acceptodds
Under review as a conference paper at ICLR 2027

Multi-Objective Protein Optimization via Modular Diffusion Steering

Abstract

Protein engineering often requires improving several properties simultaneously, motivating the search for high-performing sequences that represent useful fitness tradeoffs. Discovering these sequences is challenging because the design space is combinatorially large and experimental campaigns often permit only a few rounds of batched measurements. We introduce *MODS*, a modular framework for multi-objective protein optimization via inference-time diffusion steering. For each proposal, MODS samples augmented Chebyshev scalarization weights and a fitness-surrogate ensemble member to define a scalar steering reward. Varying the weights directs generation toward different objective tradeoffs, while varying ensemble members diversifies the predictions used for guidance. Because MODS expresses multi-objective acquisition through scalar rewards, it can be paired with different diffusion models and compatible steering procedures. After each evaluation batch, MODS updates the surrogates while keeping the generative model fixed. We evaluate MODS on six protein optimization tasks using predictive oracles trained on experimental assay data, with two diffusion models and two steering procedures. On every task, at least one MODS configuration substantially outperforms all evaluated baselines in final hypervolume, while the other configurations remain competitive. Our results demonstrate a flexible and effective approach to multi-objective protein optimization with limited feedback.

Primary area:
generative models
Keywords:
multi-objective optimization, protein design, guided diffusion, pareto optimization, generative models

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_1sVLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MODS combines randomized augmented Chebyshev scalarization with surrogate-ensemble sampling to construct scalar rewards for inference-time steering of a fixed discrete diffusion model. The surrogate ensemble is retrained each round using accumulated observations, while the generative model is kept fixed. The experiments cover six assay-derived oracle benchmarks and four model–steering combinations, each evaluated at four steering strengths. On every task, at least one MODS DAPS configuration achieves the highest final mean hypervolume among the evaluated methods.

Critical strengths
  1. The framework is clearly formulated. Sections 4.1–4.2 separate the choice of objective tradeoff and surrogate prediction from the diffusion steering procedure through a scalar reward. The experiments demonstrate this flexibility with two diffusion models and two steering procedures.

  2. The DAPS results are strong on the reported benchmarks. Table 1 and Figure 3 show that at least one MODS DAPS configuration achieves the highest final mean hypervolume on each task. Comparisons with the corresponding unsteered priors support the benefit of steering, and Table F.2 shows gains at multiple steering-strength settings, although performance depends on the model and task.

Critical weaknesses
  1. The comparison with recent inference-time multi-property guidance methods is incomplete. LaMBO-2 is included and provides a relevant multi-objective diffusion baseline. However, Section 2 also cites more recent approaches, including ProteinGuide and PepTune, that are not evaluated, and the paper does not explain whether they can be adapted to the proposed benchmark setting. Since MODS emphasizes separating multi-objective acquisition from the generative model and steering procedure, a comparison with a compatible method that also keeps the pretrained generator fixed would be particularly informative. Without such a comparison, or an explanation of the relevant incompatibilities, it is difficult to assess the practical benefit of the proposed acquisition construction relative to these alternatives.

  2. The experiments do not isolate the contribution of surrogate-member sampling. Section 4.1 states that sampling kk selects a plausible fitness landscape and that varying kk promotes exploration under predictive uncertainty. However, Table 1, Figure 3, and Table F.2 evaluate only complete MODS configurations against external baselines and unsteered priors. These comparisons do not establish whether sampling ensemble members improves optimization beyond scalarized reward steering alone. Replacing the sampled member with the ensemble mean or a fixed member, while keeping the random Chebyshev weights and steering settings unchanged, would directly test this component.

Questions for authors
  1. Can the authors ablate the ensemble-member sampling described in Section 4.1 by replacing the sampled member with the ensemble mean or a fixed surrogate member, while retaining the random Chebyshev weights? Separately, can they compare augmented Chebyshev scalarization with linear scalarization? Sections 2 and 4.1 motivate the Chebyshev form by its ability to recover tradeoffs on nonconvex Pareto fronts that weighted sums can miss, but no comparison tests whether this advantage matters on the evaluated tasks. The steering-strength sweeps in Figures E.1 and E.2 do not address either question, since they vary steering strength while leaving both acquisition choices unchanged. These ablations could use the same surrogate architecture and training protocol under the same evaluation budget.

  2. Would the authors evaluate at least one compatible recent multi-property guidance method on the same oracle benchmarks and under the same evaluation budget, or explain why such a comparison is infeasible? ProteinGuide (Xiong et al., 2026) appears particularly relevant because it also guides fixed pretrained generators. For PepTune (Tang et al., 2025), please clarify whether its multi-objective guidance procedure can be adapted from peptide SMILES to the amino acid sequence setting studied here.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 1sVL

Thank you for your thoughtful review. We appreciate your assessment that “The framework is clearly formulated” and that “The DAPS results are strong on the reported benchmarks.” Below, we present additional comparisons with recent multi-property guidance methods and ablations examining scalarization and surrogate ensemble sampling. To keep this response concise and make the additional evidence easy to find, we have collected the supporting figures and tables in Appendix G of the updated manuscript. We will incorporate your feedback, these results, and the clarifications below into the revised manuscript. If these additions address your concerns, we would sincerely appreciate your consideration of a higher score.

1. Comparisons with recent multi property guidance methods

“Would the authors evaluate at least one compatible recent multi-property guidance method […] under the same evaluation budget, or explain why such a comparison is infeasible?”

Response.

Thank you for suggesting these comparisons. We added comparisons with ProteinGuide and MosPro on the five two-objective benchmarks (Table G.3, Appendix G). On each task, at least one MODS configuration achieves higher mean hypervolume than ProteinGuide and MosPro. ProteinGuide uses the same oracle-evaluation budget as MODS. For MosPro, we report both a matched-budget evaluation with eight rounds of 100 new sequences and an increased-budget evaluation with fifteen rounds of 10,000 new sequences. These comparisons extend our evaluation beyond LaMBO-2 and include ProteinGuide, which guides a fixed pretrained generator.

PepTune requires a different adaptation. Its released pipeline operates on peptide SMILES, with a tokenizer, masking schedule, and validity constraints tailored to that representation. Our benchmarks instead concern amino acid sequences of full-length proteins. Converting these sequences to SMILES would not by itself resolve the mismatch with PepTune’s pretrained generator. Its property predictors also target therapeutic peptide properties rather than the objectives in our benchmarks and would need to be replaced with task-specific surrogates. Adapting PepTune therefore requires changes to the sequence representation, generator, and property predictors. We could not complete this adaptation within the discussion period. We will clarify these differences in the revised manuscript and include the new ProteinGuide and MosPro comparisons.

2. Surrogate ensemble sampling and scalarization ablations

“Can the authors ablate the ensemble-member sampling […] by replacing the sampled member with the ensemble mean or a fixed surrogate member […]? Separately, can they compare augmented Chebyshev scalarization with linear scalarization?”

Response.

Thank you for these suggestions. We added separate ablations of surrogate ensemble sampling and scalarization for D3PM and MDLM with DAPS across all six tasks (Figures G.8–G.9, Appendix G). Results report mean hypervolume and one standard error over five repetitions, under the same oracle evaluation budget and steering settings as the corresponding MODS configurations.

For surrogate ensemble sampling, Figure G.8 compares MODS with a single surrogate, using the same architecture and training procedure as the ensemble members. Performance is generally similar, with clearer improvements from surrogate ensemble sampling for D3PM with DAPS on KRAS and Src. We believe the single-surrogate variant remains strong partly because each candidate uses a separately sampled scalarization, which already promotes diversity. Steering with the ensemble mean would require evaluating all members at every reward evaluation, increasing the computational cost of guidance. We were unable to complete this comparison with the computational resources available during the discussion period, but we will include it in the revised manuscript. The single-surrogate results help clarify the contribution of surrogate ensemble sampling alongside randomized scalarization.

MODS is compatible with other scalarizations, including linear scalarization; augmented Chebyshev is a design choice rather than a requirement of the framework. Figure G.9 compares randomized augmented Chebyshev scalarization with randomized linear scalarization. Relative performance varies across tasks and generative models, and linear scalarization is competitive. We focused on augmented Chebyshev because it can recover tradeoffs on nonconvex Pareto fronts that linear scalarization can miss. Since the geometry of the Pareto front is generally unknown in advance, we continue to recommend augmented Chebyshev as a default. We will clarify this rationale and MODS’s compatibility with alternative scalarizations in the revised manuscript.

Reviewer_1sVLOpenReview ↗

Official Comment by Reviewer 1sVL

Thank you for the additional comparisons and ablations. Could you clarify how ProteinGuide and MosPro were adapted, particularly whether their property predictors were updated each round, and what explains ProteinGuide’s unusually low performance? Also, given that the single-surrogate variant often performs similarly to full MODS, how much of the reported improvement can really be attributed to ensemble-member sampling? Based on the provided information that linear scalarization also matches or exceeds Chebyshev in several settings, what empirical advantage do these acquisition choices demonstrate over simpler randomized scalarized steering?

Follow-up response on baseline adaptation and ablations

“Could you clarify how ProteinGuide and MosPro were adapted, particularly whether their property predictors were updated each round, and what explains ProteinGuide’s unusually low performance? […] how much of the reported improvement can really be attributed to ensemble-member sampling? […] what empirical advantage do these acquisition choices demonstrate over simpler randomized scalarized steering?”

Response.

Thank you for these follow-up questions. We appreciate the opportunity to clarify the baseline protocols and what our ablations establish. We have added further comparisons in Section G.11 to address your question about updating the baseline predictors.

In our initial comparisons, we followed the fixed-predictor setups used in the methods’ published in-silico evaluations, including additional labeled data to train their predictors before optimization. We apologize for not considering this adaptation in our initial comparison. We have now evaluated iterative adaptations of both methods on GB1 and ERBIN-PDZ1. We removed the additional predictor-training data and instead trained and updated the predictors using the campaign’s initial data and subsequently collected oracle feedback, aligning their access to labeled data with MODS.

Both iterative baselines improve, but MODS retains a substantial performance advantage on both tasks (Table G.6). We include standard MODS and a faster configuration that uses fewer sampling steps under the same oracle-evaluation budget (Section G.10). On GB1, ProteinGuide improves from 47.6 to 171.0 and MosPro from 224.2 to 459.1, compared with 781.4 for standard MODS and 671.7 for faster MODS. On ERBIN-PDZ1, ProteinGuide improves from 20.6 to 191.4 and MosPro from 197.0 to 206.8, compared with 413.8 and 380.8 for standard and faster MODS, respectively. All values are mean final hypervolumes.

The ablations show that surrogate ensemble sampling provides additional gains on some tasks, while the single-surrogate variant often performs similarly. Figure G.8 shows modest benefits from surrogate ensemble sampling, particularly for D3PM with DAPS on KRAS and Src. We believe the single-surrogate variant remains strong partly because each candidate uses a separately sampled scalarization, and the generative sampling procedure is itself stochastic. Both sources of randomness promote diversity among the proposed solutions. These results also give users the option of a simpler single-surrogate configuration while retaining strong performance.

We interpret “simpler randomized scalarized steering” as using a single surrogate with randomly sampled linear scalarization weights. This alternative can be accommodated within our framework, and we expect it to perform competitively on these benchmarks. Our ablations examine these two simplifications separately; we were unable to complete their combined evaluation within the discussion period.

In light of these results, we will explicitly present both randomized linear and augmented Chebyshev scalarization as options within MODS. Linear scalarization is competitive in our experiments (Figure G.9). We favor augmented Chebyshev as a default because it can also recover tradeoffs on nonconvex Pareto fronts that linear scalarization may miss, with little additional implementation or computational cost. This flexibility is useful when the front’s geometry is unknown in advance, although it need not translate into better performance on every task.

MODS’s contribution lies in the framework for combining randomized scalarization, surrogate-guided generation, and sequential feedback, with flexibility in the choice of surrogate and scalarization. Moreover, taken together, the experiments in the paper and the additional evaluations during the discussion provide strong evidence that MODS achieves state-of-the-art performance on the protein optimization benchmarks studied here. We would sincerely appreciate the reviewer taking this expanded evidence and these clarifications into account when reconsidering their assessment and score.

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

The paper proposes a method that guides a diffusion model to generate protein sequences with several desired properties. Small predictors learn from feedback and steer generation toward promising candidates. Experiments on six tasks show improvements in predicted property scores, but the generated proteins are not experimentally validated.

Critical strengths

The method can adapt a fixed generator using new feedback. Updating small surrogates avoids retraining the diffusion model after every optimization round. This provides a practical way to incorporate newly acquired property scores.

The experiments cover several multi-objective tasks. Testing different proteins, objectives, and diffusion models provides evidence that the approach can work across several computational settings.

Critical weaknesses
  1. The biological relevance of the gains remains uncertain. The method receives feedback from CNN oracles and is evaluated using those same oracles. Higher predicted scores may reflect exploitation of prediction errors, particularly for sequences outside the measured training distribution. Without experimental validation or stronger evidence of oracle reliability on the optimized candidates, the results do not establish improved protein performance.

  2. Surrogate reliability under limited data is insufficiently examined. Optimization starts with only 100 labeled sequences and acquires 900 in total. Guidance actively selects sequences with high predicted scores, making errors on these selected candidates especially consequential. For CG, predicting clean-sequence fitness from noisy inputs adds another difficulty. The paper does not sufficiently establish whether the surrogates remain accurate under these conditions.

  3. The computational cost of guidance is insufficiently established. DAPS involves repeated clean-sequence generation and surrogate-based updates. Its practical cost depends on how these clean sequences are obtained and how many inner sampling steps are required. A comparison based only on oracle queries does not show whether the improvements justify the additional computation.

Questions for authors

How are clean sequences obtained during DAPS guidance? At each outer step, does the method run an inner reverse diffusion trajectory from the current noisy sequence, or use a conditional mean or another point estimate? If it uses a point estimate, how is applying a nonlinear surrogate to that estimate justified? If it runs an inner trajectory, please report the inner and outer sampling steps, total diffusion and surrogate evaluations, and wall-clock time.

How reliable are the surrogates on the candidates selected by guidance? Please report their predictive accuracy on newly proposed candidates before those candidates’ oracle scores are acquired, across optimization rounds. For CG, how does accuracy change with noise level? This would help determine whether the limited data support useful guidance or lead to overly optimistic predictions.

2 comments
AuthorsOpenReview ↗

Response to Reviewer RAMo (Part 1 of 2)

Thank you for your thoughtful review. We appreciate your recognition that updating small surrogates “provides a practical way to incorporate newly acquired property scores” and that our experiments provide evidence across “several computational settings.” Below, we present Rosetta results, assess surrogate predictions on new candidates, report runtimes, and explain the DAPS sampling procedure. To keep this response concise and make the additional evidence easy to find, we have collected the supporting figures and tables in Appendix G of the updated manuscript. We will incorporate your feedback, these results, and the clarifications below into the revised manuscript. If these additions address your concerns, we would sincerely appreciate your consideration of a higher score.

1. Biological relevance and oracle reliability

“Higher predicted scores may reflect exploitation of prediction errors”

Response.

Thank you for raising this concern. Our evaluation oracles and the surrogates used for steering are distinct models: the oracles use CNN architectures, whereas the surrogates use MLPs. Moreover, learned neural-network oracles are widely used to evaluate protein optimization methods. Examples already cited in our manuscript include Yang et al. (2025a) and Luo et al. (2025). Such benchmarks enable reproducible comparisons under matched evaluation budgets.

However, to complement the CNN-based benchmarks, we conducted an additional GB1 optimization campaign with the same objectives—binding and stability—replacing the CNN oracles with Rosetta throughout the campaign. Rosetta scores provide the feedback used to update the surrogates and evaluate optimization performance. MODS achieves a hypervolume exceeding 300 with a steering strength of α = 1,024, compared with approximately 150 for SAGE-Prot and unsteered sampling (Figure G.3, Appendix G). These results demonstrate strong performance with simulation-based feedback, complementing the CNN-based benchmarks.

Wet-lab validation requires resources and time beyond the scope of this methodological study. Our evaluation uses learned oracles and Rosetta, but experimental validation remains necessary to establish improvements in biological function. We hope future experimental campaigns will demonstrate that MODS makes multi-objective protein optimization more effective and easier to use. We will include the Rosetta results and clarify this scope in the revised manuscript.

2. Surrogate reliability under limited data

“Please report their predictive accuracy on newly proposed candidates […] across optimization rounds. For CG, how does accuracy change with noise level?”

Response.

Thank you for suggesting this analysis. We assessed surrogate accuracy for DAPS and classifier guidance on newly proposed candidates for NHERF3-PDZ2 and Src at optimization rounds 4 and 8, using MDLM (Figures G.4–G.7 and Table G.1). Predictions are computed using the surrogates available before incorporating each batch’s oracle scores into training.

Prediction errors on newly proposed candidates are not specific to MODS. Surrogate-based optimization uses an approximation of the objective and selects candidates to improve performance or explore, including where predictions are less accurate. Predictions remain positively correlated with oracle scores in all reported settings, and correlations generally improve from round 4 to round 8. Together with the optimization results, these findings support the usefulness of surrogate guidance.

For classifier guidance, we additionally assess predictions of clean-sequence fitness from noisy inputs across three noise-level bins. Correlations are generally lower at higher noise levels but remain positive even in the highest-noise bin. We will include these analyses and describe the evaluation protocol in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer RAMo (Part 2 of 2)

3. Computational cost

“A comparison based only on oracle queries does not show whether the improvements justify the additional computation.”

Response.

We agree that oracle-query efficiency and computational cost are complementary aspects of practical performance. Table G.2 (Appendix G) reports average total time per optimization round across eight rounds, with 100 candidates per round, on a single NVIDIA RTX A6000 (48 GB). Including model updates and candidate generation, MODS requires approximately 11.2–12.4 minutes per round across configurations—less than half the time of ProteinGuide and less than one-third that of LaMBO-2. The corresponding baseline times are approximately 1.6 minutes for ARLM + DPO, 5.9 for SAGE-Prot, 45.2 for LaMBO-2, 3.5 for MosPro, and 27.3 for ProteinGuide. The table also provides a breakdown of generation and model-update times. Timings exclude oracle evaluations and report means ± one standard deviation. Model-update time covers the training or fine-tuning required by each method; dashes indicate no model update. Total time also includes model loading, which accounts for differences from the sum of the two components.

MODS targets campaigns where experimental measurements are costly. Additional computation can therefore be worthwhile when it improves the candidates selected for evaluation. Our experiments demonstrate improved optimization under matched oracle-query budgets, while the runtime comparison makes the associated computational overhead explicit. The generative model remains fixed throughout the campaign and only the small surrogate ensemble is retrained as new observations become available.

For DAPS, we detail the inner and outer sampling steps and the total numbers of diffusion-model and surrogate evaluations in the response below.

4. Clean sequence sampling during DAPS guidance

“Does the method run an inner reverse diffusion trajectory […] or use a conditional mean or another point estimate?”

Response.

Thank you for asking us to clarify this. At each outer DAPS step, we run an inner reverse diffusion trajectory from the current noisy sequence to obtain a clean sequence. The surrogate is evaluated on discrete clean sequences, rather than on a conditional mean or a soft representation. Working with discrete samples avoids applying the nonlinear surrogate to a conditional-mean estimate. Operating in clean-sequence space is a strength of the DAPS formulation: it separates denoising from reward-guided updates, allowing the nonlinear surrogate to be evaluated directly on clean sequences. This follows the decoupled sampling strategy of Zhang et al. (2025).

We use 50 outer steps, an inner diffusion schedule with 20 time points, and 1,000 Metropolis–Hastings updates per outer step. This requires 950 diffusion-model evaluations and 50,050 surrogate evaluations per generated sequence. After the reward-guided updates, we apply forward corruption to match the next outer noise level. Runtime comparisons are addressed in our response above.

Reviewer_C8dQOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MODS, a module for steering diffusion-based protein generation models at inference time. Unlike previous methods that mainly focus on optimizing generation for a single objective and often only build for a particular generative model, MODS claims to be a module that simultaneously optimizes for multiple objectives and can be flexibly paired to different base generation models. MODS achieves this through training an ensemble of surrogates, sampling a member from them, and using the sample with scalarization weights to construct steering rewards. To validate the claimed contributions, MODS is tested across six tasks using two different base generative model.

Critical strengths

Overall, the paper is well-written and supported with evidence. The problem that the authors are investigating, steering protein generative models for multiple objectives at the same time, is interesting and meaningful. Across six different protein tasks we consistently see notable improvements over the baselines. The authors are also being honest about their limitations.

Critical weaknesses

One of the central claimed contribution of MODS is general applicability to different diffusion models. This can be partially justified by experiments on two different models, MDLM and D3PM. However, both remain discrete sequence diffusion models trained/fine-tuned on data specific to each tasks, casting doubts on whether MODS can be used to steer a broadly pretrained model without task-specific fine-tuning.

Questions for authors

Clarification requests:

  1. Is it possible to complete the six protein tasks with a broadly pretrained protein generation model or a retraining/fine-tuning is a must? If possible, would MODS work well on steering a pretrained model without task-specific training/fine-tuning for the base model?
  2. Does MODS works well in single objective tasks too? Can it still show notable improvement over existing methods?

Future work suggestions:

  1. Can MODS be applied to protein structure generation models (e.g. MultiFlow, FrameFlow) in addition to the sequence generation models discussed here?
3 comments
AuthorsOpenReview ↗

Response to Reviewer C8dQ

Thank you for your thoughtful review. We appreciate your assessment that the problem is “interesting and meaningful” and that “Across six different protein tasks we consistently see notable improvements over the baselines.” In response to your questions, we added experiments using a broadly pretrained generator without task-specific fine-tuning and a single-objective evaluation. We summarize these results below and explain how existing steering procedures could extend MODS to structure generation. To keep this response concise and make the additional evidence easy to find, we have collected the supporting figures and tables in Appendix G of the updated manuscript. We will incorporate your feedback, these results, and the clarifications below into the revised manuscript. If these additions address your concerns, we would sincerely appreciate your consideration of a higher score.

1. Applicability without task-specific generator training

“Would MODS work well on steering a pretrained model without task-specific training/fine-tuning for the base model?”

Response.

Thank you for this suggestion. Task-specific training or fine-tuning of the generator is not a requirement of MODS. We evaluated this across all six protein optimization tasks using the pretrained EvoDiff 38M model without task-specific fine-tuning. This is the checkpoint used to initialize the task-specific D3PM models in our original experiments.

Figure G.1 in Appendix G shows that MODS improves over unsteered generation with both pretrained and fine-tuned generators. MODS achieves its strongest overall performance with the fine-tuned generator, as expected from adapting the generator to the target protein family. Fine-tuning is therefore beneficial, but MODS can still guide optimization effectively without it.

2. Single-objective optimization

“Does MODS work well in single-objective tasks too? Can it still show notable improvement over existing methods?”

Response.

Thank you for suggesting this evaluation. We focused primarily on multi-objective optimization because protein design often requires balancing several properties, such as binding, stability, and abundance.

MODS supports single-objective optimization as a special case. With one objective, the scalarization weight is w₁ = 1, and augmented Chebyshev scalarization reduces to (1 + ρ) times the predicted objective. This constant factor can be absorbed into the steering strength. The resulting procedure uses a sampled surrogate ensemble member to steer the fixed generator, with the surrogates updated as new observations become available.

We evaluated MODS on three single-objective tasks: GB1 binding to IgG-Fc, ERBIN-PDZ1 binding to the CRIPT peptide, and KRAS binding to RAF1. We compared against APEX Generative Optimization using a larger budget of 80,100 evaluations, as APEX did not achieve significant improvements within 8 rounds. Averaged over five runs, MODS achieves higher best-found objective values on all three tasks, even though it uses a much smaller budget of 900 evaluations (Figure G.2, Appendix G).

3. Extension to structure generation

“Can MODS be applied to protein structure generation models (e.g. MultiFlow, FrameFlow)?”

Response.

MODS can be combined with structure-generating models whenever a compatible single-objective steering procedure is available. For example, FrameFlow has already been used with gradient-based guidance for structural objectives (Winkler et al., 2026). This provides a concrete route to applying MODS: supply its scalarized surrogate reward to the guidance procedure while keeping the pretrained generator fixed.

We could not evaluate this extension within the discussion period, but we hope MODS will enable multi-objective structure generation. We will discuss this extension and the existing steering mechanisms that make it possible in the revised manuscript.

Thanks for the rebuttal. I have read all the new results and my questions are answered with clear evidence of improvement over existing methods. I will raise my rating to 3.

We want to thank you sincerely for taking the time to review the additional results and for raising your rating. We are glad that the new evidence addresses your questions, and we will incorporate these results and clarifications into the revised manuscript. Thank you again for your contribution to the review process and for helping us strengthen our work.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global Response to the Reviewers and Area Chair

We thank the reviewers and the Area Chair for the time and care devoted to our paper. We appreciate the reviewers’ recognition of MODS’s clear formulation and strong empirical results. The main questions concern its scope, evaluation, and practical use, rather than novelty or methodological correctness. We have added experiments and analyses in Appendix G of the updated manuscript, with references to the relevant figures and tables in our responses below. We will incorporate the feedback, these results, and the clarifications into the revised manuscript. The main additions are:

  • Broader applicability: MODS improves a broadly pretrained protein generator without requiring task-specific fine-tuning.

  • Evidence beyond learned predictors: We ran an additional protein-optimization campaign using Rosetta simulations to provide feedback throughout optimization. MODS outperforms both baselines, showing that its gains extend beyond benchmarks scored by neural networks.

  • Stronger baseline comparisons: We evaluated two additional protein-design methods, ProteinGuide and MosPro. On each of the five two-objective tasks, at least one MODS configuration achieves better joint optimization performance, measured by mean hypervolume.

  • Understanding the method and its cost: We tested the contributions of surrogate ensemble sampling and scalarization separately, checked prediction accuracy on newly proposed proteins, and compared computation time per optimization round across methods. These analyses clarify how MODS works and what it costs to use.

These additions reinforce MODS’s central contribution: a reusable framework for turning single-objective steering procedures into multi-objective optimization methods while keeping the pretrained generator fixed. Researchers can build on advances in generative models and steering methods without designing a new multi-objective algorithm for each application. Our evaluation across six tasks, two diffusion models, and two steering procedures demonstrates the value of this approach.

We hope these additions give the reviewers greater confidence in the paper and help the Area Chair assess how the concerns have been addressed. If the responses address your concerns, we would sincerely appreciate an increase in your scores to reflect this updated assessment.

AuthorsOpenReview ↗

Closing Summary and Thanks

As the rebuttal period draws to a close, we sincerely thank the reviewers for their constructive feedback and engagement, and the Area Chair for overseeing the review process.

The discussion helped us clarify MODS’s contribution: a modular framework for turning a protein generative model and a compatible single-objective steering procedure into a multi-objective protein optimization method. This allows researchers to build on advances in generators and steering techniques with minimal additional implementation effort.

In response to the reviews, we added:

  • Broader evaluation: experiments using a pretrained generator without task-specific fine-tuning, single-objective optimization, and a campaign with Rosetta feedback.
  • Additional baseline comparisons: ProteinGuide and MosPro, including iterative adaptations that update their predictors using newly collected feedback.
  • Computational comparisons: experiments with a larger generator, increased baseline compute budgets, and a faster MODS configuration.
  • Component analysis: ablations of surrogate ensemble sampling and scalarization, alongside assessments of surrogate prediction accuracy.

Together, the original experiments and additional evaluations provide strong evidence that MODS achieves state-of-the-art performance on the protein optimization benchmarks studied here, including against iteratively adapted baselines and baselines given additional compute. The faster configuration further demonstrates a practical tradeoff between computational cost and optimization performance. Supporting results are collected in Appendix G and will be incorporated into the revised manuscript alongside the reviewers’ feedback.

We are grateful to Reviewers C8dQ and RAMo for increasing their scores after reviewing the additional evidence. Throughout the discussion, we also provided experiments and clarifications addressing Reviewer 1sVL’s questions. In their latest response, they confirmed that the iterative baseline comparisons resolved their concern about feedback use and acknowledged MODS’s continued strong performance on the tasks tested. Their remaining reservation concerns the added value of ensemble sampling and augmented Chebyshev scalarization relative to simplified variants of MODS evaluated in our ablations. We have clarified how these variants fit within the framework and highlighted its modularity and overall empirical performance.

Thank you again for your time and for helping us strengthen our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.