acceptodds
Under review as a conference paper at ICLR 2027

Atomic Structure Modeling with Sparse Voxel Representations in Cryo-EM

Abstract

High-resolution cryo-electron microscopy maps provide rich stereochemical cues for proteins, nucleic acids, and ligands, but preserving this information throughout atomic modeling remains challenging. Existing learned builders rely on local map crops, patch tokens, or density samples at predicted atoms, without retaining voxel-wise semantics through coordinate generation. We present CryoASTRA, which computes on the 1 Å grid only within the macromolecular region and retains a semantic voxel representation through atomic modeling. It introduces (i) a two-scale representation: EM points aligned with the input sequences, and a semantic voxel representation from the voxel encoder for atomic modeling; (ii) Heterogeneous BallTree Attention, a sparse cross-attention where atoms query nearby voxels with linear attention cost for a fixed leaf size; and (iii) a GPU sparse pipeline for whole-map inference without sliding windows. On 103 sequence-novel maps from 2026 (1.5–3.9 Å), our model places on average 92.5% of protein atoms within 3 Å of deposited positions, has the highest TM-score among the evaluated builders on proteins and nucleic acids, and models the input ligands.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cryo-EM, biomolecular structure modeling, large-scale 3D learning, sparse 3D representations, efficient attention

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_kk76OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CryoASTRA, a framework for building atomic models from cryo-EM maps, polymer sequences, and ligand identities. It combines sparse semantic voxel features, sequence-aligned EM points, and an atomic diffusion model with local atom-to-voxel attention.

Critical strengths
  • The framework integrates protein, nucleic-acid, and ligand modeling and supports whole-map encoding without sliding windows.
  • Evaluation on 103 temporally held-out, sequence-filtered maps demonstrates improved protein model completeness and a statistically significant TM-score gain over CryoAtom2.
Critical weaknesses
  • Missing core ablations. The paper does not isolate the contributions of semantic voxel conditioning and Heterogeneous BallTree Attention. Consequently, the improvements cannot be clearly attributed to these proposed designs versus pretrained structure-prediction components, MSA information, or other architectural changes.
  • Incomplete validation of atomic model quality. The reported backbone and side-chain accuracy and RMSD results are informative, but complementary assessments of stereochemical quality and model–map agreement are missing. Representative geometry and density-fit measures, such as MolProbity and Q-score or local map–model correlation, would strengthen the evidence for improved atomic model quality. Such comparisons should use consistent refinement protocols across methods.
  • Limited validation beyond proteins. Mean Accuracy@3 Å is approximately 47% for both nucleic acids and ligands, indicating substantial remaining errors. The absence of comparisons with dedicated ligand-fitting methods further limits conclusions about the practical advantage of unified modeling.
Questions for authors
  1. With the pretrained backbone and training setup held fixed, how does performance change when semantic voxel features are removed or replaced with raw density, and when Heterogeneous BallTree Attention is replaced with simpler local attention?
  2. Can the authors report stereochemical quality, local map–model agreement, and stricter atom-position accuracy metrics, including comparisons on commonly modeled atoms, to distinguish improvements in completeness from improvements in coordinate accuracy?
4 comments

Thank you for the clear and actionable review. To stay well within OpenReview's 5,000-character limit per comment, we split our response into Part 1 (W1, Q1) and Part 2 (W2, Q2, W3).

1. HBA Benchmarks & Sequence-Prior Controls (W1, Q1)

"Benchmark HBA directly against GPU kkNN and global attention, and isolate map conditioning from pretrained sequence priors."

  • HBA vs. GPU kkNN & global attention (§4.5, Table 3; App. I.2, Tables 16–17): Across 10 real maps (4,942–35,021 atoms; 42,686–345,181 active voxels) with identical 32-basis Gaussian RBF distance biases, 3D RoPE, and BF16 SDPA across two blocks:
MethodGraph + Layout (ms) ↓Forward (ms) ↓F+B Time (ms) ↓F+B Peak (Δ MiB) ↓Completed
HBA (ℓav=64\ell_{av}=64)2.39–3.132.20–12.626.57–45.74722–5,23410 / 10
HBA (ℓav=256\ell_{av}=256, default)2.09–2.852.40–12.496.20–37.57692–5,02810 / 10
PyG kkNN (k=64k=64)296–1,7823.49–24.858.99–60.671,235–8,77910 / 10
Torch kkNN (k=256k=256)5.30–490.2513.42–96.8431.85–269.274,653–26,9659 / 10 (OOM 9PHF)
Global (16-query chunked)—452–32,521OOM on all 10OOM on all 100 / 10
  • At matched capacity (ℓav=256\ell_{av}=256, averaging 225–233 valid voxel neighbors per atom across two partitions), HBA builds graphs 2.5–172× faster than exact Torch GPU kkNN (k=256k=256), runs forward 5.58–7.91× faster, runs forward+backward 5.14–7.36× faster, and uses 5.22–6.72× less peak memory, completing all 10 maps while Torch kkNN OOMs on 9PHF during autograd. Against PyG GPU kkNN (k=64k=64, max PyG CUDA kernel is 100), HBA (ℓav=64\ell_{av}=64) builds graphs 124–569× faster and uses 1.68–1.71× less F+B peak memory. Unchunked explicit RBF tensors require 25.15–1,375.62 GiB per block, and 16-query chunked global cross-attention OOMs on all 10 maps in backward.
  • Map conditioning vs. sequence priors (§4.1, §4.4, Table 1; App. H.1.1, H.3, Tables 9–11): Even after US-align rigid fitting onto the deposited PDB across all 103 targets, map-free AlphaFold 3 and Protenix-Mini reach only 68.43% and 63.04% protein Accuracy@3 Å (Q-scores 0.2083, 0.1794) and 18.07% / 11.64% on nucleic acids, vs. 92.49% (0.4691 Q-score) and 47.39% for unaligned CryoASTRA (MSA). Conversely, removing MSAs via single-sequence ESM2-3B embeddings (CryoASTRA ESM) keeps protein Accuracy@3 Å at 89.07% (0.948 TM-score, 0.4775 Q-score) and improves nucleic-acid Accuracy@3 Å to 51.90% (0.608 TM-score). As noted in §5, raw density ρv\rho_v is already in dv=[ρv;pv]\mathbf{d}_v = [\rho_v; \mathbf{p}_v] while pv\mathbf{p}_v selects V\mathcal{V}, so retaining ev\mathbf{e}_v is presented as a design property rather than an isolated voxel-encoder ablation.

2. Common-Atom Accuracy, Q-score & Stereochemistry (W2, Q2)

"Evaluate accuracy on commonly modeled atoms, map fit (Q-score), and stereochemistry / real-space refinement."

  • Commonly modeled atoms (§4.2; App. H.5, Table 12): On the intersection of reference protein heavy atoms modeled by both methods in the map frame:
    • All 101 protein targets vs. CryoAtom2 (98.26% atom coverage): Common-atom Accuracy@3 Å is 93.16% vs. 91.84% (Δ=+1.32\Delta = +1.32 pp, 95% CI [−1.72, 4.64]), and common-atom RMSD is significantly lower for CryoASTRA (3.20 Å vs. 6.64 Å, Δ=−3.44\Delta = -3.44 Å, CI [−6.63, −0.72]), while CryoASTRA builds 100.00% vs. 92.48% of polymer residues (TM-score 0.972 vs. 0.855).
    • 50 CryoACE-completed targets (Table 12): Common-atom Accuracy@3 Å is 95.00% vs. 95.63% against CryoAtom2 (RMSD 2.40 vs. 2.45 Å), 96.21% vs. 94.59% against EMProt (RMSD 2.02 vs. 2.66 Å), and 94.05% vs. 85.54% against CryoACE (Δ=+8.52\Delta = +8.52 pp, CI [2.50, 15.62], RMSD 2.78 vs. 4.72 Å).
  • Q-score, active-voxel filtering & Phenix refinement (§4.2, §4.4, Table 1; App. H.3, Table 10; App. H.6, Tables 13–14):
    • Unfiltered CryoASTRA (MSA) places 100% of sequence residues (including disordered loops without map density), giving full-protein Q-score 0.4691 vs. 0.5230 for CryoAtom2, which narrows to 0.5089 vs. 0.5350 on commonly modeled atoms (Table 13). Filtering to atoms within ≤1\le 1 Å of an active voxel center (91.53% completeness; Tables 1 & 10) lifts full-protein Q-score to 0.5198 (MSA) / 0.5318 (ESM) and conditional Accuracy@3 Å to 94.92%, while lowering RMSD from 3.47 → 2.73 Å.
    • Before refinement on 101 targets (Table 13), raw diffusion outputs have higher clashscore (211.68 vs. 32.23), Ramachandran outliers (5.47% vs. 0.58%), and rotamer outliers (19.71% vs. 4.51%) than CryoAtom2. Running paired phenix.real_space_refine across 9 targets (45 CryoASTRA samples and 9 CryoAtom2 models; Table 14) reduces CryoASTRA's mean clashscore from 237.0 → 21.5 and raises Q-score from 0.424 → 0.456, while keeping Accuracy@3 Å at 87.3% → 87.2%.

3. Nucleic Acids and Ligands (W3)

"Provide deeper analysis of nucleic-acid and ligand performance and tone down overclaims."

We toned down wording in the Abstract and Conclusion ("models nucleic acids and the input ligands") and added sub-component and reference-aligned evaluations in Table 1, App. H.3 (Table 10), and App. H.4 (Table 11):

  • Nucleic acids (14 targets, Tables 1 & 11): CryoASTRA builds 100% of nucleotides (vs. 66.62% CryoAtom2, 50.49% ModelAngelo) with the highest joint NA TM-score (0.561 MSA, 0.608 ESM, vs. 0.424 CryoAtom2, 0.274 ModelAngelo, 0.143 CryoREAD, 0.133 EM2NA), and leads all baselines in Accuracy@3 Å across backbone (46.17% / 50.99% vs. 41.78% CryoAtom2, 16.86% AF3†, 8.93% CryoREAD, 8.69% EM2NA), phosphate (44.35% / 49.46%), sugar (46.80% / 51.52%), and base (48.90% / 53.02%; Table 11).
  • Ligands (66 targets, Tables 1 & 10): Among the evaluated builders, only CryoASTRA models proteins, nucleic acids, and input ligands jointly, reaching 47.35% Accuracy@3 Å (53.42% at ≤1\le 1 Å filtering) and 0.3312 Q-score (0.4121 filtered; 0.4270 ESM filtered), ahead of PDB-aligned AF3 (40.72%, Q-score 0.2026) and Protenix-Mini (31.53%, Q-score 0.1349). As noted in §4.2 and §5, ligand accuracy remains limited by lower local resolution and the single ligand semantic channel, so placements should undergo downstream refinement.

Thank you for the additional experiments. The HBA benchmarks clarify its efficiency advantage, and the common-atom comparisons, Q-scores, and stereochemical checks show both the strengths of the protein models and where refinement is needed. These results largely address my concerns, and I will raise my rating from Weak Reject to Weak Accept. Please retain one caveat in the revision: the experiments do not isolate whether retaining voxel-level semantic features during coordinate generation improves accuracy over using density alone.

Thank you very much for re-evaluating our work, for the helpful suggestions, and for raising your score. We agree completely and will ensure that this caveat—that our experiments do not isolate the accuracy gain of retaining voxel-level semantic features during coordinate generation over density alone.

Reviewer_huyfOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents CryoASTRA, an automated framework for building protein, nucleic-acid, and ligand structures from cryo-EM density maps and molecular sequences. The method combines residue-level EM points with a sparse semantic voxel representation, which conditions an AlphaFold-style trunk and an atomic diffusion model. The paper also introduces GPU-based Overlapping BallTree construction and Heterogeneous BallTree Attention to enable atom-voxel interactions. The method achieves strong performance on a prospective set of 103 sequence-novel cryo-EM maps.

Critical strengths
  • Cryo-EM maps are highly sparse in voxel space, yet retaining a fine voxel grid is important for preserving structural information. The sparse voxel representation and optimized sparse U-Net pipeline therefore address an important computational bottleneck, particularly GPU memory consumption.

  • Retaining informative voxel-level features throughout atomic modeling is intuitive and well motivated. CryoASTRA also achieves strong experimental results, outperforming most baselines across the main benchmarks.

  • CryoASTRA provides a relatively comprehensive modeling framework that supports proteins, nucleic acids, and input ligands, whereas most baselines support only a subset of these molecule types.

  • The GPU engineering contributions, including the accelerated MeanShift++ implementation, GPU BallTree construction, and reusable attention primitives, may also benefit other applications involving large sparse 3D data.

Critical weaknesses
  • The most important weakness is the lack of ablation studies for the semantic voxel representation, which is the central methodological contribution of the paper. CryoASTRA differs from the baselines in several respects, including MSA conditioning, a pretrained structure-prediction backbone, and atomic diffusion. A controlled comparison using the same architecture with (i) EM points only, (ii) raw-density voxels without semantic channels, and (iii) the full semantic voxel representation is needed to determine whether retaining voxel-wise semantics actually accounts for the reported improvement. Although the authors acknowledge this limitation, it remains essential for supporting the paper’s main claim.
Questions for authors
  • Can the authors report end-to-end runtime and peak memory for CryoASTRA and the main baselines on the same targets? The current efficiency evaluation focuses on individual operators and does not establish the computational cost of the complete pipeline.
2 comments

Thank you for the positive evaluation and helpful suggestions. Both requested analyses are now in the revised paper.

1. Semantic Voxel Representation vs. Raw Density & Sequence-Prior Controls (§4.1, §4.4, Table 1; App. H.1.1, H.3)

"Can you clarify or test the contribution of semantic voxel features vs. raw density and sequence priors?"

  • Raw density & active-voxel selection (§3.1, §3.4, §5): Raw density ρv\rho_v is not discarded: the voxel encoder takes dv=[ρv;pv]∈R40\mathbf{d}_v = [\rho_v; \mathbf{p}_v] \in \mathbb{R}^{40}, concatenating normalized raw density with the 39 continuous semantic probabilities on every active voxel v∈Vv \in \mathcal{V} to produce the 128-channel semantic voxel representation ev\mathbf{e}_v. Meanwhile, the sparse U-Net probabilities pv\mathbf{p}_v select V\mathcal{V} (4.3×1044.3\times 10^4 to 3.5×1053.5\times 10^5 voxels) out of 10710^7 to 10810^8 grid voxels (in 13,030 EMDB maps at ≤4\le 4 Å, a median of only 0.54% of voxels lie at or above contour; App. C, Fig. 5). Attending to unthresholded full-grid density is far beyond GPU memory (App. I.2), while restricting ρv\rho_v to V\mathcal{V} still relies on pv\mathbf{p}_v to define V\mathcal{V}. In §5 (Limitations), we explicitly state that retaining ev\mathbf{e}_v is a design property rather than claiming an isolated ablation.
  • Map conditioning vs. sequence priors (Table 1; App. H.1.1, H.3, Tables 9–11):
    1. Map-free AF3 & Protenix-Mini with oracle PDB alignment: Even after US-align rigid fitting onto the deposited PDB across all 103 targets, AlphaFold 3 and Protenix-Mini reach only 68.43% and 63.04% protein Accuracy@3 Å (Q-scores 0.2083 and 0.1794) and 18.07% / 11.64% on nucleic acids, vs. 92.49% (Q-score 0.4691) and 47.39% for unaligned CryoASTRA (MSA).
    2. Single-sequence CryoASTRA (ESM): Replacing protein MSAs with single-sequence ESM2-3B embeddings keeps protein Accuracy@3 Å at 89.07% (91.66% with ≤1\le 1 Å active-voxel filtering; 0.948 TM-score, 0.4775 Q-score) and improves nucleic-acid Accuracy@3 Å to 51.90% (0.608 TM-score).

2. End-to-End Runtime & GPU Memory Profiling (§4.5; App. I.5, Tables 20–21, Fig. 11)

"Provide end-to-end runtime and GPU memory comparisons across maps of varying sizes."

We profiled CryoASTRA (MSA and ESM, 5 samples × 5 Euler steps) against all 7 map-building baselines across 10 benchmark maps (247 to 5,271 tokens; 42,679 to 345,227 active voxels) on a single NVIDIA A100 80 GB GPU (full per-map table in App. I.5, Tables 20–21 & Fig. 11):

  • CryoASTRA (MSA, excluding sequence search) runs in 8.55 s (8,257 MiB peak) on 9N3J (247 tokens) up to 1,725.50 s (75,669 MiB peak) on 9PHF (5,271 tokens); CryoASTRA (ESM) runs in 8.85 s to 1,728.15 s including ESM2-3B embedding generation (0.00–1.04 s).
  • Up to 2,520 tokens (9M4C), CryoASTRA is faster than CryoAtom2 (8.55–202.11 s vs. 74.3–405.2 s), ModelAngelo (155.8–224.6 s), EMProt (180.3–282.4 s), and CryoACE (7,567.6–12,958.8 s). Above 2,520 tokens, sliding-window builders keep peak memory flat (8.3–13.4 GiB) via local crops while CryoASTRA scales with EMTrunk's token-pair layers (O(R2)O(R^2)–O(R3)O(R^3)), yet CryoASTRA completes all 10 maps on one 80 GB GPU while CryoACE OOMs on all 5 targets at ≥2,520\ge 2,520 tokens.

I thank the authors for the added experiments and the response. I will keep my rating.

Reviewer_o5YfOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces CryoASTRA, a map-conditioned atomic model builder for proteins, nucleic acids, and supplied ligands. Sparse U-Nets classify voxels on a 1 Å grid; EM points provide a coarse representation for aligning sequence and density, while semantic voxel features are retained during atomic diffusion. GPU Overlapping BallTrees support local self-attention and atom-to-voxel Heterogeneous BallTree Attention. An AlphaFold-style trunk is initialized from Protenix-Mini, while the atomic diffusion module is trained from scratch. On 103 sequence-novel 2026 maps, the model reports a protein TM-score of 0.972 and 92.49% protein atom accuracy within 3 Å in the map frame. Nucleic-acid and ligand accuracy is appreciably lower. Operator benchmarks and an atom-error classifier complement the modeling experiments.

Critical strengths
  1. The evaluation is unusually attentive to coverage, absolute placement, and denominator integrity. The prospective 103-map set applies sequence-similarity exclusions and pre-window MSA databases (Section 4.1; Appendix I.1). Samples are averaged without ground-truth selection; map-frame metrics count unmodeled reference atoms as misses and are accompanied by coverage/completeness. The paper also reports a common 50-target subset for CryoACE rather than silently dropping its failures. The protein TM-score gain over CryoAtom2 is supported by paired uncertainty estimates, while the non-significant Accuracy@3 Å gain is explicitly acknowledged.

  2. The system addresses a real computational limitation of map-conditioned atomic modeling. Retaining a fine voxel representation while routing atom queries through local sparse attention is a plausible way to preserve map detail without dense atom–voxel attention. The GPU partitioning, sparse pipeline, and detailed pseudocode constitute substantial engineering work. Completing whole assemblies and retaining input ligand components broaden the potential utility beyond protein-only tracing.

Critical weaknesses
  1. The contributions of EM points, semantic voxel features, and token–point pair representations are not isolated experimentally.

The final system changes the pretrained trunk, point alignment, atomic generator, voxel features, and local attention together. There are no controlled end-to-end ablations that remove or replace the EM-point pathway, replace semantic voxel inputs with density-only features, remove the fine voxel pathway, or replace the token–point pair stream with simpler point-to-token conditioning. Section 4.3 and Table 2 assess EM-point localization before structure generation; this is a useful component-quality probe, but does not establish how much those points contribute to the final atomic models. Likewise, operator speed benchmarks do not isolate a representation's contribution to modeling accuracy. Section 5 explicitly acknowledges the need for semantic voxel ablations. Consequently, the benchmark establishes the performance of the integrated system but cannot identify which input representation or interaction supplies the decisive structural information, or whether the semantic voxel and token–point pathways provide complementary gains. This matters especially because the protein Accuracy@3 Å improvement over CryoAtom2 is only 2.2 points and is not significant (p = 0.43), even though the TM-score/completeness results are much stronger.

2 The missing structure-prediction control also leaves the source of the gains unresolved. The evaluation lacks a direct AlphaFold3 comparison and, more importantly for attribution, a sequence-only control using the same Protenix-Mini prior that initializes EMTrunk. These controls should distinguish intrinsic structural quality from map-frame placement: report aligned structural metrics for sequence-only predictions, and raw map-frame metrics after docking/fitting using only the input map, without reference-structure alignment or reference-based candidate selection. Otherwise, the current evidence cannot establish how much the proposed representations improve on a strong structure prior coupled to conventional map fitting.

  1. The evidence does not yet establish a sufficiently distinct methodological advance or its practical significance. The paper's own related-work discussion identifies sparse convolution, MeanShift++, AlphaFold-style trunks, and rotated BallTree partitions as existing ingredients (Section 2). The proposed retained semantic pathway and heterogeneous atom–voxel attention may constitute a useful advance, but their necessity and advantages over simpler local cross-attention are not established by matched accuracy/cost comparisons. An implementation-level GPU acceleration is valuable, but isolated operator timings do not demonstrate a new representation-learning principle or an improved end-to-end accuracy–cost tradeoff. In particular, the remaining quadratic-to-cubic token-pair bottleneck limits the extent to which linear local attention establishes whole-model scalability.

  2. The application results also qualify the significance claim. The protein Accuracy@3 Å gain over CryoAtom2 is not significant, and part of the gain reflects complete sequence-conditioned generation rather than improved placement of the same atoms (Section 4.2). On the common 50-target subset, CryoAtom2 and EMProt have lower matched RMSDs over the residues they model; this does not negate CryoASTRA's completeness advantage, but calls for a common-atom comparison to isolate local accuracy. For nucleic acids and ligands, only about 47% of atoms are within 3 Å, with map-frame RMSDs of approximately 18 Å and 15 Å, respectively (Table 1). A supplied ligand list and high modeled-instance completeness do not establish correct chemical placement, and no dedicated ligand-fitting comparison or comprehensive stereochemical validation accompanies these outputs. Thus, the general all-atom scope is demonstrated at the level of producing components, but reliable atomic modeling across molecule types is not yet established. Together, these gaps leave it unclear whether the work delivers a generalizable methodological advance beyond a well-engineered integrated builder.

Questions for authors
  1. Causal ablation: Can you compare the full system with matched retrained variants that (a) remove or replace EM-point conditioning, (b) use density-only rather than density-plus-semantic voxel inputs, (c) remove atom-to-voxel conditioning, and (d) replace the token–point pair stream with simpler point-to-token conditioning? Please keep trunk initialization, training data/budget, and sampling budget comparable, and report final map-frame RMSD, Accuracy@3 Å, and coverage. For variants that retain EM points, please hold the predicted active voxels and point extraction fixed when testing voxel semantics, so that the effect of semantic features is not confounded with a change in spatial support. A compact set of controlled comparisons would clarify which pathways supply useful information beyond the pretrained sequence/structure prior and whether their contributions are complementary.

  2. Atomic validity: What are clash, bond/angle, chirality, and torsion statistics before and after any refinement, particularly for nucleic acids and ligands? Can you compare ligand placement with an applicable dedicated fitter using the same supplied chemical identities and report successes at the ligand-instance level?

  3. Practical cost: What are end-to-end wall time and peak GPU memory versus residue count and active voxel count, including all preprocessing, five diffusion samples, and the error head? Please report the error head's five out-of-memory failures separately from the successful base-model inference and show the common completed subset for timing comparisons.

  4. AlphaFold3 and prior-only controls: Please include AlphaFold3 predictions, as well as a sequence-only control from the same Protenix-Mini prior used by CryoASTRA. Compare aligned structural accuracy before fitting and raw map-frame RMSD/Accuracy@3 Å after a clearly specified map-only docking/fitting procedure. Use the same input sequences, molecular components, and comparable MSA/template information where possible; document weight/data cutoffs and candidate budgets. Neither docking nor sample selection should use the reference structure. Do the proposed map representations still improve substantially over these controls, especially on targets whose experimental conformation differs from the prior prediction?

  5. Distinct contribution and useful accuracy: Can you compare Heterogeneous BallTree Attention with a simpler GPU spatial-neighborhood cross-attention under comparable accuracy and resource budgets, and report common-atom protein accuracy alongside full-reference coverage? These comparisons would clarify whether the contribution is a better learning mechanism, an implementation improvement, or primarily more complete generation, and which benefit justifies the additional architectural complexity.

3 comments

Thank you for the thorough review. Due to OpenReview's 5,000-character limit per comment, we split our response into Part 1 (Points 1–2) and Part 2 (Points 3–4).

1. Map Conditioning vs. Sequence Priors & Component Design (W1, Q1, Q4)

"How much performance comes from the pretrained trunk and MSA vs. the map? Can you ablate EM points or semantic probabilities vs. raw density?"

  • Map-free AF3 & Protenix-Mini controls (§4.1, §4.4, Table 1; App. H.1.1, Tables 9 & 11): We ran AlphaFold 3 (v3.0.1) and Protenix-Mini (protenix_mini_default_v0.5.0, the checkpoint initializing EMTrunk) on all 103 targets with identical inputs and rigidly aligned each prediction onto the deposited PDB via US-align.
  • Single-sequence ESM variant (§4.4, Table 1; App. H.3, Table 10): Replacing protein MSAs with single-sequence ESM2-3B embeddings (CryoASTRA ESM) tests accuracy without sequence search:
Component & MethodTargetsTM-score ↑Acc@3 Å (%) ↑RMSD (Å) ↓Q-score ↑
Protein: AF3† / Protenix-Mini† (PDB-aligned)101 / 950.875 / 0.82668.43 / 63.048.41 / 12.250.2083 / 0.1794
Protein: CryoASTRA (MSA / ESM, unaligned)1010.972 / 0.94892.49 / 89.073.47 / 5.560.4691 / 0.4775
Nucleic Acid: AF3† / Protenix-Mini†14 / 130.422 / 0.28118.07 / 11.6431.05 / 42.740.0576 / 0.0105
Nucleic Acid: CryoASTRA (MSA / ESM)140.561 / 0.60847.39 / 51.9018.02 / 15.760.2402 / 0.2814
Ligand: AF3† / Protenix-Mini† / Ours (MSA)66 / 63 / 66—40.72 / 31.53 / 47.3515.14 / 18.64 / 15.080.2026 / 0.1349 / 0.3312
  • Role of EM points & semantic probabilities (§3.4–3.6, §4.3, Table 2, §5): Sequence tokens have no 3D coordinates, so EMTrunk requires EM points yp∈R3\mathbf{y}_p \in \mathbb{R}^3 (P≈RP \approx R) to anchor tokens to the map frame. Our EM-point localization probe (§4.3, Table 2) evaluates extracted EM points before voxel encoding or diffusion: they match deposited disto atoms at 0.61 Å RMSD (94.62% recall) on proteins and 0.64 Å RMSD (90.78% recall) on nucleic acids. Moreover, raw density ρv\rho_v is already in the voxel encoder input dv=[ρv;pv]∈R40\mathbf{d}_v = [\rho_v; \mathbf{p}_v] \in \mathbb{R}^{40}, while pv\mathbf{p}_v selects the active voxel set V\mathcal{V} (4.3×1044.3\times 10^4 to 3.5×1053.5\times 10^5 out of 10710^7 to 10810^8 grid voxels; median contour occupancy is 0.54%, App. C Fig. 5). Full-grid attention exceeds GPU memory (App. I.2), and restricting ρv\rho_v to V\mathcal{V} still relies on pv\mathbf{p}_v to define V\mathcal{V} (discussed as a design property in §5).

2. Completeness vs. Precision, Q-score, Stereochemistry & Ligands (Q2)

"How does CryoASTRA compare on commonly modeled atoms, Q-score, stereochemistry, real-space refinement, and ligands?"

  • Common-atom accuracy (§4.2; App. H.5, Table 12): On the intersection of reference protein heavy atoms modeled by both methods across all 101 targets (98.26% CryoAtom2 coverage), CryoASTRA (MSA) matches CryoAtom2 in Accuracy@3 Å (93.16% vs. 91.84%, 95% CI [−1.72, 4.64]) with significantly lower RMSD (3.20 Å vs. 6.64 Å, CI [−6.63, −0.72]), while modeling all remaining residues (TM-score 0.972 vs. 0.855). On the 50 CryoACE-completed targets (Table 12), common-atom Accuracy@3 Å is 95.00% vs. 95.63% against CryoAtom2, 96.21% vs. 94.59% against EMProt, and 94.05% vs. 85.54% against CryoACE (Δ=+8.52\Delta = +8.52 pp).
  • Q-score & active-voxel distance filtering (§4.2, §4.4, Table 1; App. H.3, Table 10; App. H.6, Table 13): Unfiltered full-protein Q-score is 0.4691 vs. 0.5230 for CryoAtom2 (92.48% completeness), narrowing to 0.5089 vs. 0.5350 on common atoms (Table 13). Filtering to atoms within ≤1\le 1 Å of an active voxel center (91.53% completeness; Tables 1 & 10) raises conditional protein Accuracy@3 Å to 94.92% (MSA) / 91.66% (ESM), drops RMSD from 3.47 → 2.73 Å, and lifts Q-score to 0.5198 (MSA) / 0.5318 (ESM) (and 53.42% Accuracy@3 Å, 0.4121 Q-score on ligands).
  • Stereochemistry & Phenix refinement (§4.2; App. H.6, Tables 13–14): Unrelaxed diffusion samples have local clashes (mean clashscore 211.68, Ramachandran outliers 5.47%, rotamer outliers 19.71% across 101 targets; Table 13). One pass of phenix.real_space_refine across 9 targets (45 CryoASTRA samples and 9 CryoAtom2 models; Table 14) cuts CryoASTRA's clashscore by >11× (237.0 → 21.5) and raises Q-score (0.424 → 0.456) while preserving Accuracy@3 Å (87.3% → 87.2%).

3. Heterogeneous BallTree Attention vs. kkNN & Global Attention (W3, Q5)

"How does HBA compare against standard GPU kkNN and global cross-attention?"

In §4.5 (Table 3) and App. I.2 (Tables 16–17), we benchmarked two atom–voxel cross-attention blocks on real coordinates from 10 maps (4,942–35,021 atoms; 42,686–345,181 active voxels) with identical 32-basis Gaussian RBF distance biases, 3D RoPE, and BF16 SDPA:

MethodGraph + Layout (ms) ↓Forward (ms) ↓F+B Time (ms) ↓F+B Peak (Δ MiB) ↓Completed
HBA (ℓav=64\ell_{av}=64)2.39–3.132.20–12.626.57–45.74722–5,23410 / 10
HBA (ℓav=256\ell_{av}=256, default)2.09–2.852.40–12.496.20–37.57692–5,02810 / 10
PyG kkNN (k=64k=64)296–1,7823.49–24.858.99–60.671,235–8,77910 / 10
Torch kkNN (k=256k=256)5.30–490.2513.42–96.8431.85–269.274,653–26,9659 / 10 (OOM 9PHF)
Global (16-query chunked)—452–32,521OOM on all 10OOM on all 100 / 10

Whereas kkNN gathers and duplicates kk voxel keys/values and RBF tensors per atom query, HBA shares one set of voxel keys/values per mixed leaf (ℓav=256\ell_{av}=256, averaging 225–233 valid voxel neighbors per atom across the two partitions). Consequently, compared to exact Torch GPU kkNN (k=256k=256), HBA builds graphs 2.5–172× faster, runs forward+backward 5.14–7.36× faster, uses 5.22–6.72× less peak memory, and finishes 9PHF where Torch kkNN OOMs. Unchunked global RBF tensors require 25.15–1,375.62 GiB per block, and 16-query chunked global attention OOMs on all 10 maps during autograd.

4. End-to-End Runtime, Memory Scaling & Error-Head OOMs (Q3)

"What are the end-to-end runtime and peak memory costs vs. baselines, and why did 5 targets OOM in confidence evaluation?"

  • End-to-end profiling (§4.5; App. I.5, Tables 20–21, Fig. 11): We profiled end-to-end inference (5 samples, 5 Euler steps) on a single NVIDIA A100 80 GB GPU across the same 10 maps against all 7 map-building baselines (full table in App. Table 21):
PDBTokensActive VoxelsCryoASTRA (MSA) s / MiBCryoASTRA (ESM)* s / MiBCryoAtom2 s / MiBModelAngelo s / MiBEMProt s / MiBCryoACE s / MiB
9N3J24742,6798.55 / 8,2578.85 / 2,48174.3 / 10,487N/AN/AN/A
10UJ84565,68518.83 / 5,81719.29 / 11,671174.0 / 13,373155.8 / 9,075198.5 / 8,2597,567.6 / 29,565
9M4C2,520157,151202.11 / 19,085198.24 / 16,761405.2 / 13,373215.8 / 9,075225.7 / 8,261OOM / 77,663
9WAE3,948317,855752.07 / 46,409742.78 / 39,503940.1 / 13,373708.7 / 9,075489.3 / 8,261OOM / 80,565
9PHF5,271329,5321,725.50 / 75,6691,728.15 / 75,0411,008.0 / 13,373496.0 / 9,075526.4 / 8,261OOM / 75,827

*ESM includes ESM2-3B embedding generation (0.00–1.04 s); MSA excludes sequence search. All 10 maps are in App. Table 21 & Fig. 11.

Up to 2,520 tokens, CryoASTRA is faster than CryoAtom2, ModelAngelo, EMProt, and CryoACE. On larger assemblies, sliding-window builders keep memory flat (8.3–13.4 GiB) via local crops while CryoASTRA scales with EMTrunk's token-pair layers (O(R2)O(R^2)–O(R3)O(R^3)), yet CryoASTRA finishes all 10 maps on one 80 GB GPU while CryoACE OOMs on all 5 targets at ≥2,520\ge 2,520 tokens.

  • Atom error head OOMs (§4.6; App. J, Table 22): All 103 base structure predictions succeeded (Table 1). During confidence scoring on the 5 largest assemblies, holding EMTrunk's pair tensor zii′\mathbf{z}_{ii'} alongside both the 8-block atom error head and 4-block AlphaFold-style confidence head exceeded 80 GB on one GPU.
Reviewer_o5YfOpenReview ↗

Steric Violations and Unresolved Attribution

Thanks for the detailed response and the extra experiments. The common-atom evaluation, Q-score analysis, refinement results, and end-to-end runtime profiling do make the empirical evaluation stronger.

That said, a few central concerns are still unresolved. The response doesn't give true end-to-end ablations of EM-point conditioning, semantic voxel features, atom-to-voxel conditioning, or token–point pairs, so it's still unclear how much each representation actually contributes. The AlphaFold3 and Protenix-Mini comparisons are reported after rigid alignment to the deposited PDB, not after map-only docking with raw map-frame metrics. As a result, they don't fully establish an advantage over a strong sequence prior combined with conventional map fitting.

More seriously, the unrefined generated structures have severe geometric violations: the reported mean clashscore is 211.68, with substantial Ramachandran and rotamer outliers. A structure with this level of steric and geometric error can't be regarded as an acceptable atomic model without reliable refinement. The refinement result is shown only for a small subset of targets and doesn't establish that the full set of generated structures is geometrically valid.

Overall, the response strengthens the empirical analysis, but the main concerns about causal attribution, methodological novelty, and the validity of the generated atomic structures remain. My recommendation remains unchanged.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Thanks to the Area Chair and Reviewers o5Yf, huyf, and kk76 for the helpful reviews. We are glad that the reviewers found the whole-map sparse setup, the custom BallTree kernels, the 2026 prospective test set, and the multi-component modeling promising.

In response to the reviews, we added five sets of experiments and updated the 10-page main text and appendix:

  1. Map-free sequence priors and single-sequence ESM ablation (§4.1, §4.4, Table 1; Appendix H.1.1, Tables 9–11): We ran AlphaFold 3 (v3.0.1) and Protenix-Mini (v0.5.0) on all 103 targets with identical inputs and rigidly aligned them onto the deposited PDB structures via US-align. Even with reference alignment, AF3 and Protenix-Mini reach only 68.43% and 63.04% protein Accuracy@3 Å (vs. 92.49% for unaligned CryoASTRA MSA). Replacing protein MSAs with single-sequence ESM2-3B embeddings keeps protein Accuracy@3 Å at 89.07% (0.948 TM-score) and improves nucleic-acid Accuracy@3 Å to 51.90% (0.608 TM-score).
  2. Heterogeneous BallTree Attention (HBA) vs. GPU kNN and global attention (§4.5, Table 3; Appendix I.2, Tables 16–17): Across 10 real maps with identical 32-basis RBF distance biases, HBA (ℓav=256\ell_{av}=256) builds graphs 2.5–172× faster than exact Torch GPU kNN (k=256k=256), runs forward+backward 5.14–7.36× faster with 5.22–6.72× less peak memory, and finishes all 10 maps (Torch kNN OOMs on 9PHF; 16-query chunked global cross-attention OOMs on all 10 maps during autograd).
  3. Common-atom accuracy vs. completeness (§4.2; Appendix H.5, Table 12): On the intersection of reference protein heavy atoms modeled by both CryoASTRA (MSA) and CryoAtom2 across all 101 targets, CryoASTRA reaches 93.16% vs. 91.84% Accuracy@3 Å and lower common-atom RMSD (3.20 Å vs. 6.64 Å), while modeling 100.00% vs. 92.48% of polymer residues (TM-score 0.972 vs. 0.855).
  4. Q-score, active-voxel filtering, MolProbity, and Phenix refinement (§4.2, §4.4, Table 1; Appendix H.3, Table 10; Appendix H.6, Tables 13–14): We added Q-scores across all tables, an optional ≤1 Å active-voxel distance filter (raising protein Q-score to 0.5198 for MSA / 0.5318 for ESM and conditional Accuracy@3 Å to 94.92% / 91.66%), full-assembly MolProbity checks on 101 targets, and paired phenix.real_space_refine runs on 9 targets (dropping CryoASTRA's clashscore from 237.0 to 21.5 and raising Q-score from 0.424 to 0.456 with Accuracy@3 Å at 87.3% → 87.2%).
  5. End-to-end runtime and memory profiling (§4.5; Appendix I.5, Tables 20–21, Fig. 11): We profiled CryoASTRA (MSA and ESM) against all 7 map-building baselines across 10 maps (247–5,271 tokens) on one A100 80 GB GPU.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.