acceptodds
Under review as a conference paper at ICLR 2027

Guiding protein ensemble generators with cryo-EM particles, without maps or poses

Abstract

Understanding a protein's function requires knowing its conformational distribution: what forms it takes and how often. Cryo-electron microscopy provides this distribution in principle, since every particle images a single molecule, but the averaging needed to address the experiment's low signal-to-noise ratio throws away the very heterogeneity we seek. Pretrained generative models of protein ensembles could restore part of what is lost if individual particles steered their sampling. However, this steering would require a likelihood of the observed images given a generated structure, and constructing this term requires knowledge of the particle's pose. We build this likelihood in a low-dimensional code shared by a frozen pretrained generative model for protein conformations and a frozen cryo-EM image encoder, and fit the code with simulated structure–image pairs. Because the code holds only information available in both models' representations, it suppresses imaging parameters such as defocus. The scatter that imaging, pose included, still leaves between images of the same conformation sets the likelihood's covariance. Using this likelihood, we guide generation with only 128 particles and no density map or pose estimates toward rare states and PDB structures. This is possible because the prior fills in high-resolution details, while the particles constrain large-scale rearrangements. The same code also recovers free-energy landscapes by deconvolving imaging scatter, resolving barriers that particle counting cannot. We test the approach on simulated particles of a GroEL subunit, lysenin and DppA.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cryo-EM, protein conformational ensembles, diffusion models, training-free guidance, inverse problems, heterogeneous reconstruction, free energy landscapes, structural biology

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

This paper connects a frozen protein ensemble generator, BioEmu, and a frozen cryo-EM image encoder, Cryo-IEF, through a learned low-dimensional shared representation. It constructs an uncertainty-weighted likelihood for guiding conformational sampling and uses simulated candidate-image distributions to estimate conformational populations. The approach is evaluated using simulated particles generated from BioEmu conformations, PDB structures, and an independent molecular-dynamics ensemble. The experiments demonstrate shifts toward target conformations and recovery of imposed population landscapes under the tested simulation conditions.

Critical strengths
  1. The shared representation provides an interesting interface between pretrained structural and image models. Using the same code for both conditional generation and population estimation gives the approach a coherent methodological foundation and allows the pretrained models to remain frozen.

  2. The statistical modeling is carefully developed. The guidance accounts separately for image variability and time-dependent uncertainty in the generator representation, while population estimation explicitly models imaging-induced scatter. These design choices address substantive inference issues and constitute the main technical strength of the work.

Critical weaknesses
  1. The guidance task has an unresolved dependency on the conformational identification it is intended to support. Section 3.3 requires particles from a single conformation, which the synthetic experiments obtain directly from known structure identities. The discussion proposes using 3D classification or heterogeneous reconstruction to supply these particles. However, if existing methods cannot reliably identify and isolate a rare state, they cannot provide the required input; if they can, three-dimensional structural information is already available, and the proposed method must demonstrate an additional benefit over reconstruction, refinement, or structural modeling using that information. The paper does not establish this intermediate regime. Furthermore, a conformational class in continuously heterogeneous data is not equivalent to the experimental setup of 128 images generated from exactly one structure. Selecting a narrow conformational neighborhood introduces a trade-off between homogeneity, particle count, and viewing coverage, but the required homogeneity and tolerance to residual variation are not evaluated. These issues leave the practical applicability of the guidance task unresolved, even before considering simulation-to-experiment transfer.

  2. The population estimates lack a demonstrated basis for interpretation as reliable scientific measurements. All particle-image experiments are simulated, including the PDB and MD evaluations. Recovering imposed populations under the chosen simulator does not establish that fitted weights represent physical populations in experimental data. Missing candidate structures, inaccurate imaging distributions, and biases in the learned representations can all potentially be absorbed into those weights. A good fit in the shared feature space therefore does not independently distinguish correct populations from compensation for model mismatch. The paper does not establish how this distinction would be made or how incorrect estimates would be detected. The independent MD experiment illustrates this concern: Appendix I.5 reports that the displayed landscape uses a code-offset correction measured using the ensemble’s known structural coordinate, reducing the error from 0.57 to 0.10 kBTk_BT. The availability of equivalent information in the intended application is unclear. My concern is therefore that the current evaluation does not establish a credible route from successful synthetic recovery to trustworthy inference on experimental data.

Questions for authors
  1. Practical workflow and added value: Beyond the proposed use of upstream 3D classification, what concrete experimental setting allows a rare conformational class to be reliably isolated while leaving a structural problem that the proposed guidance demonstrably improves? What structural information would already be available, and how would the method compare with reconstruction, refinement, or candidate selection using comparable information and computation?

  2. Validity of the input assumption: What degree of within-class conformational variation can the guidance tolerate? Is there evidence that particles selected from a realistic conformational neighborhood, rather than rendered from one fixed structure, provide a sufficiently accurate target code without introducing misleading averaging effects?

  3. Population validation: What independent validation or diagnostic would distinguish genuine population changes from compensation for missing candidate states, imaging mismatch, or representation bias? Please clarify what evidence currently supports this distinction and what remains a proposed future validation strategy.

  4. MD calibration: Appendix I.5 describes a code-offset correction measured using the MD ensemble’s known structural coordinate. How could this correction be estimated when the target conformations are unknown, and how should the displayed MD result be interpreted if that information is unavailable?

7 comments

Thank you for raising two points on which the method's usefulness depends: where its input comes from and whether its population estimates can be trusted.

W1, Q1. What guidance adds to classification.

The paper does not establish this intermediate regime.

The dilemma is well put, and we tested both of its cases.

When classification works. The transporter's first-round classes (8.6 Å resolution) come from the depositors' pose-based processing, so a low-resolution map of each class exists. From about 1,000 particles of one such class, guidance produces atomic models within 0.25–0.43 nm of the deposited inward-facing models (unguided: 0.65–0.71 nm).

When classification cannot isolate the state. We simulated 4,000 particles of the 57 kDa GroEL subunit at the paper's signal-to-noise ratio. They come from two equally populated conformational clusters, 1.1 nm apart along the opening coordinate. Even given the true poses, RELION and cryoDRGN separate the two clusters with only 52–61% purity (chance: 50%). Our population model assigns the same particles with 93% purity and 72% recall, and guidance weighted by its assignments is within 0.03 nm of guidance on perfectly sorted particles. It can do so because it compares the particles with the generator's candidate structures, which RELION and cryoDRGN do not have.

Comparison with reconstruction. In simulation with the true poses, a map from 32 particles identifies the closest candidate structure more accurately than guidance from 128. We therefore see two uses for guidance. One is the first case above, where guidance provides an atomic model for each coarse class without high-resolution refinement. The other is for particles whose poses cannot be determined, as in the subunit above, where RELION's refinement did not recover the poses (median error 88–124°; random: 132°).

W1, Q2. Variation within a class.

What degree of within-class conformational variation can the guidance tolerate?

In simulation, the classes do not have to be narrow. A class of 128 different conformations that follow the population density gives a median core RMSD within 0.03 nm of 128 images of its central member. Guidance improves on unguided sampling in all twelve classes, with members within a median of 0.36 nm of the central member, compared with 0.41 nm for the widest class we tested.

The target code is the mean over the class. A class that mixes two distinct states (1 nm apart, 90 and 38 particles) therefore gives samples pulled toward the majority state, as a consensus map would be, and such data should be split first.

Four to eight particles of a narrow class already reduce the distance to the extended GroEL target from 1.82 to 0.56–0.70 nm. Viewing coverage has a larger effect, because top views alone give 1.53 nm and side views alone 0.69 nm (all views: 0.53 nm).

W2, Q3. Populations, or compensation for mismatch?

What independent validation or diagnostic would distinguish genuine population changes from compensation …

On experimental particles, the estimates follow changes in composition. As the reviewer says, a good fit alone cannot show this.

Current evidence. We therefore mixed experimental particles of two deposited classes at known ratios from 0 to 100% and estimated the share without the labels. All three kinds of mismatch the reviewer lists are present but identical in every mixture; only the composition changes. On both proteins, the estimated share rises at every step, with a slope of about 0.8 on the transporter and about 0.3 on the phytochrome (1 would be ideal). Mismatch thus compresses the scale and adds an offset, but the direction of change is recovered. We therefore claim only changes in population between datasets processed in the same way.

Future validation. We still need datasets whose state fraction is set biochemically and measured independently, such as a ligand titration.

W2, Q4. The MD calibration.

How could this correction be estimated when the target conformations are unknown …

It can now be estimated from the particles and the generator's samples, without the MD structures. We fit a single offset of the particle embeddings jointly with the populations. One component of this offset changes the embeddings in the same way as moving the population along the coordinate, and the joint fit cannot separate the two. We set this component with a linear map from the embedding to the coordinate, learned on the generator's samples. The landscape error is then 0.12 kBT RMS, compared with 0.57 kBT uncorrected and 0.10 kBT with the true structures. The test covers one MD ensemble, so for experimental data we still rely on the mixture test above.

The changes are collected in the general response. If the new experiments settle these two points, we would ask you to consider updating your assessment. Where they fall short, we would be glad to discuss the point further.

I appreciate the authors' efforts to address my concerns, particularly by including experiments on real data. However, my major concerns remain: the existing experiments do not demonstrate a complete workflow on real data that would establish the method's applicability to real cryo-EM tasks. The experiments on real data still bypass the classification issue, directly using well-analyzed classification results from published datasets, which were obtained using conventional pose-search-based reconstruction methods. More importantly, the limitations and potential biases associated with pose estimation and density reconstruction are inherited by the downstream guidance pipeline through these pre-classified particles. In other words, pose estimation and density reconstruction are implicitly performed during upstream classification, which undermines the claim that the method can provide guidance without maps or poses.

Moreover, the results on real data are still unsatisfactory. Substantial mismatches with the deposited atomic models can be observed in the qualitative comparisons. Quantitative results also show relatively large differences, especially for PaBphP. Besides, the population results further indicate that the bias introduced by the proposed conformational population inference cannot be ignored.

At this stage, I would consider the proposed method to be an interesting idea, yet it appears unsuitable for practical cryo-EM workflows. I would like to see the authors demonstrate how the proposed method can be applied to real data, for example, by conducting the same workflow used in Appendix E.2 on TmrAB and PaBphP.

Thank you for the concrete request. In the revised App. J we guide both proteins by unsorted particles, without the classes of the depositors (Table S19). The guided samples also lie closer to the deposited models, now that we model radiation damage in the simulated images and correct the offset between the embeddings of experimental particles and simulated particles (Table S18, Figure S1).

Guidance without classes. We take the deposited particles without their class labels and fit the population model of §3.4 to them. We then guide each sample toward a conformation drawn from the fitted population. We do not guide by the unsorted particles directly, because a set that mixes two states shifts the samples toward the majority state (App. E.2).

For phytochrome, we guided 24 samples toward an open conformation, and all became open, while no unguided sample became open. These 24 samples lie at a Cα RMSD of 0.54 nm from the deposited open model, and the unguided samples that start from the same noise lie at 0.79 nm. For the transporter, 35 samples were guided toward an inward-facing conformation, and 31 became inward-facing, compared with 2 of the unguided samples that started from the same noise.

Bias inherited from classification. In these runs we assign no particle to a class, estimate no orientation and build no map. However, the runs still rely on the depositors for picking, cleaning, and centring the particles, and for the defocus values, and App. J lists this as a limit.

Agreement with the deposited models. The agreement is closer in the revision. Samples that are guided by the open class of the phytochrome lie 0.35–0.48 nm from the deposited open model, and unguided samples lie 0.8 nm from it. Samples that are guided by the inward-facing class of the transporter lie 0.24–0.26 nm from the deposited inward-facing model, and unguided samples lie 0.7 nm from it. Figure S1 shows the guided samples next to the deposited models.

Populations. We agree that the absolute shares are biased. In the runs without classes, the fitted share of the smaller state is 16% for the phytochrome and 25% for the transporter, and the depositors assigned 36% of the particles to it. The method recovers changes in share better than the shares themselves. On mixtures of known composition, the estimated share rises with the true share at a slope of about 0.8 on both proteins. We therefore limit our claims about populations to changes between the particle sets of one dataset.

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

This methods guides a pretrained protein structure predictor to target conformations directly using cryo-EM images. It learns a shared latent space between the predictor's diffusion network and a pretrained cryo-EM image encoder, enabling pose-free comparisons of conformations between cryo-EM images and diffusion rollouts, thus steering the latter. The method is validated on three synthetic datasets.

Critical strengths

The idea of generating atomic ensembles directly from cryo-EM images (without going through 3D maps) is of interest in the field. Robust cryo-EM likelihoods are necessary for this, but physics-based likelihoods often lack sensitivity with noisy cryo-EM data. Thus the approach of learning a pseudo-likelihood through a shared embedding space between the images and structure predictor intermediates is novel and can enable interesting applications.

Critical weaknesses

The task setting feels contrived.

  • The training structures come from BioEmu itself, so we must assume that it already has sufficient coverage of the conformational space, which is often not the case (see BioEmu's performance at multi-conformation prediction in the ConforNets benchmark [1] for example). There is no evidence that this method can drive the diffusion to distant states.
  • In inference in the GroEL experiment, 128 images of the same conformation are used, which is possible as they are synthetically generated from a single structure. In a real experimental setting, conformation labels are unknown, and can only be imperfectly estimated using a heterogeneous cryo-EM reconstruction algorithm. Doing this kind of analysis is possible but does in most cases require upstream estimated poses, which undercuts the pose-free benefits of the proposed method. Pose and conformation must still be disentangled elsewhere in the pipeline even if the learned embedding space is pose-invariant.

Additionally, the results even in this setting are not compelling.

  • All 3 demonstrated systems use synthetic test images which are in-distribution with the simulated training images by construction. It is not clear if imperfect experimental images would be embedded consistently with the synthetic training images.
  • One of these 3 is reported to not steer well with guidance, instead requiring SMC on pre-sampled trajectories.
  • While core RMSD does improve compared to the unguided baseline, it is still quite high, e.g. 12.2 A (=1.22nm) for GroEL extended and 16.3 A for lysenin extended.

[1] Lee et al. ConforNets: Latents-Based Conformational Control in OpenFold3. arXiv, 2026.

Questions for authors
  • Can your method steer to conformations outside the support of the unguided BioEmu ensemble?
  • Does the shared embedding space transfer to real experimental cryo-EM images once trained on synthetic ones?
4 comments

Thank you for the frank review. Both of your questions concern whether the method is practical, and we now answer both on experimental particles.

W1, Q1. States outside BioEmu's samples.

Can your method steer to conformations outside the support of the unguided BioEmu ensemble?

Yes, within the range of the structures used to fit the likelihood (training structures), which need not be BioEmu's own samples.

BioEmu rarely samples the extended GroEL target, and guidance reduces the median distance to it from 1.82 to 0.53 nm.

BioEmu2 never samples the phytochrome's open state, so for this protein the training structures come from sampling with a bias that moves the two PHY domains apart (general response, point 2). Guided by experimental particles of the open class, and with radiation damage modelled in the simulated images, the samples have PHY domains 0.7 nm farther apart than when guided by the closed class (the deposited states differ by 2 nm).

The limit is therefore the coverage of the training structures, which can be wider than the unguided ensemble. We agree that BioEmu's own coverage is limited, and we now cite the ConforNets benchmark.

W2. One conformation, labels and poses.

Pose and conformation must still be disentangled elsewhere in the pipeline …

Guidance needs neither images of a single conformation nor fine classes, and in simulation it works without classes.

A simulated class of 32 different conformations, four views of each, gives a median core RMSD within 0.03 nm of that from 128 images of one conformation.

For the experimental data, we used the depositors' classes; for the transporter, these are the first-round classes (8.6 Å resolution). As the reviewer says, they come from pose-based processing, but only from its first and coarsest step, after which we estimate no orientations and build no map.

We can also replace classes with the population model. Fitted to the whole mixed dataset, it assigns each particle a weight for every state, and we use these weights to guide toward each state without orientation estimates. In simulation, the results are within 0.03 nm of guidance on perfectly labelled particles, even for a state that makes up 10% of the data.

W3, Q2. Transfer to experimental images.

Does the shared embedding space transfer to real experimental cryo-EM images …

Yes on the transporter, and partly on the phytochrome. The concern was justified for the submitted simulator, because its white noise embedded the experimental particles away from the simulated ones. When we simulate the training images with noise measured on empty ice from the same datasets, much of this gap disappears.

On the transporter, samples guided by the inward-facing class are 0.25–0.43 nm from the deposited inward-facing models (unguided: 0.65–0.71 nm). Samples guided by an outward-facing class remain closest to the outward-facing models.

As a control, we randomised the Fourier phases of the particles, which destroys the image of the molecule but keeps the noise and the defocus. The samples of the two classes then differ by 0.13 nm in the distance between the nucleotide-binding domains, down from 0.88 nm. Guidance therefore distinguishes the classes mainly by the molecule in the images and little by the imaging conditions.

On the phytochrome, the samples move in the correct direction but stop well short of the deposited open state (W1 above).

W4. One system needed SMC.

One of these 3 is reported to not steer well with guidance …

This was a weakness of the submission, and we fixed it in the revision (general response, point 1). Guidance alone now gives samples a median 0.21 nm from DppA's open state (unguided: 0.53 nm), and SMC is no longer used.

W5. Core RMSD still high.

it is still quite high, e.g. 12.2 A (=1.22nm) for GroEL extended …

We agree that the submitted samples were far from GroEL's targets. With the revised sampler, the core RMSD is 5.3 Å to GroEL's extended target (unguided: 18.2 Å) and 2.6 Å to the compact target (unguided: 6.2 Å). DppA's two states are at 2.1 and 1.0 Å. The remaining error on GroEL is in domain arrangement, and each domain keeps its fold. Lysenin's extended target is worse than in the submission (18.2 Å; unguided: 23.5 Å), mainly because the new preconditioner restricts one segment, and we report this as a limitation.

We hope that the results on experimental particles and the revised sampler address the concern about practicality, and the general response provides the full list of changes. If they do, we would appreciate it if you would revisit your score. If you still see an obstacle to practical use, we would like to hear it while we can still respond.

Dear Reviewer RkSf,

The revised PDF addresses each point in your review, and the two you raised as questions are now answered in the main text. We would be grateful if you could look at it before the discussion closes.

Conformation labels and upstream poses. Guidance now starts from an unsorted stack. We estimate the populations of its particles and guide the samples using them, without classes or orientations (§3.5, §4.3).

Experimental images (Q2). Two experimental datasets are now in the main text (§4.5, Figure 4).

States outside the generator's samples (Q1). BioEmu2 does not sample the open state of the phytochrome on its own. Guided by the unsorted experimental particles of this protein, 30 of 128 samples are open, and none of the unguided samples is (§4.5).

SMC. DppA opens by guidance alone, and SMC is no longer used (§4.4).

Core RMSD. For GroEL's extended target, it is 0.53 nm, whereas the submission had 1.22 nm and unguided samples have 1.82 nm (§4.4).

If these changes answer your concerns, we would appreciate it if you reconsidered your rating. If you still have concerns, please tell us which ones, so we can respond.

I greatly appreciate the authors' efforts in their additional experiments. I think the additional theory around class-free guidance and the inclusion of experimental data enhance the applicability of the method, so I will raise my score to a 3. Looking at the RMSD column of Table S18, it appears that guidance produces a better structure for one state of each dataset, the one which is not sampled unguided, but does not move the other states. Can the authors clarify this? Would you say the method mostly drives coarse conformational changes but cannot refine higher resolution details (at least as demonstrated on experimental data)? Such a method could still be valuable, but in that case, I think the dependence on conformational coverage of the training distribution becomes a more critical bottleneck. Needing to generate a training set for the phytochrome by biasing the domain movement then undercuts the conformational steering the method intends to support.

Thank you for the careful reading and for raising your rating. Your last point led us to run a new experiment, which is now App. L.

You read Table S18 correctly. BioEmu2 already samples one state of each dataset without guidance (the closed phytochrome, the outward-facing transporter), so there is nothing for guidance to add for that state. The other state is one that BioEmu2 reaches rarely or not at all, and there guidance brings the samples much closer to the deposited model (0.35–0.48 nm instead of 0.78–0.80 nm for the open phytochrome). At another reviewer's request, we now report these RMSDs in Å in the revised PDF, and we keep nm here to match your comment.

As you suggest, what we show on experimental data are movements of whole domains. In the open phytochrome, the two PHY domains end up about 2 nm further apart, as in the deposited model. We have not tested whether guidance also improves finer detail, because we use the state that BioEmu2 already samples as the reference for our calibration. In simulation, it does, since guided samples of DppA are 0.10 nm from the closed structure and unguided samples are 0.25 nm away (§4.4).

We agree that coverage is then the bottleneck, and that the bias is a real limitation that we have not fully solved. The bias only makes the open state available during the map fitting between latent spaces, and whether the samples open still depends on the particles. With the same candidates, 1 of 32 samples is open when we guide with particles of the closed class, and 27 of 32 are open when we guide with particles of the open class (App. J). In App. L we explore how to build the bias with less prior knowledge. We compare the structures of related proteins in the PDB and find the two parts that move most against each other, which are the two PHY domains. With the bias between them, 20 of 128 guided samples are open and no unguided sample is. Our own choice of domains gives 30. So this route works only in part. In the Discussion, we describe coverage as a central constraint, and we now point to App. L there.

Thank you again. Your questions made the paper clearer about what it shows on experimental data and what it still needs.

Reviewer_6ZaKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary
  1. The shared-code likelihood is an interesting way to guide a frozen protein model with particle images. Accounting for image and denoising uncertainty gives a clear motivation for the guidance weights.

  2. The appendix reports useful failure cases and controls, including unsuccessful DppA guidance, equal-cost sample selection, imaging shifts, and limits on landscape resolution.

Critical strengths
  1. The shared-code likelihood is an interesting way to guide a frozen protein model with particle images. Accounting for image and denoising uncertainty gives a clear motivation for the guidance weights.

  2. The appendix reports useful failure cases and controls, including unsuccessful DppA guidance, equal-cost sample selection, imaging shifts, and limits on landscape resolution.

Critical weaknesses
  1. Practical validation and baseline fairness are limited. All particles are simulated without translation, and each guidance stack comes from one conformation. Several main RMSD comparisons use a repaired guided sampler but an unrepaired reference, with only 25 of 300 lysenin reference samples retained. The repaired PDB/SMC controls are helpful, but the main comparisons should use the same sampler, repair, and filtering. Please also discuss Fold'EM, which likewise uses a protein generative prior to infer structures directly from particle images. This is relevant to the claim of conditioning without map reconstruction.

  2. The independent MD result uses ground-truth information. Appendix I.5 corrects a code offset using the true structural coordinate, reducing RMS error from 0.57 to 0.10 kBTk_\mathrm{B}T. Appendix I.4 also selects the penalty using imposed landscapes. The uncertainty bands do not include candidate-library or model bias. The uncorrected result and these calibration limits should be clearer in the main text, since they affect the claim of recovering unknown landscapes.

Questions for authors
  1. What RMSD gains remain with the same sampler, repair, and filtering? Please summarize the existing matched controls.
  2. Can the MD offset be estimated without the true structural coordinate? Which landscape results do not use ground-truth calibration?
  3. Are there existing results with particle translations or mixed-conformation stacks? How would single-conformation stacks be obtained in practice?
  4. How does the method differ from Fold'EM in particle conditioning and treatment of pose uncertainty?
1 comment

Thank you for the detailed review and the concrete questions. Each point led to a change in the paper.

W1, Q3. Translations and mixed stacks.

Are there existing results with particle translations or mixed-conformation stacks?

The method now also runs on experimental particles from two public datasets (general response, point 2).

Translations. Yes. We added 1 nm shifts, a B-factor envelope, and structural noise to both the simulated particles and the training images, and guidance retains 98% of its improvement over unguided sampling on both GroEL targets.

Mixed-conformation stacks. Yes. With 10% of the particles replaced by another state or by junk particles, the guided samples are 0.36 nm from the compact GroEL target (pure stack: 0.27 nm; unguided: 0.62 nm).

W2, Q1. Matched controls.

What RMSD gains remain with the same sampler, repair, and filtering?

We agree, and the matched comparison is now the paper's main evaluation. Each guided sample is paired with an unguided sample from the same starting noise, with the same sampler, repair and scoring. We count every pair, so the filtered lysenin reference (25 of 300) is no longer used.

On all 17 targets, the guided samples have the lower core RMSD, and the 95% confidence interval of the paired difference always excludes zero (App. A). The last column of the table in the general response gives the unguided medians of the six main targets. With all settings fixed, we also ran 27 targets that we did not select by hand (20 GroEL conformations newly sampled from the generator, and seven conformations that no earlier run had used as a target). Guidance lowers the median core RMSD on 26 of them.

W3, Q4. Fold'EM.

How does the method differ from Fold'EM …

Thank you for raising Fold'EM. It appeared on arXiv on 1 October 2026, after the submission deadline. We now cite and discuss it as concurrent work and make no experimental comparison. Both methods condition a protein generative model directly on the particle images.

Particle conditioning. Fold'EM renders the current structure at each particle's pose, compares the rendering with the particle, and optimises an intermediate representation of AlphaFold3 to improve the agreement. Our method instead embeds each particle once with a frozen image encoder. We fit a likelihood of these embeddings on simulated images and add it to the generator's score.

Pose uncertainty. Fold'EM needs one pose per particle, either supplied by a consensus refinement or estimated from the particles in alternation with the structure updates. Our likelihood instead treats the viewing direction as unknown. The viewing direction explains only 5.6% of the variance of our shared code, and the likelihood accounts for the resulting scatter with a measured covariance.

W4, Q2. The MD correction.

Can the MD offset be estimated without the true structural coordinate?

Largely, yes. The estimate now uses the particles and structures sampled by the generator, and never the MD structures or their coordinate. We fit one shift of the particle embeddings together with the populations. The particles cannot determine the part of this shift that lies along the opening coordinate, because it has the same effect as a shift of the population. We therefore obtain that part from a linear predictor of the coordinate that is trained on the generator's samples. On this ensemble, the RMS error is 0.57 kBT without a correction, 0.12 kBT with this estimate and 0.10 kBT with the correction from the true structures.

W5. The penalty.

Appendix I.4 also selects the penalty using imposed landscapes.

The results are not sensitive to this choice. When we apply the penalty to one half of the landscapes and evaluate on the other half, the error is nearly the same (0.289 vs 0.287 kBT). In the revision, the primary rule for the penalty is the Laplace approximation to the model evidence, which needs no ground truth.

W6. Uncertainty bands.

The uncertainty bands do not include candidate-library or model bias.

This is a fair criticism. The submitted bands reflected only the finite number of particles. They now also include the spread from refitting on resampled sets of candidate structures, which raises coverage of the true landscape to 80% (89–98% for barriers up to 3 kBT). Higher barriers remain under-covered because their height is underestimated, a bias that resampling cannot capture.

W7. Limits in the main text.

these calibration limits should be clearer in the main text …

We agree. §4.4 now gives the uncorrected result first and states that barrier heights are underestimated.

The general response lists every change and where it appears in the paper. If the matched comparisons and the results on experimental particles resolve your main concerns, we would be grateful if you would reconsider your rating. If a point remains open, please tell us which, and we will go into more detail.

Comments on the paper

2 comments

We thank all three reviewers for their careful reading and constructive criticism. All three reviewers see the paper's strength in its likelihood, defined through a representation shared by images and structures. Their concerns are about the evidence that the method works in practice, and we group them into three points: (1) the results were not convincing and were not compared on equal terms (RkSf, 6ZaK); (2) every particle was simulated (all three); (3) guidance seemed to require particles of a single conformation (all three). We revised the sampler and added experiments that address each of the three points, while the core of the method remains unchanged.

1. The results are stronger and are compared on equal terms. In the submission, guidance alone did not open DppA, and we fell back on SMC. The main cause was the preconditioner, which smoothed the guidance along the chain and thereby damped hinge motions. In the revised sampler we build the preconditioner from the protein's residue contacts instead. With guidance alone, the samples are now a median 0.21 nm from DppA's open state, so we removed SMC from the paper.

Further, we now compare every guided sample with an unguided sample generated from the same starting noise, with the same sampler, repair and scoring (last column of the table below). On all 17 targets (the six in the table, a third lysenin target and ten GroEL chains from the PDB), the guided samples are closer to the target, and every 95% confidence interval of the paired difference excludes zero.

Median core RMSD to the target, in nm:

TargetSubmittedRevisionUnguided
GroEL, extended1.220.531.82
GroEL, compact0.440.260.62
DppA, open0.470.210.53
DppA, closed0.120.100.25
Lysenin, compact1.270.841.60
Lysenin, extended1.631.822.35

For DppA, the submitted and revised columns both show guidance without SMC.

2. On experimental particles, guidance steers the samples toward the correct state. We applied the method to two public datasets, a transporter (EMPIAR-10317) and a phytochrome (EMPIAR-12874). Both are multi-chain complexes that BioEmu cannot sample, so we used its successor, BioEmu2, released in October after submission. Each run used about 1,000 particles from one deposited class. On the transporter, samples guided by the inward-facing class are 0.25–0.43 nm from the deposited inward-facing models, whereas unguided samples are outward-facing and 0.65–0.71 nm away. BioEmu2 never samples the phytochrome's open state, so we generated training structures by biasing its sampler to move the two PHY domains apart. We chose these domains because PHY domains are known to separate in phytochromes, and we used neither the deposited open structure nor the particles. Samples guided by the open class have PHY domains 0.7 nm farther apart than samples guided by the closed class (the deposited open and closed models differ by 2 nm).

3. The particles do not need to show a single conformation. There are two ways to group the particles, without requiring pure classes. The first is to use existing classes; for the experimental data, we used the depositors' classes unchanged, and for the transporter, we used the coarse classes from the first round of classification (8.6 Å resolution). In simulation, replacing 128 images of one conformation with images of 32 different conformations (four views of each) changes the median core RMSD by at most 0.03 nm. Lastly, the second approach requires no classes because the paper's population model estimates how particles divide between candidate states. We fit it to the whole mixed dataset and obtain a weight for each particle and state. In simulation, guidance with these weights is within 0.03 nm of guidance on perfectly labelled particles, even when one state makes up 10% of the data.

The offset in the comparison with MD (§4.4) is now estimated from the particles and the generator's samples, without the MD structures (6ZaK, CUwr). The table below lists every change.

changewhere
revised sampler (preconditioner from residue contacts); SMC removed§3.3, §4.3, App. B
matched unguided controls, confidence intervals§4, App. A
experimental particles, controls, mixtures of known compositionnew appendix, §5
classes of different conformations, wide or impure§3.3, App. E
grouping particles without classes; RELION, cryoDRGN, mapsApp. H, new appendix
particle shifts; noise from empty iceApp. E, new appendix
MD offset estimated from the particles; wider uncertainty bands§4.4, App. H
BioEmu's coverage (ConforNets); a wider set of training structures§2, §5

We thank all three reviewers. Your comments have made this a better paper, and we have rewritten the main text around them. The building blocks are those of the submission: a frozen generator, a frozen image encoder, the likelihood that links them, and the population model. What is new is that the population model now provides the targets of the guidance.

Unsorted particles come first (§3.4–§3.5, §4.2–§4.3). Reviewers RkSf and CUwr pointed out that particles of one conformation are not available without upstream processing. Guidance now starts from an unsorted stack. We estimate its populations and guide the samples toward conformations with these populations, without classes, orientations or maps. The experiments of the submission are the special case (§4.4).

Experimental particles are in the main text (§4.5, Figure 4). As Reviewer CUwr asked, the unsorted experimental particles are now a main experiment, and Figure 4 shows the generated structures next to the deposited models.

The claims are scoped (§5). In response to Reviewers 6ZaK and CUwr, we state in the Discussion what is validated on experimental data and what is so far shown only in simulation.

Where a number differs from an earlier comment, the revised PDF has the current one. We hope that the revised paper answers your concerns, and we will gladly respond to any that remain before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.