acceptodds
Under review as a conference paper at ICLR 2027

Seeing the Molecule Behind the Map

12 views

Abstract

Biomolecular structure predictors can generate plausible conformations, but these may not correspond to the state observed in a particular sample. Cryo-electron microscopy (cryo-EM) density provides spatial constraints from the target experiment; using them to guide both global structural relationships and local atomic placement remains a central challenge. We introduce EviBoltz, which uses global density interactions to update the single and pair representations of the Boltz-2 Pairformer, and atom-centered, multiscale density queries continually guide diffusion-based coordinate generation. Among evaluable cases in an independent benchmark, EviBoltz raises acceptable protein–protein interface success from 27.8% to 70.4% and joint ligand-pocket success from 7.0% to 34.9% relative to Boltz-2. The protein–peptide and antibody–antigen tasks show the same direction of change, with success rates rising from 0.0% to 75.0% and from 0.0% to 66.7%, respectively. In addition, long-loop analysis shows improved backbone agreement with the input density. In a controlled, supervised experiment on two states of the same URAT1 sequence, representations induced by different target maps guide a frozen generator toward the corresponding states. These results connect improved interaction recovery and local geometry with the ability of map-conditioned representations to carry information about the observed state.

Primary area:
generative models
Keywords:
biomolecular structure prediction, cryo-EM, structure generation, protein structure, experimental structure determination, generative modeling

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

The paper presents EviBoltz, a cryo-EM density-conditioned extension of Boltz-2 for experimental atomic model building. Global density features are incorporated into the single and pair representations inside the Pairformer, while local density features are repeatedly queried around current atomic coordinates during diffusion. The model is jointly fine-tuned on paired experimental maps and structures.

The evaluation covers protein–protein, protein–ligand, protein–peptide, and antibody–antigen systems. The paper reports improvements in interface recovery, ligand placement, and local loop geometry, together with a supervised two-state URAT1 experiment examining whether map-conditioned representations can convey conformational-state information to a frozen generator. The relevant application is the interpretation of experimental density maps into atomic models, with the target map providing substantial sample-specific structural information.

Critical strengths

Combining molecular priors with cryo-EM observations is useful for recovering the structure present in a particular experimental sample. Incorporating density into both relational representations and coordinate generation is a reasonable design, and the proposed layer-wise Pairformer conditioning is a potentially useful extension of existing density-conditioned approaches.

Acceptable PPI interface recovery increases from 15/54 to 38/54 cases, and joint protein–ligand success increases from 3/43 to 15/43. The paired source-level analysis and the independently aligned loop-shape evaluation provide additional evidence that the gains extend beyond a few qualitative examples or global rigid-body placement.

Critical weaknesses

The target density is a legitimate input for experimental model building, but it supplies substantial structural information unavailable to density-free Boltz-2. Improvements over Boltz-2 therefore do not, by themselves, isolate the contribution of the proposed architecture. The paper does include density-aware baselines, which is appreciated. However, the adapted implementations, incomplete original runs, and backbone differences make these comparisons difficult to interpret. CryoACE already includes density–sequence fusion, atom-centered density profiling, and eventual end-to-end fine-tuning. The most informative control would therefore compare layer-wise Pairformer conditioning with a simpler density-fusion alternative using the same backbone, training data, and optimization budget. It should also be clarified whether the pathway ablations were independently trained or disabled only at inference.

The reported reference-based accuracy gains are encouraging, but complete output coverage does not establish that all generated regions are supported by the experimental data. In weak-density regions, a generative model may produce plausible coordinates without resolving the underlying ambiguity. A compact evaluation of local map agreement, protein/ligand geometric validity, and accuracy stratified by local density quality would help establish the practical value of the additional coverage. The scope of the task-level claims also needs clearer qualification: the 75% peptide success appears to represent only 3/4 evaluable interfaces from a cohort of 13 cases. Explicit eligibility criteria, counts, and uncertainty are needed to assess how broadly the reported improvements apply.

Questions for authors
  1. Were the pathway ablations independently trained under matched settings, or obtained by disabling components only at inference? Is there a matched-backbone and matched-data comparison between layer-wise Pairformer conditioning and post-Pairformer density fusion?

  2. What changes were made to the reproduced CryoBoltz and CryoACE pipelines? Please specify the checkpoints, initialization/alignment procedures, sampling budgets, and post-processing, and explain how reproduction fidelity was checked. Were reference coordinates used in any test-time preprocessing or prediction-selection step? A paired comparison on a common completed subset would be helpful.

  3. What evidence shows that the increased completeness corresponds to accurate, experimentally supported atoms, particularly in weak-density regions? Please report or point to available local map-agreement and geometric-validity analyses. A large prospective deployment study would be useful future work, but is not necessary to address this question.

  4. Please confirm the effective sample counts and explain why only four of the thirteen peptide cases appear to enter interface evaluation. Also, what explains the Boltz-2 ligand RMSD difference between Tables 1 and 2, namely 10.6177 Å versus 11.721 Å?

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

EviBoltz is a method using a structure prediction prior for fitting atomic structures into cryo-EM density maps. It builds global and local features from a density map, using them to modulate the sequence trunk and structure diffusion head, respectively, of Boltz-2. EviBoltz builds atomic structures consistent with the target conformations specified by density maps, outperforming de novo building methods and structure guidance approaches on a benchmark of protein interactions.

Critical strengths

Leveraging structure predictors as priors for cryo-EM model building is a highly active area. EviBoltz's modulation both the trunk and diffusion head is an interesting approach, and the ablations clearly demonstrate the value of both the global conformational control in the trunk and local structural control in the diffusion.

The results are compelling. The authors curate a sizable (120 systems) benchmark of varied protein interactions, which can be especially challenging as docking interfaces can be predicted with high uncertainty. An extensive set of existing methods are baselined, both de novo building (ModelAngelo, E3-CryoFold, CryoAtom) and prediction guidance methods (CryoBoltz, CryoACE). It is appreciated that the authors additionally reproduced CryoBoltz and CryoACE themselves, improving upon the official releases in order to process more of the benchmark. Significant improvements in success rates are shown across all benchmark categories, e.g. going from 0% success for Boltz-2 to 66.7% success for EviBoltz. The local geometry of loops and ligands appears robust despite being challenging to model in cryo-EM maps.

Critical weaknesses

None, but please address questions below.

Questions for authors
  • The curated benchmark of protein interactions is interesting and challenging, but did you also benchmark on an existing model building test set like the one used in ModelAngelo?
  • The details of the method are unclear - please properly define all variables used in equations 1 and 2.
Reviewer_nuWROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors of the paper propose a new framework called EviBoltz, a co-folding model that is conditioned on cryo-EM electrostatic potential (ESP) maps. To this end, they train a density encoder to combine density embeddings with Pairformer embeddings and fine-tune Boltz so that the target map conditions both the trunk's evolutionary representations and the diffusion generator. The model is fine-tuned jointly on a dataset of 10k+ paired structure-map sources, and evaluated on a held-out benchmark of 120 cryo-EM cases.

Critical strengths
  1. The authors propose an alternative to coordinate guidance and a novel approach for generating experimentally faithful protein structures. Moreover, the two conditioning pathways are also separately ablated (Table 2).

  2. The authors explicitly state the metrics used to evaluate their framework against native structures, and they compare its performance with relevant benchmarks.

Critical weaknesses
  1. Lack of map-to-structure metrics and geometric validation

All metrics in the paper are dependent on the native structure. However, standard map-to-structure metrics like resolution at FSC=0.143, real-space cross correlation, Q-score are missing from the table. Furthermore, the authors do not report structure-quality metrics like MolProbity, clashscore, or Ramachandran and rotamer outliers. It is important to show that the fine-tuned model does not result in unphysical structures.

  1. Concerns regarding Train / Test split

While the authors explicitly mention that the train/test split have no intersecting identifiers, they do not address the sequence similarity between the train / test split, which can play a key role in terms of structure prediction and can be an essential source of data leakage. Both sets are drawn from the same post-2023 PDB/EMDB pool, the training set contains 6,606 distinct whole-protein sequences, and no identity- or structure-based filter is applied. Multi-state and multi-construct depositions of one system are routine in cryo-EM.

  1. Numerous missing details

At each step of the diffusion process, Boltz-2 randomly rotates and translates protein structure. The generator is therefore trained to produce structures in an arbitrary global frame. However, the authors do not specify whether random augmentation is disabled in EviBoltz, whether an explicit alignment step is performed, or whether coarse placement is expected to emerge from the positional encoding in the map tokens and the global context. Also, map pre-processing details like voxel resampling, box cropping, centering, intensity normalization, or masking are either vaguely specified or not specified at all. In addition, it is not clearly specified how micelle or lipid density is masked out.

  1. The stencil geometry is not ablated

The stencil geometry is essential to controlling the high-frequency details of the sampled structure. There are numerous concerns about the hyperparameter choice. i. at typical EMDB voxel sizes, offsets much below 1 Å would often be sub-voxel; trilinear interpolation would return near-identical values and the stencil would degenerate to a 1-point read with redundant channels ii. the radii are fixed spanning 1.90–3.70 Å nominal resolution, although the supportable sensing scale differs substantially across that range.

  1. Chain assignment

Chain correspondence is an essential component for DockQ. Especially for antibodies, heavy/light chain assignment is a standard failure mode and no chain-permutation handling for homo-oligomers is mentioned. Also, the "Interface macro DockQ" column in Table 11 is never defined.

  1. Behaviour in unresolved regions is unmeasured

The metrics and losses are only computed for atoms with valid reference coordinates. However, the model is not evaluated where depositors could not resolve a structure. A self-consistency run that feeds a training map back in and compares the output to the deposited structure, particularly in unresolved regions, would show generalizability of the model.

  1. Heterogeneity

Given a single consensus map, EviBoltz produces structures near a single state: it cannot recover an ensemble from a heterogeneous map. This should be stated as an explicit limitation of the paper.

  1. Miscellaneous

Instead of reporting averaged numbers over the 120 test cases, the authors should instead show numbers over cases where the prior predicts structure correctly and cases where prior predicts the structure poorly. This way, it is easier to understand the contribution of the experimental data and also shows that the training does not come at the cost of ruining the prior. Additionally, the authors do not discuss related works like IT-Opt [1] and ROCKET [2] that similarly optimize / manipulate intermediate representations of AlphaFold2/3. Neither appears even in the extended survey of Appendix A, despite being the closest published work on acting through a folding model's internal representations rather than its coordinates.

References:

  1. Maddipatla et al., "Inference-time optimization for experiment-grounded protein ensemble generation", ICML 2026.
  2. Fadini et al., "AlphaFold as a prior: experimental structure determination conditioned on a pretrained neural network", Nature Methods 2026.
Questions for authors
  1. Can the authors specify key implementation details. These include: i. How is the predicted structure aligned to the map during sampling, and is the random rigid-body augmentation of the Boltz-2 diffusion module disabled during fine-tuning? ii. Is the map pre-centered or cropped, and if so, using what — the box, the density's center of mass, or the deposited structure's extent? Is any masking applied to remove detergent or unresolved density?

  2. What fraction of the 120 test targets have a training-set sequence above 90% / 50% / 30% identity?

  3. What are the map to structure FSC (at 0.143), Q-score, and MolProbity/clashscore statistics for EviBoltz outputs versus Boltz-2 and the deposited structures?

  4. Can the authors provide summary statistics over cases where Boltz-2 largely predicts the structure well and over cases where Boltz-2 mispredicts the structure?

  5. What happens with radii scaled to each map's nominal resolution?

  6. How is chain correspondence established for DockQ, and how are homo-oligomer permutations resolved? What is "interface macro DockQ"?

  7. If a training-set map is fed back in, does the output improve on the deposited structure where density is strong, and what is produced in unresolved regions?

  8. Since nothing is frozen during fine-tuning, how does the checkpoint perform without a map on a standard sequence-only benchmark, on specific X-ray targets, or on Boltz-2's binding-affinity prediction? I am worried about catastrophic forgetting.

1 comment
AuthorsOpenReview ↗

A Brief Follow-Up on Our Rebuttal

Thank you again for your thoughtful review. As the discussion deadline approaches, we wanted to briefly follow up on our rebuttal. If you have a chance to review our response, we would appreciate hearing whether the additional experiments and clarifications address your main concerns. We would be happy to clarify any remaining points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.