acceptodds
Under review as a conference paper at ICLR 2027

A General Harness for Protein Foundation Model Fitness Prediction

Abstract

Accurate fitness prediction is central to protein engineering and understanding sequence–function relationships. With advances in deep learning, protein foundation models (PFMs) have become widely used for this task. Recent analyses, however, show that these models share preferences reflecting their training corpora, while unreliable inputs can further distort fitness predictions. Family-specific evolutionary evidence and structural context can help address these limitations by providing complementary constraints on model scores, motivating REM-Harness a general, model-agnostic, training-free Retrieval-Enhanced Mutation harness. It fuses frozen model scores with multiple sequence alignment (MSA) evidence according to model uncertainty, then applies gated background correction and score shrinkage based on structural confidence and solvent exposure. Across assays and million measured variants from ProteinGym, VenusMutHub, and the newly curated viral benchmark ViroHub, all configurations improve Spearman correlation on all benchmarks by on average, with broad gains across metrics. Extended analyses relate retrieval gains to model–MSA preference differences, assess domain-level gains and immune-escape cases, and quantify computational speedups. Built with RH, VenusREM2 is the first to rank highest in all function, taxon, MSA-depth, and mutation-depth categories, with a ProteinGym Average Spearman of , above the prior best.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein Engineering, Zero-shot Mutation Prediction, Directed Evolution, Protein Fitness Prediction, Protein Foundation Model, Model Harness, Viral Fitness Benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7oB8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Review

This was a challenging read and I apologize if my review sounds overly negative. It will also be quite long since I have a lot of questions.

The main text appears (to me) to be heavily LLM generated. I think the introduction does a fine job of setting the stage and motivating what prior research and results influenced this paper. It is however also immediately unclear to me what the actual contributions are. Clearly, REM Harness (RH) is proposed by the authors. But then the benchmark ViroHub is referred to as “a newly curated benchmark” without a reference; looking through the appendix, I believe it is also a contribution, but is never clearly stated. The same applies to VenusREM2. In the introduction, is says “VenusREM2 combines the harness with 6 variants of ProSST” but it is also not clearly stated that this model is introduced by the authors. Perhaps this is obvious to other reviewers, but I think a crucial job of the introduction is to clearly frame the paper and its contributions and I don’t think this is the case here. This results in a wobbly introduction, where I don’t know what to expect from the paper.

My main issue with the paper and reason for rejection is Section 2. Section 2 is overly complex and simultaneously verbose and overly concise: the introduction to the section is essentially a bullet point list later sentence become unreadable: “To allocate the complementary homolog evidence used by TRANCEPTION and VENUSREM (Notin et al., 2022a; Tan et al., 2025b), we distinguish a diagnostic-based target share from its scale-adjusted coefficient.”

A flood of new terms and quantities are constantly introduced without motivation whatsoever. I struggled to understand what the authors are actually proposing. Let’s take the Section 2 subsection titles as examples “ADAPTIVE RETRIEVAL AS SCALE ALLOCATION” and “COHERENCE-GATED CALIBRATION OF THE NATIVE CHANNEL”; I really don’t know what any of this means, even after reading the paper several times. What is “scale allocation” and what is “adaptive retrieval”? It isn’t clearly defined. What is the “native channel”? Solely based on the presentation, I cannot recommend acceptance of this paper. It needs a thorough revision by humans. If all text is human written then I apologize, but my points about verbosity, complexity, and general confusion still stand.

Paper summary

Now for the actual summary. The authors make three contributions: (1) a REM Harness method, which is a training-free method of adding MSA-derived information, structural plausibility via RSA/pLDDT, (2) VenusREM2 which is an application of the harness by which an ensemble of the existing ProSST models are wrapped in the REM Harness, and (3), I think, a ViroHub benchmark. As the ViroHub benchmark is only described in the appendix, I will not include in the review.

The REM Harness, which is the main contribution of the paper, builds on existing ideas and seeks to modulate existing variant effect predictions:

  1. MSA-derived per-site retrieval, akin to the retrieval mechanism in Tranception. A parameter α\alpha is defined in terms of site uncertainties.
  2. Model scores are calibrated using to account for the model’s bias.
  3. Model scores are further calibrated given the exposure (via RSA) and structural uncertainty (via pLDDT).

While I think these components make sense, their motivation is lacking and they are mostly applied and ablated. Additionally, the ablations in table 1 show that each contribution as a marginal effect across benchmarks. This table does not contain any measure of uncertainty and the net effects on aggregated Spearman correlations range from 0.02 to 0.028. I would very much like to see statistical test for significance. Figure 13 in the appendix shows a 95 % assay-bootstrap interval and the overlap between e.g. the full RH pipeline and the effect of including MSA retrieval only does not appear significant.

Critical strengths
  • State of the art scores on known benchmarks like ProteinGym is certainly a strength.
  • Comprehensive experiments
Critical weaknesses
  • Poor presentation. See above summary for detailed description. It is very challenging to read Section 2 and get an understanding of the exact mechanism the authors propose. While numerous equations define many quantities, few are adequately introduced, motivated, and described. To me, a lot of the text appears to be heavily LLM generated. For a concrete example, see L179: "The preference perspective of Gordon et al. (2025) motivates this direction; gated and ungated variants test the rule (Section 3.4)." This sounds like its directly taken from Claude. The over-the-top complexity obfuscates the REM Harness method. I think that simplicity is positive. If a simple, well-motivated component makes a positive contribution, I would prefer to let the simplicity speak for itself.
  • Unclear contributions. It is clear that the REM Harness is the main contribution of the paper. VenusREM2 appears as a secondary contribution which is the harness wrapped around an ensemble of the existing ProSST suite. ViroHub is described as a novel benchmark but is not detailed in the text. Reading the introduction, it is not clear what the main contributions are.
  • Lack of significance. Table 1 does not include measures of uncertainty and all reported scores are highly similar. This warrants statistical tests to identify whether the improvements from the RH components are statistically significant.
  • Lack of novelty. This point is more a matter of personal taste. I find new contributions that keep hill climbing on the ProteinGym benchmark (and related benchmarks) uninspiring. While there are some neat ideas, continuous hill climbing will not solve any real problems.
Questions for authors

This might seem like a lot and I don't expect the reviewers to answer all of my questions during the rebuttal. I hope that the extent of my questions will underline the points I raised in the summary and weaknesses.

  • L073: What does “scale-aware” mean? What is a “scale-aware mixing weight”? What "scale" are we referring to?
  • L076: “ reversing mutation contrasts.”? What is a mutation contrast?
  • L098: Using the term “backbone” in this context is a bit confusing to me since backbones are often used to refer to the structural backbone of the protein in question.
  • L099: Equation 1 is confusing and I had to read this section a couple of times to understand the notation. The tf shorthand is not explained - only in the appendix. I now see that equation aims to establish three types of “score fields”; one in the encoder-only setting, one in the decoder-only setting, and one for the wild-type sequence. This can be explained better.
  • L104: What is meant by the following: “Other native fields retain interface-specific scores, including unnormalized or window-aggregated outputs”? What is a “native field”? What is a “Raw-RH pair”?
  • L107: What is a “query-mapped site”? Maybe the query is the wild-type sequence?
  • L107: Here a tokenizer is explicitly mentioned; so am I right in assuming that we’re firmly in a language modeling paradigm?
  • L107: What does “Equal row weights and no pseudocounts define aligned frequencies and a log-softmax family field; either field X uses wild-type-relative contrasts” mean? What are wild-type relative contrasts? If I had just finished reading the EVE/EVcouplings papers, this might make more sense, but a more thorough explanation of the terms you introduce would greatly help the reader.
  • L145: What is the role of the mixing weight? What exactly is it mixing?
  • L146: “To allocate the complementary homolog evidence used by TRANCEPTION and VENUSREM (Notin et al., 2022a; Tan et al., 2025b), we distinguish a diagnostic-based target share from its scale-adjusted coefficient.” What does this even mean?
  • L151: What is z, and why are we clipping it?
  • What is model–family Pearson alignment? None of the quantities or how they’re defined is motivated whatsoever.
  • L179: “The preference perspective of Gordon et al. (2025) motivates this direction; gated and ungated variants test the rule (Section 3.4).” Perhaps I’m wrong, but this sounds so much like Claude.
  • L205: What is RSA? What is pLDDT? I am well aware of the answers, but these terms need to be defined before being used. Which structure prediction model is used to get the structures and these quantities? Is it the wild-type structure? How reliable is that for the assays with very high mutation numbers?
  • L240: It has not been explicitly stated that VenusREM2 is a contribution of this paper. Is that the case?
  • L240: since ProSST is an integral part of VenusREM2, it warrants a more thorough introduction. What does it do differently? Why did you choose this model? What do the different vocabularies signify?
  • What is ViroHub? Is it also a contribution by the authors in this paper? It is not clear from the main text.
  • L305: “Across 148 ProteinGym proteins scored with ESM-2 wild-type readouts, the fraction of globally–locally separated sites correlates with retrieval gain” - what is meant by wild-type readout? What are globally-locally separate sites?
  • L320: What is adaptive integration?
  • L438: What is v_48_020? Not all readers know that this refers to a specific ProteinMPNN weight checkpoint name.
  • Why wasn’t the existing ViroBench and ViroGym benchmarks used?
4 comments

We sincerely thank the reviewer for the careful reading, detailed suggestions, and recognition of our comprehensive experiments and predictive performance. We address the presentation and technical concerns below.

Weaknesses

W1. Poor presentation.

The Method now explains each step's purpose before introducing its calculation.

Thank you for your candid feedback. We agree that a simple method should be explained plainly. We simplified the subsection titles, clarified the definitions, and removed the sentence you quoted.

For example, Section 2.3 now explains the purpose of background correction before introducing its equation:

L196–198: MSA fusion adds family information but retains shared PFM preferences, which can reflect training-data composition (Figure 1; Gordon et al., 2025). We subtract this background from the PFM contribution to emphasize preferences at individual sites.

W2. Unclear contributions.

The Introduction now presents our method, benchmark, and ensemble model as three distinct contributions (L73–85).

We appreciate your suggestion to make their roles explicit:

  1. REM-Harness is a general, training-free framework for frozen PFMs. Across three benchmarks, all 71 configurations improve Spearman correlation, with an average gain of 0.073.

  2. ViroHub is a community benchmark containing 616,690 substitutions from 89 assays across 18 virus types, supporting studies of immune escape and other viral phenotypes. Section 3.2 (L306–319) describes its scope and data curation, and Table 1 compares it with ProteinGym and VenusMutHub.

  3. VenusREM2 combines six ProSST checkpoints with the harness, reaching 0.556 Average Spearman in the official ProteinGym comparison. Section 2.5 explains the ensemble. Compared with VenusREM's loops, MSA counting and mutation scoring are 62.3× and 20.7× faster, respectively, after model inference.

W3. Lack of significance.

For ProSST-2048, the full harness significantly improves on adaptive MSA fusion on all three full benchmarks.

Thank you for requesting this analysis. We added 95% confidence intervals and two-sided p-values for paired differences between component settings. We use 20,000 paired bootstrap samples for the intervals and 20,000 paired group-wise permutations for the p-values, retaining each benchmark's aggregation.

The table below summarizes the full benchmarks. The added SaProt ablation also shows a gain from the full harness beyond adaptive fusion on full ProteinGym (0.01257, p < 0.0001; Appendix Table 29).

Component comparisonProteinGym (217)VenusMutHub (905)ViroHub (89)
Adaptive MSA vs. PFM< 0.0001 (Yes)< 0.0001 (Yes)0.00015 (Yes)
Adaptive vs. fixed MSA (α=0.8\alpha=0.8)< 0.0001 (Yes)0.7676 (No)0.8932 (No)
Background correction after MSA< 0.0001 (Yes)0.1291 (No)< 0.0001 (Yes)
Gated vs. ungated correction0.0003 (Yes)0.363 (No)0.0019 (Yes)
RSA after background correction0.0008 (Yes)0.00025 (Yes)0.00155 (Yes)
pLDDT after background correction< 0.0001 (Yes)0.2359 (No)0.1368 (No)
pLDDT added after RSA0.0647 (No)0.5131 (No)0.476 (No)
RSA added after pLDDT0.2605 (No)0.00015 (Yes)0.00725 (Yes)
Full harness vs. MSA fusion< 0.0001 (Yes)< 0.0001 (Yes)0.00175 (Yes)

Each entry gives the two-sided p-value (Yes if p < 0.05; No otherwise). Score differences and 95% confidence intervals are reported in Appendix Table 27 for ProSST and Table 29 for SaProt.

W4. Lack of novelty.

The work contributes a reusable prediction framework, a viral dataset, and biological analysis alongside benchmark gains.

Thank you for raising the question of scientific value beyond benchmark scores.

  1. We view reproducible progress on shared benchmarks as a useful part of method development, as illustrated by ImageNet in computer vision and CASP in protein structure prediction.
  2. We provide a training-free framework across PFM families, curated viral protein data, and faster scoring, together with code, evaluation scripts, experiment configurations, and usage instructions.
  3. Section 3.6 examines immune-escape cases in which mutations predicted to impair protein function can still help viruses evade antibody recognition. These cases explain why antibody or serum information matters for predicting immune escape.

Questions

Q1. "Scale" is the spread of substitution scores, measured by standard deviations sRs_R and sEs_E.

Thank you for asking us to define it more clearly. We have replaced "scale-aware mixing weight" with a direct explanation in Section 2.2(iv):

L183–186: The same mixing weight gives more influence to scores that vary more. We therefore adjust ρMSA\rho_{\mathrm{MSA}} from Equation 6 for the variation in PFM and MSA scores. Mutation scores compare each replacement with the WT residue. Across all possible substitutions, the sample standard deviation of these score differences is sRs_R for the PFM and sEs_E for the MSA.

Equation (7) gives the resulting mixing coefficient (L187–190).

Q2. We replaced "mutation contrast" with the explicit difference between a replacement score and the WT score.

Thank you for pointing out the unclear wording. A structural weight between zero and one leaves the substitution score unchanged or shrinks it toward zero without reversing its sign (Section 2.4, L246–258; Appendix E.7, L2156–2159).

Q3. We meant the frozen PFM and have replaced the ambiguous use of "backbone."

Thank you for pointing out the possible confusion with protein structure. The text now uses "PFM" or "original model," and we updated the labels in Figure 5b.

Q4. Section 2.1 now explains the three scoring methods and defines teacher-forced scoring (tf).

We appreciate your suggestions for clarifying the notation. The revised definitions are:

  1. L100–101: "Masked scoring replaces each WT residue xix_i in turn with a mask token to form x∖ix_{\setminus i} and predicts amino-acid probabilities at that position."
  2. L101–102: "WT scoring predicts probabilities at every position from the full unmasked WT sequence in one pass."
  3. L103–104: "Teacher-forced scoring (tf) predicts them from preceding WT residues x<ix_{< i} in the decoding order."

Equation (1) (L104–107) defines how these probabilities give the log-probability scores.

Q5. We replaced "native field" with "PFM scores" and now describe the before-and-after comparison directly.

We appreciate your request for clearer descriptions. Section 2.1 (L098–099) now defines the PFM score matrix. Section 3.1 (L300–303) compares predictions before and after applying the harness to the same configuration. Appendix E.4 (L1887–1889) explains that ProtSSN returns logits and long ProSST sequences use log probabilities averaged over overlapping windows.

Q6. We replaced "query-mapped sites" with a direct description of how the MSA aligns to the WT sequence.

Thank you for pointing out the ambiguity. Yes, the query is the WT sequence. A "query-mapped site" is an MSA position aligned to a WT residue. Section 2.1 (L136–140) now identifies the WT sequence as the alignment reference. Appendix E.5 (L2000–2003) defines the positions used when MSA coverage is partial.

Q7. The tokenizer maps amino-acid symbols to columns of the PFM score matrix.

Thank you for raising this question. The MSA calculation itself is frequency counting. We now describe this step directly using amino-acid counts and frequencies in Section 2.1 (L136–142), with vocabulary and gap handling in Appendix E.4.

Q8. We separated MSA frequency counting from substitution scoring and removed the compressed sentence.

Thank you for the careful reading. We count each aligned sequence once, divide amino-acid counts by the number of sequences, and apply log-softmax (L136–142, Section 2.1, Equation (2)). "WT-relative contrast" meant the replacement amino-acid score minus the WT score at the same position. We now state this directly when introducing mutation scoring in Section 2.4 (L253–258).

Q9. α\alpha controls the balance between PFM scores RR and MSA scores EE.

We appreciate your request to clarify what is being mixed. Equation (3) (L147–150) states

F=(1−α)R+αE.F=(1-\alpha)R+\alpha E.

A larger α\alpha gives more weight to the MSA scores. Section 2.2 introduces this relationship before explaining how we calculate α\alpha for each protein and PFM.

Q10. We removed the unclear sentence and now explain the MSA weight calculation step by step.

Thank you for highlighting how difficult the sentence was to follow. We intended to describe two steps:

  1. We first set the MSA weight from WT-residue probabilities and the correlation between PFM and MSA scores, using prediction entropy to combine the two measures.

  2. We then adjust for the different standard deviations of their substitution scores to obtain α\alpha.

Section 2.2 now introduces the weighted average and explains these steps in Equations (3)–(7) (L147–193).

Q11. Clipping keeps the WT-probability term ρWT\rho_{\mathrm{WT}} in the interval [0, 1].

Thank you for noticing that zz was undefined. It denotes the input to

clip⁡(z,0,1)=min⁡1,max⁡0,z.\operatorname{clip}(z,0,1)=\min\\{1,\max\\{0,z\\}\\}.

Very low WT probabilities can make the mean negative log⁡20\log_{20} probability exceed one, so clipping caps the resulting weight. Section 2.2 (L156–162) and Appendix E.5 (L2007–2010) give the definition.

Q12. We compare the PFM and MSA vectors of 20 amino-acid scores at each position.

We appreciate your request for a definition and motivation. We calculate their Pearson correlation, then average across positions. Lower correlation suggests that the MSA may supply amino-acid preferences not reflected in the PFM scores. Section 2.2(ii) and Equation (5) (L164–171) explain this directly.

Q13. We removed the unclear sentence and explain the background correction directly.

Thank you for the feedback. Section 2.3 (L196–215) now describes how we estimate amino-acid preferences repeated across PFM positions and determine how much to subtract.

Q14. We now define RSA and pLDDT before using them and identify the structure source.

Thank you for pointing out these omissions. We should have introduced the abbreviations before using them.

  1. Definitions and inputs (L241–245, Section 2.4). The revised text states:

    After fusion and background correction, we use two measurements from the WT structure predicted by AlphaFold2 (Jumper et al., 2021): (i) pLDDT (predicted local distance difference test score) estimates confidence in the local structure. Low pLDDT indicates less reliable structural information (Zhang et al., 2024). (ii) RSA (relative solvent accessibility) measures residue exposure to solvent. Exposed residues often tolerate substitutions better than buried residues (Tsishyn et al., 2025).

  2. Other abbreviations. We also spell out SARS-CoV-2, RBD, ZIKV, and E in the immune-escape analysis (L428–431, Section 3.6).

  3. Variants with many substitutions. The scores use the WT sequence and structure and sum the individual substitution scores (L253–258). They therefore cannot fully capture mutation-induced structural changes or interactions. Figure 3 reports performance by mutation count, including five or more substitutions.

Q15. We now explicitly identify VenusREM2 as a contribution of this paper.

We are grateful that you flagged this omission. VenusREM2 combines six ProSST models with the harness and is now listed as the third contribution in the Introduction (L081–085). Section 2.5 (L263–287) explains the ensemble, and Section 3.3 reports its results.

Q16. We added an introduction to ProSST, our reason for choosing it, and the meaning of its vocabulary sizes.

We appreciate your request for a fuller introduction.

  1. ProSST [1] combines amino-acid tokens with quantized local-structure tokens through disentangled attention.
  2. We chose it for its mutation prediction performance and to build directly on VenusREM.
  3. The vocabulary sizes (20, 128, 512, 1024, 2048, and 4096) specify the number of structure-token categories.

The Introduction now introduces ProSST alongside VenusREM:

L053–055: Similarly, VenusREM combines MSA amino-acid frequencies with scores from ProSST, a sequence–structure Transformer with disentangled attention (Li et al., 2024a; Tan et al., 2025a).

Section 2.5 (L263–267) lists the six structure vocabularies, with inference details in Appendix E.4.

[1] Mingchen Li et al. ProSST: Protein Language Modeling with Quantized Structure and Disentangled Attention. NeurIPS 2024.

Q17. We now explicitly present ViroHub as a dataset contributed by this work.

Thank you for pointing out that its role was unclear. It contains 616,690 single-amino-acid substitutions from 89 assays, covering 18 virus types, 52 WT sequences, and seven phenotypes. The Introduction lists it as the second contribution, and Section 3.2 (L306–319) describes it. Appendix A.3 documents the sources and processing.

Q18. This analysis compares site-specific MSA scores with protein-average PFM preferences and log UniProt amino-acid frequencies.

We appreciate your careful reading. "WT readout" meant PFM scores from the unmasked WT sequence. A separated site has Pearson r<0.3r<0.3 between its MSA profile and at least one of two references: PFM scores averaged across positions in the same protein or log UniProt amino-acid frequencies. This differs from the site-specific PFM–MSA correlation used to calculate α\alpha. Section 3.4 and Appendix C.1 define the analysis.

Q19. We now use "adaptive MSA fusion" consistently and refer to its definition in Section 2.2.

Thank you for flagging the inconsistent terminology. This is the weighted average of PFM and MSA scores, with α\alpha calculated for each protein and PFM (L152–193). Section 3.4 evaluates its effect on prediction.

Q20. We replaced the unexplained checkpoint name with "ProteinMPNN" in the main text.

Thank you for flagging this. The checkpoint details for ProteinMPNN [2] are now explained in Appendix G.3:

L2287–2289: In the identifier v_48_020, 48 specifies the graph neighborhood size and 020 denotes the standard deviation (0.20 Å) of Gaussian noise added to backbone coordinates during training (Dauparas et al., 2022).

[2] Dauparas et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science, 2022.

Q21. We removed the ViroBench citation and clarified how ViroHub expands on ViroGym.

We appreciate your question about these benchmarks.

  1. ViroBench [3] concerns nucleic acids. We removed its citation from the paper to avoid confusion.
  2. Compared with ViroGym [4], ViroHub provides broader viral protein mutation coverage: 89 assays across 18 virus types, including 21 immune-escape and 18 cell-entry assays, versus 14 and 11 in ViroGym. Section 3.2 (L309–312) states this comparison, and Appendix A.3 gives the sources and processing details.

[3] Dongxin Ye et al. ViroBench: Benchmarking Nucleotide Foundation Models on Viral Genomics Tasks. KDD 2026. [4] Yichen Zhou et al. ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins. arXiv, 2026.

Reviewer_tQ79OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents a training-free framework REM-HARNESS (RH) for improving zero-shot protein fitness prediction. RH enhances frozen protein foundation models through three components, which are adaptive MSA fusion, coherence-gated calibration, and structural attenuation based on RSA and pLDDT. The authors evaluate RH on 71 model configurations across ProteinGym, VenusMutHub, and a newly curated viral benchmark ViroHub. They report consistent improvements across different model architectures and evaluation metrics. The authors also build VenusREM2 by applying RH to six ProSST variants and combining their predictions. VenusREM2 achieves an Average Spearman correlation of 0.556 on ProteinGym, outperforming the strongest baseline in their comparison by 0.038.

Critical strengths
  1. Instead of using fixed MSA mixing weights like Tranception and VenusREM, RH computes protein-specific weights on the fly to combine frozen backbone predictions with family-specific evolutionary information. Model uncertainty determines the relative importance of WT support and model–MSA disagreement when estimating the MSA contribution. And the method also considers differences in score scales to balance the two sources. This allows protein-specific evidence to be incorporated without retraining the backbone.

  2. The paper evaluates RH on 71 model configurations across three benchmarks and four backbone families. And each Raw–RH pair uses the same frozen checkpoint and scoring interface. All 71 configurations show improvements in Spearman correlation on all three benchmarks (Figure 4), which supports the general applicability of RH across different model architectures.

Critical weaknesses
  1. The comparison in Figure 3 mainly follows the public ProteinGym leaderboard, but does not include the results reported by Truong and Bepler (NeurIPS 2025). Their ensemble of PoET-2 and VenusREM achieves a reported Spearman correlation of 0.543 on ProteinGym, close to VenusREM2's 0.556, with higher scores in some strata such as organismal fitness and virus. As the authors claim that VenusREM2 ranks first across all 17 strata, clarifying whether the evaluation protocols are directly comparable would help establish the scope of the reported SOTA results.

  2. The advantage of adaptive MSA weighting over fixed mixing is less clear. Table 1 shows only small differences between entropy-weighted fusion and a fixed coefficient of α=0.8\alpha=0.8 at the retrieval stage. Since Section 2.2 argues that fixed weights can lead to different effective MSA contributions across backbones, a matched comparison across backbone families would help establish its contribution.

Questions for authors
  1. Could the authors clarify whether the ProteinGym results reported by Truong and Bepler (NeurIPS 2025) are directly comparable with those in Figure 3? If so, how would including their ensemble affect the rankings across the 17 strata?

  2. Does adaptive MSA weighting consistently outperform a fixed mixing coefficient across different backbone families?

4 comments

We appreciate the reviewer's recognition of the framework and its general applicability. Thank you for identifying the missing comparison and requesting a broader ablation.

Weaknesses

W1. Comparison with PoET-2.

We have added the published results and clarified the scope of our ranking claim.

Thank you for drawing our attention to this comparison.

  1. Comparison scope. Figure 3 originally used the official ProteinGym leaderboard baselines. Appendix Table 63 now includes PoET-2 and PoET-2 + VenusREM from Truong and Bepler (NeurIPS 2025).

  2. Ensemble architecture. PoET-2 + VenusREM combines different model architectures, while VenusREM2 combines six checkpoints from the ProSST family.

  3. Evaluation and rankings. Both studies evaluate ProteinGym substitutions using Spearman correlation. Table 63 compares published scores rather than results recomputed with an identical evaluation pipeline. Including these results, VenusREM2 leads overall (0.556 vs. 0.543) and in 15 of 17 strata, ranking second in organismal fitness and virus. We have limited the claim of leading all 17 strata to the official baselines in Figure 3. The detailed comparison is provided under Q1 below.

W2. Ablation study.

We have added a matched SaProt ablation to test adaptive MSA weighting beyond ProSST.

Thank you for pointing out the need for a broader comparison.

  1. Experimental setup. We selected ProSST for its strong mutation prediction performance and to allow a direct comparison with VenusREM, which also uses ProSST. We have now repeated the component ablation with SaProt-650M-AF2 and masked scoring on the same ProteinGym validation and full sets.

  2. Adaptive versus fixed MSA weighting. On full ProteinGym, adaptive fusion improves over fixed α=0.8\alpha=0.8 by 0.00407 for ProSST (p < 0.0001) and 0.00318 for SaProt (p = 0.1622). The gain is statistically significant for ProSST, but not for SaProt (Appendix Tables 27–29).

  3. Full harness. On SaProt, the full harness reaches 0.4868, compared with 0.4575 for the original model. Background correction and structural weighting together improve on adaptive MSA fusion by 0.01257 (p < 0.0001). Appendix Tables 28–29 report the ablation scores and paired statistical comparisons.

ComponentSettingPG validation (21): ScorePG validation: ΔRawPG full (217): ScorePG full: ΔRaw
RawPFM scores0.4953—0.4575—
MSA fusionFixed MSA weight (α=0.8\alpha=0.8)0.5150+0.0200.4710+0.014
MSA fusionAdaptive MSA weight0.5185+0.0230.4742+0.017
Model correctionWithout gate0.5203+0.0250.4741+0.017
Model correctionWith gate0.5240+0.0290.4778+0.020
Structural weightingRSA only (M+C+R)0.5347+0.0390.4862+0.029
Structural weightingpLDDT only (M+C+P)0.5289+0.0340.4798+0.022
REM-HarnessFull method (M+C+R+P)0.5369+0.0420.4868+0.029

PG validation reports mean assay Spearman; PG full reports official Average Spearman. ΔRaw is the change from the original model. M: adaptive MSA fusion; C: gated background correction; R: RSA weighting; P: pLDDT weighting.

Questions

Q1. The task and metric match, but Table 63 compares reported scores rather than results recomputed with an identical evaluation pipeline.

Thank you for asking us to clarify comparability.

  1. Evaluation comparability. Both studies evaluate ProteinGym substitutions using Spearman correlation. Appendix J now states:

    L3625–3628: We retain each source’s reported scores. Shared baselines differ slightly: VenusREM has overall Spearman 0.519 in their paper and 0.518 in the official summary used for Figure 3. Thus, the task and metric match, but the table compares published values rather than results recomputed with an identical evaluation pipeline.

  2. Updated rankings. Including their ensemble leaves VenusREM2 first overall (0.556 vs. 0.543) and in 15 of 17 strata, with second place in organismal fitness and virus. The original claim of leading all 17 strata now explicitly refers to the official leaderboard baselines in Figure 3. Appendix Table 63 (L3631–3659) gives the following comparison, with the ranking summary at L3661–3665.

CategoryVenusREM2PoET-2PoET-2 + VenusREMΔVenusREM2 rank
Overall0.5560.5000.543+0.0131
Function: Activity0.5410.5080.538+0.0031
Function: Binding0.4950.4230.475+0.0201
Function: Expression0.5570.5030.552+0.0051
Function: Organismal fitness0.4940.4820.505-0.0112
Function: Stability0.6910.5820.644+0.0471
Taxon: Human0.5620.5060.548+0.0141
Taxon: Other eukaryotes0.6150.5690.604+0.0111
Taxon: Prokaryotes0.6010.5070.562+0.0391
Taxon: Virus0.5400.5280.551-0.0112
MSA depth: Low0.5530.4880.528+0.0251
MSA depth: Medium0.5570.5070.550+0.0071
MSA depth: High0.6230.5550.593+0.0301
Mutation depth: 10.5760.5060.556+0.0201
Mutation depth: 20.4730.3570.402+0.0711
Mutation depth: 30.4680.4440.442+0.0261
Mutation depth: 40.4240.4190.411+0.0131
Mutation depth: 5+0.4730.4470.441+0.0321

Δ is the VenusREM2 score minus the PoET-2 + VenusREM score.

Ranks include the Figure 3 baselines and the two added models. PoET-2 results follow the appendix tables in [1].

[1] Truong Jr, T. F., and Bepler, T. Understanding protein function with a multimodal retrieval-augmented foundation model. NeurIPS 2025.

Q2. On full ProteinGym, adaptive fusion scores higher on both tested PFMs, with statistical significance only for ProSST.

We appreciate the request to distinguish this component's effect from the full harness.

  1. ProSST coefficient sweep. For ProSST-2048, we tested fixed MSA weights from 0 to 1 in increments of 0.1 on the 21-assay validation set. Adaptive fusion matches or exceeds every tested fixed weight, and averaging over all ten folds preserves this pattern (Appendix Figure 12).

  2. Comparison on two PFMs. On full ProteinGym, adaptive fusion improves over fixed α=0.8\alpha=0.8 by 0.00407 for ProSST (p < 0.0001) and 0.00318 for SaProt (p = 0.1622). The improvement is statistically significant for ProSST, but not for SaProt (Appendix Tables 27–29).

Thanks for the additional experiments and clarifications.

The comparison with PoET-2 + VenusREM supports VenusREM2's higher reported overall score (0.556 vs. 0.543).

The authors have also clarified that VenusREM2 ranks first in 15 of the 17 strata when this comparison is included, and have limited the original claim of leading all 17 strata to the official leaderboard baselines. The new SaProt ablation provides further evidence for adaptive MSA weighting, although its improvement over fixed weighting is not statistically significant (p = 0.1622).

Considering the newly added experiments and clarifications above, I will keep my original score unchanged.

We sincerely thank Reviewer tQ79 for reviewing our rebuttal and acknowledging the newly added experiments and clarifications.

We deeply appreciate your constructive suggestions, particularly regarding the comparison with the PoET-2 ensemble and the broader SaProt ablation study. Your feedback has been invaluable in helping us refine the scope of our claims and significantly strengthen the manuscript.

Thank you again for your time, effort, and support throughout the review process.

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

The paper proposes Rem-Harness, a wrapper around a frozen PLM which uses MSAs and structural features to improve fitness prediction. Rem-Harness has no trainable parameters (it is a hardcoded recipe). Applying Rem-Harness to an ensemble of 6 ProSST models with different vocab sizes achieves SoTA results on multiple benchmarks, for instance a 0.556 average spearman on ProteinGym, 0.038 above prior best. This ensemble is called VenusREM2.

Critical strengths
  • Strong SoTA performance on multiple benchmarks.
  • Rem-Harness has no trainable parameters, so it can be readily applied to any PLM.
Critical weaknesses
  • The paper is plagued with contrived phrases which make it hard to follow. This submission it not suitable for publication with its current writing style. The description of the Rem-Harness itself reads like a recipe, with little motivation as to why things are defined the way they are defined. The method reads like a hacky, ad-hoc procedure that achieves SoTA results but is otherwise uninteresting.
  • There is not much discussion of how this work relates to prior work. For example, it is unclear from the writing itself what the contributions over VenusREM are; the Related Work section is remarkably short.
Questions for authors

Please clarify the differences between VenusREM and this work. It is not clear in the writing what the main contributions of this work are; the Related Work section is exceptionally short. In addition, the paper is plagued with contrived phrasing at many places which alone render it unsuitable for publication. For example:

  • "Runtime analyses further show 20.7× faster mutation lookup and 62.3× faster MSA construction through native-field reuse." => what does "native-field reuse" mean. Please explain.
  • "Other native fields retain interface-specific scores" => What does this phrase mean?
  • "Let I contain n query-mapped sites with standard wild-type residues" => Define I more clearly/formally. It is a set of what exactly?
  • "Equal row weights and no pseudocounts define aligned frequencies and a log-softmax family field; either field X uses wild-type-relative contrasts" => long, dense, unreadable phrasing.
  • Eq. (3): The use of X here was a bit confusing. State in the equation itself (not in the previous long, dense, unreadable phrasing), that X is in {f, E}. Or, just use two equations instead, one for f, one for E.
  • "The site-wise log-normalizer cancels in contrasts" => Please explain this phrase, it is vague and cryptic.
  • "we distinguish a diagnostic-based target share from its scale-adjusted coefficient" => Please explain this phrase, it is vague and cryptic.
  • Eq. (6b): "corr_{a \in A}(R_i, E_i)" => Did you mean "corr_{a \in A}(R_{i,a}, E_{i,a})"?
  • "We set the target family share to" => what does this mean? "family share" of what?
  • "Low-entropy fields emphasize wild-type support; high-entropy fields emphasize model–family disagreement." => Cryptic phrase, please clarify.
  • "its conversion to a mixing coefficient follows from the scale constraint below" => Cryptic phrase, please clarify.
  • "Requiring the family channel to supply a fraction ρMSA of the summed weighted scales gives" => long convolved phrasing. What does this mean?
  • "Calibration adjusts the shared native background according to its coherence with site-level preferences, preserving family evidence." => what does this mean?
  • Eq. (11) => R_{i,a}?
  • "calibration removes the protein-wide direction only from the native channel" => what is the "protein-wide direction"? Non-standard, cryptic phrasing.
  • "Only above-mean exposure or uncertainty attenuates a site, and neither factor can reverse a contrast" => What does it mean to "reverse a contrast". Please explain this phrase. Provide a formalization and proof if this is relevant, instead of being vague.
  • "The complete readout is" => "readout"? Why not just call this "score"? Please use standard phrasing.
  • "Spearman is primary; NDCG, AUC, MCC, and top-variant recall assess complementary outcomes." => Do you simply mean that spearman is the main metric? This reads like a riddle.
6 comments

We sincerely thank the reviewer for the accurate summary and for recognizing the predictive performance and training-free framework. We address each concern below.

Weaknesses

W1. Poor writing.

We revised the writing and made the motivation for each component explicit.

Thank you for pointing this out. We reorganized the Introduction and rewrote Section 2 to explain why we add family information, correct shared PFM preferences, and use structural confidence and exposure. Sections 3.4–3.5 and Table 2 examine these choices through comparisons and ablations. Major revisions are marked in red.

L196–198, Section 2.3: MSA fusion adds family information but retains shared PFM preferences, which can reflect training-data composition (Figure 1; Gordon et al., 2025). We subtract this background from the PFM contribution to emphasize preferences at individual sites.

W2. Limited related work.

We clarified the additions to VenusREM and expanded the Related Work.

We appreciate this suggestion. Sections 2.2–2.4 now explain that REM-Harness replaces VenusREM's fixed MSA weight with one calculated for each protein and PFM, and adds gated background correction and structural weighting.

L152, Section 2.2: VenusREM fixes α=0.8\alpha=0.8 using validation data. We instead calculate α\alpha for each protein and PFM.

The Introduction now explicitly lists REM-Harness, ViroHub, and VenusREM2 as our three contributions (L073–085). We also expanded Section 4 with material from the appendix and additional references, covering PFMs, post-hoc calibration and post-training, and mutation fitness benchmarks (L505–528).

Questions

Q1. We replaced "native-field reuse" with an explicit description of score caching and array operations.

Thank you for pointing out the unclear wording. We reuse each protein's PFM score matrix to look up and sum substitution scores. The MSA speedup concerns counting amino acids in an existing alignment. Section 3.8 now states:

L494–496: After ProSST inference, cached mutation lookup is 20.7× faster and MSA amino-acid counting with array operations is 62.3× faster than the corresponding loops (Appendix Figure 14; Appendix G.1).

Q2. We removed "native fields" and now define the PFM score matrix directly.

Thank you for flagging this unclear phrase. The revised text states:

L098–099, Section 2.1: We store PFM log probabilities or logits in R∈RL×20R\in\mathbb{R}^{L\times20}, with row Ri,⋅R_{i,\cdot} containing the 20 scores at position ii.

We also clarified that ProtSSN uses logits, while long ProSST sequences use log probabilities averaged across overlapping windows (Appendix E.4, L1887–1889).

Q3. II is the set of WT positions covered by the MSA and containing standard amino acids.

We appreciate your request for a precise definition. Its elements are position indices, and nn is their number. The revised definition is:

L2000–2003, Appendix E.5: Section 2.2 assumes an MSA covering the full WT sequence. For partial coverage, let A\mathcal A be the 20 standard amino acids and I\mathcal I the covered positions with xi∈Ax_i\in\mathcal A, with n=∣I∣n=|\mathcal I|. Equations 4–6 then average over these nn positions.

Q4. We split the dense sentence into the MSA calculation and the later substitution-scoring step.

Thank you for flagging this dense sentence. Section 2.1 now explains the steps in order: count amino acids at each aligned position, divide the counts by the number of aligned sequences, and apply log-softmax (L136–142, Equation (2)). Subtracting the WT score is introduced later with mutation scoring in Section 2.4, Equation (12).

Q5. We removed the generic XX notation and now state the WT-score subtraction explicitly.

Thank you for suggesting clearer notation. Section 2.4 introduces this step using the fused and background-corrected scores GG (Equation (12), L253–258):

Using GG from Equation 10, we subtract the WT score from the replacement score.

Q6. Log-softmax subtracts a shared constant, which cancels when we subtract the WT score.

Thank you for asking us to explain this step. At each position, log-softmax subtracts the same normalization term from every amino-acid frequency. Therefore,

Ei,a−Ei,xi=fi,a−fi,xi.E_{i,a}-E_{i,x_i}=f_{i,a}-f_{i,x_i}.

Appendix E.4 (L1883–1885) now gives this explanation.

Q7. We replaced the abstract wording with the two steps used to calculate the MSA coefficient.

Thank you for pointing out the unclear wording. Section 2.2 now explains two steps:

  1. Calculate the MSA weight ρMSA\rho_{\mathrm{MSA}} from WT probabilities, entropy, and PFM–MSA score correlation (L156–182).
  2. Adjust this weight for the variation in PFM and MSA scores to obtain α\alpha (L183–193), so that greater numerical variation alone does not give one source more influence.

Q8. The correlation is between two vectors of 20 amino-acid scores at each position.

Thank you for pointing out the unclear notation. Section 2.1 (L097–099) defines the row-vector notation. The correlation term in revised Equation (5) (L169–171) is:

rˉR,E=1L∑i=1Lcorr⁡(Ri,⋅,Ei,⋅).\bar r_{R,E}=\frac{1}{L}\sum_{i=1}^{L}\operatorname{corr}(R_{i,\cdot},E_{i,\cdot}).

Thus, each correlation uses two rows, and we average across positions.

Q9. We replaced "family share" with "MSA weight" and clarified its meaning.

Thank you for pointing out this undefined term. It means the weight assigned to MSA substitution scores when combining them with PFM substitution scores. These weights apply after each set of substitution scores is divided by its own standard deviation (Appendix E.5, L2025–2029). Equation (7) gives the corresponding α\alpha for the original scores.

Q10. Entropy distinguishes concentrated predictions from probabilities spread across many amino acids.

We appreciate your request to explain its role. The revised text first distinguishes the two cases:

L172–174, Section 2.2: A low WT probability can mean that the PFM strongly favors other residues or assigns similar probabilities to many residues. Entropy helps distinguish these cases by measuring how spread out all 20 probabilities are.

  1. At low entropy, the PFM favors a few amino acids, so a low WT probability is informative and we give more weight to ρWT\rho_{\mathrm{WT}}.
  2. At high entropy, the probabilities are more similar and a low WT probability is less informative. We then give more weight to ρcorr\rho_{\mathrm{corr}}, which compares the PFM and MSA scores.

Section 2.2, Equation (6), defines entropy and how it weights these two terms.

Q11. We divide each weight by its score standard deviation, then normalize the two weights.

Thank you for highlighting this missing explanation. Let sRs_R and sEs_E be the standard deviations of the differences between replacement and WT scores for the PFM and MSA, respectively. We divide each weight by its corresponding standard deviation, then normalize the two weights to sum to one. The resulting MSA coefficient is

α=ρMSA/sE(1−ρMSA)/sR+ρMSA/sE.\alpha=\frac{\rho_{\mathrm{MSA}}/s_E}{(1-\rho_{\mathrm{MSA}})/s_R+\rho_{\mathrm{MSA}}/s_E}.

Section 2.2(iv), Equation (7) (L185–190), now explains this calculation.

Q12. We replaced the compressed wording with an explicit definition of the proportion.

Thank you for highlighting the unclear wording. The weighted PFM and MSA substitution scores have standard deviations (1−α)sR(1-\alpha)s_R and αsE\alpha s_E. The condition makes the MSA standard deviation a fraction ρMSA\rho_{\mathrm{MSA}} of their sum. Section 2.2(iv) now explains the adjustment directly (L183–193), with the precise condition and derivation in Appendix E.5 (L2025–2044).

Q13. We subtract preferences repeated across positions, with stronger correction when the position scores resemble the background.

Thank you for your close reading. We summarize the PFM's average amino-acid preferences in a vector of 20 values (L200–205). Section 2.3 then explains the gate:

L206–208: We apply stronger correction when positions repeatedly favor the same amino acids as the background. We measure this using the mean Pearson correlation cc between each PFM row Ri,⋅R_{i,\cdot} and b\mathbf b from Equation 8.

The gate is κ(R)=max⁡(0,c)\kappa(R)=\max(0,c) (Equation (9), L209–211). Equation (10) applies this correction to the PFM contribution while keeping the MSA term unchanged (L212–215).

Q14. Revised Equation (8) explicitly uses Ri,aR_{i,a} for an individual PFM score.

Thank you for catching the missing index. Ri,aR_{i,a} is the score for amino acid aa at position ii. Section 2.1 (L097–099) defines the indices and row-vector notation. Equation (8) is at L203–205.

Q15. We removed "protein-wide direction" and describe the subtracted amino-acid preferences directly.

We appreciate your suggestion to use standard terminology. Equation (10) now shows how we subtract preferences shared across positions from the PFM contribution (L212–215, Section 2.3; Appendix E.6).

Q16. Structural weights shrink each substitution score toward zero without reversing its sign.

Thank you for asking for a precise statement and proof. Section 2.4 explains the intended effect:

L246–248: We represent greater mutational tolerance by smaller predicted changes from WT. We also reduce effect estimates when structural confidence is low. Both adjustments shrink scores at sites with above-average exposure or uncertainty.

The structural weight λi\lambda_i is a product of two factors in [0,1][0,1] (Equation (11)). For one substitution score ss,

\quad\Longrightarrow\quad \lvert\lambda_i s\rvert\le\lvert s\rvert.$$ The score $\lambda_i s$ has the same sign as $s$ or is zero. Appendix E.7 (L2156–2159) gives the proof. Q17. **We now call this the "variant score."** Thank you for the suggestion. We replaced "readout" with this term and define the score in Section 2.4, Equation (12) (L253–258). Q18. **Spearman correlation is our main metric.** Thank you for pointing this out. The revised text states this directly: > **L294–297, Section 3.1:** Spearman correlation is the primary metric. Normalized discounted cumulative gain (NDCG) and top-variant recall measure ranking quality. Area under the receiver operating characteristic curve (AUC) and Matthews correlation coefficient (MCC) measure binary classification performance.

Comments on the paper

2 comments

Global Response to All Reviewers and Area Chair

We sincerely thank reviewers HLiX, tQ79, and 7oB8 and the Area Chair for their time and constructive feedback. We appreciate the recognition of our comprehensive experiments, strong benchmark performance, and training-free design. We have revised the manuscript and added experiments and statistical analyses, with substantive changes highlighted in red. Below we summarize the main revisions and the concerns they address.

1. Clarification of Contributions

In response to HLiX and 7oB8, the Introduction now states three contributions explicitly:

  1. REM-Harness: a general, model-agnostic, training-free harness for protein foundation models (PFMs). It combines information from multiple sequence alignments (MSAs), correction of repeated amino-acid preferences in PFM scores, and weighting by structural confidence and solvent exposure. Across ProteinGym, VenusMutHub, and ViroHub (1,211 assays and 3.1 million measured variants), all 71 configurations improve Spearman correlation on each benchmark, by 0.073 on average. We provide code, data, experiment configurations, and usage instructions to support reuse.
  2. ViroHub: a community benchmark for viral protein research. We curate 616,690 substitutions from 89 assays across 18 virus types, 52 wild-type (WT) sequences, and seven phenotypes. Its coverage of viral function, cell entry, and immune escape supports evaluation relevant to virus and antibody research. Section 3.2 now presents its scope, curation, and complementarity to existing benchmarks.
  3. VenusREM2: improved prediction and efficient scoring. Applying the harness to six ProSST checkpoints gives ProteinGym Average Spearman 0.556, exceeding the strongest official leaderboard baseline in Figure 3 by 0.038. After model inference, MSA count-matrix construction and mutation scoring are 62.3× and 20.7× faster than VenusREM's corresponding loop-based implementations.

2. Major Experimental Updates

  • Statistical support for gains beyond MSA fusion (7oB8). Appendix Table 27 now reports 95% confidence intervals and paired permutation tests for nine ProSST-2048 component comparisons. On the full benchmarks, the harness improves on adaptive MSA fusion on ProteinGym and VenusMutHub (both p < 0.0001) and ViroHub (p = 0.00175). These comparisons directly test the additional benefit of background correction and structural weighting together.
  • Ablation on a second PFM (tQ79). We added component ablations for SaProt-650M-AF2 with masked scoring (Appendix Tables 28–29). On full ProteinGym, the full harness improves Spearman correlation over adaptive MSA fusion by 0.01257 (p < 0.0001), reaching 0.4868. These results provide additional evidence for the benefit of background correction and structural weighting.
  • Comparison with published PoET-2 results (tQ79). Appendix Table 63 adds PoET-2 and the PoET-2 + VenusREM ensemble reported by Truong and Bepler (NeurIPS 2025). VenusREM2 remains first overall (0.556 vs. 0.543 for their ensemble) and in 15 of 17 strata, placing second in organismal fitness and virus. We have clarified that the original claim of leading all 17 strata refers to the official leaderboard baselines in Figure 3.

3. Writing and Method Clarifications

  • Method motivation and notation (HLiX, 7oB8). We apologize that the original presentation made the method difficult to follow. Section 2 now explains the purpose of each component before introducing its calculation. We define the PFM and MSA score matrices, explain the fusion weight step by step, and clarify the background correction and structural weights. Terms such as "native field" and "readout" have been replaced with "PFM scores" and "mutation scores."
  • Relationship to prior work (HLiX). Section 2.2 explains the change from VenusREM's fixed α=0.8\alpha=0.8 to weights computed for each protein and PFM. Sections 2.3–2.4 describe the added background correction and structural weighting. We expanded Section 4 with references covering PFMs, post-hoc calibration and post-training, and mutation fitness benchmarks.
  • Model, dataset, and biological context (7oB8). Section 2.5 introduces ProSST and explains the six-checkpoint ensemble. Section 3.2 and Table 1 make ViroHub's contribution explicit. Section 3.6 discusses the immune-escape cases, explaining why mutations predicted to impair protein fitness can still escape antibody recognition.

Thank you again for the careful reading and specific suggestions. We hope the revised manuscript and point-by-point responses make the method, evidence, and scope of our contributions clearer.

AuthorsOpenReview ↗

Final Remarks

We sincerely thank the reviewers and Area Chair for their valuable time, thoughtful feedback, and consideration.

We have carefully addressed the concerns raised through clearer explanations, improved presentation, and additional analyses. The core scientific contributions, methodology, and main findings remain unchanged. We hope these revisions and our point-by-point responses will be considered in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.