8780.7481.5700.545
CryoNet.Density170.7860.2440.3050.8730.8030.9240.690

CryoNet.Density leads in six of seven metrics; ChimeraX molmap has the highest SSIM. Relative to CryoNet molmap, FSC AUC increases 0.206 → 0.244, PCC 0.545 → 0.690, and log-RAPS MAE decreases 1.570 → 0.924. Excluding the two labels also present in validation (4.57 and 7.82 Å) leaves 15 cases. The same six-metric lead remains: FSC AUC is 0.223 and PCC 0.704, compared with 0.186 and 0.546 for CryoNet molmap. Added to Appendix I.3.2, Table 18 (lines 1946–1978).

AuthorsOpenReview ↗

ModelAngelo evaluation and downstream training

Comment 3: What are the ModelAngelo results over all 1,524 eligible cases? The reported results cover only the 711 cases where every input yields a nonzero sequence match. Could the authors report each method’s success rate, and aggregate performance over the full set with failures included?

Response: We re-evaluated all 1,524 eligible cases after correcting an evaluation error. Unavailable Phenix scores had been recorded as zero, so the previous analysis retained only 711 cases. The other 813 cases had at least one failed Phenix evaluation. We corrected missing-score handling and structure-file formatting/chain identifiers, then repeated the comparison with the same 3 Å matching radius. Model weights, predictions, and coordinates were unchanged.

All 1,524 cases now have valid evaluations for each input. Success is defined as a nonempty prediction with positive Phenix sequence match: all four inputs yield a nonempty prediction with a positive sequence match for every case; this does not establish complete or accurate structure recovery. The case-equal means are:

Input mapResidue accuracy ↑Sequence match (%) ↑Sequence score ↑
Experimental0.58885.10.786
ChimeraX molmap0.47883.00.824
CryoNet molmap0.49684.10.835
CryoNet.Density0.62992.80.912

CryoNet.Density remains highest in all three metrics. Scores are accepted only after successful Phenix execution and parsing; unresolved evaluations remain missing and block aggregation. Eight previously successful control comparisons retained identical sequence match and sequence score.

Resived in the manuscript: Section 4.4, Table 3 (lines 415–427), and the corrected protocol in Appendix I.4 (lines 2015–2027).

Comment 4: The value of the CryoNet.Density Database is asserted but not demonstrated. No experiment trains or augments a downstream model with the generated maps. Showing improved performance on held-out experimental data after such training would be much stronger evidence of the resource’s impact.

We fine-tuned ModelAngelo using CryoNet.Density maps and compared it with fine-tuning on physically rendered maps. Both models were evaluated on the same 60 held-out experimental maps at 3–6 Å.

Both training sets contain 863 structures at seven resolutions (3–6 Å in 0.5 Å intervals), giving 6,041 maps per set. They use identical atomic coordinates and grids. We initialized both models from the official ModelAngelo checkpoint and fine-tuned the full GNN for 10,000 AdamW steps at a learning rate of 10⁻⁶, using the same loss and number of training graph crops. CNN and ESM-1b parameters remained fixed. Evaluation used the same CNN candidate points, ESM features, three GNN iterations, sequence assignment, and pruning.

The official model is the common baseline (O). Values are percentages; Δ denotes percentage-point changes calculated before rounding. The first four metrics are averaged across cases; the remaining metrics combine residue counts across cases.

Metric ↑Official (O)Generated (G)ΔG−OPhysical (P)ΔP−O
Cα precision88.4593.51+5.0787.84-0.61
Cα recall42.3145.97+3.6733.23-9.08
Cα F150.1053.71+3.6040.23-9.87
Sequence recovery40.3243.83+3.5131.22-9.10
Pre-pruning residue accuracy53.0756.60+3.5343.40-9.67
Final end-to-end macro-F150.5653.69+3.1339.18-11.38

Training on CryoNet.Density maps increased Cα F1 from 50.10% to 53.71%, sequence recovery from 40.32% to 43.83%, and end-to-end macro-F1 from 50.56% to 53.69%. Training on physically rendered maps gave lower scores. These results show that the generated maps improve ModelAngelo's performance on experimental data. Residue accuracy measures amino-acid classification among matched residues; end-to-end macro-F1 also accounts for missing and incorrectly predicted residues.

The test set contains 60 cases: 39 at 3–<4 Å, 18 at 4–<5 Å, and three at 5–6 Å. We required CC_mask ≥ 0.5, Cα coverage ≥ 0.95, and compatible map sampling, and excluded sequence overlap with the fine-tuning data. The 60 PDB entries also have no ID overlap with ModelAngelo training set. Predicted and reference Cα atoms were matched one-to-one within 3 Å without rigid-body fitting; valid empty predictions received zero recovery.

Revised in the manuscript: Section 4.4, Table 4 (lines 447–462); training and evaluation details in Appendix I.4.1, Tables 19–20 (lines 2029–2115). The improvement is also stated in the Abstract (lines 27–28) and Conclusion (lines 534–535).

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

CryoNet.Density translates an analytically rendered density map into an experimental-like cryo-EM map using a resolution-conditioned 3D Brownian Bridge Diffusion Model. Requested resolution controls both the renderer and Fourier-FiLM conditioning in the denoiser. Training includes bridge states and the clean rendered boundary; default inference uses one stochastic transition and two evaluations of the same denoiser. The paper reports improved spatial and Fourier-domain agreement on a 1,923-case test set, resolution sweeps, frozen downstream recognition experiments, and a database of over 3.6 million generated maps.

Critical strengths
  1. The physical-to-experimental formulation is well matched to the task, and the reported gains are substantial across complementary measurements. Table 1 improves mean masked FSC AUC from 0.304 for the strongest baseline to 0.449, and PCC from 0.662 to 0.751. Section 5 also reports comparisons that separate resolution conditioning, bridge training, and sampling. This helps motivate the individual design choices rather than attributing all improvement to diffusion.

  2. The paper tests utility beyond map similarity using frozen Emap2sec+ and ModelAngelo models, and examines resolution response across 96 structures rather than relying only on a single illustration. Taken together, these are promising steps toward a useful synthetic-data tool, even though the downstream experiments do not yet demonstrate improved learning from synthetic training data.

Critical weaknesses
  1. The dataset construction and separation need a clearer justification. Section 4.1 combines archive maps with multiple reconstructions from subsampled particle sets, then reports a split based on PDB IDs and sequence clusters. Reconstructions from a common experimental source are correlated, so it is important to establish that parent particle sets, as well as related structures, remain within a single split. A supporting concern from the dataset appendix is that its source counts are difficult to reconcile with the reported 19,005 distinct PDB IDs: many reconstructions are described as reusing their parent atomic models. This may be a reporting issue, but it affects confidence in the benchmark on which the main fidelity claims rest. A concise reconciliation of sources, independent structures, and split units is needed.

  2. The added value over physical rendering needs a stronger demonstration. Producing an idealized density map from atomic coordinates is already straightforward, and the paper appropriately includes analytical renderers as baselines. The substantive contribution is therefore the learned correction toward experimental reconstructions, not map generation itself. Table 1 shows meaningful gains in reference-map agreement, but an atomic model and a nominal resolution do not uniquely determine occupancy, conformational heterogeneity, imaging conditions, or reconstruction artifacts. Section 5 acknowledges that the learned residual combines several of these effects. It remains unclear which corrections are useful experimental signal and which reflect the processing conventions of the training maps. Section 4.3 establishes bandwidth control, and the paper correctly distinguishes this from independently measured experimental resolution. The frozen downstream tests are useful, but they do not demonstrate that training on the generated database improves performance on unseen experimental maps. Moreover, Section 4.4 retains only 711 of 1,524 eligible ModelAngelo cases, so the reported utility excludes more than half of that cohort. A focused synthetic-training-to-real-test experiment, or another direct test of the database's intended use, would make the practical contribution more convincing.

For context, the authors could consider CryoDRGN-AI [Levy et al., 2025], which infers heterogeneous structures from particle images, and CryoBench [Jeon et al., 2024], which studies evaluation of heterogeneous reconstruction. These address a different task and are not required direct baselines. They would help distinguish structure-conditioned map synthesis from recovering experimentally supported structural variability, and clarify which uses of the proposed database are actually tested. The ModelAngelo reference should also be updated to its published version [Jamali et al., 2024].

References to consider or update: [Levy et al., 2025] Axel Levy et al. "CryoDRGN-AI: neural ab initio reconstruction of challenging cryo-EM and cryo-ET datasets." Nature Methods, 22:1486–1494, 2025.

[Jeon et al., 2024] Minkyu Jeon et al. "CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM." Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2024.

[Jamali et al., 2024] Kiarash Jamali et al. "Automated model building and protein identification in cryo-EM maps." Nature, 628:450–457, 2024.

Questions for authors
  1. Clarification: Please reconcile the source counts with the number of independent structures and explain how parent experimental sources are handled in the split. Are reconstructions from a common particle source always confined to one split? How much cross-split overlap remains when all polymer entities are considered?

  2. Assessment-changing analysis: What are ModelAngelo's per-input failure rates and results over all 1,524 eligible cases under a prespecified treatment of failed correspondences? Does the ranking remain the same?

  3. Assessment-changing utility test: Does training a downstream method on CryoNet.Density maps improve its performance on held-out experimental maps compared with training on the same atomic structures rendered analytically, under matched training conditions? This would directly test the benefit of the learned correction for the proposed database.

  4. Clarification: Which aspects of experimental variability are intended to be modeled at fixed atomic coordinates and nominal resolution, and what evidence validates those aspects beyond spectral bandwidth and agreement with a single reference map?

6 comments
AuthorsOpenReview ↗

Overview and dataset construction

OVERALL RESPONSE

Thank you for the your suggestions on data separation and downstream evaluation, and source accounting. We have clarified the dataset construction and added experiments to assess the practical value of our generated maps. The main updates are (line numbers refer to the revised manuscripts):

1. CryoNet.Density generated maps improve downstream training performance. We fine-tuned ModelAngelo using CryoNet.Density maps and evaluated it on 60 held-out experimental maps. Compared with the original ModelAngelo, Cα F1 improved from 50.10% to 53.71%, and sequence recovery improved from 40.32% to 43.83% (Question 3; Table 4, lines 447–462).

2. CryoNet.Density performs best across the full test cohort. After correcting an evaluation error, we re-evaluated all 1,524 eligible cases. Maps generated by CryoNet.Density achieved the highest residue accuracy (0.629), sequence match (92.8%), and sequence score (0.912) among the evaluated inputs (Question 2; Table 3, lines 415–427).

3. We clarified dataset composition and data separation. The dataset contains 19,273 map–model pairs from 19,005 PDB entries. For each of the seven structures with particle-subsampled maps, all maps derived from the same parent dataset were assigned to the same data split; none of these subsampled maps was included in the test set (Question 1; Section 4.1, lines 272–284; Appendix F.1, lines 1117–1137).

4. The learned density difference captures spatial density patterns around the molecule. We analyzed the differences between experimental maps and physically rendered maps. These differences show spatial patterns aligned with the molecular structure that cannot be explained by the power spectrum alone (Question 4; Tables 30–31, lines 2722–2759).

We first clarify the dataset composition and splitting procedure, then explain how the full-cohort evaluation and fine-tuning experiment test the utility of the generated maps.

Comment 1: Clarification: Please reconcile the source counts with the number of independent structures and explain how parent experimental sources are handled in the split. Are reconstructions from a common particle source always confined to one split? How much cross-split overlap remains when all polymer entities are considered?

Response: Thank you for requesting this audit. We checked the dataset records and corrected the source counts. The retained dataset is:

SourceTrainingValidationTestHomology-excludedTotal pairsDistinct PDB IDs
EMDB archive13,4561,8941,9231,75519,02819,005
Particle-subsampled reconstructions20342002457
Combined13,6591,9361,9231,75519,27319,005

All seven subsampling parent PDB IDs already occur in the archive source, yielding 19,273 pairs from 19,005 distinct PDB IDs when counting distinct PDB IDs. Multiple maps of one PDB count as multiple pairs but one structure ID. Training, validation, and test contain 17,518 pairs from 17,255 PDB IDs; the remainder is homology-excluded.

For every subsampling parent, all derived maps and its archive record belong to the same split:

Parent PDB IDParent EMDB IDSubsampled mapsPartition of subsampled maps and archive record
PDB-3j7hEMD-59951Training
PDB-6knfEMD-073042Validation
PDB-6vrsEMD-2137150Training
PDB-6xdcEMD-221366Training
PDB-6zbjEMD-1115647Training
PDB-6zblEMD-1115749Training
PDB-7c7lEMD-3029950Training

The 203 maps from six structures are in training, and the 42 maps from PDB-6knf are in validation. None is in the test set. Parent groups are defined by recorded PDB/EMDB identifiers; acquisition identifiers across separate archive entries are incompletely recorded.

All records with the same PDB ID receive one assignment. For sequence grouping, we select entity 1 from each PDB entry, or the lowest-numbered entity if entity 1 is absent. We use the selected entities' RCSB 40% sequence clusters to construct the splits and check that these selected-entity clusters do not overlap between training, validation, and test. The original cluster snapshot covers 18,888 PDB entries; the remaining 117 are individually grouped in training.

Revised in the manuscript: Section 4.1 (lines 272–284); source counts and parent assignments in Appendix F.1, Table 6 (lines 1117–1137); split protocol and audit in Appendix F.3, Table 7 (lines 1244–1277).

AuthorsOpenReview ↗

Full-cohort evaluation and downstream training

Comment 2: Assessment-changing analysis: What are ModelAngelo's per-input failure rates and results over all 1,524 eligible cases under a prespecified treatment of failed correspondences? Does the ranking remain the same?

Response: We re-evaluated all 1,524 eligible cases after correcting an evaluation error. Unavailable Phenix scores had been recorded as zero, so the previous analysis retained only 711 cases. The other 813 cases had at least one failed Phenix evaluation. We corrected missing-score handling and structure-file formatting/chain identifiers, then repeated the comparison with the same 3 Å matching radius. Model weights, predictions, and coordinates were unchanged.

All 1,524 cases now have valid evaluations for each input. Success is defined as a nonempty prediction with positive Phenix sequence match: all four inputs yield a nonempty prediction with a positive sequence match for every case; this does not establish complete or accurate structure recovery. The case-equal means are:

Input mapResidue accuracy ↑Sequence match (%) ↑Sequence score ↑
Experimental0.58885.10.786
ChimeraX molmap0.47883.00.824
CryoNet molmap0.49684.10.835
CryoNet.Density0.62992.80.912

CryoNet.Density remains highest in all three metrics. Scores are accepted only after successful Phenix execution and parsing; unresolved evaluations remain missing and block aggregation. Eight previously successful control comparisons retained identical sequence match and sequence score.

Incorporated into the manuscript: Section 4.4, Table 3 (lines 415–427), and the corrected protocol in Appendix I.4 (lines 2015–2027).

Comment 3: Assessment-changing utility test: Does training a downstream method on CryoNet.Density maps improve its performance on held-out experimental maps compared with training on the same atomic structures rendered analytically, under matched training conditions? This would directly test the benefit of the learned density difference for the proposed database.

Response: We fine-tuned ModelAngelo using CryoNet.Density maps and compared it with fine-tuning on physically rendered maps. Both models were evaluated on the same 60 held-out experimental maps at 3–6 Å.

Both training sets contain 863 structures at seven resolutions (3–6 Å in 0.5 Å intervals), giving 6,041 maps per set. They use identical atomic coordinates and grids. We initialized both models from the official ModelAngelo checkpoint and fine-tuned the full GNN for 10,000 AdamW steps at a learning rate of 10⁻⁶, using the same loss and number of training graph crops. CNN and ESM-1b parameters remained fixed. Evaluation used the same CNN candidate points, ESM features, three GNN iterations, sequence assignment, and pruning.

The official model is the common baseline (O). Values are percentages; Δ denotes percentage-point changes calculated before rounding. The first four metrics are averaged across cases; the remaining metrics combine residue counts across cases.

Metric ↑Official (O)Generated (G)ΔG−OPhysical (P)ΔP−O
Cα precision88.4593.51+5.0787.84-0.61
Cα recall42.3145.97+3.6733.23-9.08
Cα F150.1053.71+3.6040.23-9.87
Sequence recovery40.3243.83+3.5131.22-9.10
Pre-pruning residue accuracy53.0756.60+3.5343.40-9.67
Final end-to-end macro-F150.5653.69+3.1339.18-11.38

Training on CryoNet.Density maps increased Cα F1 from 50.10% to 53.71%, sequence recovery from 40.32% to 43.83%, and end-to-end macro-F1 from 50.56% to 53.69%. Training on physically rendered maps gave lower scores. These results show that the generated maps improve ModelAngelo's performance on experimental data. Residue accuracy measures amino-acid classification among matched residues; end-to-end macro-F1 also accounts for missing and incorrectly predicted residues.

The test set contains 60 cases: 39 at 3–<4 Å, 18 at 4–<5 Å, and three at 5–6 Å. We required CC_mask ≥ 0.5, Cα coverage ≥ 0.95, and compatible map sampling, and excluded sequence overlap with the fine-tuning data. The 60 PDB entries also have no ID overlap with ModelAngelo training set. Predicted and reference Cα atoms were matched one-to-one within 3 Å without rigid-body fitting; valid empty predictions received zero recovery.

Revised in the manuscript: Section 4.4, Table 4 (lines 447–462); training and evaluation details in Appendix I.4.1, Tables 18–19 (lines 2029–2113). The improvement is also stated in the Abstract (lines 27–28) and Conclusion (lines 534–535).

AuthorsOpenReview ↗

Experimental variability and related work

Comment 4: Which aspects of experimental variability are intended to be modeled at fixed atomic coordinates and nominal resolution, and what evidence validates these aspects beyond spectral bandwidth and agreement with a single reference map?

Response: Experimental variability. Experimental maps depend on imaging conditions, particle sampling, reconstruction, local resolution, and structural heterogeneity, which are not determined by atomic coordinates and nominal resolution. CryoNet.Density learns density differences from paired maps, but it cannot specify these factors for an individual sample or uniquely attribute differences between generated and physically rendered maps to one physical process. Variation across random seeds does not directly establish real conformational heterogeneity or a validated estimate of experimental uncertainty.

Evidence. (1) Spatial autocorrelation and power-spectrum analyses on 1,923 cases show that differences between experimental and physically rendered maps have spatial structure. Phase-randomization controls on 96 structures remove alignment with molecular features while preserving the power spectrum (Tables 30–31, lines 2722–2759; Figure 23).

(2) Frozen Emap2sec+ and ModelAngelo tests on 136 and 1,524 cases show that generated maps contain structural features that models trained on experimental maps can use (Tables 2–3, lines 408–427). Five samples for each of three structures demonstrate conditional sample diversity, with similar overall density and local differences around the molecule (Table 28, lines 2533–2544; Figure 21).

(3) For five repeated experimental reconstructions and 20 generated maps, the mean maps have PCC 0.801 (Table 29, lines 2594–2618; Figure 22). Fine-tuning results on 60 held-out experimental maps further support the use of generated maps for training (Question 3; Table 4, lines 447–462).

Revised in the manuscript: Section 3.1 describes sampling at fixed coordinates and resolution (lines 177–179); supporting results are included at the locations above.

Comment 5: For context, the authors could consider CryoDRGN-AI [Levy et al., 2025], which infers heterogeneous structures from particle images, and CryoBench [Jeon et al., 2024], which studies evaluation of heterogeneous reconstruction. These address a different task and are not required direct baselines. They would help distinguish structure-conditioned map synthesis from recovering experimentally supported structural variability, and clarify which uses of the proposed database are actually tested. The ModelAngelo reference should also be updated to its published version [Jamali et al., 2024].

Response: We added the recommended studies to clarify their relationship to our task. CryoDRGN-AI (Levy et al., 2025) reconstructs heterogeneous three-dimensional structures from particle images, and CryoBench (Jeon et al., 2024) provides datasets and evaluation protocols for heterogeneous reconstruction. CryoNet.Density generates maps from a supplied atomic structure and nominal resolution; different samples vary in density around the same atomic coordinates.

We also updated the ModelAngelo citation to its published version: Jamali et al. (2024), “Automated model building and protein identification in cryo-EM maps,” Nature 628:450–457.

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

The authors presents CryoNet.Density, a framework for generating resolution-controllable cryo-EM density maps from atomic structures. The method combines a differentiable physics-based renderer with a conditional Brownian bridge diffusion model to transform simulated densities into experimental-like maps. It is trained on 19,273 structure-map pairs and outperforms several analytical and learning-based baselines across map similarity metrics and structure analysis tasks. The paper also presents a large database containing over 3.6 million generated density maps.

Critical strengths
  • CryoNet.Density achieves good results across multiple evaluation metrics compared to the baselines. Additionally, the large database of resolution-controlled density maps is a valuable contribution and could facilitate future data-driven method development in cryo-EM.
Critical weaknesses
  • Lack of motivation. The paper argues that “bridging these two representations ... is important for developing, training, and evaluating computational methods for cryo-EM structure analysis” (L40-42), but this is not sufficiently validated. The experiments mainly demonstrate map similarity and compatibility with frozen downstream models, rather than showing that the proposed bridge improves the development or training of such methods.

  • It is well-known that synthetic and and experimental maps look different, and the discrepancy arises from many case-specific factors, including structural heterogeneity, spatially varying resolution, reconstruction artifacts, and imperfect data processing. Many of these effects cannot generally be inferred from an atomic model and a nominal resolution alone. More fundamentally, why do we need to train a generative model to reproduce these degradations and dataset-specific biases?

  • The work therefore feels incomplete as a whole story. The method and database provide large-scale “experimental-like” structure-density pairs, whose most natural application is data augmentation. However, the paper does not show any downstream method with these generated data or demonstrate a resulting improvement. Without such validation, the key advantage of learning this density-generation process remains unclear.

Questions for authors

The authors select EMDB entries with reported resolutions between 1.5 and 15 Å and retain cases with a map-model CC_mask >= 0.5. Could the authors show some examples with relatively low reported resolution, e.g. those worse than 5 Å? At such resolutions, the map itself may not support accurate atomic modeling, and the deposited model may instead have been obtained from other sources and docked into the map. Treating these cases as ground truth atomic model-density pairs during training may not be appropriate.

5 comments
AuthorsOpenReview ↗

Overview and downstream training

OVERALL RESPONSE:

Your suggestions helped us test whether generated maps improve downstream training and clarify the use of low-resolution data. We added the following results (line numbers refer to the revised manuscript):

1. Training on generated maps improves ModelAngelo on experimental maps. We fine-tuned ModelAngelo on CryoNet.Density maps and evaluated it on 60 held-out experimental maps. Cα F1 increased from 50.10% to 53.71%, anxd sequence recovery from 40.32% to 43.83% (Weaknesses 1 and 3; Table 4, lines 447–462).

2. Show more low-resolution examples clarify map-model agreement. We added examples showing how the CC_mask ≥ 0.5 criterion removes poorly matched map–model pairs. The 1,340 retained pairs with resolutions worse than 5 Å have a median CC_mask of 0.6844 (Question; Appendix F.2, lines 1163–1172; Figure 10).

3. The density difference contains spatial patterns associated with the molecule. We analyzed the differences between experimental and physically rendered maps. The differences are spatially aligned with the molecular structure and cannot be explained by the power spectrum alone (Weakness 2; Tables 30–31, lines 2722–2759).

We first describe how training on generated maps improves ModelAngelo on experimental data, then explain the density differences and the use of low-resolution training pairs.

Comment 1: The work therefore feels incomplete as a whole story. The method and database provide large-scale “experimental-like” structure-density pairs, whose most natural application is data augmentation. However, the paper does not show any downstream method with these generated data or demonstrate any resulting improvement. Without such validation, the key advantage of learning this density-generation process remains unclear.

Response: We fine-tuned ModelAngelo using CryoNet.Density maps and compared it with fine-tuning on physically rendered maps. Both models were evaluated on the same 60 held-out experimental maps at 3–6 Å.

Both training sets contain 863 structures at seven resolutions (3–6 Å in 0.5 Å intervals), giving 6,041 maps per set. They use identical atomic coordinates and grids. We initialized both models from the official ModelAngelo checkpoint and fine-tuned the full GNN for 10,000 AdamW steps at a learning rate of 10⁻⁶, using the same loss and number of training graph crops. CNN and ESM-1b parameters remained fixed. Evaluation used the same CNN candidate points, ESM features, three GNN iterations, sequence assignment, and pruning.

The official model is the common baseline (O). Values are percentages; Δ denotes percentage-point changes calculated before rounding. The first four metrics are averaged across cases; the remaining metrics combine residue counts across cases.

Metric ↑Official (O)Trained with CryoNet.Density Generated maps (G)ΔG−OTrained with Physical Generated maps (P)ΔP−O
Cα precision88.4593.51+5.0787.84-0.61
Cα recall42.3145.97+3.6733.23-9.08
Cα F150.1053.71+3.6040.23-9.87
Sequence recovery40.3243.83+3.5131.22-9.10
Pre-pruning residue accuracy53.0756.60+3.5343.40-9.67
Final end-to-end macro-F150.5653.69+3.1339.18-11.38

Training on CryoNet.Density maps increased Cα F1 from 50.10% to 53.71%, sequence recovery from 40.32% to 43.83%, and end-to-end macro-F1 from 50.56% to 53.69%. Training on physically rendered maps gave lower scores. These results show that the generated maps improve ModelAngelo's performance on experimental data. Residue accuracy measures amino-acid classification among matched residues; end-to-end macro-F1 also accounts for missing and incorrectly predicted residues.

The test set contains 60 cases: 39 at 3–<4 Å, 18 at 4–<5 Å, and three at 5–6 Å. We required CC_mask ≥ 0.5, Cα coverage ≥ 0.95, and compatible map sampling, and excluded sequence overlap with the fine-tuning data. The 60 PDB entries also have no ID overlap with ModelAngelo training set. Predicted and reference Cα atoms were matched one-to-one within 3 Å without rigid-body fitting; valid empty predictions received zero recovery.

Revised in the manuscript: Section 4.4, Table 4 (lines 447–462); training and evaluation details in Appendix I.4.1, Tables 18–19 (lines 2029–2113). The improvement is also stated in the Abstract (lines 27–28) and Conclusion (lines 534–535).

AuthorsOpenReview ↗

Density differences and low-resolution examples

Comment 2: It is well-known that synthetic and experimental maps look different, and the discrepancy arises from many case-specific factors, including structural heterogeneity, spatially varying resolution, reconstruction artifacts, and imperfect data processing. Many of these effects cannot generally be inferred from an atomic model and a nominal resolution alone. More fundamentally, why do we need to train a generative model to reproduce these degradations and dataset-specific biases?

Response: The purpose is to generate training data that contain density features encountered in experimental maps. CryoNet.Density learns differences in local density intensity, boundary sharpness, and spatial texture between physical renderings and experimental maps of the same structure. These differences reflect the combined effects of data collection, reconstruction, occupancy, and structural heterogeneity. The fine-tuning experiment in Part 1 directly tests their usefulness: training on generated maps improves ModelAngelo on experimental maps.

We also analyzed the experimental map minus the physical rendering. Across 1,923 cases, the median correlation between neighboring residual voxels is 0.635, and the median fraction of squared residual density within the molecular region is 0.845. In a control on 96 structures, randomizing Fourier phases while preserving the power spectrum reduced alignment with density gradients from 0.4611 to −0.0002 and with curvature from 0.5512 to −0.0001. Thus, the map differences contain spatial information associated with molecular geometry beyond spectral bandwidth.

We further compared five experimental reconstructions at 4.57 Å with 20 generated samples. The mean maps had PCC 0.801, but generated voxel variance was 3.85% of the experimental value and the variance-map PCC was 0.090. This quantifies both agreement in mean density and the smaller variation among generated samples.

Revised in the manuscript: the interpretation in Section 5 (lines 506–517), residual controls in Appendix J.5, Tables 30–31 (lines 2722–2759), and repeated-reconstruction results in Appendix J.4.1, Table 29 (lines 2594–2618). These accompany the training-utility result in Table 4.

Comment 3: The authors select EMDB entries with reported resolutions between 1.5 and 15 Å and retain cases with a map-model CC_mask >= 0.5. Could the authors show some examples with relatively low reported resolution, e.g., those worse than 5 Å? At such resolutions, the map itself may not support accurate atomic modeling, and the deposited model may instead have been obtained from other sources and docked into the map. Treating these cases as ground truth atomic model-density pairs during training may not be appropriate.

Response: Following your suggestion, we added examples of deposited models and experimental maps at resolutions worse than 5 Å. The model supplies the atomic coordinates used as input, and the experimental map is the training target. We use CC_mask ≥ 0.5 to retain pairs with consistent molecular arrangements. This checks map–model agreement before dataset splitting; it is not a criterion for assigning training and test sets.

Of 1,950 geometrically valid pairs above 5 Å, 600 had CC_mask < 0.5 and nine lacked scores. Of the 1,341 passing pairs, one PDB-7kod pair was reserved for the external resolution series, leaving 1,340 pairs with median CC_mask 0.6844.

Excluded pairResolution (Å)CC_maskRetained pairResolution (Å)CC_mask
PDB-6nme / EMD-04495.670.0460PDB-7pfw / EMD-133825.200.8660
PDB-6eny / EMD-39065.800.3946PDB-6v9i / EMD-211215.200.5231
PDB-6sxb / EMD-103387.900.0451PDB-7pez / EMD-133627.900.8311

The retained examples follow the experimental density, whereas excluded examples contain regions that do not match it. Figure 10 shows two orthogonal views per case in the deposited coordinate frame, using the recommended map contours and the same opacity. A model fitted from another source can provide suitable input geometry when its molecular arrangement agrees with the map; this filter assesses that agreement, not the accuracy of every atomic coordinate.

Revised in the manuscript: Appendix F.2 (lines 1163–1172) and Figure 10, including the examples and their role in conditional density training.

We sincerely hope that our responses and revisions have addressed your concerns. We would be very grateful if you could let us know whether the issues you raised have been adequately resolved. If so, we would be deeply grateful if you would consider raising your evaluation of the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of manuscript revisions and further discussion

We thank all reviewers for their constructive comments, which have helped us improve the manuscript. We appreciate Reviewer Sxot’s positive reassessment and Reviewer bxBG’s recognition of the additional experiments and clarifications.

To facilitate further discussion, we have incorporated the following additional results and clarifications into the revised manuscript:

  • Section 4.4 and Appendix I.4.1 — Downstream training and evaluation. Following the reviewers’ suggestions, we fine-tuned ModelAngelo on generated maps and evaluated it on 60 held-out experimental maps. Cα F1 improved from 50.10% to 53.71%, and sequence recovery from 40.32% to 43.83%. We also expanded the original ModelAngelo evaluation to all 1,524 eligible cases.
  • Appendices F.2 and J.5 — Low-resolution examples and density differences. In response to Reviewer xWut, we added low-resolution map–model examples illustrating agreement in the retained pairs. Analyses of 1,923 cases, together with phase-randomization controls on 96 structures, show that differences between experimental and physically rendered maps contain spatial patterns associated with molecular structure.
  • Appendices F.3 and I.4.1 — Data independence. In response to Reviewer bxBG’s follow-up question, we clarified the dataset partition and the protein-chain sequence comparisons used to exclude related structures. We also clarified that the 60 downstream test maps were held out from training both the map generator and the downstream models.
  • Appendices I.3.1, J.4.1, and the diffusion comparison — Additional validation requested by Reviewer Sxot. We added comparisons with repeated experimental reconstructions, evaluated local map agreement and intermediate target resolutions, and compared our method with standard conditional diffusion under matched training conditions. These experiments provide further evidence concerning conditional density variation, resolution control, and the Brownian bridge formulation.

We understand the demands on reviewers’ time. We would be grateful if Reviewer xWut could consider the added results and if Reviewer bxBG could let us know whether our follow-up clarification addresses the remaining concern about data independence. We welcome further discussion and sincerely thank all reviewers for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.