acceptodds
Under review as a conference paper at ICLR 2027

AReUReDi: Annealed Rectified Updates for Refining Discrete Flows with Multi-Objective Guidance

Abstract

Biomolecular design requires individual candidates to combine strong target activity with favorable developability. Multi-objective generation must therefore address both trade-off exploration and the weakest properties within each candidate. We introduce **AReUReDi** (**A**nnealed **Re**ctified **U**pdates for **Re**fining **Di**screte Flows), an inference-time method for preference-controlled refinement with frozen discrete generators. AReUReDi combines generator-informed token proposals with annealed Tchebycheff guidance to target limiting properties and a weighted-utility acceptance rule to preserve aggregate quality throughout refinement. Across five predicted binding and developability objectives, AReUReDi outperformed all five comparators in candidate-level balance on both peptide benchmarks, achieving stronger worst-objective scores and more even property profiles. Preference variation enabled controllable affinity, and its leading balance on PPP5 persisted under predictor refitting. Experiments across amino-acid sequences and chemically modified peptide SMILES demonstrated broad applicability to peptide design. AReUReDi provides a unified refinement framework for designing peptide candidates with balanced predicted binding and developability profiles and tunable property priorities.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
discrete flow models, multi-objective optimization, Pareto optimization, biomolecular sequence design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KB37OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AReUReDi, a training-free multi-objective peptide sequence optimization method. It uses a frozen Rectified Discrete Flow model to propose amino acid substitutions, combines annealed Tchebycheff guidance to prioritize weaker properties, and employs a monotone weighted-utility acceptance rule to prevent decreases in aggregate scores. The method is evaluated across multiple peptide design tasks, reporting improvements in predicted binding affinity, solubility, non-hemolysis, non-fouling, and half-life. The authors also demonstrate its applicability to both amino acid sequences and chemically modified peptide SMILES representations.

Critical strengths

1. A relevant problem formulation focusing on candidate-level multi-objective optimization.

The paper emphasizes balancing multiple properties within individual candidates rather than focusing solely on Pareto front coverage. Since peptide candidates must simultaneously satisfy binding and developability requirements, considering trade-offs at the individual-candidate level is a reasonable research direction. Combining the improvement of weaker properties with aggregate utility optimization is a potentially useful approach to multi-objective molecular design.

2. No generator retraining is required, with relatively broad experimental coverage.

The method supports adjustable optimization preferences using frozen generators and can be applied to different molecular sequence representations. The authors provide comparisons across multiple targets, component ablations, preference-control experiments, and predictor-refitting analyses, offering some evidence of effectiveness under the adopted scoring framework.

Critical weaknesses

1. The necessity of the proposed search algorithm and its methodological novelty are insufficiently demonstrated.

The core algorithm combines established techniques, including Tchebycheff scalarization, Barker-balanced proposals, annealed guidance, and greedy weighted-utility acceptance. Although this combination is reasonable, the overall approach remains closely related to existing multi-objective guided generation and local search frameworks. For example, MOG-DFM has explored multi-objective biological sequence optimization using discrete flow models, while PepTune and ParetoFlow have investigated the integration of generative models with multi-objective optimization. The paper therefore needs to better clarify its distinctive methodological contribution.

More importantly, the method already evaluates multiple amino acid substitutions using property predictors at each refinement step. This suggests that simpler search strategies, such as directly selecting the substitution with the highest weighted utility or performing greedy local search based on the worst-objective score, could potentially achieve similar results. Although the authors provide ablations on proposal distributions, annealing schedules, and acceptance rules, they do not directly compare against such simpler, complete search algorithms. It therefore remains unclear whether the proposed guidance and sampling mechanisms provide substantial benefits beyond straightforward local search, or whether the main improvements primarily arise from repeated property evaluations and candidate selection.

Furthermore, the theoretical results concerning invariance and concentration for an exact Metropolis–Hastings sampler do not directly apply to the monotone acceptance algorithm used in the experiments. The nondecreasing weighted-utility guarantee of the implemented algorithm follows directly from its acceptance rule, which further limits the significance of the theoretical contribution.

2. The connection between the primary evaluation metrics and practical candidate quality is insufficiently established.

Table 1 primarily evaluates the worst-objective score (Min.), objective range (Range), and dispersion (Dispersion). These metrics characterize closely related aspects of within-candidate balance and align strongly with the proposed weakest-property guidance mechanism. Consequently, leading on all three metrics does not necessarily demonstrate improvements across independent dimensions of candidate quality.

More importantly, the paper maps heterogeneous properties onto normalized scales using classifier outputs, affinity-score scaling, and transformations of predicted half-life. However, numerical normalization does not establish biological comparability across different properties. The lowest normalized score does not necessarily correspond to the most important practical limitation, and alternative reasonable normalization schemes could change the optimization priorities and method rankings.

Table S2 illustrates this concern. On 1B8Q, AReUReDi substantially improves several predicted developability properties compared with NSGA-III, but its predicted affinity score is lower (5.7130 versus 7.2178). This demonstrates a different property trade-off, but higher worst-objective scores or lower dispersion alone do not establish that the resulting candidates are globally preferable. Their practical value depends on whether individual properties meet acceptable thresholds and on the significance of the corresponding trade-offs.

Although the authors provide objective-level analyses and Pareto coverage results, they do not systematically evaluate the proportion of candidates simultaneously satisfying multiple practical design constraints. Moreover, AReUReDi does not consistently outperform competing methods in Pareto coverage, and its candidate-level balance advantage on 1B8Q disappears after predictor refitting. Therefore, the current evidence mainly supports improved balance under the chosen normalization and scoring framework, rather than a convincing improvement in overall candidate quality or practical developability.

3. The experimental comparisons do not sufficiently isolate the benefits of the proposed optimization mechanism.

The main experiments use method-specific generation budgets rather than matching the number of property predictor evaluations or computational budgets. Table S2 shows that AReUReDi requires substantially more time per candidate than several baselines. Consequently, it remains unclear how much of its performance advantage can be attributed to the proposed optimization mechanism rather than greater computational expenditure.

Although some ablations use matched evaluation budgets, the main baseline comparisons do not systematically characterize the trade-off between optimization performance and computational cost. Together with the absence of direct comparisons against simple greedy search strategies, the experiments do not sufficiently isolate the benefits of the proposed search mechanism.

Questions for authors

1. On the necessity and novelty of the search algorithm:

Since the method already evaluates multiple candidate substitutions at each refinement step, why are Barker-balanced proposals and annealed Tchebycheff guidance necessary instead of simpler greedy or random local search strategies? Could the authors compare against greedy weighted-utility maximization and worst-objective-guided local search under the same generator, property predictors, and evaluation budget, reporting both optimization performance and candidate diversity? Such comparisons would help establish the distinctive benefits of the proposed search mechanism.

2. On evaluation metrics and practical applicability:

How do the authors justify the comparability of normalized scores across heterogeneous properties? Would the method rankings remain stable under alternative reasonable normalization schemes? Furthermore, could the authors report the proportion of candidates simultaneously satisfying predefined thresholds for all objectives, and evaluate this joint feasibility rate under different thresholds and preference settings? This would provide more direct evidence that improved property balance translates into practically useful candidates.

6 comments

We sincerely thank the reviewer for the detailed and constructive feedback, particularly for recognizing the importance of candidate-level multi-objective optimization and the benefits of refinement without generator retraining. Your questions regarding methodological novelty, practical candidate quality, and computational fairness motivated several additional experiments, which we present below.

Weakness 1 and Question 1

The methodological novelty and necessity of AReUReDi remain insufficiently established without direct greedy and random-search comparisons and theory applicable to the implemented algorithm.

Thank you for challenging the necessity of our refinement algorithm. A central challenge in multi-objective peptide design is that maximizing predicted properties does not necessarily preserve biologically meaningful sequence patterns. Property predictors are imperfect surrogates, and aggressive optimization can favor high-scoring sequences poorly represented by peptide training data. We therefore designed AReUReDi to jointly address three requirements: improving limiting properties, retaining the learned peptide sequence prior, and preserving aggregate utility. This separation of proposal guidance and acceptance establishes a distinct optimization principle beyond preference-guided generation in MOG-DFM and multi-objective tree search in PepTune.

The stochastic proposal has an exact KL-regularized variational interpretation:

\Bigl\lbrace \mathbb E_{y\sim q}[\log\sigma(\eta\Delta S_y)] -D_{\mathrm{KL}}(q\Vert b_C) \Bigr\rbrace,$$ where $b_C$ denotes the filtered generator distribution and $\Delta S_y$ the bottleneck-score change. This construction balances property guidance against deviation from the generator's local probabilities and improves the expected proposed bottleneck-score change relative to the unweighted proposal. The acceptance rule independently guarantees nondecreasing weighted utility. **We deliberately use this non-equilibrium refinement rule because exact Metropolis–Hastings sampling targets distributional invariance, requires sequence-density ratios, and generally permits utility-decreasing transitions.** Its equilibrium guarantees therefore address a different objective from our finite-budget refinement task. To test the contribution of this design, we compared AReUReDi with Utility-Greedy, worst-objective greedy, random gated search, and the $\eta=0$ ablation under controlled conditions (Table R1). Bottleneck reweighting improved both endpoint utility and worst-property scores relative to $\eta=0$ on both targets, establishing a contribution beyond generator sampling and monotone acceptance alone. **More importantly, the experiments reveal the practical value of preserving the generator prior.** Compared with Utility-Greedy, AReUReDi achieved 9.4% and 9.1% lower PepReDi3 endpoint cross-entropy, indicating stronger agreement with peptide sequence patterns learned by the pretrained generator. Under the separate PeptiVerse predictor panel, AReUReDi also achieved higher worst-property scores on both targets, retaining its advantage across 82/100 and 97/100 predictor combinations (Table R2). These results support generator-informed stochastic refinement as a means of balancing surrogate optimization with learned sequence preferences, with stronger performance under alternative evaluators. Together, these theoretical and empirical results establish the contribution of AReUReDi: a generator-regularized stochastic refinement mechanism that improves limiting properties while preserving aggregate utility, with measurable advantages in learned-prior alignment and cross-evaluator candidate quality beyond direct surrogate maximization. **Table R1. Controlled local-search comparisons (Table S29 in the revised manuscript).** |Method|1B8Q Utility ↑|1B8Q Hamming Diversity ↑|PPP5 Utility ↑|PPP5 Hamming Diversity ↑| |-|-|-|-|-| |AReUReDi|0.7792|0.8618|0.8351|0.8592| |No bottleneck reweighting (η=0)|0.7657|0.8709|0.8323|0.8568| |Worst-objective greedy + utility gate|0.7445|0.8861|0.7812|**0.9136**| |Random proposals + utility gate|0.7726|**0.8916**|0.8147|0.9116| |Utility-Greedy|**0.8214**|0.8570|**0.8666**|0.8891| **Table R2. Generator alignment and independent-predictor evaluation (Table S30 in the revised manuscript).** |Target|Method|Search-Time Mean Min. ↑|Search-Time Utility ↑|Hamming Diversity ↑|PepReDi3 CE ↓|PeptiVerse-Panel Mean Min. ↑| |-|-|-|-|-|-|-| |1B8Q|AReUReDi|0.4735|0.7792|**0.8618**|**2.5786**|**0.4663**| |1B8Q|Utility-Greedy|**0.5552**|**0.8214**|0.8570|2.8453|0.4369| |PPP5|AReUReDi|0.5927|0.8351|0.8592|**2.6093**|**0.4109**| |PPP5|Utility-Greedy|**0.6481**|**0.8666**|**0.8891**|2.8707|0.3828|

Weakness 2 and Question 2

Normalized balance metrics may not reflect practical candidate quality, motivating sensitivity analyses of score normalization and evaluation under joint feasibility thresholds.

We appreciate the reviewer's questions on score interpretation and candidate quality. In AReUReDi, score transformations and objective weights jointly define user-specified optimization preferences. For positive rescaling hn′=anhnh'_n=a_nh_n, adjusting the weights as ωn′=(ωn/an)/Z\omega'_n=(\omega_n/a_n)/Z and guidance strength as ηk′=Zηk\eta'_k=Z\eta_k, where Z=∑jωj/ajZ=\sum_j\omega_j/a_j, preserves the proposal distribution and utility acceptance rule under fixed candidate support. Nonlinear transformations encode different preferences and may change method rankings. We therefore evaluate methods under common, explicitly specified scoring conventions without assuming universal biological comparability across properties.

Importantly, our task is continuous, preference-controlled multi-objective optimization, not threshold-constrained optimization. The benchmark specifies no minimum acceptable values or experimentally calibrated feasibility thresholds for the five objectives. Consequently, joint feasibility rate is not a well-defined metric for this task. Introducing arbitrary thresholds would impose additional application-specific constraints and change the optimization problem being evaluated, without establishing biological feasibility. We instead assess predicted candidate quality through complementary metrics, including individual property scores, worst-property performance, aggregate utility, sequence diversity, generator alignment, and Pareto coverage. Our controlled preference experiments further demonstrate that AReUReDi can systematically adjust candidate-level balance and trade-off coverage through user-specified preferences (Table R3). Together, these evaluations directly support the continuous multi-objective design capability that AReUReDi is intended to provide.

Table R3. Controlled preference comparison (Table S32 in the revised manuscript). Each setting returned 108 candidates per target using identical generators, predictors, score transformations, and refinement schedules.

TargetPreferenceMean Min. ↑HV ↑IGD+ ↓
1B8QEqual weights0.4700.5870.0203
1B8QDirichlet(1) sweep0.3840.5780.0105
PPP5Equal weights0.5870.6760.0121
PPP5Dirichlet(1) sweep0.4440.6870.0067

Weakness 3

Method-specific evaluation budgets and the absence of direct local-search baselines make it difficult to isolate AReUReDi's algorithmic benefits from additional computation.

Thank you for emphasizing the need for budget-controlled comparisons. Under matched property-predictor query budgets, AReUReDi outperformed NSGA-III, SMS-EMOA, and SPEA2 across all five candidate-quality metrics on both targets, achieving stronger worst-property scores, better within-candidate balance, higher aggregate utility, and greater sequence diversity (Table R4). It also recorded the shortest observed search time, although the shared-GPU execution conditions limit direct runtime comparisons.

As discussed in W1/Q1, preserving the learned peptide sequence prior is central to AReUReDi's refinement strategy, particularly when optimizing imperfect property predictors. Although MOPSO achieved higher optimization-time worst-property scores, AReUReDi maintained greater sequence diversity and achieved the lowest PepReDi3 endpoint cross-entropy among all four evolutionary baselines on both targets, including reductions of 13.2% and 10.9% relative to MOPSO. This confirms that the stronger generator alignment observed against Utility-Greedy also extends to MOPSO under matched query budgets. We further evaluated Utility-Greedy, worst-objective greedy, random gated search, and the η=0\eta=0 ablation under shared local-search conditions, as detailed in W1/Q1. Together, these controlled comparisons establish AReUReDi's contribution beyond additional predictor evaluations: effective multi-objective refinement with substantially stronger preservation of the learned peptide sequence prior.

Table R4. Matched-query external comparisons (Tables S26–S28 in the revised manuscript).

TargetMethodMean Min. ↑Range ↓Dispersion ↓Mean Utility ↑Hamming Diversity ↑Search Time (min) ↓PepReDi3 CE ↓
1B8QAReUReDi0.4735 ± 0.01630.4775 ± 0.01120.1807 ± 0.00520.7792 ± 0.01010.8618 ± 0.01143.9 ± 0.22.5786 ± 0.0196
1B8QMOPSO0.6727 ± 0.02270.2792 ± 0.02660.1033 ± 0.01080.8165 ± 0.01500.6179 ± 0.07508.6 ± 1.32.9712 ± 0.1251
1B8QNSGA-III0.4140 ± 0.01590.5310 ± 0.01320.1955 ± 0.00430.7326 ± 0.00920.6083 ± 0.05118.0 ± 0.62.9462 ± 0.0635
1B8QSMS-EMOA0.2913 ± 0.02620.6417 ± 0.02460.2353 ± 0.00850.6801 ± 0.00770.6267 ± 0.069912.5 ± 1.93.0956 ± 0.0805
1B8QSPEA20.3297 ± 0.01590.6001 ± 0.01130.2231 ± 0.00520.6535 ± 0.00920.7179 ± 0.07457.9 ± 2.23.0554 ± 0.0669
PPP5AReUReDi0.5927 ± 0.01910.3812 ± 0.01540.1389 ± 0.00650.8351 ± 0.00760.8592 ± 0.009616.7 ± 0.22.6093 ± 0.0179
PPP5MOPSO0.6979 ± 0.01410.2611 ± 0.01210.0961 ± 0.00440.8296 ± 0.01280.7859 ± 0.030337.3 ± 3.42.9292 ± 0.0363
PPP5NSGA-III0.5247 ± 0.08360.4473 ± 0.07750.1683 ± 0.03400.7942 ± 0.05450.4007 ± 0.013735.1 ± 5.62.8007 ± 0.0702
PPP5SMS-EMOA0.5180 ± 0.06710.4586 ± 0.06400.1689 ± 0.02100.8019 ± 0.02930.3728 ± 0.041044.5 ± 1.22.7987 ± 0.1822
PPP5SPEA20.3386 ± 0.01540.6115 ± 0.02170.2302 ± 0.01050.6629 ± 0.01730.5864 ± 0.070836.4 ± 5.62.9184 ± 0.2166

We hope the new greedy and random-search controls, generator-prior and cross-evaluator analyses, and matched-budget comparisons substantively address your concerns about algorithmic necessity, candidate quality, and computational cost. Together, these results strengthen the evidence for AReUReDi's distinctive contribution to multi-objective peptide design. We would be grateful if you would reconsider your score and confidence in the submission in light of these findings.

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

This paper studies multi-property peptide design: given a target protein and property preferences, it generates amino-acid sequences or chemically modified peptide SMILES that jointly achieve favorable predicted binding affinity and developability.

The motivation is that good Pareto-set coverage does not guarantee a satisfactory weakest property for each individual candidate.

Building on existing discrete flow models and ReDi generators, the method combines bottleneck-guided proposals with a utility-preserving acceptance rule. At each iteration, it first proposes an edit by applying annealed Barker reweighting to the frozen generator’s probabilities over local substitution candidates, favoring improvements in the minimum weighted property score. It then accepts the proposed edit only if weighted aggregate utility does not decrease. The proposal score and acceptance utility use the same preference weights (§1–3, Eqs. 1–4).

The main experiments report improved candidate-level balance on two amino-acid benchmarks, component ablations support the roles of the acceptance gate and learned proposals, and preference and cross-representation experiments demonstrate property control and applicability to modified peptides. Rescoring frozen candidates preserves the advantage on PPP5 but reverses the ranking on 1B8Q, indicating target-dependent sensitivity to the evaluators (Table 1; §4.3–4.5; Appendix H.3).

Critical strengths

1. The optimization objective and division of responsibilities are clear. Bottleneck-guided proposals and aggregate-utility acceptance address candidate balance and overall property utility, respectively, with a nondecreasing-utility guarantee for the implemented acceptance rule (§3, Eqs. 1–4; Proposition A.4). Under the reported evaluation configurations, the method achieves mean worst-property scores of 0.516 and 0.613 on 1B8Q and PPP5, together with smaller property ranges and dispersion, supporting its candidate-level balance objective (Table 1).

2. The ablations provide concrete evidence for several components. At the same refinement-step budget, the acceptance gate improves all five mean property scores on both 8CN1 and 4EBP2. Objective-removal experiments show that omitting a property from guidance can harm that property. Dedicated comparisons also examine annealing and learned proposals, supporting several components under the tested conditions (Tables S6–S10, S12–S13).

3. Preference control and cross-representation applicability are empirically supported. Increasing the affinity weight monotonically increases mean normalized affinity on both main targets (Table S18). The method also operates on chemically modified peptide SMILES, improving predicted affinity and non-fouling across all five tested targets (§4.4; Figures 3, S4–S5).

Critical weaknesses

W1. The empirical increment over the closest related method remains unclear. MOG-DFM (arXiv:2505.07086, Section 3) already performs multi-objective local updates and preference guidance using a pretrained generator. The specific increment here lies primarily in combining bottleneck-guided Barker proposals with aggregate-utility acceptance, but the paper does not directly compare against MOG-DFM. The authors explicitly disclose method-specific budgets and report runtime differences (Section 4; Table S2; Appendices F and J). These results support performance under the reported system configurations, but do not clearly separate the contributions of the new update rules, the generator, and search expenditure. This limits assessment of the method's benefits over its closest existing alternative. (Q1 addresses this concern.)

W2. The endpoint contribution of bottleneck reweighting within the full method has not been isolated. The equations explain how bottleneck scores influence local proposals, and the paper includes experiments on acceptance, objective removal, annealing, and learned proposals (Section 3; Tables S6-S10 and S12-S13). What remains unclear is whether bottleneck reweighting further improves candidate balance after a finite refinement budget when generator-based proposals and the aggregate-utility gate are retained. The smallest fixed guidance strength in Table S10 is eta = 1, which still applies property-dependent reweighting and therefore does not directly answer this question. This limits attribution of the observed balance gains to bottleneck reweighting. (Q2 addresses this concern.)

Questions for authors

Please prioritize Questions 1&2. The answer of them can serve as the response to the corresponding weaknesses; there is no need to answer both Weaknesses sections and Questions section separately.

Q1. Could the authors clarify the key algorithmic differences from MOG-DFM, explain which existing evidence supports the relative benefits, and provide the relevant budget configurations or objective-evaluation counts? If a compatible implementation is available, a small comparison on one main target using a common scoring protocol, comparable preference objectives, and similar objective-evaluation budgets would help assess the relative increment. If the backbones differ or there are substantive adaptation barriers, please explain how these conditions affect the conclusions that can be drawn.

Q2. Could the authors provide existing results, or a small comparison between the full method and a variant with ηk=0\eta_k=0 throughout? The latter reduces the proposal to the generator probabilities over the filtered candidates while retaining the aggregate-utility gate. On one main target, a small set of shared initial sequences, with the generator, scorers, candidate-filtering rules, and refinement budget held fixed, would allow a direct assessment of the additional benefit of bottleneck reweighting. Endpoint worst-property scores and aggregate utility, together with their variability, would address the question; acceptance rates or existing trajectories could aid interpretation.


Questions 3–5 are just secondary concerning about reporting, implementation details, and wording. Brief clarifications should generally suffice, with corrections to affected values or claims where necessary.

Q3. Correspondence between the two tables. For PPP5–MOPSO, Table 1 reports a mean worst-property score of 0.440, whereas Table S2 reports mean solubility of 0.4255. If both tables use the same candidates and solubility evaluation scale, the mean minimum should not exceed mean solubility (Appendix H.1). Please briefly clarify the correspondence between candidate batches and score transformations, and check the values if necessary.

Q4. Generator interface and scope of the uniform-proposal ablation. Please specify the model time tkt_k and actual sequence input used to obtain bk,ib_{k,i} in §3.1. For Tables S12–S13, are probabilities made uniform after candidate filtering, or are uniform probabilities used to construct a new candidate set? This determines whether the ablation changes only probability weights or also the available substitutions. Brief pseudocode or a corresponding code location in the supplementary package would suffice (§3.1; Algorithm 1; Appendix F).

Q5. Attribution of coverage gains to preference sweeps. Appendix H.1 already states that Table S17 involves changes in candidate counts, generation configurations, and scoring configurations. The coverage claims in §4.3 and related summaries should reflect this qualification, avoiding attribution of the cross-configuration differences solely to preference variation. A wording revision should suffice; the existing preference-control evidence in Tables S18–S19 can be retained.

4 comments

We sincerely thank the reviewer for the thoughtful review and for recognizing the clear distinction between bottleneck-guided proposals and utility-preserving acceptance, the informative ablations, and the demonstrated preference control across peptide representations. We have conducted additional experiments and made the requested clarifications to address each of your questions below.

Weakness 1 and Question 1

The advantages over MOG-DFM remain unclear without a direct comparison and consideration of differences in generator backbones and evaluation budgets.

Thank you for requesting a direct comparison with MOG-DFM. Our central insight is that multi-objective peptide refinement must jointly address three challenges: (1) improving the limiting property of each candidate, (2) preserving aggregate utility under competing objectives, and (3) maintaining consistency with the learned peptide sequence prior to mitigate overoptimization of imperfect property predictors. MOG-DFM employs rank-directional guidance and adaptive hypercone filtering to explore preferred Pareto trade-offs. AReUReDi addresses these three requirements through a unified stochastic refinement rule. The Barker proposal admits a local KL-regularized interpretation that balances bottleneck improvement with generator alignment, while the acceptance rule guarantees nondecreasing weighted utility. This coupling establishes a distinct candidate-level optimization principle with complementary control over property balance, aggregate quality, and learned sequence preferences.

We directly compared both methods on 1B8Q and PPP5 using identical property predictors and score transformations. AReUReDi achieved higher worst-property scores and lower range and dispersion on both targets (Table R1), supporting its candidate-level balance objective. Importantly, AReUReDi also achieved lower endpoint cross-entropy under both frozen PepReDi3 and PepDFM evaluators, including MOG-DFM's own generator backbone. This cross-generator agreement supports stronger preservation of learned peptide sequence preferences beyond optimization-time property scores. MOG-DFM retained its PepDFM backbone and native schedule, using 81,278 and 127,063 five-property evaluations on 1B8Q and PPP5, respectively. AReUReDi used PepReDi3 and its original refinement schedule, so this complete-method comparison does not isolate the effect of the update rule from backbone and budget differences.

Together, these theoretical and empirical results support AReUReDi's distinctive contribution to multi-objective peptide design: improving candidate-level property balance while preserving aggregate utility and stronger alignment with learned peptide sequence priors.

Table R1. Comparison with MOG-DFM under common property predictors and cross-generator evaluation (Table S25 in the revised manuscript).

TargetMethodMean Min. ↑Range ↓Dispersion ↓PepReDi3 CE ↓PepDFM CE ↓
1B8QAReUReDi0.5160.4830.1862.55970.8119
1B8QMOG-DFM0.3990.5600.2082.77940.9180
PPP5AReUReDi0.6130.3870.1482.62040.7186
PPP5MOG-DFM0.4160.5540.2062.85070.8681

Weakness 2 and Question 2

Existing ablations do not isolate whether bottleneck reweighting improves final candidate balance beyond generator-based proposals and monotone utility acceptance.

Thank you for suggesting the direct ηk=0\eta_k=0 ablation. We conducted this experiment on both 1B8Q and PPP5, removing bottleneck-dependent Barker reweighting while keeping other settings unchanged. At ηk=0\eta_k=0, the proposal reduces to the normalized generator probabilities over the filtered candidates, directly isolating the contribution of bottleneck guidance. Across three paired seeds with 100 outputs per seed, AReUReDi improved both endpoint worst-property scores and aggregate utility on both targets, with positive utility gains across all three seeds (Table R2). These results directly establish that bottleneck reweighting improves finite-budget multi-objective refinement beyond the generator prior and utility-preserving acceptance alone.

Table R2. Controlled ablation of Barker bottleneck reweighting (Table S31 in the revised manuscript).

TargetFull Utility ↑ηk=0\eta_k=0 Utility ↑Δ Utility ↑Δ Mean Min. ↑Positive Paired Seeds
1B8Q0.77920.7657+0.0135+0.02033/3
PPP50.83510.8323+0.0028+0.02123/3

Question 3

The PPP5–MOPSO worst-property score in Table 1 appears inconsistent with the corresponding solubility score in Table S2.

Thank you for catching this inconsistency. The correct PPP5–MOPSO mean Min. is 0.407, instead of 0.440. We have corrected this value in Table 1 of the revised manuscript. This correction does not affect any of our conclusions or method rankings.

Question 4

The generator input and model-time convention require clarification, particularly whether uniform-proposal ablations change only proposal probabilities or also the candidate sets.

We appreciate the request for implementation details. In the verified amino-acid refinement implementation, the frozen generator receives the full, unmasked current sequence with BOS/EOS tokens at model time tk=k/Tt_k=k/T, where T=20LT=20L; we use its token probabilities at the selected position.

For both amino-acid and SMILES ablations, we first apply the original forbidden-token, nucleus, and applicable top-KK filters, then assign uniform probabilities over the same retained candidate set before Barker guidance. Thus, Tables S12–S13 change the proposal weights without changing which substitutions are evaluated. We have stated the sequence input, time convention, and filtering order in §3.1 and Appendix F.


Question 5

The reported preference-sweep coverage gains may be confounded by differences in candidate counts and scoring configurations, requiring clearer attribution.

Thank you for identifying the configuration differences in Table S17. We conducted a new controlled comparison between equal weighting and Dirichlet(1) preference sweeping, using identical generators, predictors, score transformations, refinement schedules, and 108 candidates per setting. Under these matched conditions, preference sweeping improved IGD+ on both targets and increased hypervolume on PPP5 (Table R3), providing direct evidence that preference variation improves empirical trade-off coverage independently of the original configuration differences. We have revised §4.3 and Appendix H.1 accordingly. Together with the affinity-control results in Table S18, these findings establish AReUReDi's ability to systematically control property priorities and explore diverse multi-objective trade-offs through user-specified preferences.

Table R3. Candidate-level balance and empirical coverage under controlled preference variation (Table S32 in the revised manuscript).

TargetPreferenceMean Min. ↑HV ↑IGD+ ↓
1B8QEqual weights0.4700.5870.0203
1B8QDirichlet(1) sweep0.3840.5780.0105
PPP5Equal weights0.5870.6760.0121
PPP5Dirichlet(1) sweep0.4440.6870.0067

We hope the direct MOG-DFM comparison, cross-generator evaluations, controlled ηk=0\eta_k=0 ablation, and revised preference experiments provide convincing evidence for AReUReDi's distinct methodological contribution and address your concerns. We would be grateful if you would reconsider your score and confidence in the submission in light of these new findings.

Thank you for the detailed response and additional experiments. The zero-guidance ablation addresses my main concern about bottleneck reweighting, and the MOG-DFM comparison is helpful, with the backbone and budget differences clearly acknowledged. The correction and clarifications also resolve my other questions. I am satisfied with the response and will keep my Rating.

Dear Reviewer rVZm,

Thank you very much for your thoughtful follow-up. We are glad that the additional experiments and clarifications addressed your concerns, and we sincerely appreciate your constructive feedback, which helped us strengthen the manuscript.

Thank you again!

The Authors

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

This paper proposes an inference-time method for multi-objective peptide refinement with frozen discrete generators. Generator probabilities inform candidate token substitutions, annealed Tchebycheff guidance emphasizes limiting weighted properties, and a weighted-sum acceptance rule retains substitutions that do not reduce aggregate utility. The paper evaluates within-candidate property balance on two main peptide targets, along with amino-acid and chemically modified peptide SMILES experiments, ablations, preference sweeps, and predictor-refit analyses. All optimized biological properties are predicted rather than newly measured in physical assays.

Critical strengths
  1. The paper clearly distinguishes within-candidate balance from coverage of trade-offs across a set. In peptide design, a broad frontier does not guarantee that any individual candidate meets several requirements at once. The proposal and acceptance rules are built around this concrete design objective.
  2. Under the reference predictors, the method improves candidate balance on both main targets. The evaluation goes beyond endpoint scores by examining generator choice, acceptance, guidance schedules, and predictor refitting. These analyses help identify both the sources of the gains and their dependence on the scoring models.
Critical weaknesses
  1. The main comparison does not rule out the contribution of additional computation. Table S2 reports substantially longer per-candidate runtimes for AReUReDi than for the main alternatives. Table S24 matches evaluation counts, but only for internal proposal variants. I would like to see the principal external baselines compared under a matched property-predictor evaluation budget, with runtime reported alongside quality. In addition, some comparators primarily target frontier coverage. A comparison using a common balance objective or a consistent final-candidate selection rule would make the contribution of the search method easier to assess.
  2. The conclusions are sensitive to both the scoring models and the score coordinates. In Table S21, MOPSO exceeds AReUReDi in mean worst-objective score on 1B8Q after predictor refitting, while AReUReDi remains ahead on PPP5. The equal-weight and preference-sweep results in Table S17 also use different predictor configurations and half-life transformations, as acknowledged in Appendix H.1, and return different numbers of candidates. Their hypervolume difference therefore cannot be attributed to preference variation alone. The disclosures are helpful, but a common rescoring protocol is still needed to support the main comparisons more firmly.
Questions for authors
  1. How does the principal balance comparison change when the external baselines receive the same total property-predictor evaluation budget? For methods primarily designed for frontier coverage, can the final candidates also be compared under a common balance objective or selection rule?
  2. If the equal-weight outputs, preference-sweep outputs, and baseline candidates are evaluated with the same predictors and transformations, and with matched candidate counts where appropriate, do the balance and coverage conclusions remain? Please relate the answer to the target-specific ranking reversal in Table S21.
3 comments

We sincerely thank the reviewer for the careful and constructive assessment, particularly for recognizing the importance of candidate-level balance beyond Pareto coverage, as well as the value of our ablation and predictor-sensitivity analyses. We appreciate the questions regarding computational fairness and scoring consistency, which we address individually below.

Weakness 1 and Question 1

The benefits of AReUReDi remain unclear without query-matched comparisons against external baselines and consistent candidate-level evaluation criteria.

Thank you for requesting a query-matched comparison. Under matched property-predictor query budgets, AReUReDi outperformed NSGA-III, SMS-EMOA, and SPEA2 across all five evaluation metrics on both targets, achieving stronger candidate-level balance, higher aggregate utility, and greater sequence diversity. It also achieved the highest diversity among all four evolutionary baselines with shortest runtime (Table R1).

MOPSO achieved higher optimization-time worst-property scores, but AReUReDi achieved the lowest PepReDi3 cross-entropy among all four evolutionary baselines on both targets, including reductions of 13.2% on 1B8Q and 10.9% on PPP5 relative to MOPSO. This distinction is particularly relevant to biological sequence design, where aggressive optimization of imperfect property predictors can favor sequences outside the patterns learned from peptide data. By retaining generator probabilities during refinement, AReUReDi incorporates learned peptide sequence preferences that help counterbalance this risk.

Together, these results establish the value of AReUReDi under matched evaluation budgets: strong multi-objective performance with greater preservation of the learned peptide sequence prior, an important consideration beyond surrogate optimization scores alone.

Table R1. External comparison under matched property-predictor query budgets (Tables S26–S28 in the revised manuscript).

TargetMethodMin. ↑Range ↓Dispersion ↓Mean Utility ↑Hamming Diversity ↑Search Time (min) ↓PepReDi3 CE ↓
1B8QAReUReDi0.4735 ± 0.01630.4775 ± 0.01120.1807 ± 0.00520.7792 ± 0.01010.8618 ± 0.01143.9 ± 0.22.5786 ± 0.0196
1B8QMOPSO0.6727 ± 0.02270.2792 ± 0.02660.1033 ± 0.01080.8165 ± 0.01500.6179 ± 0.07508.6 ± 1.32.9712 ± 0.1251
1B8QNSGA-III0.4140 ± 0.01590.5310 ± 0.01320.1955 ± 0.00430.7326 ± 0.00920.6083 ± 0.05118.0 ± 0.62.9462 ± 0.0635
1B8QSMS-EMOA0.2913 ± 0.02620.6417 ± 0.02460.2353 ± 0.00850.6801 ± 0.00770.6267 ± 0.069912.5 ± 1.93.0956 ± 0.0805
1B8QSPEA20.3297 ± 0.01590.6001 ± 0.01130.2231 ± 0.00520.6535 ± 0.00920.7179 ± 0.07457.9 ± 2.23.0554 ± 0.0669
PPP5AReUReDi0.5927 ± 0.01910.3812 ± 0.01540.1389 ± 0.00650.8351 ± 0.00760.8592 ± 0.009616.7 ± 0.22.6093 ± 0.0179
PPP5MOPSO0.6979 ± 0.01410.2611 ± 0.01210.0961 ± 0.00440.8296 ± 0.01280.7859 ± 0.030337.3 ± 3.42.9292 ± 0.0363
PPP5NSGA-III0.5247 ± 0.08360.4473 ± 0.07750.1683 ± 0.03400.7942 ± 0.05450.4007 ± 0.013735.1 ± 5.62.8007 ± 0.0702
PPP5SMS-EMOA0.5180 ± 0.06710.4586 ± 0.06400.1689 ± 0.02100.8019 ± 0.02930.3728 ± 0.041044.5 ± 1.22.7987 ± 0.1822
PPP5SPEA20.3386 ± 0.01540.6115 ± 0.02170.2302 ± 0.01050.6629 ± 0.01730.5864 ± 0.070836.4 ± 5.62.9184 ± 0.2166

Weakness 2 and Question 2

Differences in scoring configurations and candidate counts confound the reported preference-sweep gains, while predictor refitting raises concerns about the stability of method rankings.

We thank the reviewer for highlighting the scoring inconsistencies. We repeated the equal-weight and Dirichlet(1) preference experiments using identical generators, predictors, score transformations, refinement schedules, and 108 candidates per setting. Under these controlled conditions, AReUReDi achieved stronger candidate-level balance with equal weighting, while preference sweeping consistently improved IGD+ on both targets and increased hypervolume on PPP5 (Table R2). These results establish effective preference control over the balance–coverage trade-off, independent of the scoring and candidate-count differences in the original Table S17. We have clarified this distinction in §4.3 and Appendix H.1.

We also evaluated predictor sensitivity separately. Under the 100 refitted predictor combinations in Table S21, AReUReDi retained the strongest candidate-level balance on PPP5, with a target-dependent ranking reversal against MOPSO on 1B8Q. The new matched-query baseline comparisons further use common scoring definitions within their evaluation protocol. Together, these controlled evaluations reinforce AReUReDi's effectiveness in multi-objective peptide refinement, demonstrating its ability to produce balanced candidates and systematically adjust trade-off coverage through user-specified preferences.

Table R2. Controlled equal-weight and preference-sweep comparison (Table S32 in the revised manuscript).

TargetPreferenceMean Min. ↑HV ↑IGD+ ↓
1B8QEqual weights0.4700.5870.0203
1B8QDirichlet(1) sweep0.3840.5780.0105
PPP5Equal weights0.5870.6760.0121
PPP5Dirichlet(1) sweep0.4440.6870.0067

We hope the new matched-query baseline comparisons, controlled preference experiments, and generator-alignment analyses substantially address your concerns and strengthen the evidence for AReUReDi's multi-objective refinement capabilities. We would be grateful if you would reconsider your score and confidence in the submission in light of these additional results.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer NtuD,

Thank you again for your thoughtful review and constructive feedback.

Our additional experiments provide substantially stronger evidence for AReUReDi's effectiveness. Under matched property-predictor evaluation budgets, AReUReDi achieves better candidate-level balance, higher utility, and greater sequence diversity than NSGA-III, SMS-EMOA, and SPEA2 on both targets, while maintaining stronger generator alignment than MOPSO. We also conducted fully controlled preference experiments with identical predictors, score transformations, refinement schedules, and candidate counts, confirming that preference sweeping improves IGD+ on both targets independently of the original configuration differences. Together with our clarification of target-dependent predictor sensitivity, these results directly strengthen the evidence for AReUReDi's multi-objective refinement and preference-control capabilities.

We hope these new results address your concerns and would greatly appreciate your reconsideration of the rating in light of the additional evidence. We welcome any further questions and would be happy to address them during the discussion period.

The Authors

Comments on the paper

1 comment

Dear Area Chair,

As discussion closes, we summarize AReUReDi's contributions and rebuttal evidence. We thank reviewers for their constructive feedback. Reviewer rVZm confirmed that our additional results resolved their concerns and maintained a weak-accept rating. Reviewer KB37 acknowledged the new evaluations but maintained rejection, while Reviewer NtuD has not responded to our follow-up.


Strengths Recognized by the Reviewers

Reviewers highlighted the following strengths:

  • Practical biological motivation. Optimizing multiple properties within individual peptide candidates rather than focusing solely on Pareto-front coverage.
  • Distinct refinement formulation. Bottleneck-guided proposals and utility-preserving acceptance provide a principled mechanism for multi-objective refinement.
  • Flexible and training-free optimization. Refinement of frozen generators, preference control, and support for amino-acid sequences and chemically modified peptide SMILES.
  • Experimental breadth. Multiple peptide targets, component ablations, preference-control experiments, and predictor-sensitivity evaluations.

Key Contributions and Additional Evidence

1. A principled approach beyond surrogate-score maximization.

A central challenge in biomolecular design is that higher predicted property scores do not necessarily imply better biological candidates, particularly with imperfect surrogate predictors. AReUReDi addresses this through generator-informed stochastic refinement, combining bottleneck improvement, preservation of learned sequence preferences, and monotone aggregate-utility acceptance. Its KL-regularized proposal interpretation formalizes the balance between property guidance and generator alignment.

2. Direct evidence for the proposed optimization mechanism.

Controlled comparisons against greedy and random search, together with the zero-guidance ablation, establish that bottleneck reweighting contributes to finite-budget refinement beyond generator sampling and utility acceptance alone. The direct MOG-DFM comparison further supports AReUReDi's distinct approach to candidate-level multi-objective optimization.

3. Stronger generator alignment and independent-predictor performance.

Although Utility-Greedy and MOPSO achieve higher optimization-time scores on certain metrics, AReUReDi demonstrates stronger generator alignment and, against Utility-Greedy, better worst-property performance under independent predictors. These findings reinforce an important distinction: improving surrogate scores is not equivalent to retaining favorable performance under independent evaluation. This matters especially in biological sequence design, where predictor exploitation remains a fundamental challenge.

4. Rigorous evaluation and broader applicability.

Matched-budget experiments establish competitive multi-objective performance, stronger diversity and generator alignment against MOPSO, and advantages over NSGA-III, SMS-EMOA, and SPEA2. Controlled preference experiments demonstrate adjustable candidate-level balance and Pareto coverage under consistent evaluation settings. Chemically modified peptide results further support flexibility across preferences and molecular representations.


Remaining Concerns and Our Perspective

We acknowledge Reviewer KB37's observation that simpler greedy strategies and MOPSO achieve higher scores on some optimization-time objectives. This qualifies claims of universal superiority in candidate-level balance, and we are happy to clarify the scope of our comparisons.

However, we respectfully emphasize that surrogate optimization, generator alignment, and cross-predictor robustness should be considered together when evaluating biomolecular design methods. Our new evidence shows that stronger surrogate optimization does not necessarily yield stronger performance under these complementary measures. While alternative generator-regularized search strategies remain valuable directions for future investigation, the current theoretical and empirical results establish a distinct and experimentally supported contribution from AReUReDi.


Final Request

We ask the Area Chair to consider the full evidence, including our new experiments addressing reviewer concerns and Reviewer rVZm's confirmation that their concerns were resolved.

We believe AReUReDi makes a meaningful methodological contribution to multi-objective biomolecular design by moving beyond surrogate-score maximization toward generator-informed refinement that jointly considers property improvement, learned sequence preferences, and robustness across predictors.

We sincerely hope these contributions and the strengthened evidence will be recognized in the final assessment, and respectfully ask for your support for acceptance.

Thank you very much for your time and careful consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.