acceptodds
Under review as a conference paper at ICLR 2027

LatentFold: Continuous MSA Representations without Discrete Homolog Generation

Abstract

Multiple sequence alignments (MSAs) support accurate protein structure prediction by exposing conservation and coevolutionary signals. Existing MSA generative methods largely instantiate these signals as discrete homologous sequences, even though downstream folding models ultimately consume learned MSA representations. We ask whether useful MSA-derived information must take this discrete form. We propose LatentFold, a fully differentiable framework containing LatentMSA, a sequence-conditioned generator that instead represents structure-relevant MSA signals as a collection of continuous latent states. LatentMSA outputs an MSA-shaped latent tensor rather than discrete homologous sequences, producing distinct probe-conditioned states around a dynamic sequence-specific anchor. It is trained with a permutation-invariant statistical alignment objective that matches selected depth-aggregated moments and residue-pair similarity patterns of natural MSA representations, avoiding explicit correspondence between unordered homolog rows. Experiments on FoldBench and CASP15 show that LatentFold outperforms the evaluated discrete MSA-generation baselines across most concat-MSA-depth comparisons, with the clearest shallow-depth gains on FB_PROTEIN and more metric-dependent improvements on CASP15, while requiring no target-specific homology retrieval at inference time. Its differentiable design allows structural-loss gradients to pass through the frozen folding pipeline and update LatentMSA while the PLM and folding backbone remain frozen. These results support continuous latent MSA generation as a trainable alternative to discrete homolog generation for supplying structure-relevant MSA signals. The code is available at: https://anonymous.4open.science/r/LatentFold-2C68

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
MSA generation, Moment matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xJZ1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LatentFold, a protein structure prediction framework that replaces discrete homolog generation with continuous latent MSA representations. The method uses a sequence-conditioned generator, probe-specific attention, and permutation-invariant statistical matching to learn MSA-like features from protein language model embeddings. Experiments on FoldBench and CASP15 report improvements over several discrete MSA-generation baselines, particularly after structure-supervised adaptation through a frozen folding backbone.

Critical strengths

The paper challenges the assumption that evolutionary information must be represented through discrete homologous sequences. The proposed continuous representation provides a differentiable interface to existing folding models and removes the need for target-specific MSA retrieval during inference.

The evaluation includes multiple sequence-depth settings, several protein language model sizes, and analyses of individual loss components.

Critical weaknesses

The comparison does not isolate the benefit of continuous MSA representations because the strongest LatentMSA(E2E) model receives structural supervision unavailable to the discrete generation baselines, and the methods use different pretrained representation models.

The proposed coupling evaluation relies on distances between projected residue embeddings rather than direct biological coupling measurements, making it difficult to distinguish genuine coevolutionary information from residue similarity and conservation patterns.

Questions for authors

Can the authors compare LatentFold against a simpler continuous MSA adapter trained with the same structural supervision and frozen folding backbone to isolate the contribution of the proposed architecture?

Does the generated representation improve agreement with experimentally observed residue contacts or established coevolution measures, rather than only matching the natural-MSA embedding space used during training?

2 comments
AuthorsOpenReview ↗

Response to Reviewer xJZ1

Thank you for your careful review. Below is our step-by-step response to your comments.


Response to Q1

We added this comparison in Table 4. The adapter uses the same data interface, training data, and training steps, with offline and E2E variants evaluated through the same frozen Protenix backbone. LatentMSA improves overall lDDT from 0.739 to 0.793 offline and from 0.775 to 0.823 after E2E training; TM-score rises from 0.833 to 0.853 and from 0.870 to 0.885. The adapter architecture and training settings are in our anonymous repository.

Table 4. LatentMSA versus a simpler adapter on FoldBench. Oracle results; mean ± standard deviation across targets, shown as chain lDDT / TM-score.

MethodDepth [1, 100), N=33Depth [100, 500), N=47Depth ≥500, N=254Overall, N=334
LatentMSA, offline0.678 ± 0.221 / 0.739 ± 0.2500.715 ± 0.198 / 0.753 ± 0.2410.822 ± 0.122 / 0.886 ± 0.1350.793 ± 0.156 / 0.853 ± 0.177
Adapter, offline0.675 ± 0.216 / 0.749 ± 0.2520.605 ± 0.199 / 0.678 ± 0.2430.773 ± 0.137 / 0.873 ± 0.1540.739 ± 0.167 / 0.833 ± 0.194
LatentMSA, E2E0.733 ± 0.208 / 0.797 ± 0.2310.746 ± 0.181 / 0.789 ± 0.2290.849 ± 0.102 / 0.913 ± 0.1120.823 ± 0.137 / 0.885 ± 0.157
Adapter, E2E0.701 ± 0.204 / 0.771 ± 0.2370.674 ± 0.198 / 0.762 ± 0.2260.803 ± 0.107 / 0.902 ± 0.1190.775 ± 0.144 / 0.870 ± 0.163

Response to Q2

We agree that feature-space similarity alone does not establish recovery of biological coevolution. We therefore added a contact-ranking analysis using latentMSA, projected natural-MSA features, ESM features, and natural-MSA MI-APC scores.

Contacts are within-chain residue pairs with Cβ distance below 8 Å; we use Cα for glycine and other residues lacking Cβ. We rank pairs by negative L2 distance for feature representations and by MI-APC for the natural MSA, then report precision among the top L/5, L/2, and L pairs.

Table 5. Contact-ranking precision. Feature distances and MI-APC use different scoring functions, so this is an exploratory comparison.

ScoreP@L/5P@L/2P@L
LatentMSA features, negative L20.0930.0810.073
Natural MSA, MI-APC0.1690.1310.108
Projected natural-MSA features, negative L20.0920.0800.070
ESM features, negative L20.0600.0510.045
Random reference0.0250.0250.025

LatentMSA has similar contact precision to projected natural-MSA features and higher precision than raw ESM features, though it remains below MI-APC. This supports its agreement with the learned feature space. However, it does not establish a direct correspondence between the latent features and the contact map.

These analyses address important aspects of my concerns. I will maintain my current positive score.

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

This paper proposes LatentFold, a framework that generates continuous MSA-like representations from pretrained ESM-2 embeddings to improve protein structure prediction. Instead of generating discrete homologous sequences, the model learns to recover statistical properties of natural MSAs in latent space, which are subsequently used by a frozen AlphaFold3-style structure prediction model (Protenix). The authors demonstrate improvements over several discrete MSA-generation baselines on monomer structure prediction benchmarks.

The general direction of recovering evolutionary information from pretrained protein language models is worth exploring. However, I have significant concerns regarding the novelty, evaluation scope, and practical relevance of the proposed method.

Critical strengths
  • The idea of directly generating continuous MSA representations, rather than discrete homologous sequences, is interesting and allows differentiable integration with downstream structure prediction models.

  • The method demonstrates improvements over existing MSA-generation approaches, including EvoGen, EvoDiff, and MSAGPT.

  • The framework avoids target-specific homology retrieval at inference time, potentially reducing computational overhead.

Critical weaknesses
  1. Limited evaluation scope and practical relevance

My primary concern is that the paper evaluates the proposed method exclusively on monomeric protein structure prediction.

Although improving monomer structure prediction remains valuable, much of the current interest in biomolecular structure prediction has shifted toward more challenging problems, including protein–protein interactions, protein–ligand binding, and protein–nucleic acid complexes. Recent models such as AlphaFold3, Boltz, Chai-1, and Protenix are designed to address these problems.

Given that the proposed framework is built upon Protenix, an AlphaFold3-style model, I would expect the authors to investigate whether the generated latent MSA representations can improve complex structure prediction as well. Without such evaluations, it is difficult to assess the practical usefulness and broader applicability of this approach.

  1. Limited competitiveness with existing frontier models

Even within monomer structure prediction, the reported performance raises concerns about the practical competitiveness of the proposed method.

For reference, previously published FoldBench protein-monomer results reported by SeedFold include:

ModelMonomer LDDT ↑
SeedFold0.889
AlphaFold30.880
Protenix-0.50.877
Boltz-10.870
Chai-10.870

Source: https://seedfold.github.io/

In comparison, LatentFold reports monomer LDDT values of 0.733, 0.746, and 0.849 across the shallow-, medium-, and deep-MSA subsets, respectively, after structure-supervised adaptation.

Although these numbers are not directly comparable because of differences in evaluation protocols, model selection, and potentially target subsets, they raise concerns about whether the proposed framework is competitive with established frontier structure prediction models.

I acknowledge that the authors show improvements over discrete MSA-generation baselines. However, outperforming these methods does not necessarily establish that the proposed approach is practically useful for modern structure prediction. A more comprehensive comparison with stronger contemporary models, under matched evaluation conditions, would substantially strengthen the paper.

Questions for authors
  • Can the proposed method be extended to protein–protein, protein–ligand, or protein–nucleic acid complexes? Are there any preliminary results demonstrating benefits beyond monomer folding?

  • How does LatentFold compare with stronger AF3-style models under identical evaluation protocols, particularly on the FoldBench monomer subset?

1 comment
AuthorsOpenReview ↗

Response to Reviewer WRJT

Thank you for your careful review. Below is our step-by-step response to your comments.


Response to Q1

Our main contribution is learning useful continuous representations for a fixed folding model, not training a new AF3-like predictor with the highest accuracy across all tasks. Monomer prediction directly tests this question.

As an AlphaFold 3 author noted, complex prediction may work without explicit MSA pairing (AF3_issue437). We therefore apply the current LatentMSA to complex prediction without modification. Learning continuous representations of paired MSAs could further improve performance and remains worth exploring. Below, we provide a preliminary comparison of LatentMSA and MSAGPT on FoldBench's antibody–antigen subset, using the same Protenix backbone.

Table 2. DockQ success rates on FoldBench's antibody–antigen subset.

SelectionProtenix + LatentMSAProtenix + MSAGPT
Oracle0.4086020.161290
Top-10.2857140.109890
Top-50.3260870.130435

Response to Q2

Our method outperforms the evaluated MSA-generation methods under a shared backbone. This tests the value of the learned representations, whereas full-input frontier scores also reflect different models and input information. We do not claim to replace the best full-MSA predictor.

A more direct question is how much predictive value latentMSA provides relative to a natural MSA. We therefore compared it with 8-, 16-, 32-, and 64-sequence MSAs using the same Protenix backbone.

Table 3. Natural-MSA-depth comparison on FoldBench. Oracle results; mean ± standard deviation across targets, shown as chain lDDT / TM-score.

InputDepth [1, 100), N=33Depth [100, 500), N=47Depth ≥500, N=254Overall, N=334
No MSA0.676 ± 0.239 / 0.712 ± 0.2740.531 ± 0.205 / 0.465 ± 0.2260.531 ± 0.184 / 0.423 ± 0.1760.545 ± 0.197 / 0.457 ± 0.212
Natural MSA, 8 sequences0.822 ± 0.147 / 0.896 ± 0.1400.739 ± 0.172 / 0.758 ± 0.2350.746 ± 0.180 / 0.773 ± 0.2290.752 ± 0.177 / 0.783 ± 0.225
Natural MSA, 16 sequences0.830 ± 0.148 / 0.902 ± 0.1490.834 ± 0.089 / 0.876 ± 0.1730.807 ± 0.156 / 0.855 ± 0.1930.813 ± 0.148 / 0.862 ± 0.186
Natural MSA, 32 sequences0.833 ± 0.147 / 0.903 ± 0.1500.860 ± 0.069 / 0.908 ± 0.1550.852 ± 0.123 / 0.907 ± 0.1590.852 ± 0.120 / 0.906 ± 0.157
Natural MSA, 64 sequences0.841 ± 0.134 / 0.909 ± 0.1370.869 ± 0.062 / 0.927 ± 0.1210.874 ± 0.092 / 0.934 ± 0.1100.870 ± 0.093 / 0.930 ± 0.115
LatentMSA, offline, 3B0.678 ± 0.221 / 0.739 ± 0.2500.715 ± 0.198 / 0.753 ± 0.2410.822 ± 0.122 / 0.886 ± 0.1350.793 ± 0.156 / 0.853 ± 0.177
LatentMSA, E2E, 2k steps0.733 ± 0.208 / 0.797 ± 0.2310.746 ± 0.181 / 0.789 ± 0.2290.849 ± 0.102 / 0.913 ± 0.1120.823 ± 0.137 / 0.885 ± 0.157

Overall, offline LatentMSA performs between the 8- and 16-sequence references. After 2k E2E steps, it slightly exceeds the 16-sequence reference in mean lDDT and TM-score, while remaining below the 32-sequence reference. The gains vary by group: it trails 16-sequence MSA in the shallow and medium groups but exceeds it in the deep group.

This gives a concrete measure of what the proposed LatentMSA can achieve. It does not imply identical biological information, but shows that useful folding signals can be supplied without searching for homologs.

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

The paper investigates whether MSA-derived information can support protein structure prediction without explicitly generating discrete homologous sequences. It introduces LatentFold, built on LatentMSA, which uses a sequence-conditioned anchor and multiple probes to generate continuous latent features supervised by natural-MSA statistics. These features are integrated with a frozen Protenix structure predictor. The paper also describes structure-supervised adaptation by backpropagating through Protenix into LatentMSA. Evaluation is conducted on FoldBench (FB_PROTEIN) and CASP15, with ablations examining the proposed representation and architectural choices.

Critical strengths
  • Generating latent features unconstrained by residue-probability distributions is a potentially valuable direction. The sequence-conditioned anchor and multiple probes provide a concrete mechanism for representing this information, making the approach conceptually interesting beyond conventional sequence generation.

  • Testing on FoldBench and CASP15 connects the learned representation to the intended application, rather than assessing it solely through reconstruction of MSA statistics. This is an appropriate validation strategy for determining whether the generated features contain structurally useful information.

Critical weaknesses
  • The novelty is overstated, and the central design is not adequately isolated. EvoGen [1] already demonstrates differentiable MSA generation without feeding discretized generated sequences into AlphaFold2, while allowing structural losses to backpropagate into the generator. Consequently, the Introduction’s assertion that useful MSA-derived signals must first be instantiated as discrete homolog-like sequences does not accurately characterize this prior work. The more defensible contribution is the unconstrained latent representation and its anchor/probe construction. However, the current ablations change multiple factors simultaneously and therefore do not establish whether this construction improves structural accuracy over a simpler sequence-conditioned latent generator. A matched comparison without anchors and probes, with the MSA-statistics supervision controlled separately, is essential to substantiate the proposed mechanism.

  • The evaluation and released implementation leave important claims unresolved. The motivation emphasizes orphan proteins, but the MSA-depth <100 group does not directly establish performance on proteins with no detectable homologs. Missing top-1 results for ESMFold and SimpleFold also limit interpretation of the best-of-25 comparisons, while MSA-based predictors would provide useful context for the low-homology setting.

  • Reproducibility is a further concern: the released code does not appear to include the full Protenix integration or an entry point for the described structure-supervised fine-tuning. Differences in token encoding, training schedules, and loss definitions additionally leave uncertainty about which implementation produced the reported results.

[1] Zhang J, Liu S, Chen M, et al. Unsupervisedly prompting AlphaFold2 for accurate few-shot protein structure prediction. Journal of Chemical Theory and Computation, 2023.

Questions for authors
  • Given EvoGen’s differentiable interface to AlphaFold2, what precisely distinguishes the proposed representation from prior continuous MSA-generation approaches? Can the authors provide a matched comparison against a sequence-conditioned latent generator without anchors or probes, holding the folding interface, training budget, and model capacity approximately fixed? Separately controlling MSA-statistics supervision would help distinguish the architectural contribution from the supervision signal.

  • Where are the implementation and training entry point for backpropagating structural losses through frozen Protenix into LatentMSA? Which reported results use this stage, and which use only MSA-statistics supervision?

  • Can the authors report top-1 results for ESMFold and SimpleFold alongside the best-of-25 results? Please clarify whether “top-1” denotes a single prediction or the highest-confidence prediction among multiple samples, how candidates are selected, and whether sampling and selection budgets are comparable across methods.

  • Does either benchmark contain a subset with no detectable non-query homologs under the stated MSA-search protocol? If so, can the authors report its size and performance separately, including relevant MSA-based predictors under the same available-information conditions? The existing depth <100 analysis does not resolve the paper’s orphan-protein motivation.

  • The code appears to one-hot encode HHblits token IDs directly before passing them to the Protenix projection, although the two encodings assign different channels to many symbols. Is there an explicit remapping elsewhere in the pipeline? If not, how does this affect the reported results?

  • Which configurations and checkpoints produced the reported tables? The manuscript states 60k optimization steps, whereas the scripts specify 120k for the main run and 12k for each of three ablations. Please also clarify the use of std.detach() in correlation normalization and SmoothL1 for the spatial loss, where the manuscript describes normalization without a stop-gradient operation and an L1 loss. Were the reported experiments run with these implementation choices?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Ccck

Thank you for your careful review. Below is our step-by-step response to your comments.


Response to Q1

We agree that differentiability alone is not new and will acknowledge EvoGen more clearly. Our approach differs in the following ways:

  • Representation: LatentMSA directly generates features after the folding model's MSA input projection, rather than homologous-sequence probabilities.
  • Profile prediction: A separate head predicts the profile from the generated latent features.
  • Architecture and supervision: Our anchor/probe design and permutation-invariant losses do not require matching generated rows to individual homologs.
  • Training data: We can learn from natural MSAs of different depths.

We also trained a simpler adapter using the same interface, data, and training steps, and compared both offline and E2E variants with the same Protenix backbone.

Table 1. LatentMSA versus a simpler adapter on FoldBench. Oracle results; mean ± standard deviation across targets, shown as chain lDDT / TM-score.

MethodDepth [1, 100), N=33Depth [100, 500), N=47Depth ≥500, N=254Overall, N=334
LatentMSA, offline0.678 ± 0.221 / 0.739 ± 0.2500.715 ± 0.198 / 0.753 ± 0.2410.822 ± 0.122 / 0.886 ± 0.1350.793 ± 0.156 / 0.853 ± 0.177
Adapter, offline0.675 ± 0.216 / 0.749 ± 0.2520.605 ± 0.199 / 0.678 ± 0.2430.773 ± 0.137 / 0.873 ± 0.1540.739 ± 0.167 / 0.833 ± 0.194
LatentMSA, E2E0.733 ± 0.208 / 0.797 ± 0.2310.746 ± 0.181 / 0.789 ± 0.2290.849 ± 0.102 / 0.913 ± 0.1120.823 ± 0.137 / 0.885 ± 0.157
Adapter, E2E0.701 ± 0.204 / 0.771 ± 0.2370.674 ± 0.198 / 0.762 ± 0.2260.803 ± 0.107 / 0.902 ± 0.1190.775 ± 0.144 / 0.870 ± 0.163

LatentMSA improves overall lDDT by 0.054 offline and 0.048 after E2E adaptation. The adapter architecture and training settings are available in our anonymous repository.


Response to Q2

We have updated the anonymous repository with the full Protenix integration and entry points for offline training, E2E training, and the adapter comparison. The code is not yet ready to run out of the box; upon acceptance, we will publicly release the full datasets and detailed training guidelines. E2E training backpropagates structural losses through frozen Protenix into LatentMSA; the PLM also stays frozen. Both stages use the same corpus, with a September 30, 2021 cutoff. Offline results use MSA-statistics supervision, while E2E results include structural supervision.

Top-1 is the prediction with the highest ranking score, not a single sample. Both oracle and Top-1 use the same 25 predictions: five seeds × five samples. Oracle uses reference-structure accuracy; Top-1 uses Protenix v0.5's ranking score.

Our Top-1 comparison focuses on Protenix-based methods to evaluate their selected structures under the same scoring function. ESMFold and SimpleFold use different scoring functions, so they were not included in that comparison. We also found little variation in ESMFold structures across seeds, making its Top-1 and oracle results nearly identical. We will make this scope and the selection rules clearer.


Response to Q3

We agree that shallow MSAs and absence of detectable homologs are different settings. We mention “orphan proteins” only to discuss earlier work, not to label our shallow group as an orphan benchmark.

The term also needs an operational definition: detectable homology depends on the database, search method, and thresholds. Some examples we examined in earlier work are short peptides or simple folds that already perform well without MSA. Our no-MSA reference likewise performs better in the shallow group than in the other groups, so MSA depth alone should not be treated as structural difficulty. We will clarify this distinction. Our main question is how to extract structurally useful information from PLMs without inference-time search; we do not claim a separate evaluation on proteins with no detectable homologs.


Response to Q4

Thank you for catching this. The earlier code upload did contain this mismatch. We corrected the mapping before running the reported experiments, so all results use Protenix's channel ordering. The anonymous repository now contains the corrected implementation and complete training pipelines.


Response to Q5

Offline training uses 120k steps; the 60k statement is a manuscript error. We will correct it and clarify the configurations used for each result.

We intentionally detach the standard deviation in the correlation-loss denominator. This lets the correlation loss update the feature patterns without differentiating through the normalization scale, while the separate standard-deviation loss trains that scale. Removing the stop-gradient made training less stable in our experiments.

Thanks for the response with additional experiments, some of my concerns have been addressed. The new adapter results are good, and I appreciate the corrected token mapping, added E2E implementation, and clarification of the orphan-protein scope. However, the adapter comparison still changes both architecture and supervision, top-1 comparisons with ESMFold and SimpleFold remain missing, and the training settings are still inconsistent between the paper and code. I will therefore keep my current score.

Thank you for the follow-up. We would like to clarify three points:

  • Architecture comparison. The proposed LatentMSA model architecture is itself one of our core contributions. The simpler-adapter comparison is designed to test that contribution, so changing the architecture is intentional, not a weakness in itself. The specific difference in supervision remains unclear, and identifying it would allow us to address the concern directly.
  • Top-1 evaluation. We hold the backbone and confidence head fixed to isolate our method's contribution. Our method substantially outperforms the evaluated methods in the same category. ESMFold and SimpleFold were included for broader context, not as matched controls, since their backbones and scoring functions differ. Their omission from the Top-1 comparison does not diminish this contribution. Moreover, our method is competitive with both and outperforms them in the reported oracle evaluation.
  • Training settings. We have explained the implementation choices and acknowledged that the manuscript's 60k-step statement should be corrected to 120k. Beyond this correction, the remaining paper–code discrepancies are not specified; concrete examples are needed for us to address them.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary

Rebuttal Summary

We thank Ccck, xJZ1, and WRJT for their careful reviews and helpful suggestions.

Ccck and xJZ1 both asked whether a simpler MSE-trained adapter performs similarly under the same budget. We added this comparison with matched interfaces, data, and training steps. LatentMSA achieves higher overall accuracy in both offline and E2E settings. The experiments are included in the revised manuscript and individual responses, and the adapter code is available in our anonymous repository.

Ccck also noted that EvoGen supports end-to-end training. We acknowledge this and have clarified the distinction in the manuscript: LatentMSA directly learns features after the MSA input projection rather than generating homologous-sequence probabilities, predicts the profile through a separate head, and explicitly accounts for MSA row-permutation symmetry in both its architecture and supervision. We especially appreciate Ccck's detailed code checks. The earliest uploaded version did contain implementation issues, but the reported experiments used the corrected code. We have updated the anonymous repository, although the code is not yet ready to run out of the box and step-by-step training guidelines are not yet available.

Regarding Ccck's orphan-protein concern, finding no homologs under a particular search setup does not necessarily establish that a protein is truly an orphan, nor does it determine folding difficulty. We believe the definition and evaluation of orphan proteins warrant more careful study.

For xJZ1's contact-map question, we added a contact-ranking analysis. LatentMSA has similar precision to projected natural-MSA features and higher precision than ESM embeddings, but remains below natural-MSA MI-APC. These results do not establish a direct correspondence with the contact map or recovery of biological coevolution. Folding models process MSA features through complex learned modules; our main evidence is that the learned latent features closely match those obtained by encoding natural MSAs and improve downstream structure prediction.

WRJT focused on frontier-model competitiveness and complex prediction. Our goal is not to train a new AF3-like predictor claiming universal SOTA performance, but to explore retrieval-free prediction with a fixed backbone. Although many high-accuracy predictors rely on MSA information, accessing useful signals without inference-time search remains an open question. Our added natural-MSA comparison provides an accuracy reference, while preliminary FoldBench antibody–antigen results show that the current LatentMSA supports complex prediction without modification and outperforms MSAGPT in this setting.

We thank xJZ1 for maintaining their positive assessment. We would appreciate hearing whether our responses have partly or fully addressed Ccck's and WRJT's concerns, and welcome further discussion with both reviewers to help us improve the work.

Dear reviewers,

Please be aware that little time remains for the author-reviewer discussion. As the deadline is approaching, we kindly request that you refrain from asking for additional experiments. Please summarize your final position concisely.

Sincerely, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.