acceptodds
Under review as a conference paper at ICLR 2027

OmniPepBuilder: Unified Peptide Binder Design via Systematic Exploration of Binding Landscapes

Abstract

Peptides offer a versatile modality for modulating protein–protein interactions, but scarce complex structures constrain design across diverse cyclization chemistries. Training-free methods repurpose binding-related knowledge encoded in pretrained structure predictors, yet peptide structure prediction and scoring remain insufficiently benchmarked, and efficient exploration of diverse poses remains challenging on difficult targets. We begin with rigorous structural and binding benchmarks to improve prediction and reassess scoring, finding that confidence and Rosetta scores are more informative in controlled mutation series than in heterogeneous screening. We therefore introduce OmniPepBuilder, a unified, training-free two-stage framework for linear and cyclic peptide binder design. To address limited pose diversity and sampling efficiency, global proposal combines state replay, dynamic structural cell partitioning and revisit penalties for adaptive, systematic exploration of binding basins. Local refinement then uses physics-based scoring to optimize interfaces within selected basins. Computational benchmarks show pronounced gains in pose discovery on difficult targets, while local refinement further improves predicted interactions in already confident complexes. Encouraged by this exploration capability, we extend the framework to multi-target binders and macrocyclic molecular glues, where a single peptide must satisfy multiple interface constraints. The putative designs accommodate distinct target pockets in separate complexes or bridge protein partners in diverse ternary poses, demonstrating efficient exploration under the joint geometric constraints of these demanding applications.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
peptide binder design, cyclic peptides, inference-time search, binding landscape exploration, physics-guided optimization, multi-target design, macrocyclic glues

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cYmVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper presents OmniPepBuilder, a training-free framework for designing linear and cyclic peptide binders. The proposed framework consists of two stages: global exploration of peptide–protein binding poses and local sequence refinement within selected pose basins. The main methodological contribution is a Go-Explore-inspired search strategy that combines state replay, dynamic structural cell partitioning, and revisit penalties to improve exploration efficiency and pose diversity. Local refinement subsequently employs ProteinMPNN-based mutation proposals and Rosetta interface energy to optimize predicted interactions while preserving the binding pose.

The paper also includes preliminary benchmarks of cyclic peptide structure representations and binding-affinity scoring metrics. The proposed method is evaluated across eight protein targets and five peptide topologies against several existing binder design approaches. Additional computational applications to multi-target peptide binders and macrocyclic molecular glues are presented.

Overall, the paper covers a broad range of computational experiments, but the connection between these experiments and the central methodological contributions is not always sufficiently established.

Critical strengths

The paper addresses an important problem in peptide binder design and presents a practical framework for improving the exploration of diverse binding poses. The experiments cover multiple protein targets, peptide topologies, and existing design methods. The structural representation and binding-scoring benchmarks provide useful empirical insights, and the implementation details are generally well documented.

Critical weaknesses

Despite the extensive experiments, the validation is not sufficiently focused on the central methodological contributions. The proposed exploration strategy mainly combines existing techniques, while its advantages over simpler search alternatives remain unclear. The main performance comparisons rely on linearly extrapolated pose-cluster counts, and he global-exploration benchmark and its policy ablations rely on single-seed trajectories, without repeated runs or statistical analysis. Furthermore, the reported improvements largely depend on computational confidence metrics, with limited independent validation of binding plausibility. The local refinement and multi-target applications are also not systematically evaluated as a complete framework. Consequently, the title's emphasis on Unified Peptide Binder Design may overstate the demonstrated scope.

Questions for authors
  1. While the no-replay ablation demonstrates the value of reusing intermediate states, the benefit of the specific allocation strategy remains unclear. Could the authors compare against simpler alternatives, such as uniform or confidence-based replay, under the same computational budget?
  2. Figure 4 relies on linearly extrapolated pose-cluster throughput, although the number of distinct clusters may not scale linearly with sample size. Could the authors report the observed results without extrapolation and clarify whether the main conclusions remain consistent under matched evaluation budgets?
  3. Appendix C.4 reports substantial disagreement between AF3 and Protenix-v2, particularly for isopeptide designs. Could the authors clarify how much of this disagreement arises from cyclization geometry versus prediction confidence, and provide further evidence supporting the structural validity of these designs?
  4. The global-exploration ablations rely on single-seed trajectories, and the single-cell ablation modifies multiple selection criteria. Could the authors provide results across multiple independent seeds and more isolated ablations to establish the robustness of the proposed exploration mechanisms?
4 comments
AuthorsOpenReview ↗

Response to Reviewer cYmV (1/2)

We thank the reviewer for the detailed feedback on evaluation and the exploration policy. We address the main concerns in four points below.

1. Full-pool evaluation without linear extrapolation

Figure 4 relies on linearly extrapolated pose-cluster throughput, although the number of distinct clusters may not scale linearly with sample size. Could the authors report the observed results without extrapolation and clarify whether the main conclusions remain consistent under matched evaluation budgets?

We thank the reviewer for identifying the extrapolation issue. Previously, resource constraints led us to evaluate subsampled pools for RFpeptides, BoltzGen and AnewOmni. Because additional candidates can belong to existing pose groups, linear extrapolation overestimated the baselines' diversity. We now evaluate their complete output pools and report observed counts. The updated full-pool benchmark preserves a 2.5–9.7-fold pooled-throughput advantage over all five external baselines, on each comparator's supported cases (Fig. 4a). All methods use the common 12-hour, single-GPU generation allowance, with throughput calculated from observed accepted pose groups divided by actual generation GPU-hours (Appendix C.3.3).

2. Controlled ablations and the exploration policy

While the no-replay ablation demonstrates the value of reusing intermediate states, the benefit of the specific allocation strategy remains unclear. Could the authors compare against simpler alternatives, such as uniform or confidence-based replay, under the same computational budget?

Confidence-based replay was already included as the original single-cell ablation; we have rewritten its description to clarify this comparison. We have also added Uniform replay, with all control settings summarized in Table 4. Both simpler replay policies yield significantly fewer accepted pose groups than our full policy: under the same 12-hour generation budget, they reduce median yield by 29.9% and 12.0%, respectively, across 14 cases spanning seven targets and two topologies (Fig. 5). The exploratory paired-test p-values are 0.000244 and 0.00537 (Appendix C.4.4). These results show that replay alone is insufficient: how states are retained and replay effort is allocated matters.

The proposed exploration strategy mainly combines existing techniques, while its advantages over simpler search alternatives remain unclear.

We would like to argue that our contribution is a new pose-aware cell-partitioning and replay-allocation scheme, well motivated by the need to discover diverse binding basins. Confidence-only optimization can concentrate sampling in readily accessible regions. We therefore compare candidates within local pose neighborhoods, retaining promising alternatives across basins and dynamically grouping them into cells (Section 3.1; Appendix C.1). Unlike the single representative retained per cell in original Go-Explore [1], every retained local Pareto state remains available for replay. Our allocation then balances proximity to confidence thresholds with cell size and return history: densely populated cells do not gain priority solely through multiplicity, and repeated returns shift relative priority toward less explored regions.

The global-exploration ablations rely on single-seed trajectories, and the single-cell ablation modifies multiple selection criteria. Could the authors provide results across multiple independent seeds and more isolated ablations to establish the robustness of the proposed exploration mechanisms?

We would like to note that we have added exact paired significance tests across the evaluated target–topology cases, supporting systematic trends across multiple systems (Appendix C.4.4). All global-policy ablations used the fixed seed 2026, with no manual cherry-picking. Due to computational resource limitations, multi-seed runs are still ongoing; we will add their results when completed.

Regarding the request for more isolated ablations, all comparisons use the same prediction models, acceptance criteria and computational budget (Table 4; Appendix C.3.4). In the original single-cell control, Pareto selection becomes global and the common cell-size and history factors cancel, yielding confidence-based replay. This tests the contribution of spatial organization. No revisit penalty separately isolates the history term while preserving local Pareto retention, cell balancing and confidence weighting. Its aggregate effect is small, but retaining the penalty yields 25–35% more pose groups on both KRAS G12D topologies and head-to-tail CXCR4 (Fig. 18).

[1] Ecoffet et al., First return, then explore. Nature, 2021.

AuthorsOpenReview ↗

Response to Reviewer cYmV (2/2)

3. Orthogonal prediction and structural validity

Appendix C.4 reports substantial disagreement between AF3 and Protenix-v2, particularly for isopeptide designs. Could the authors clarify how much of this disagreement arises from cyclization geometry versus prediction confidence, and provide further evidence supporting the structural validity of these designs?

Furthermore, the reported improvements largely depend on computational confidence metrics, with limited independent validation of binding plausibility.

We would like to note that OmniPepBuilder retains its advantage under orthogonal prediction and structural validity checks. Requiring acceptance by both AF3 and Protenix v2 preserves a 3.1–12.7-fold pooled-throughput advantage on matched cases (Fig. 15a). Separately, PoseBusters-style clash screening followed by Rosetta peptide-energy screening preserves a 2.6–9.4-fold advantage (Fig. 15b). To assess binding plausibility beyond predictor confidence, we also evaluated Rosetta separated-interface energies, which are more favorable than those of external baselines in most matched cases (Fig. 17; Appendix C.4.3).

For isopeptides, linkage geometry accounts for most Protenix rejections. Among 1,738 AF3-accepted designs across generators, 1,728 fail Protenix linkage-distance checks and 826 fail confidence thresholds (Appendix C.4.3). The issue extends across generators, including 580/581 BoltzGen isopeptides. Our native-complex benchmark also finds that none of the 30 selected Protenix v2 predictions meets the isopeptide distance-closure criterion (Fig. 12; Appendix A.3). Together, these findings suggest that limited isopeptide-closure accuracy in Protenix v2 contributes to the disagreement, which is not specific to OmniPepBuilder.

We acknowledge that these metrics do not yet provide a well-calibrated estimate of experimental binding success. We are therefore pursuing experimental binding validation.

4. Local refinement and multi-target applications

The local refinement and multi-target applications are also not systematically evaluated as a complete framework.

Both local refinement and multi-target/glue design have been evaluated computationally. Local refinement has a matched benchmark across six targets (Section 3.2; Appendix D.2), and Fig. 6 shows refinement of AF3-generated candidates. Section 4 and Appendix E.1 evaluate the global-proposal extension to multi-target binders and glues across seven campaigns (Table 6), using shared sequences, joint pose neighborhoods and acceptance at every required interface.

We acknowledge that these computational evaluations do not yet establish experimental binding or activity. As discussed in the Introduction, peptide design still lacks broadly validated metrics and filters across therapeutic targets. We therefore call on the community to build large, consistently measured peptide-design datasets, following efforts in minibinder [2] and antibody design [3], to calibrate screening against experimental outcomes. For E3-recruiting glues, predicted ternary compatibility alone does not establish productive ubiquitination or degradation; we are pursuing wet-lab validation of these functional outcomes.

We did not include further baseline comparisons because relevant approaches, including Odin-Multi [4] and EvoBind-multimer [5], emerged concurrently with our work; we plan to benchmark them in follow-up work. We do note that EvoBind-multimer has a substantial per-trajectory computational cost: its reported 1,000-step trajectory takes 8–11 hours on one A100 40GB GPU. Odin-Multi reports a median of 23.3 minutes per trajectory for its two-toxin task on one T4. For context, our MDM2/MDMX and VHL–BRD4 campaigns yield 134 accepted designs in 42 pose groups and 73 in 19 groups, respectively, in 5.34 and 11.31 hours on four L40S GPUs (Table 6). These campaign outcomes demonstrate efficient discovery of diverse predicted poses in the evaluated multi-target and glue applications.

[2] Cao et al., Design of protein-binding proteins from the target structure alone. Nature, 2022.

[3] Bennett et al., Atomically accurate de novo design of antibodies with RFdiffusion. Nature, 2026.

[4] Brasas et al., Specificity-driven protein binder design with Odin-Multi. bioRxiv, 2026.

[5] Brunner et al., De novo Design of Macrocyclic Molecular Glues. bioRxiv, 2026.

Thank you for the detailed rebuttal and substantial additional experiments. I appreciate the full-pool evaluation without linear extrapolation, the new comparisons with simpler replay strategies, and the additional structural validity checks. These results address several of my main concerns and provide stronger evidence for the proposed exploration strategy.

I still have some reservations about the statistical robustness of the single-seed exploration experiments and the limited end-to-end validation of the complete framework. However, I acknowledge that the revised manuscript presents a clearer methodological contribution with more convincing experimental support. I am therefore inclined to raise my score from 2 to 3.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you for your thoughtful reassessment and for raising your score. We appreciate your constructive feedback, which helped us strengthen the evaluation and clarify our methodological contribution.

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

This paper propose a comprehensive pipeline for designing cyclic peptides based on co-folding models. The authors first explore the effects of different featurizers for structure prediction of different cyclic topology. Then the authors adopt the hallucination-based design style, and propose global proposal mechanism with state replay, pose clustering, and revisit penalty. After the first stage, a second stage of local refinement is introduced to optimize the binding interaction. The model is benchmarked on 8 targets and 5 cyclic topolgies, with throughput of success candidates filtered by AF3 higher than strong baselines. The authors further extend the framework to multi-target peptide and macrocyclic molecular glues.

Critical strengths

The paper is comprehensive and well-structured, proposing a systematic framework for cyclic peptide design. The logics are quite complete, from representation, scoring models, to go-explore methodologies. The benchmarks include a wide range of state-of-the-art baselines across diverse targets and peptide topolgies. Overall, this is a relatively complete and substantial piece of work.

Critical weaknesses

The most problematic thing is that the generator and the evaluator share the same co-folding models in Figure 4, which is the setting that the proposed method show great improvement overall baselines. Such advantage disappears when changing the evaluator to an orthogonal one, as in Figure 14. Therefore, the improvement might originate from hacking the evaluator during the generative process instead of from the proposed methodology itself.

Questions for authors
  1. Why are intermediate candidates from HalluDesign excluded from evaluation, while OmniPepBuilder retains successful intermediate states throughout exploration?
  2. Why is the number of discovered pose clusters extrapolated linearly from a subsampled candidate pool? Would this overestimate diversity, given that newly generated candidates may belong to existing clusters?
  3. Does replaying a previously discovered state actually encourage exploration within its original binding basin? Could the authors quantify the probability of remaining in or transitioning between basins after replay?
4 comments
AuthorsOpenReview ↗

Response to Reviewer VHTB

We thank the reviewer for the careful assessment and the questions on evaluation and replay behavior.

1. Full-pool evaluation without linear extrapolation

Why is the number of discovered pose clusters extrapolated linearly from a subsampled candidate pool? Would this overestimate diversity, given that newly generated candidates may belong to existing clusters?

We thank the reviewer for pointing this out. Resource constraints previously led us to evaluate subsampled pools for RFpeptides, BoltzGen and AnewOmni. Because additional candidates can enter existing pose groups, linear extrapolation overestimated the baselines' pose diversity. We now report observed counts from complete pools. Under the common 12-hour generation allowance, throughput is accepted pose groups per actual generation GPU-hour (Appendix C.3.3). Fig. 4a shows that OmniPepBuilder substantially outperforms all five external baselines, with 2.5–9.7-fold higher pooled throughput on each comparator's supported cases.

2. Evaluation beyond AF3 confidence

The most problematic thing is that the generator and the evaluator share the same co-folding models in Figure 4, which is the setting that the proposed method show great improvement overall baselines. Such advantage disappears when changing the evaluator to an orthogonal one, as in Figure 14. Therefore, the improvement might originate from hacking the evaluator during the generative process instead of from the proposed methodology itself.

With complete output pools, joint AF3/Protenix v2 acceptance preserves a 3.1–12.7-fold pooled-throughput advantage over external baselines on matched cases (Fig. 15a). Separately, PoseBusters-style clash screening followed by Rosetta peptide-energy screening preserves a 2.6–9.4-fold advantage (Fig. 15b). Rosetta separated-interface energies, which were not optimized during global proposal, are more favorable than those of external baselines in most matched cases (Fig. 17; Appendix C.4.3).

The original orthogonal evaluation used Protenix alone, which was more permissive on confidence metrics for the evaluated baseline candidates. OmniPepBuilder had already incurred stringent AF3 filtering costs during generation. We therefore apply the same joint AF3/Protenix requirement to every method (Appendix C.3.3), consistent with recent multi-predictor minibinder assessment [1].

3. Replay behavior and the role of intermediate states

Does replaying a previously discovered state actually encourage exploration within its original binding basin? Could the authors quantify the probability of remaining in or transitioning between basins after replay?

We thank the reviewer for this insightful suggestion. We systematically analyzed 6,507 replays across twelve randomly selected cases spanning different levels of search difficulty. We observed source-local Pareto gains (improvements to retained candidates within the replay source’s pose neighborhood), showing that replay supports local exploration. Interestingly, non-local Pareto gains (improvements in other pose neighborhoods) occur approximately twice as often as source-local gains (26.1% versus 13.1% of replays; Fig. 14). For immediate source–first-child transitions, 43.9% of 5,244 observed pairs remain source-local and 56.1% are non-local (Appendix C.4.2).

We hypothesize that sequence redesign preserves a local binding anchor while the surrounding backbone reorganizes. Supporting this hypothesis, the 1,697 source–non-local-gain pairs show a median target-frame backbone RMSD of 6.08 Å, while descendants retain 53.1% of the source hotspot–peptide-position contacts on average (Appendix C.4.2). Descendants can improve the candidate collection in another pose cell even when their confidence is lower than that of the replay source. Intermediate states thus support systematic landscape exploration beyond isolated trajectory optimization.

Why are intermediate candidates from HalluDesign excluded from evaluation, while OmniPepBuilder retains successful intermediate states throughout exploration?

Following the discussion above, we therefore save accepted designs as they arise and reuse eligible intermediate states to launch alternative continuations as the design choice (Section 3.1).

For reproducibility, we used the public HalluDesign implementation without further modifications (Appendix C.3.2). Regarding the concern for a fair comparison, our No replay control follows the same independent AF3–ProteinMPNN trajectory principle within our optimized infrastructure (Appendix C.2). The full policy achieves 26% higher pooled throughput than No replay (Section 3.1; Fig. 4a; Fig. 5; Table 4). Both policies retain accepted candidates at any cycle, isolating replay's benefit beyond retaining intermediate successes.

[1] Claude Science and Amir Shanehsazzadeh, Autonomous de novo protein binder design with Claude. Anthropic Technical Report, 2026.

AuthorsOpenReview ↗

Gentle Follow-up on Our Response

Thank you again for your thoughtful review. With the discussion deadline approaching, we wanted to gently follow up on our response. We would be grateful for your feedback on whether our revisions address your concerns, and would be happy to clarify any remaining points.

Reviewer_VHTBOpenReview ↗

Thanks

Thanks for the responses! I think they are helpful and within expectation, therefore I will maintain my positive recommendation.

AuthorsOpenReview ↗

Thank you for your feedback

We are glad that our responses addressed your concerns. Thank you for your insightful comments, which helped us further refine and strengthen the paper.

Reviewer_bVY7OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposed OmniPepBuilder, a two-stage test-time exploring workflow for target-conditioned peptide (including cyclic peptide) design. The method input topological information as control for cyclization modes. In first stage, AF3 and MPNN iteratively update sequence to update confidence, pose-aware replay, structural clustering, and revisit penalties are also involved to decide which trajectory to explore. Then in second stage rosetta is involved to select low energy candidates.

Critical strengths
  1. Successfully transmitted go-explore mind into hallucination design workflow, optimized computational budgets allocation and significantly raised scores in complex targets.
  2. The workflow is training free and covered different cyclization modes. It may also be capable of transmitted to similar folding models.
Critical weaknesses
  1. "Success" is highly relied on single confidence predictor (AF3), and isopeptide fails in Protenix evaluation. In main experiment, clash rejection is not considered, it may result in false positive candidates.
  2. In ablation, each case is evaluated using only a single seed trajectory. Repeated evaluations with more random seeds are needed to improve the robustness of reported results.
Questions for authors
  1. After enabling clash rejection and applying Proteinix simultaneously verifications, how many valid candidates and diverse outputs remain? Does the method's advantages still hold true?
  2. How should we initiate sequence? Since AF3 generation itself (not workflow) is not guided by hotspots, does this risk cold-start failure or waste computational resources in warm-up steps?
3 comments
AuthorsOpenReview ↗

Response to Reviewer bVY7

We thank the reviewer for recognizing the value of adaptive exploration and for the suggestions on validation and initialization.

1. Orthogonal prediction and structural validity checks

"Success" is highly relied on single confidence predictor (AF3), and isopeptide fails in Protenix evaluation. In main experiment, clash rejection is not considered, it may result in false positive candidates.

After enabling clash rejection and applying Proteinix simultaneously verifications, how many valid candidates and diverse outputs remain? Does the method's advantages still hold true?

Additional Protenix v2 acceptance preserves a 3.1–12.7-fold pooled-throughput advantage over external baselines on matched cases (Section 3.1; Fig. 15a). Separately, PoseBusters-style clash screening followed by Rosetta peptide-energy screening preserves a 2.6–9.4-fold advantage (Fig. 15b; Appendix C.3.3).

For isopeptides, Appendix C.4.3 reports linkage-distance failures in 1,728 of 1,738 AF3-accepted designs across generators. Our native-complex benchmark (Fig. 12; Appendix A.3) similarly found that none of the 30 selected Protenix v2 predictions met the isopeptide distance-closure criterion. Together, these findings suggest that limited isopeptide-closure accuracy in Protenix v2 contributes to the disagreement, which is not specific to OmniPepBuilder.

2. Sequence initialization and computational overhead

How should we initiate sequence? Since AF3 generation itself (not workflow) is not guided by hotspots, does this risk cold-start failure or waste computational resources in warm-up steps?

We initialize exploration with random canonical sequences of uniformly sampled lengths (8–18 residues), excluding unconstrained cysteine and fixing topology-defining residues. Early tests with Protenix and AF3 supported both random and unknown-residue (X) starts followed by ProteinMPNN design. The historical Protenix pilot yielded 84 and 67 accepted designs from random and X starts, respectively, with approximately 36% higher yield per prediction call for random initialization. We therefore use random initialization as the default. Active learning or Bayesian optimization could guide future initialization toward less explored sequence regions.

We use early stopping to limit wasted evaluations: if none of the five AF3 samples passes both hotspot and topology (cyclization bond-length) checks, the lineage terminates after that evaluation. Our audit of all 40 benchmark cases (Appendix C.4.1) shows modest overall failed-start overhead and a short wait before replay: first-round hotspot and topology failures together account for 2.44% of generation evaluations, varying from 0% to 15.79% across cases, and a selectable replay archive appears within 0.73–5.04 minutes in every case.

Constraint-guided sampling could further reduce overhead and improve geometric constraint satisfaction. Inference-time steering and search have improved constraint satisfaction or design quality in Boltz-2 [3], Proteína-Complexa [1] and RFD4-Proteina [2]. Such guidance can be combined with our replay policy: it guides individual trajectories, while replay allocates exploration across retained states. We are actively developing this integration. However, strong guidance may concentrate sampling and reduce diversity [4], requiring a balance with exploration across pose basins.

3. Cross-case consistency and multi-seed evaluation

In ablation, each case is evaluated using only a single seed trajectory. Repeated evaluations with more random seeds are needed to improve the robustness of reported results.

The added statistical analysis shows systematic trends across multiple systems. The 14 cases with positive full-policy yield span seven targets and two topologies. Confidence-based and uniform replay reduce median yield by 29.9% and 12.0% relative to the full policy, with exploratory paired-test p-values of 0.000244 and 0.00537 (Fig. 5; Table 4; Appendix C.4.4). Retaining revisit penalties yields 25–35% more pose groups for both KRAS G12D topologies and head-to-tail CXCR4 (Fig. 18). These results support pose cell partitioning and replay scheduling across the tested systems.

We emphasize that all global-policy ablation runs used the same fixed seed, 2026, with no manual cherry-picking. Due to resource limitations, multi-seed runs are still ongoing; we will add their results when completed.

[1] Didi et al., Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute. ICLR, 2026.

[2] Corley et al., Biomolecular design through latent reasoning with RFD4-Proteina. Preprint, 2026.

[3] Passaro et al., Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv, 2025.

[4] Kalisz et al., How to Spend Your Oracle Budget: Practical Guidance for Protein Structure Prediction Models. arXiv:2608.12192, 2026.

I thank the authors for the thorough rebuttal and the impressive additional validations (full-pool evaluation, Protenix v2 checks, clash filtering, and cold-start audits), which have well addressed my methodological concerns; accordingly, I have raised my Soundness score from 2 to 3. I hold this work in very high regard and am maintaining my overall positive rating of 3 (Weak Acceptance) only because the multi-seed ablation runs are still ongoing and experimental validation remains future work. Additionally, when I attempted to follow the described workflow for disulfide-bonded peptides, I noticed that several predicted structures struggled to form clean S–S bonds without Rosetta relaxation; therefore, I strongly encourage the authors to publicly release the complete codebase and configuration scripts to ensure seamless reproduction and benefit the community.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you for your timely response and for raising the Soundness score. Your insightful comments have prompted additional analyses that helped us better understand the search behavior and strengthen the evaluation.

For disulfide-bonded peptides, we have indeed also observed unsuccessful cyclization at initialization and early stages: 280/1,526 starting sequences (18.3%) failed the S–S distance check in all five AF3 samples in the first round. These failures presumably reflect suboptimal peptide–target interactions that do not yet sufficiently constrain the binding pose. Nevertheless, 893/900 generation-accepted disulfide designs (99.2%) retained valid S–S distances upon AF3 reprediction, before any Rosetta relaxation, supporting the workflow’s ability to identify optimal designs that preserve disulfide closure.

Upon acceptance, we will publicly release the complete codebase, configuration files and raw data for the paper to support reproducibility and community reuse.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response by Authors

We sincerely thank all the reviewers for their careful evaluation and constructive feedback. We have revised the manuscript and expanded the evaluation, with the following main updates:

  1. Full-pool benchmark evaluation. We thank VHTB and cYmV for questioning the previous linear extrapolation of pose-group counts. We previously evaluated subsampled output pools for RFpeptides, BoltzGen and AnewOmni due to computational resource limits. We have since revisited this procedure and acknowledge that extrapolation overestimated the baselines' pose diversity, to varying degrees. We have now evaluated the complete pools and replaced extrapolated counts with observed results. OmniPepBuilder substantially outperforms all five external baselines, achieving state-of-the-art accepted pose-group throughput, with 2.5–9.7-fold higher pooled throughput on each comparator's supported cases. See Section 3.1, Fig. 4a and Appendix C.3.3.

  2. Orthogonal prediction and structural validity checks. We thank bVY7, VHTB and cYmV for raising concerns about predictor hacking and structural validity when relying on a single predictor. We have added an orthogonal acceptance filter using Protenix v2 and assessed structural validity with a PoseBusters-style clash filter followed by a Rosetta peptide-energy filter. OmniPepBuilder retains its leading performance under both assessments: 3.1–12.7-fold higher pooled throughput with additional Protenix acceptance and 2.6–9.4-fold higher throughput after both validity filters, relative to external baselines on matched cases. See Section 3.1, Fig. 15 and Appendix C.4.3.

  3. Expanded ablations and statistical analysis. We thank bVY7 and cYmV for their suggestions on the exploration-policy ablations. We added uniform replay for comparison, rewrote the ablation discussion to clarify the four controls, and introduced significance tests across 14 target–topology cases. Confidence-based and uniform replay yield 29.9% and 12.0% fewer accepted pose groups in the median, respectively, with exploratory paired-test p-values of 0.000244 and 0.00537. These comparisons support the contribution of pose cell partitioning and replay scheduling beyond replay alone. See Section 3.1, Fig. 5, Table 4, Fig. 18 and Appendices C.3.4 and C.4.4.

  4. More detailed TNFα stress-test analysis. We rewrote the TNFα analysis to more clearly explain the two receptor-contacting sites, stricter acceptance criteria and extended-search outcomes. We emphasize that TNFα is particularly challenging: neither AlphaProteo nor PXDesign obtained experimental minibinder hits in their campaigns, and no method produced a candidate passing common AF3 evaluation in our benchmark. We therefore selected it for extended stress tests, where four-GPU campaigns yielded accepted disulfide and head-to-tail designs at both sites while requiring contacts with at least three hotspot residues. Confidence and allocation histories show how initially subthreshold pose regions develop accepted designs and how revisit penalties shift relative replay priority toward less explored regions. See Section 3.1, Fig. 4b–e, Fig. 19 and Appendix C.5.

  5. Further analysis of initialization and replay dynamics. In response to bVY7, we quantify failed-start overhead and the time to the first selectable replay archive. See Appendix C.4.1. In response to VHTB, we analyze 6,507 replays across twelve randomly selected cases spanning different levels of search difficulty, distinguishing new-pose discovery, source-local and non-local Pareto gains, and hard-gate termination. Interestingly, non-local gains occur approximately twice as often as source-local gains, with substantial backbone displacement coexisting with partial contact retention. These results show that reusing intermediate states improves the candidate collection across pose cells, supporting systematic landscape exploration beyond isolated trajectory optimization. See Appendix C.4.2 and Fig. 14.

  6. Other updates. We have systematically revised the manuscript to improve logical flow and readability. We have also reorganized the appendix into more clearly delineated subsections, making the methods, additional results and supporting analyses easier to locate.

AuthorsOpenReview ↗

Remark on Multi-Seed Ablations

In further response to reviewers bVY7 and cYmV, we ran the Full policy at two additional generation seeds to test whether its original advantage depended on a lucky run. We also repeated No revisit penalty at the same seeds to further clarify its mixed, target-dependent effects. We kept all other settings as in Fig. 5, including the 12-GPU-hour generation budget and common AF3 reprediction. The three simpler controls use their original seed-2026 runs.

1. The Full policy’s yield advantage over simpler replay strategies persisted across seeds

Across the benchmark, the Full policy averaged 57.00, 51.93 and 59.07 accepted pose groups per case at the three seeds. Each mean exceeded those of confidence-based replay (40.29), uniform replay (49.71) and no replay (50.36). Under the Fig. 5 eligibility rule, the Full policy significantly outperformed confidence-based replay and no replay at all three seeds, and uniform replay at two (Table 1). The higher mean yield at both additional seeds gives us confidence that the original advantage of pose-aware replay scheduling did not depend on a single lucky Full policy run.

Table 1. Exact, two-sided, unadjusted paired p-values by generation seed, restricted to cases where the Full policy yielded at least one accepted group as in Fig. 5. ns p ≥ 0.05, * p < 0.05, ** p < 0.01, *** p < 0.001.

ComparisonSeed 2026Seed 1002026Seed 2002026
Full policy vs No revisit penalty0.444 ns0.460 ns0.315 ns
Full policy vs confidence-based replay0.000244 ***0.000244 ***0.000122 ***
Full policy vs uniform replay0.00537 **0.0801 ns0.000488 ***
Full policy vs no replay0.0178 *0.0459 *0.000244 ***
No revisit penalty vs confidence-based replay0.000122 ***0.000244 ***0.000244 ***
No revisit penalty vs uniform replay0.0248 *0.0474 *0.00439 **

2. The revisit penalty consistently increased pose discovery across seeds

As in the original single-seed ablation, the effect of the revisit penalty on the number of accepted pose groups at 12 hours varied by target. The benefit was particularly clear on KRAS G12D, a relatively challenging target, where the penalty yielded 8, 5 and 12 additional head-to-tail groups across the three seeds and 7, 4 and 0 additional linear groups. For head-to-tail CXCR4 which is also challenging, the difference changed sign (+9, +7, −4 groups), while head-to-tail MCL1 yielded more groups without the penalty at every seed. Across the benchmark, mean yield was similar with and without the penalty (56.00 vs 55.52 groups per campaign).

We used the multi-seed runs to further assess how the penalty affected exploration by tracking where replays started and whether their rollouts reached a backbone pose previously unseen in the same campaign. With the penalty, more replays started from a previously unreplayed geometry (53.66% vs 44.41%), and more rollouts discovered a new pose (44.43% vs 40.09%). The pose discovery rate increased at all three seeds and in 13 of 14 target–topology cases when averaged across seeds. We argue that some of these newly explored regions may require more than 12 hours to yield accepted designs. In the extended TNFα searches, Fig. 4b,d shows replay priority shifting toward less explored regions that subsequently yielded accepted designs. Broader exploration may also benefit multi-target and molecular-glue searches, where a candidate must satisfy several interface constraints.

If accepted, we will include these results in the camera-ready version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.