acceptodds
Under review as a conference paper at ICLR 2027

PEGASUS: Query-Efficient Multi-Objective Biomolecular Design With Discrete Flow Maps

Abstract

Discrete flow maps compress generation into one or a few model evaluations, but multi-objective optimization can still require many expensive oracle evaluations because intermediate states provide limited information about reachable terminal quality. To address this, we introduce **PEGASUS** (**P**areto-**E**fficient **G**uidance for **A**ction **S**election **U**sing Future-Observability **S**cores), a query-efficient inference-time optimizer for pretrained discrete flow maps. PEGASUS introduces *future observability*, the conditional terminal objective potential of an intermediate state under the generator's continuation distribution. It estimates this potential through objective-independent finite-time Koopman transport and an online task readout, combining it with multiscale Hamming actions and Pareto Decision Observability (PDO) to focus exact evaluations on decision-relevant uncertainty. Under an exact decision-linear response model, the PDO acquisition bound depends on the intrinsic unresolved decision dimension and is independent of the candidate count. Empirically, PEGASUS achieves competitive or superior terminal Pareto quality with up to fewer unique oracle queries than state-of-the-art discrete multi-objective generative baselines. Overall, PEGASUS provides a powerful, sequence-based framework for query-efficient multi-property biomolecule design.

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

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_XZ1gOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PEGASUS, a query-efficient inference-time optimizer for multi-objective biomolecular sequence design with frozen discrete flow models. The method combines future observability, implemented through objective-independent finite-time Koopman transport and an online task-specific readout, with multiscale Hamming-space actions and Pareto Decision Observability (PDO). PDO selectively acquires exact objective evaluations only when unresolved uncertainty may change the action decision. The paper also derives a query bound for PDO under an exact decision-linear response model.

Experiments on six-objective peptide design and three-objective DNA design show that PEGASUS achieves competitive or better Pareto quality with substantially fewer oracle queries than several generative and optimization baselines. The paper also includes component ablations, generator substitutions, synthetic validation of the PDO theory, and structural analyses.

Critical strengths
  1. PEGASUS shows strong query-efficiency. It achieves Pareto quality comparable to or better than strong discrete generative baselines while using dramatically fewer unique objective evaluations. The advantage is especially clear for the longer peptide setting, where PEGASUS achieves better hypervolume with roughly two orders of magnitude fewer queries. This directly supports the main motivation of the paper.

  2. The empirical evaluation is thorough. The paper includes useful component ablations, query accounting, generator substitutions, peptide and DNA experiments, multiscale-action analyses, and synthetic experiments validating the PDO bound in the exact-linear regime. The experiments are generally careful about distinguishing theoretical guarantees from empirical behavior.

Critical weaknesses
  1. The role of the Koopman formulation is not fully convincing. Table S12 shows that a dimension-matched generic finite-time linear observer performs better than the proposed Koopman observer on several downstream metrics, especially for L=40, while using similar query counts. This suggests that the main benefit may come from the broader idea of future observability rather than from the Koopman realization itself. The current emphasis on Koopman transport therefore seems somewhat stronger than what the experiments support.

  2. The strongest query-efficiency comparison is not fully budget-matched. MOG-DFM and AReUReDi operate at much larger native query budgets, so the “up to 99× fewer queries” result does not directly establish that PEGASUS is better at the same oracle budget. A quality-versus-query comparison or budget-matched versions of the strongest baselines would make the central claim more convincing.

Questions for authors
  1. Table S12 shows that a generic dimension-matched finite-time linear observer can outperform the Koopman observer, especially for L=40. What practical advantage justifies treating Koopman transport as a central methodological contribution rather than one possible implementation of future observability?

  2. Can the authors provide budget-matched results or quality-versus-query curves for MOG-DFM and AReUReDi at query budgets comparable to PEGASUS?

  3. The practical PDO implementation uses approximate response models and a query cap, while the theorem assumes exact decision-linear responses. To what extent do the authors believe the intrinsic-dimension argument explains the empirical query savings observed in PEGASUS?

5 comments

We sincerely thank the reviewer for their thorough and constructive evaluation, particularly for recognizing PEGASUS's strong query efficiency, extensive experimental validation, and careful distinction between theoretical guarantees and empirical behavior. We appreciate the insightful questions regarding Koopman transport, query-budget fairness, and PDO's practical efficiency, which we address with additional experiments and analyses below.


W1 + Q1. Koopman transport versus generic linear observers

The specific advantage of Koopman transport remains unclear because a generic linear observer achieves better downstream optimization performance on several metrics, particularly at L = 40.

What practical advantage justifies using Koopman transport rather than a generic linear observer for future observability?

Thank you for this important question regarding the specific contribution of Koopman transport. Koopman provides an objective-independent mechanism for transporting future objective-relevant information through the frozen generative dynamics. Its finite-time operators are calibrated without downstream objective labels, while task-specific readouts are updated using sparse exact evaluations. This separation enables prospective terminal-quality estimation directly from intermediate states without repeated rollout-based scoring.

To directly assess Koopman's advantage over a generic linear observer, we conducted an additional held-out task-projected transport evaluation, measuring the accuracy of future objective-relevant information rather than full feature reconstruction. As shown in Table W1, Koopman reduced task-projected error by 32.6% at L = 12 and 15.8% at L = 40 relative to the dimension-matched generic linear observer, with positive paired improvements whose 95% confidence intervals exclude zero. Without finite-time transport, errors increased substantially to 0.663 and 0.816, respectively.

While Table S12 evaluates downstream optimization involving multiple interacting decision mechanisms, the new experiment directly isolates the predictive contribution of Koopman transport. These results establish a clear, quantitative advantage in preserving terminal-objective-relevant information, supporting Koopman as a principled and empirically validated realization of PEGASUS's Future Observability mechanism.

Table W1: Task-Projected Transport Error of Future Observers (Table S35 in the revised manuscript).

LengthKoopman ↓Generic Linear ↓No Transport ↓Error Reduction ↑Paired ΔError [95% CI] ↑
L = 120.05730.08500.663032.6%0.0277 [0.0276, 0.0278]
L = 400.03580.04250.816015.8%0.0068 [0.0065, 0.0071]

W2 + Q2. Comparable-query single-candidate comparison

The query-efficiency gains over MOG-DFM and AReUReDi are not established under matched budgets, requiring budget-matched or quality-versus-query tests.

We thank the reviewer for this valuable suggestion. PEGASUS generates a complete library of 100 terminal peptides using approximately the same number of oracle queries that MOG-DFM and AReUReDi require to produce their first completed peptide. We therefore compared PEGASUS's highest-utility terminal peptide with the first completed peptide from each baseline, preserving their native generation procedures (Table W2).

PEGASUS achieved substantially higher Top-1 utility than both baselines at both sequence lengths, with a particularly pronounced advantage at L = 40, where PEGASUS also attained the highest mean scores across all six objectives. These results establish PEGASUS's strong best-available-candidate quality at comparable oracle-query expenditures, with candidate-selection opportunities reflecting the number of peptides each method could complete at that cost.

Together with our original fixed-100 comparison demonstrating comparable or superior Pareto quality with up to 99× fewer queries, these results establish PEGASUS's strong quality-efficiency advantage at both individual-candidate and terminal-library levels.

Table W2: Approximately budget-matched Top-1 peptide optimization (Table S37 in the revised manuscript).

LengthMethodNon-Hemolysis ↑Non-Fouling ↑Solubility ↑Permeability ↑Half-Life ↑Affinity ↑Top-1 Utility ↑Queries ↓
L = 12PEGASUS0.9887 ± 0.01090.8569 ± 0.04250.9964 ± 0.00870.8161 ± 0.03900.9177 ± 0.06640.8014 ± 0.0287−0.0343 ± 0.00491303.8 ± 67.6
MOG-DFM0.9728 ± 0.01260.9654 ± 0.02821.0000 ± 0.00000.9517 ± 0.03090.6016 ± 0.05720.5848 ± 0.0545−0.0733 ± 0.00791045.2 ± 205.2
AReUReDi0.9756 ± 0.01370.9417 ± 0.03960.9995 ± 0.00070.9498 ± 0.02840.6705 ± 0.15990.6341 ± 0.0433−0.0743 ± 0.01521061.2 ± 192.9
L = 40PEGASUS0.9320 ± 0.04820.7990 ± 0.04320.9119 ± 0.08320.7470 ± 0.04710.7869 ± 0.05360.7452 ± 0.0419−0.0469 ± 0.00521568.2 ± 118.7
MOG-DFM0.9168 ± 0.03870.1479 ± 0.21960.8952 ± 0.23870.3806 ± 0.25030.7320 ± 0.10510.6834 ± 0.0925−0.1540 ± 0.02631509.2 ± 89.7
AReUReDi0.8473 ± 0.06670.2463 ± 0.33460.9110 ± 0.07700.5828 ± 0.10060.6358 ± 0.09700.6412 ± 0.0816−0.1452 ± 0.03311736.3 ± 83.8

Q3. PDO intrinsic dimension, query caps, and empirical stopping

To what extent does the exact-linear intrinsic-dimension theorem explain PDO's query savings when its implementation uses approximate response models and a query cap?

Thank you for this insightful question. The theorem identifies a fundamental source of query savings: under exact decision-linear responses, query complexity depends on unresolved decision-relevant directions rather than the number of candidate actions. Practical PDO approximates this principle through committee-based decision resolution. To assess whether its observed efficiency primarily arises from this mechanism or hard query-cap truncation, we conducted additional paired capped and uncapped replays on 96 frozen Hamming-1 decision states (Table Q3).

Over 95% of decisions terminated through committee-based stopping before reaching the query cap. Removing the cap preserved the selected action in 94/96 cases, with median uncapped query demand remaining at just one. Furthermore, our existing candidate-set scaling experiment (Table S20) showed that increasing the Hamming-1 candidate pool eightfold, from 64 to 512, increased mean fresh queries only from 2.52 to 3.38. PDO also improved accepted fine-stage gain under matched query allowances (Table S19).

These results show that practical PDO's sparse querying is primarily driven by decision-based stopping, with the query cap serving as a safeguard. The intrinsic-dimension theorem characterizes this principle under exact-linear assumptions, while our empirical results establish its practical relevance without assuming an exact dimension-based scaling law for the approximate implementation.

Table Q3: PDO Query-Cap Sensitivity on 96 Frozen Decision States (Table S38 in the revised manuscript).

Replay SettingCommittee-Resolved Stops ↑Mean Queries (Capped / Uncapped) ↓Uncapped Median Queries ↓Decisions Changed ↓
Production Prescreen92/96 (95.8%)1.22 / 1.7212/96
Full Hamming-193/96 (96.9%)1.20 / 2.2012/96

We hope that our new task-projected Koopman evaluation, comparable-budget comparison with MOG-DFM and AReUReDi, and PDO query-cap sensitivity experiments address the reviewer's concerns and further substantiate PEGASUS's contributions. We welcome any further questions during the discussion period and would greatly appreciate your support for acceptance!

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer XZ1g,

Thank you again for your detailed review and helpful suggestions.

Our additional experiments directly address your three questions. First, the new task-projected evaluation demonstrated Koopman's advantage over a generic linear observer, reducing prediction error by 32.6% and 15.8% at L = 12 and L = 40, respectively (Table S35). Second, at approximately comparable oracle-query expenditures, PEGASUS achieved substantially higher Top-1 utility than both MOG-DFM and AReUReDi at both sequence lengths, strengthening our original quality-efficiency findings (Table S37). Third, our PDO sensitivity experiments confirmed that over 95% of decisions terminate before reaching the query cap, with the selected action unchanged in 94/96 cases after removing the cap. This supports decision-based stopping as the primary source of practical query savings (Table S38).

We hope these new results address your questions and further support PEGASUS's contributions. We look forward to hearing your thoughts and would be happy to address any remaining concerns.

The Authors

Reviewer_V6YdOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PEGASUS, a multi-objective inference-time optimization framework for pretrained discrete flow maps. It estimates the terminal potential of intermediate states through future observability and combines multiscale sequence editing with Pareto Decision Observability to reduce expensive objective evaluations. The results demonstrate promising generation quality and query efficiency. However, the novelty relative to existing general-purpose optimization techniques remains unclear, and the experimental coverage does not sufficiently establish its general applicability.

Critical strengths
  1. Solid mathematical foundation. The method integrates finite-time Koopman representations, terminal objective estimation, and decision-oriented query selection into a coherent framework. Its query complexity analysis under explicit assumptions provides theoretical support for the algorithm.

  2. Practically relevant motivation. The paper distinguishes generation costs from objective evaluation costs and addresses an important limitation: few-step generation does not necessarily imply query-efficient optimization.

  3. Promising quality–efficiency trade-off. The reported results show substantial reductions in oracle queries and actual runtime while maintaining or improving terminal Pareto quality.

Critical weaknesses
  1. The novelty and attribution of the core components are insufficiently clear. Future observability has conceptual connections to terminal value estimation; Koopman transport and task readouts relate to dynamics representation learning; PDO relates to decision-oriented active querying; and multiscale editing resembles neighborhood search. It remains unclear which mechanisms are newly introduced, which are inherited from other settings, and which are adaptations to discrete flow generation. Clearer mathematical comparisons and controlled experiments are needed to establish the distinct contributions.

  2. General applicability is not sufficiently validated. The framework does not appear to rely on biology-specific knowledge, yet the experiments focus on biological sequences. At least one non-biological discrete generation task, such as multi-objective program generation or discrete design generation, would help assess its effectiveness under different constraints, objective structures, and generative dynamics.

  3. The coverage of biological tasks remains limited. Peptide and DNA experiments provide useful cross-task evidence, but broader biological applicability would benefit from tasks with substantially different representations or constraints, such as long protein sequence design, RNA design, or molecular graph generation. These would test robustness to sequence length, structural constraints, and edit validity.

  4. More direct optimization baselines and component-level controls are needed. Comparisons against conventional value-guided search, surrogate-assisted optimization, and active query selection should use the same generative backbone and query budget. Ablations should also separate the contributions of Koopman representations, PDO, multiscale actions, and the generative backbone.

Questions for authors

None

5 comments

We sincerely thank the reviewer for their careful and insightful review, particularly for recognizing PEGASUS's mathematical foundation, practical motivation, and strong quality-efficiency tradeoffs. We appreciate the constructive questions regarding methodological novelty, general applicability, and component attribution, which have motivated further clarification and additional experiments detailed below.


W1. Methodological novelty and component attribution

The novelty and individual contributions of PEGASUS's core components remain unclear relative to existing methods, requiring stronger mathematical distinctions and controlled experiments.

Thank you for encouraging a clearer articulation of PEGASUS's contributions. We identify three fundamental, interconnected challenges in query-efficient discrete multi-objective optimization: predicting reachable terminal quality from intermediate generative states, selecting effective actions across discrete scales, and allocating exact objective queries according to decision-relevant uncertainty. These challenges are closely coupled: prospective evaluation determines which actions are worth exploring, multiscale actions expand the range of possible improvements, and decision-aware acquisition determines which candidates require expensive evaluation. Addressing these challenges independently leaves important dependencies unresolved under limited oracle budgets. PEGASUS introduces a unified decision-centered optimization framework that identifies and jointly addresses these three challenges, coordinating future observability, multiscale actions, and Pareto Decision Observability within a single optimization procedure. To our knowledge, prior discrete MOO approaches have not explicitly identified and addressed these interconnected challenges as a unified optimization problem.

Future value estimation, Koopman dynamics, and multiscale search have established foundations in other fields, which PEGASUS builds upon and integrates for discrete MOO. PDO introduces a new mathematical formulation of decision observability, including an intrinsic-dimension exact-query bound (Theorem 3.3). Controlled ablations in Tables S9–S11 isolate the contributions of future observability, multiscale actions, and PDO, respectively, while our additional task-projected transport comparison evaluates Koopman's contribution (Table S35 in the revised manuscript). Section 2 (Related Work) further establishes the distinctions from prior methods.

Together, these theoretical and empirical results establish PEGASUS's central contribution: a unified decision-centered approach that identifies and coordinates three complementary capabilities for query-efficient multi-objective optimization.

W2. Non-biological discrete design: EDA

PEGASUS's general applicability remains insufficiently validated because its experiments focus on biological sequences without testing non-biological discrete optimization tasks.

Thank you for suggesting a non-biological evaluation. To assess cross-domain applicability, we evaluated PEGASUS on multi-objective electronic design automation (EDA) logic synthesis using three EPFL benchmark circuits: priority (priority encoder), max (maximum-selection circuit), and sin (sine-function circuit). Each candidate is a 20-step sequence of seven possible logic synthesis operations, evaluated on four hardware objectives: area, delay, switching, and fanout. This task introduces distinct discrete transformation dynamics, circuit constraints, and objective relationships compared with biological sequence design (Appendix I).

PEGASUS achieved the highest mean utility and top-10 utility across all three circuits, and the highest terminal-100 hypervolume in two of the three settings, with competitive runtime (Table W2). These results demonstrate that PEGASUS's optimization mechanisms transfer effectively beyond biological sequence design to non-biological discrete optimization with distinct action semantics and objective structures, supporting its broader applicability to query-efficient multi-objective generation.

Table W2: Multi-Objective EDA Logic Synthesis under Comparable Oracle-Query Budgets Across Three Seeds (Table S39 in the revised manuscript).

CircuitMethodTerminal-100 HV ↑Mean Utility ↑Top-10 Utility ↑Runtime (min) ↓Unique Queries ↓
priorityCB-EVO3.185 ± 0.056−0.00962 ± 0.006010.00122 ± 0.0000726.97 ± 0.851,150 ± 55
priorityFlowTune3.002 ± 0.165−0.01219 ± 0.00086−0.00114 ± 0.0019117.43 ± 0.971,290 ± 74
priorityLKF Best-of-N2.925 ± 0.112−0.01463 ± 0.00168−0.00197 ± 0.0018211.29 ± 0.401,150 ± 55
priorityPEGASUS3.189 ± 0.0750.00075 ± 0.000070.00125 ± 0.0000623.02 ± 0.701,150 ± 55
maxCB-EVO2.082 ± 0.084−0.11707 ± 0.00078−0.07021 ± 0.00116106.19 ± 11.841,304 ± 128
maxFlowTune1.771 ± 0.103−0.10326 ± 0.00296−0.06504 ± 0.0045891.59 ± 0.141,204 ± 0
maxLKF Best-of-N1.741 ± 0.112−0.10519 ± 0.00262−0.07187 ± 0.0079268.38 ± 6.721,304 ± 128
maxPEGASUS1.749 ± 0.166−0.08253 ± 0.00557−0.05855 ± 0.00128110.81 ± 10.461,304 ± 128
sinCB-EVO1.810 ± 0.167−0.16317 ± 0.00676−0.08418 ± 0.00814248.37 ± 5.901,143 ± 35
sinFlowTune2.666 ± 0.365−0.09197 ± 0.00300−0.05418 ± 0.00907338.20 ± 0.741,204 ± 0
sinLKF Best-of-N2.641 ± 0.135−0.09040 ± 0.00120−0.05519 ± 0.00290232.35 ± 6.991,143 ± 35
sinPEGASUS2.674 ± 0.198−0.04871 ± 0.00266−0.02971 ± 0.00339317.81 ± 10.261,143 ± 35

W3. Molecular graph generation and cross-backbone transfer

The peptide and DNA experiments provide limited evidence of generalization to biological modalities with substantially different representations, structural constraints, and edit validity requirements.

Thank you for encouraging us to evaluate PEGASUS on a broader biological design modality. We selected molecular graph generation because explicit atom–bond connectivity and chemical validity constraints introduce substantially different representations and challenges from peptide and DNA sequences. We evaluated PEGASUS with two graph-native generators, CatFM (categorical flow maps) and DeFoG (discrete flow matching), across three optimization settings, with experimental details provided in Appendix J.

Under matched oracle-query budgets, both PEGASUS implementations outperformed SVDD-PM and SMC across all four optimization-quality metrics in all three settings (Table W3), with additional improvements over their respective same-backbone Best-of-N controls. DeFoG + PEGASUS achieved the highest archive HV and HV-AUC, while CatFM + PEGASUS achieved the highest mean utility across all three settings. DeFoG + PEGASUS had longer estimated runtime than CatFM + PEGASUS, consistent with the greater model-side computation required by DeFoG's multi-step flow-matching sampler. The ZINC benchmarks additionally include computationally expensive protein–ligand docking objectives. By achieving superior Pareto quality and utility within the same limited oracle-query budgets, PEGASUS demonstrated its ability to make more effective use of costly evaluations across distinct molecular generators. Together, these results establish PEGASUS's effectiveness across molecular graph representations, target-conditioned objectives, and distinct generative dynamics.

Table W3: Query-Matched Multi-Objective Molecular Graph Optimization at 1,000 Oracle Queries per Setting (Table S43 in the revised manuscript).

SettingMethodArchive HV ↑HV-AUC ↑Mean Utility ↑Top-10 Utility ↑Runtime (min) ↓
QM9SVDD-PM0.565 ± 0.0070.539 ± 0.009−0.1684 ± 0.0037−0.1492 ± 0.003534.07 ± 9.16
QM9SMC0.575 ± 0.0040.534 ± 0.021−0.1682 ± 0.0029−0.1485 ± 0.001111.31 ± 11.53
QM9CatFM Best-of-N0.558 ± 0.0110.529 ± 0.009−0.1787 ± 0.0095−0.1709 ± 0.01780.51 ± 0.03
QM9DeFoG Best-of-N0.583 ± 0.0270.540 ± 0.016−0.1863 ± 0.0086−0.1491 ± 0.00155.09 ± 0.06
QM9CatFM + PEGASUS0.638 ± 0.0110.601 ± 0.007−0.1450 ± 0.0024−0.1210 ± 0.003310.93 ± 0.79
QM9DeFoG + PEGASUS0.657 ± 0.0010.612 ± 0.005−0.1455 ± 0.0033−0.1273 ± 0.002922.27 ± 4.23
ZINC/PARP1SVDD-PM0.471 ± 0.0170.465 ± 0.015−0.1795 ± 0.0059−0.1646 ± 0.0065334.63 ± 28.39
ZINC/PARP1SMC0.467 ± 0.0060.460 ± 0.007−0.1825 ± 0.0055−0.1680 ± 0.0065336.92 ± 27.86
ZINC/PARP1CatFM Best-of-N0.481 ± 0.0010.469 ± 0.002−0.1796 ± 0.0016−0.1585 ± 0.0012210.81 ± 8.92
ZINC/PARP1DeFoG Best-of-N0.504 ± 0.0100.481 ± 0.005−0.1812 ± 0.0026−0.1573 ± 0.0025319.91 ± 3.17
ZINC/PARP1CatFM + PEGASUS0.521 ± 0.0220.496 ± 0.014−0.1671 ± 0.0031−0.1426 ± 0.0060308.26 ± 57.23
ZINC/PARP1DeFoG + PEGASUS0.538 ± 0.0080.511 ± 0.009−0.1682 ± 0.0047−0.1423 ± 0.0009489.68 ± 48.04
ZINC/JAK2SVDD-PM0.441 ± 0.0200.425 ± 0.013−0.1850 ± 0.0042−0.1712 ± 0.0047430.52 ± 12.08
ZINC/JAK2SMC0.437 ± 0.0120.417 ± 0.008−0.1907 ± 0.0025−0.1824 ± 0.0032380.00 ± 67.01
ZINC/JAK2CatFM Best-of-N0.448 ± 0.0050.434 ± 0.005−0.1869 ± 0.0024−0.1694 ± 0.0022211.17 ± 6.47
ZINC/JAK2DeFoG Best-of-N0.472 ± 0.0030.448 ± 0.001−0.1903 ± 0.0013−0.1704 ± 0.0004487.70 ± 305.13
ZINC/JAK2CatFM + PEGASUS0.485 ± 0.0280.456 ± 0.021−0.1782 ± 0.0091−0.1541 ± 0.0107309.48 ± 81.33
ZINC/JAK2DeFoG + PEGASUS0.508 ± 0.0170.478 ± 0.014−0.1805 ± 0.0033−0.1573 ± 0.0028456.76 ± 47.63
Reviewer_t2kDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces PEGASUS, an inference-time framework for efficient, multi-objective biological sequence optimization of discrete generative models. The method combines predictions of terminal objective values from intermediate generation states, sequence edits at multiple scales, and selective property evaluations. The authors evaluate the approach on peptide and DNA design tasks. PEGASUS achieves competitive or improved predicted Pareto quality with substantially fewer property-model evaluations than the compared baselines.

Critical strengths

PEGASUS substantially reduces the number of oracle queries needed to achieve competitive or improved multi-objective performance, addressing an important bottleneck in biological sequence design. I also appreciated the extensive experiments, including comparisons with multiple optimization baselines, component ablations, and evaluations on both peptide and DNA design tasks.

Critical weaknesses

The contribution of the specific Koopman formulation is unclear, since a simpler linear observer performs competitively and sometimes better.

Also, the biological validation remains computational, with mixed agreement across independent property predictors.

Questions for authors

What practical advantage does the Koopman observer offer over the simpler linear observer?

2 comments

We sincerely thank the reviewer for their thoughtful and constructive assessment, particularly for recognizing PEGASUS's substantial improvements in query efficiency and its extensive evaluations across peptide and DNA design. We appreciate the questions regarding Koopman's specific contribution and biological validation, which we address individually below.


W1 + Q1. Koopman formulation and finite-time transport

The contribution of the specific Koopman formulation is unclear, since a simpler linear observer performs competitively and sometimes better.

Thank you for highlighting the importance of distinguishing Koopman transport from a generic linear observer. The contribution of our formulation is not linearity itself, but objective-independent finite-time transport in learned Koopman coordinates, with task-specific readouts applied afterward. Since end-to-end optimization metrics also depend on candidate selection and query allocation, we additionally evaluated held-out task-projected transport error to directly assess the practical contribution of future objective-relevant predictions. As shown in Table W1, Koopman reduces this error relative to the generic linear observer by 32.6% at L = 12 and 15.8% at L = 40, with positive paired improvements whose 95% confidence intervals exclude zero. Without transport, errors increase substantially to 0.663 and 0.816, respectively. These results establish a concrete predictive advantage of our Koopman formulation in transporting terminal-objective-relevant information, validating its role in PEGASUS's Future Observability framework while preserving the separation between reusable, objective-independent dynamics and task-specific prediction.

Table W1: Task-Projected Transport Error of Future Observers (Table S35 in the revised manuscript).

LengthKoopman ↓Generic Linear ↓No Transport ↓Error Reduction ↑Paired ΔError [95% CI] ↑
L = 120.05730.08500.663032.6%0.0277 [0.0276, 0.0278]
L = 400.03580.04250.816015.8%0.0068 [0.0065, 0.0071]

W2. Independent biological-property validation

Also, the biological validation remains computational, with mixed agreement across independent property predictors.

Thank you for emphasizing the importance of biological validation. To assess whether PEGASUS designs retain favorable properties beyond the optimization oracles, we performed orthogonal computational evaluations using independent property predictors excluded from optimization (Table S7). PEGASUS designs achieved 2.08× top-quartile enrichment for non-fouling under PeptideBERT (L = 12) and 1.60×/1.92× enrichment for penetrance under PERSEUcpp (L = 12/40), with additional favorable agreement for non-hemolysis under HemoPI2. We further assessed peptide–target interactions using five independent AlphaFold3 predictions per complex and complementary HADDOCK docking (Figs. 2 and S1). Representative AMHR2 and MYC complexes both achieved ipTM scores of 0.81, supporting plausible peptide–target interfaces.

Cross-predictor agreement varied across properties, reflecting differences in predictor formulations, training distributions, and biological endpoints. In particular, CamSol evaluates intrinsic physicochemical solubility using a different formulation from PeptiVerse, while PLifePred2 is not calibrated to the human-serum half-life endpoint used in our optimization. These differences help explain the weaker agreement observed for solubility and half-life. Together, the independent property evaluations and structural analyses provide additional computational support for PEGASUS-designed peptides, particularly for non-fouling, penetrance, and plausible target interactions. Experimental characterization remains necessary to establish biological activity, and we distinguish this validation requirement from PEGASUS's demonstrated contributions to query-efficient multi-objective optimization.


We hope that our additional task-projected transport experiments and independent biological property evaluations address the reviewer's concerns and further strengthen the evidence supporting PEGASUS's methodological contributions. We welcome any further questions during the discussion period and would greatly appreciate your support for acceptance!

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer t2kD,

Thank you again for your careful review and constructive feedback.

Our new task-projected transport evaluation directly establishes Koopman's predictive advantage over a generic linear observer, reducing prediction error by 32.6% at L = 12 and 15.8% at L = 40 (Table S35). This provides quantitative evidence for Koopman's specific contribution to future observability. Regarding biological validation, our independent property evaluations showed favorable agreement for non-fouling, penetrance, and non-hemolysis, complemented by structural analyses supporting plausible peptide–target interactions (Table S7; Figs. 2, S1).

We hope these additional results address your concerns and strengthen the evidence supporting PEGASUS. We look forward to hearing your thoughts and would be happy to address any remaining questions.

The Authors

Comments on the paper

2 comments

Dear Reviewers,

We are nearing the end of the discussion phase. If you haven't already done so, please read the authors' responses, reply to acknowledge their feedback, and indicate whether your initial concerns have been addressed (along with any score adjustments, if warranted).

Thank you for your dedicated efforts and valuable insights!

Best,

Your AC

Dear Area Chair,

As the rebuttal period closes, Reviewer XZ1g has confirmed that our additional experiments substantially addressed their concerns and has maintained a Weak Accept rating. We subsequently provided a further clarification addressing their remaining question about the Koopman versus linear observer comparison in Table S12. Reviewers t2kD and V6Yd have not yet responded to our detailed rebuttals and follow-up comments. We appreciate all reviewers' time and summarize their positive assessments and the evidence addressing their concerns.


Strengths noted by the reviewers

  • Query efficiency and Pareto quality. All reviewers highlighted the substantial reduction in oracle queries while maintaining competitive or better Pareto quality, especially for longer peptides (t2kD, V6Yd, XZ1g).
  • Practical motivation. The reviewers recognized expensive objective evaluations as an important bottleneck and appreciated the distinction between fast generation and query-efficient optimization (t2kD, V6Yd).
  • Mathematical foundation. V6Yd praised the coherent combination of Koopman transport, terminal objective estimation, decision-oriented querying, and query-complexity analysis.
  • Experimental breadth. t2kD and XZ1g noted the extensive peptide and DNA benchmarks, baseline comparisons, ablations, generator substitutions, and query accounting.
  • Soundness and clarity. t2kD and V6Yd found no substantive soundness issues; V6Yd and XZ1g praised the paper's clarity.

Revisions addressing the reviewers' concerns

  1. Methodological novelty and component attribution (V6Yd). PEGASUS identifies three fundamental, interconnected challenges in query-efficient discrete MOO: predicting reachable terminal quality, selecting effective actions across discrete scales, and allocating exact queries according to decision-relevant uncertainty. Our central contribution is a unified decision-centered optimization framework that jointly addresses all three challenges, supported by a new mathematical formulation of PDO and controlled component-level evidence.

  2. Koopman's contribution (t2kD, XZ1g). Our dimension-matched task-projected transport experiments directly establish Koopman's predictive advantage over a generic linear observer. We further clarified that Table S12's end-to-end comparison used unmatched feature dimensions and representation-dependent readouts, preventing it from isolating Koopman transport's contribution. The No-Future ablation independently supports the optimization benefit of Koopman-based future observability.

  3. Baseline fairness (V6Yd, XZ1g). Same-backbone, query-matched comparisons isolate PEGASUS's gains beyond the generator and conventional surrogate optimization. Comparable-query tests against MOG-DFM and AReUReDi confirm superior best-available-candidate quality.

  4. Applicability across domains and representations (V6Yd). Non-biological EDA benchmarks demonstrate PEGASUS's advantages beyond biomolecular design. Molecular graph results support effectiveness across chemical constraints and generative backbones.

  5. Practical PDO efficiency (XZ1g). Our query-cap sensitivity experiments confirm that PDO's practical query savings arise primarily from decision-based stopping, with the query cap serving as a safeguard. These findings support the practical relevance of decision-selective acquisition and clarify its relationship to the intrinsic-dimension theory.

  6. Independent biological assessment (t2kD). Independent property evaluations and structural analyses provide complementary computational evidence supporting PEGASUS-designed peptides and motivate future collaborative experimental studies to assess their biological activity.


Final Request

We respectfully ask that the final assessment give full consideration to the new evidence addressing all three reviews, particularly the theoretical and experimental results addressing V6Yd's four concerns. We understand that reviewer availability may be limited and welcome further feedback. We hope the absence of follow-up from t2kD and V6Yd will not diminish consideration of our completed rebuttal.

We believe our additional experiments and clarifications have substantially strengthened the evidence for PEGASUS's contributions, and sincerely hope the final assessment will support its acceptance.

Thank you 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.