acceptodds
Under review as a conference paper at ICLR 2027

Energy-Guided Test-Time Refinement for Full-Atom Peptide Binder Design

Abstract

Peptide binder design generates binder sequences and structures conditioned on a target receptor. Across model architectures, scaling test-time sampling has become the dominant strategy to improve binding quality, yet it wastes computation on unpromising samples and overlooks residue-level signal. To address both, we introduce Energy-guided Test-Time Refinement (E-TTR), which reallocates the inference budget to high-potential candidates and refines them at the residue level. Concretely, given a receptor, we sample an initial pool of peptides and score each with Rosetta to obtain a binding energy together with its per-residue decomposition. For the best candidate, we re-noise only the high-energy residues and denoise them conditioned on the rest, producing a new candidate pool. This loop replaces parallel sampling with several sequential rounds, progressively improving binding quality within the same inference budget. Because standard denoisers share one timestep across the entire peptide, we post-train them via Diffusion Forcing to assign per-residue timesteps, making residue-level edits far more effective. Experiments across three benchmarks and two generative backbones demonstrate that E-TTR largely outperforms competing test-time scaling strategies in binding affinity and success rate, with up to 28% lower binding energy at matched sampling budgets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
peptide design, sequence-structure co-design, diffusion models, test-time scaling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tWftOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Energy-Guided Test-Time Refinement (E-TTR), a framework for improving full-atom peptide binder design through iterative refinement at inference time. Instead of relying solely on independent sampling, the method uses energy feedback to select promising candidates and refine unfavorable residues. The approach shows improvements in Rosetta binding energy compared with several existing test-time scaling strategies. Additional ablation studies investigate the contributions of the proposed components and different refinement settings.

Critical strengths

The paper is well organized, clearly written, and easy to follow, with effective figures illustrating the overall framework. The proposed approach is intuitive and well motivated, offering a reasonable strategy for improving peptide binder design through energy-guided refinement. The experimental evaluation is relatively comprehensive, covering multiple benchmarks and representative baselines, with ablation studies supporting the effectiveness of the proposed method.

Critical weaknesses

The main concern is that Rosetta energy is used both to guide refinement and to evaluate the generated designs. The independent-seed rescoring in Appendix E helps, but it is still Rosetta scoring Rosetta-optimized designs. Some non-Rosetta evaluation is needed before the dG improvements can be read as better binding. Second, while the candidate budget is matched across methods, E-TTR's refinement rounds are strictly sequential — each round must wait for Rosetta scoring of the previous one — whereas BoN sampling is fully parallel, and no wall-clock or compute measurement is reported anywhere in the paper, leaving the practical efficiency unclear. Third, I accept the authors' point that RMSD and AAR measure recovery of the native peptide rather than binding, but no measure of interface quality is reported in their place, so the consistent degradation of both metrics is left without evidence that the lower dG corresponds to physically plausible binding conformations. Finally, the paper reports no seed replication, variance, or confidence intervals, which makes the thinner margins hard to assess.

Questions for authors
  1. The proposed method uses Rosetta energy for both guiding refinement and evaluating the generated peptides. Could the authors provide an independent evaluation to verify that the improvements in dG reflect better binding quality rather than optimization toward the Rosetta scoring function?
  2. Could the authors report the actual inference time and computational cost of each method? Does E-TTR retain its advantage under a matched wall-clock time budget?
  3. Table 1 shows that E-TTR improves binding energy while yielding worse RMSD and AAR, particularly on PepBDB, where RMSD increases from 11.34 Å to 15.47 Å. Could the authors provide additional evaluations of binding interface quality to clarify whether the lower dG corresponds to physically plausible binding conformations?
  4. The main results do not report variability across independent runs or confidence intervals. In particular, the improvement in median dG on PepMerge is only 1.6% over BoN. Could the authors provide results across multiple random seeds or paired confidence intervals to assess the robustness of these improvements?
5 comments
AuthorsOpenReview ↗

Response to Reviewer tWft

Thank you for the helpful comments!

W1 & Q1: Is improved binding quality supported by an evaluation independent of Rosetta?

A1: Our additional DeepPpIScore and MM/GBSA evaluations assess binding quality beyond Rosetta. Please see Section 2 of our general response for details.

W2 & Q2: Does E-TTR retain its advantage under a matched wall-clock time budget?

A2: We further compare all methods under matched time and NFE budgets at multiple cutoffs. Please see Section 3 of our general response and Figure 8 of the revised manuscript for the results.

W3 & Q3: Does lower dG indicate physically plausible binding despite worse RMSD and AAR?

A3: We evaluate the interface quality of all six methods on the three datasets using two metrics:

  • CMS (contact molecular surface): Measures distance-weighted contact area, with higher values indicating more extensive close contact [1].
  • Energy/area: Normalizes Rosetta dG by the buried solvent-accessible surface area (100×dG/ΔSASA100\times\mathrm{dG}/\Delta\mathrm{SASA}), with lower values indicating more favorable energy per unit area [2].

For each method and pocket, we compute both metrics on the candidate with the lowest dG and report medians across pockets with valid measurements.

PepBench / PepGLADPepBDB / PepGLADPepMerge / PepFlow
MethodCMS (Ų) ↑Energy/area ↓CMS (Ų) ↑Energy/area ↓CMS (Ų) ↑Energy/area ↓
BoN268.66-1.951228.63-1.945270.17-2.683
IPR239.75-1.998230.04-1.984276.95-2.723
Beam258.94-2.011246.56−2.011‾\underline{-2.011}259.96−2.731‾\underline{-2.731}
FK264.75-1.958248.96‾\underline{248.96}-1.968262.13-2.645
MCTS285.45‾\underline{285.45}−2.065‾\underline{-2.065}223.76-2.005255.65-2.564
E-TTR292.29-2.181270.79-2.076273.08‾\underline{273.08}-2.836
  • Interface contacts. E-TTR achieves the highest CMS on PepBench and PepBDB, exceeding the strongest baseline by 6.84 and 21.83 Ų, respectively, and ranks second on PepMerge. These results indicate that its low-dG candidates also form extensive close contacts at the interface.

  • Area-normalized interface energy. E-TTR achieves the lowest Energy/area on all three datasets, outperforming the strongest baseline by 0.116, 0.065, and 0.105 on PepBench, PepBDB, and PepMerge, respectively. Its energy advantage therefore remains after normalization by buried surface area.

Together, these results support the plausibility of E-TTR's binding interfaces, showing that reduced native recovery can coexist with favorable interface contacts and area-normalized energy.

[1] Design of protein-binding proteins from the target structure alone. Nature 2022.

[2] InterfaceAnalyzer (https://docs.rosettacommons.org/docs/latest/application_documentation/analysis/interface-analyzer). Rosetta documentation.

W4 & Q4: Are the reported gains statistically robust across independent runs?

A4: We repeat the final evaluation using five random seeds and report the mean ± standard deviation.

Dataset / modelMethodmean dGmedian dGSuccess rate (%) ↑
PepBench / PepGLADBoN-110-19.86 ± 0.32−21.38±0.40‾\underline{-21.38 ± 0.40}91.42 ± 0.62
IPR-18.90 ± 0.49-19.52 ± 0.5489.68 ± 0.89
Beam-20.00 ± 0.23-20.88 ± 0.8391.44 ± 0.61
FK-19.69 ± 0.31-19.46 ± 0.7491.12 ± 0.77
MCTS−21.10±0.38‾\underline{-21.10 ± 0.38}-21.28 ± 0.3692.99±0.82‾\underline{92.99 ± 0.82}
E-TTR-24.09 ± 0.53-23.75 ± 0.3495.42 ± 0.41
PepBDB / PepGLADBoN-110-18.71 ± 0.29-18.23 ± 0.5284.14 ± 0.59
IPR-18.22 ± 0.21-18.12 ± 0.5483.88 ± 0.61
Beam-21.29 ± 1.19−19.18±0.56‾\underline{-19.18 ± 0.56}86.14 ± 0.91
FK-18.78 ± 0.48-18.31 ± 0.3383.49 ± 1.09
MCTS-19.17 ± 0.24-18.89 ± 0.6085.04 ± 0.77
E-TTR−20.03±0.50‾\underline{-20.03 ± 0.50}-20.41 ± 0.3185.41±0.83‾\underline{85.41 ± 0.83}
PepMerge / PepFlowBoN-110−28.63±0.12‾\underline{-28.63 ± 0.12}−27.52±0.29‾\underline{-27.52 ± 0.29}98.82 ± 0.28
IPR-28.46 ± 0.12-27.15 ± 0.0998.40 ± 0.21
Beam-28.26 ± 0.13-26.84 ± 0.1898.47 ± 0.22
FK-27.37 ± 0.13-26.20 ± 0.1398.16 ± 0.20
MCTS-26.54 ± 0.13-26.23 ± 0.1196.58 ± 0.26
E-TTR-30.45 ± 0.20-28.00 ± 0.4198.60±0.28‾\underline{98.60 ± 0.28}

We can see that E-TTR achieves the lowest median dG on all three datasets and the lowest mean dG on PepBench and PepMerge, ranking second on PepBDB. These results support the stability of E-TTR's energy improvements across evaluation seeds.

Thank you for the detailed response and additional experiments. The new results have addressed most of my concerns and significantly strengthened the paper.

I have one minor clarification regarding Q4: Do the five random seeds correspond to independent runs of the entire generation and refinement process, or only to repeated Rosetta rescoring of a fixed candidate pool? This would help clarify the robustness of the reported improvements.

Thank you for your positive feedback and the follow-up question. The results reported in A4 correspond to repeated final Rosetta evaluations, with generation and refinement held fixed. Given the time required for full multi-seed generation and refinement, we were only able to complete the repeated final evaluations before our initial response.

We are now conducting additional independent runs of generation, refinement, and evaluation, and expect to report the results within 1–2 days to further assess robustness across the entire pipeline.

If our responses and additional experiments address your concerns, we would be grateful if you could consider increasing your score accordingly.

Reviewer_tJx1OpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Energy-guided Test-Time Refinement (E-TTR) for full-atom peptide binder design. It iteratively selects a low-energy candidate and regenerates residues with unfavorable Rosetta energy contributions. Diffusion Forcing post-training enables selective sequence–structure editing through residue-specific noise levels. Experiments across two backbones and three datasets demonstrate improved median Rosetta energy under matched candidate budgets. Component ablations distinguish post-training from refinement, while randomized selection controls support the usefulness of both candidate-level and residue-level energy feedback.

Critical strengths
  1. The ablations substantiate the proposed mechanism. Table 2 separates DF post-training from refinement, and Table 3 independently replaces energy-guided candidate and residue selection with random selection. These controls provide meaningful evidence that the gains arise from the proposed refinement mechanism rather than merely additional training.

  2. Energy improvements recur across architectures and datasets. E-TTR achieves the best median Rosetta dG in all three evaluated settings. Detailed backbone adaptations, documented baseline procedures, and independently seeded rescoring strengthen the reproducibility and credibility of this computational result.

Critical weaknesses
  1. Appendix E constructs Table 1’s diversity set from nine initialization samples and only one final-round candidate. Comparable diversity may therefore largely reflect initialization rather than the refined designs. Figure 4 evaluates the complete search archive, but does not establish diversity among high-quality outputs. This leaves the claimed quality–diversity trade-off insufficiently supported.

  2. Rosetta guides refinement, supplies the principal energy metric, and defines success through negative rescored dG. Independent-seed rescoring addresses stochastic scoring variation but does not provide independent corroboration against systematic scoring errors. The results therefore establish improved Rosetta-score optimization more clearly than improved binding affinity.

  3. Different strategies generate candidates through different amounts of denoising, intermediate-state reuse, and sequential processing. Consequently, matching 110 candidates per target does not ensure comparable computational cost. Without accounting for denoiser calls, scoring overhead, and runtime, it remains unclear whether the reported advantages persist under a matched computational budget.

Questions for authors
  1. Can you compare diversity among equally sized sets of high-quality refined outputs, excluding retained initialization samples, to establish whether energy improvements preserve useful design diversity?

  2. Is there evidence independent of the Rosetta objective supporting improved binding quality, or should the affinity and success claims be explicitly restricted to Rosetta-defined computational endpoints?

  3. How do denoiser calls, scoring costs, and runtime compare across methods, and do E-TTR’s energy improvements persist when matching computational cost rather than candidate count alone?

3 comments
AuthorsOpenReview ↗

Response to Reviewer tJx1

Thank you for the helpful comments!

W1 & Q1: Do the energy improvements preserve diversity among high-quality refined outputs?

A1: We compare diversity among equally sized sets of terminal and high-quality candidates. Please see Section 1 of our general response and Figure 7 of the revised manuscript for the results and trade-off analysis.

W2 & Q2: Is improved binding quality supported by evidence independent of Rosetta?

A2: Our additional DeepPpIScore and MM/GBSA evaluations assess binding quality beyond Rosetta. Please see Section 2 of our general response for details.

W3 & Q3: Do E-TTR's gains persist when computational cost is matched?

A3: We further compare all methods under matched time and compute budgets at multiple cutoffs. Please see Section 3 of our general response and Figure 8 of the revised manuscript for the results.

Thank you for clarifying how the existing results address my concerns. Figure 7 helps clarify the quality–diversity trade-off, although the default configuration shows substantial diversity loss among refined outputs. Table 1 measures diversity using nine initialization samples and only one final-round candidate, so it does not adequately characterize diversity among the refined outputs. This distinction should be made clearer. Figure 8 partially addresses computational fairness, although the shared absolute time budget, NFE definition, and inclusion of Rosetta scoring overhead remain unclear. Regarding binding affinity, please clarify whether the non-Rosetta evaluations consistently support the claimed improvements across metrics and settings, and qualify the conclusions where the evidence is mixed or limited. Based on the current response, I do not yet find sufficient grounds to raise my assessment of the paper and will therefore maintain my rating of 2 (weak rejection).

Thank you for the follow-up.

  1. Diversity. Figure 7 of the revised manuscript shows that E-TTR supports an adjustable quality–diversity trade-off, with configurations that dominate IPR, Beam, FK, and MCTS.
  2. Computational fairness. The time comparison includes generation and Rosetta scoring, with the time limit set to 3 hours. The compute budget is 11,000 NFE per pocket, where one NFE denotes one diffusion-network forward pass. Experimental infrastructure is described in Appendix C of the main paper.
  3. Additional evaluations. E-TTR leads on both DeepPpIScore and MM/GBSA comparison.

We sincerely appreciate your time and thoughtful feedback, which have helped us improve the clarity and evaluation of our work.

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

The authors use diffusion forcing to train a denoiser that admits per-residue diffusion levels. They then use this model to generate peptide binders. They do inference time search by measuring energies of the complex, splitting the energy into per-residue components, and noising the 3 highest energy residues, while keeping all other residues fixed (sequence and structure). This inference time search leads to designed binders for which the total binding energy is lower.

Critical strengths
  • Inference time search is very timely, and has been leveraged in many domains, including protein design, to boost the performance of multiple models.
  • The method leverages protein structures and rewards used in the space, which can be split across residues, leveraging this for the proposed inference time search method to determine which residues to re-noise and de-noise.
Critical weaknesses
  • Evaluation pipeline. As I understand, designed complexes are relaxed before energies are measured. Success is defined by measuring the energy of the bound complex, and the energy of the individual target and binder. If the bound complex has a lower energy, the sample is considered a success. The relaxation for this metric is done for the complex, right? For the chains in isolation, the side-chains are repacked, but the structures are not relaxed. Would it make sense to relax the isolated structures too? Otherwise the energies are being measured in isolation for backbones that were relaxed in the bound complex, which may be an unfavorable comparison.
  • Related to the point above, having a lower energy in the bound complex than the chains in isolation, while good, does not directly indicate a successful binder. Does the Rosetta dG translate into better binders?
  • Scope of evaluation. It would be interesting to extend this to other scenarios. For instance, one could consider using this methodology for binder design (not peptides, using e.g. Proteina-Complexa, or another open model). How does searching for Rosetta energies using this method affect in-silico success rates (from folding model metrics)? Using folding model rewards, would it be possible to identify specific residues to re-noise and de-noise (e.g. PLDDT is per-residue, PAE involves pairs of residues, etc).
  • The diversity metric in some of the tables seems strange. The work measures diversity by taking 9 of the initial 10 designs (removing the best one of the batch), and then the lowest energy design obtained with the search procedure. The search procedure might be collapsing, and this metric would not capture that. Am I understanding this correctly?

Overall, I find the methodological novelty and practical value somewhat limited, in the paper current's form.

Questions for authors
  • Only 3 residues re-designed at a time, with the rest of the protein fixed. Could this be actually limiting the method? For a bigger protein with some structure, the scaffold involving all residues except the 3 chosen ones through some criteria may pretty much completely determine the structure of the 3 redesigned residues. Maybe for peptides, which could be less structured, this is less of an issue? This is related to the point above of testing in other tasks.
3 comments
AuthorsOpenReview ↗

Response to Reviewer MEYe (1 / 2)

Thank you for the helpful comments!

W1: Should the separated structures also be relaxed when evaluating dG?

A1: We add evaluations that relax the separated structures, as suggested, and compare all six methods on the three datasets under four relaxation protocols:

  • R0 (original): We locally relax the complex and only repack side chains after separating the binding partners.
  • R1: We additionally relax the separated structures using the same settings as for the complex.
  • R2: Starting from R0, we also allow the receptor interface backbone to relax in the complex; the separated structures still undergo side-chain repacking only.
  • R3: Starting from R1, we also allow the receptor interface backbone to relax in both the complex and separated structures.
dG
DatasetMethodR0 (original)R1R2R3
PepBenchBoN-19.56-9.49−25.53‾\underline{-25.53}-7.39
IPR-19.98-8.88-24.84-6.51
Beam-21.55-8.49-23.90−7.61‾\underline{-7.61}
FK-19.39-9.76-23.19-5.84
MCTS−21.65‾\underline{-21.65}−9.96‾\underline{-9.96}-24.87-7.32
E-TTR-25.10-10.08-27.42-8.31
PepBDBBoN-18.35-6.41-21.03-6.49
IPR-18.33-6.33-21.26-5.45
Beam-19.86−8.16‾\underline{-8.16}-21.85-9.80
FK-18.78-7.05-20.75-7.81
MCTS−20.94‾\underline{-20.94}-7.25-22.59-8.00
E-TTR-21.91-8.72−21.99‾\underline{-21.99}−8.05‾\underline{-8.05}
PepMergeBoN−27.74‾\underline{-27.74}-21.30-29.08−17.71‾\underline{-17.71}
IPR-27.15-20.91−31.29‾\underline{-31.29}-17.04
Beam-26.80−21.38‾\underline{-21.38}-29.40-17.05
FK-26.40-20.12-28.72-17.04
MCTS-26.40-19.18-31.78-16.11
E-TTR-28.19-22.29-30.77-19.21
  • Relaxation of separated structures. E-TTR achieves the lowest median dG on all three datasets under R1, outperforming the strongest baseline on PepBench, PepBDB, and PepMerge by 0.12, 0.56, and 0.91, respectively. Its energy advantage therefore remains when the separated structures are relaxed using the same settings as the complex.

  • Additional receptor backbone flexibility. Under R3, E-TTR achieves the lowest median dG on PepBench and PepMerge, outperforming the strongest baseline by 0.70 and 1.50, respectively. This shows that E-TTR retains its advantage on these datasets when receptor interface backbone relaxation is allowed in both the complex and separated structures.

Together, these results support E-TTR's effectiveness beyond the original relaxation protocol.

W2: Does lower Rosetta dG indicate better binding?

A2: Rosetta dG is the energy difference dG=Ebound−Eseparated\mathrm{dG}=E_{\mathrm{bound}}-E_{\mathrm{separated}}, so dG<0\mathrm{dG}<0 indicates that the complex is energetically more favorable than the separated partners under the Rosetta scoring function. We therefore use dG as a computational measure of binding affinity, with lower values indicating more favorable predicted binding. This interpretation follows existing peptide and antibody design studies, including PepGLAD [1], CPSea [2], PepTri [3], PepFlow [4], UniMoMo [5], and AbDPO [6]. Please see Section 2 of our general response for additional binding-quality evaluations using DeepPpIScore and MM/GBSA.

[1] Full-Atom Peptide Design with Geometric Latent Diffusion. NeurIPS 2024.

[2] CPSea: Large-scale cyclic peptide-protein complex dataset for machine learning in cyclic peptide design. NeurIPS 2025.

[3] PepTri: Tri-Guided All-Atom Diffusion for Peptide Design via Physics, Evolution, and Mutual Information. ICLR 2026.

[4] Full-Atom Peptide Design based on Multi-modal Flow Matching. ICML 2024.

[5] UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design. ICML 2025.

[6] Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization. NeurIPS 2024.

AuthorsOpenReview ↗

Response to Reviewer MEYe (2 / 2)

W3: Does E-TTR generalize to non-peptide binder design and folding-model rewards?

A3: We extend E-TTR to protein binder design using pretrained Genie 3 [7] and compare all six methods with 110 candidates per target. The table reports experimental results on the binder-design problem set released with Genie 3 [8], with the binder length fixed at 80 residues. We use cross-chain PAE from AlphaFold2 [9, 10] to evaluate candidates and derive residue-level scores to guide local refinement. Due to time constraints, we do not include Diffusion Forcing (DF) post-training [11] in this experiment.

We evaluate the generated designs using four metrics:

  • iPAE: Predicted error in the relative positioning of the binder and target; lower is better [9, 10].
  • ipTM: Confidence in the predicted binder-target interface; higher is better [10].
  • Binder pLDDT: Local structural confidence within the binder; higher is better [9].
  • Geometric validity: Percentage of candidates passing Cα continuity and clash checks; higher is better.
MethodiPAE (Å) ↓ipTM ↑Binder pLDDT ↑Geometric validity (%) ↑
BoN18.7170.511455.4557.50
IPR23.1470.483942.814.77
Beam17.272‾\underline{17.272}0.5244‾\underline{0.5244}58.9668.86‾\underline{68.86}
FK19.7110.516150.7458.07
MCTS18.4120.507653.8862.73
E-TTR17.0340.548656.98‾\underline{56.98}78.75
  • Interface quality. E-TTR achieves the lowest iPAE and the highest ipTM. Compared with Beam, the strongest baseline on both metrics, E-TTR reduces iPAE from 17.272 to 17.034 and increases ipTM from 0.5244 to 0.5486. These results support improved predicted interface quality under folding-model rewards.

  • Structural quality. E-TTR achieves the highest geometric validity at 78.75%, exceeding Beam by 9.89 percentage points. Thus, more generated candidates satisfy the basic continuity and clash checks. E-TTR also ranks second in binder pLDDT, which measures local structural confidence within the binder [9]. Since our refinement focuses on the relative positioning of the binder and target, improvements in interface metrics need not yield the highest binder pLDDT. The absence of DF post-training [11] may also limit performance in this setting.

These results support the applicability of E-TTR to the tested non-peptide binder design setting with folding-model rewards.

[7] Fast and Ultra-Capable Protein Design: Advancing the Frontier Through Atomistic SE(3)-Equivariance with Genie 3. bioRxiv 2026.

[8] aqlaboratory/genie3. Construction of binder design problem set (https://github.com/aqlaboratory/genie3/blob/main/README.md#construction-of-binder-design-problem-set). GitHub repository.

[9] Highly accurate protein structure prediction with AlphaFold. Nature 2021.

[10] Protein complex prediction with AlphaFold-Multimer. bioRxiv 2021.

[11] Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. NeurIPS 2024.

W4: Does the diversity metric capture potential collapse during refinement?

A4: Please see Section 1 of our general response for diversity analysis. We also include the corresponding analysis in Section 4.5 and Figure 7 of the revision.

Q1: Does editing only three residues limit refinement of larger proteins?

A5: The number of residues edited per round is a tunable hyperparameter, with three used as the default for peptide design. In Table 6 of the main paper, we compare editing one, three, and five residues on PepGLAD/PepBench under the same evaluation budget; editing three achieves the lowest mean and median dG. For larger proteins, the editing region can be expanded; in the protein binder experiment reported in our response to W3, we edit 32 residues per round.

Thank you again for your thoughtful feedback.

As the discussion phase closes soon, we wanted to follow up on our response, particularly the new comparisons across relaxation protocols and the analysis of diversity among refined designs. We would appreciate hearing whether these results address your concerns and would be happy to clarify any remaining questions. If you feel that our response and the additional experiments resolve your concerns, we would be grateful if you would consider revisiting your score.

We sincerely appreciate your time and consideration.

Comments on the paper

5 comments
AuthorsOpenReview ↗

General Response From Authors (3 / 3)

3. Comparisons under matched evaluation, time, and compute budgets

We thank the reviewers for suggesting comparisons under matched time and compute budgets. We consider three budget settings:

  • Evaluation budget: the same budget of candidate designs for evaluation per pocket.
  • Time budget: matched time allowances for generation and scoring.
  • Compute budget (NFE): the same cumulative number of denoising network forward passes per pocket.

In the original paper, we fix the evaluation budget at 110 candidates per pocket for all methods. This setting measures the quality of designs obtained with a limited number of candidate evaluations. Computational evaluations and laboratory assays can be costly, making this comparison practically relevant for molecular and protein design in practice[3, 4, 5]. To further assess computational efficiency, we compare all six methods on PepGLAD/PepBench under matched time and compute budgets.

For the time comparison, we use 3 hours as the duration for the full PepBench task. The table reports mean dG at five time budgets: 0.6h, 1.2h, 1.8h, 2.4h, and 3h. Please see Appendix C of the main paper for details of the experimental infrastructure.

MethoddG
0.6h1.2h1.8h2.4h3h
BoN-14.73-17.03-17.82-18.53-18.83
IPR−13.98‾\underline{-13.98}−17.04‾\underline{-17.04}−18.48‾\underline{-18.48}−18.94‾\underline{-18.94}−19.67‾\underline{-19.67}
Beam-13.93-16.12-17.40-18.04-18.45
FK-13.25-14.77-15.80-17.54-18.08
MCTS-13.02-13.59-14.07-14.68-15.10
E-TTR-13.04-17.58-20.23-21.67-22.33
  • Quality under matched time budgets. E-TTR achieves the lowest mean dG at 1.2h, 1.8h, 2.4h, and 3h. IPR is the strongest baseline at all four cutoffs, and E-TTR's advantage over IPR is larger at 3h than at 1.2h, showing that the benefit extends as more time is available.

  • Quality at matched performance. At the 1.8h cutoff, E-TTR already achieves lower dG than every baseline at 3h, including IPR. It continues to improve at 2.4h and 3h, further increasing the gap over the baselines' full-budget results, as illustrated in Figure 8 (page 9) of the revised manuscript.

For the compute comparison, we set the maximum budget to 11,000 NFE per pocket, including initialization, and evaluate at 2,200, 4,400, 6,600, 8,800, and 11,000 NFE. The table reports mean search dG.

MethoddG
2,200 NFE4,400 NFE6,600 NFE8,800 NFE11,000 NFE
BoN-15.81-18.25-19.24-20.15-20.73
IPR-15.40-18.36−19.88‾\underline{-19.88}-20.53-20.95
Beam-14.40-17.77-19.04-20.11-22.93
FK−16.27‾\underline{-16.27}-18.10-19.53-20.68-22.85
MCTS-15.26−18.44‾\underline{-18.44}-19.86−21.42‾\underline{-21.42}−25.02‾\underline{-25.02}
E-TTR-18.02-21.43-23.12-24.44-25.88
  • Quality under matched NFE budgets. E-TTR achieves the lowest mean search dG at all five budgets, from 2,200 to 11,000 NFE. The strongest baseline changes among FK, IPR, and MCTS across these budgets, while E-TTR consistently ranks first. Its advantage therefore holds under both limited and larger compute budgets.

  • Quality at matched performance. With 4,400 NFE, E-TTR outperforms BoN and IPR at their full 11,000 NFE budget. With 6,600 NFE, it also outperforms Beam and FK at 11,000 NFE. These comparisons in Figure 8 (page 9) of the revised manuscript show that E-TTR can achieve better candidate scores than these four baselines with fewer denoising network evaluations.

Together with the results under the original evaluation budget, these comparisons show that E-TTR retains its dG advantage under matched time and compute budgets.

[3] Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization. NeurIPS 2022.

[4] Accelerating Bayesian Optimization for Biological Sequence Design with Denoising Autoencoders. ICML 2022.

[5] Self-driving laboratories to autonomously navigate the protein fitness landscape. Nature Chemical Engineering 2024.

AuthorsOpenReview ↗

General Response From Authors (2 / 3)

2. Binding-quality assessment beyond Rosetta

We thank the reviewers for suggesting an evaluation of binding quality beyond Rosetta. We add two complementary evaluations that are not used for refinement guidance:

  • DeepPpIScore [1] evaluates protein interfaces using a learned statistical potential. Higher scores are favorable.
  • MM/GBSA [2] estimates binding energy from molecular mechanics and solvation terms. Lower energies are favorable.

We evaluate the existing PepGLAD/PepBench candidates, while E-TTR continues to use Rosetta for refinement guidance. For each pocket, we compute the median and best valid candidate scores, then average these statistics across all pockets.

MethodDeepPpIScoreMM/GBSA
Pocket medianPocket bestPocket medianPocket best
BoN-11013.1518.02-18.79−32.51‾\underline{-32.51}
IPR13.35‾\underline{13.35}17.75-17.83-30.36
Beam13.1117.71-18.52-30.08
FK13.3418.05‾\underline{18.05}-19.38-31.67
MCTS13.3117.53−19.42‾\underline{-19.42}-30.14
E-TTR14.1118.22-23.03-32.53
  • DeepPpIScore. E-TTR ranks first on both statistics, outperforming IPR, the strongest baseline for Pocket median, and FK, the strongest baseline for Pocket best. These gains are obtained under a learned interface scoring model that is not used to guide refinement.

  • MM/GBSA. E-TTR achieves the lowest Pocket median and Pocket best energies, outperforming the strongest baselines for these statistics, MCTS and BoN, respectively.

These two additional metrics further demonstrate E-TTR's superiority in the reported comparisons.

[1] Harnessing deep statistical potential for biophysical scoring of protein-peptide interactions. Acta Pharmacologica Sinica 2026.

[2] MMPBSA.py: An Efficient Program for End-State Free Energy Calculations. Journal of Chemical Theory and Computation 2012.

AuthorsOpenReview ↗

General Response From Authors (1 / 3)

We thank all reviewers for their valuable feedback and suggestions. We have addressed the questions and suggestions and made amendments in the revised paper (highlighted in blue color). We respond to the common concerns here and reply to individual reviewers for other questions.

1. Diversity of high-quality refined designs

We thank the reviewers for suggesting a clearer evaluation of diversity. We evaluate diversity on three candidate sets and examine the quality-diversity trade-off under a matched candidate budget in the PepGLAD/PepBench setting.

We report the following three diversity metrics:

  • DIVall\mathrm{DIV}_{\mathrm{all}} measures diversity across all candidates generated during the search.
  • DIVfinal\mathrm{DIV}_{\mathrm{final}} measures diversity among 10 terminal candidates, directly assessing the outputs after refinement.
  • DIVtop\mathrm{DIV}_{\mathrm{top}} measures diversity among the 10 lowest-energy candidates from the full search history.

We also conduct experiments to examine the diversity-quality trade-off by varying the search width and depth. We use W to denote the number of selected initial candidates and D to denote the number of refinement rounds for each selected candidate. We compare W1/D10 (default), W2/D5, W5/D2, and W10/D1.

Method / configurationAllFinalTop
DIVall\mathrm{DIV}_{\mathrm{all}} ↑dGDIVfinal\mathrm{DIV}_{\mathrm{final}} ↑dGDIVtop\mathrm{DIV}_{\mathrm{top}} ↑dG
BoN0.4178-19.280.7140-11.980.6780-19.28
IPR0.0412-18.940.6374-11.350.4539-18.94
Beam0.2289-20.370.2072-15.880.4566-20.37
FK0.2323‾\underline{0.2323}-19.520.1537-14.950.4552-19.52
MCTS0.1884-21.540.1000-16.940.3328-21.54
E-TTR W1/D10 (default)0.0562-24.180.1099-21.080.1547-24.18
E-TTR W2/D50.0538−23.93‾\underline{-23.93}0.2009−19.46‾\underline{-19.46}0.2087−23.93‾\underline{-23.93}
E-TTR W5/D20.0511-23.500.3885-16.920.3044-23.50
E-TTR W10/D10.0480-22.140.6608‾\underline{0.6608}-15.270.4572‾\underline{0.4572}-22.14
  • Diversity Analysis. Under the default W1/D10 configuration, E-TTR produces terminal outputs with diversity comparable to that of MCTS, as measured by DIVfinal\mathrm{DIV}_{\mathrm{final}}. The lower DIVall\mathrm{DIV}_{\mathrm{all}} is consistent with the refinement dynamics in Figure 4 (page 8) of the main paper: diversity decreases as dG improves over successive rounds, indicating convergence toward lower-energy candidates.

  • Diversity-quality trade-offs. Adjusting W and D provides different balances between diversity and dG under the same candidate budget, as shown in Figure 7 (page 9) of the revised manuscript. Increasing W while reducing D improves both DIVfinal\mathrm{DIV}_{\mathrm{final}} and DIVtop\mathrm{DIV}_{\mathrm{top}}, with W10/D1 ranking second only to BoN on both metrics. The terminal-output results support two comparisons:

    • Lower dG at comparable diversity. W2/D5 achieves DIVfinal\mathrm{DIV}_{\mathrm{final}} comparable to Beam with substantially lower dG. W10/D1 also provides slightly higher diversity than IPR with substantially lower dG. These comparisons show that E-TTR can improve dG while maintaining a similar level of output diversity.

    • Higher diversity at comparable dG. W5/D2 achieves nearly the same dG as MCTS while producing substantially higher DIVfinal\mathrm{DIV}_{\mathrm{final}}. This configuration therefore offers a more diverse set of refined designs at a similar energy level, providing an option when output diversity is a priority.

These results show that E-TTR allows users to balance energy optimization and output diversity by adjusting W and D at the same candidate budget.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.