acceptodds
Under review as a conference paper at ICLR 2027

Peptide Design and Optimization via Structured Full-Atom Editing

Abstract

Peptide lead optimization seeks to improve molecular properties through selective modifications while retaining target engagement. The development of semaglutide from glucagon-like peptide-1 (GLP-1) illustrates this principle. Applying pocket-conditioned sequence–structure generation to lead optimization requires control over both which residues are modified and which molecular variables may change. Here we introduce EditFlow, a flow-matching model for structured full-atom editing of bound peptides. Residue-variable masks specify the scope of each edit, enabling a shared conditional model to regenerate selected amino acid identities, backbone geometry, and side-chain conformations while preserving the receptor and all unselected peptide variables. Coupled with external property evaluation and search, the same frozen editor supports iterative optimization through successive local modifications. Extensive evaluations on protein-peptide complexes demonstrate state-of-the-art performance in joint sequence-structure redesign alongside competitive full-generation performance. Across single- and multi-objective tasks, EditFlow outperforms strong baselines in improving molecular properties and achieving favorable trade-offs among competing objectives. These results establish structured full-atom editing as a unified approach to peptide design and lead optimization. The source code will be available upon acceptance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide Design.Flow Matching. Multi-objective Optimization

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 5 reviews, above 85% of the others.
Reviewer_GhZrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This work proposes a model for generating peptide drug candidates by changing some parts of the chain while leaving the rest exactly as is. The positions and properties to change are defined by the user so the binding region can be protected. The model returns a complete complex that any external scorer can evaluate. This allows the user to make edits and keep the best ones and change the objective without retraining the model. Some parts need to be further addressed.

Critical strengths
  • The user is able to choose which residues to change and which variables to release.
  • The user can change the objective without retraining the model.
  • The model performs multiple operations with one set of weights.
  • The authors prove that retained parts stay exactly as they were.
Critical weaknesses
  • The paper reports results under two different sampling methods. The reader has to track which numbers belong to which.
  • The baseline starts from a random backbone so the score is as low as random guessing. EditFlow uses a learned backbone, and the design also goes from sequential to joint. Both change at once so it is not clear which one causes the gap from 4.68% to 40.78%. Table 7 suggests it is mostly the backbone.
  • Table 15 compares the methods on the same complexes, but the main comparison does not do this. Therefore, the error bars stay wide and the two methods overlap.
Questions for authors
  • What is the percentage of test receptors that have a homolog in the training set?
  • Why is PepFlow missing from Table 2?
  • What is the error of the Caco-2 predictor? The difference between the methods is very small.
  • Can you run ProteinMPNN on your own backbones with the same regions as in Table 2 to check if the gap stays?
4 comments
AuthorsOpenReview ↗

To Reviewer GhZr

We thank you for your positive assessment of EditFlow's design and for your constructive comments, which helped us improve the clarity and rigor of our manuscript. In the revision, we clarified the sampling configurations across tasks (W1), added shared-backbone MPNN controls and PepFlow/PepHAR baselines in Table 2 to separate backbone quality from joint design (W2, Q2, Q4), performed complex-level paired analyses for the main comparison (W3), and reported the receptor homology rate (Q1) and Caco-2 predictor error with external validation (Q3). We address your concerns and questions as follows.

W1: Clarifying the two sampling methods across experiments.

A: All EditFlow evaluations share a single set of trained weights and adopt task-specific sampling configurations: joint integration for de novo generation and single-pass local redesign (Tables 1–2), and window refinement for sparse-anchor completion (Table 3) and multi-round lead optimization to iteratively refine local edits given retained context (Section 3.3, alongside a matched joint-integration control in Figure 4g). To make this easy to track, we now describe the sampling configurations in the corresponding experiments and summarize all task setups in Appendix G.2 (Table 6).

W2: Separating backbone quality from joint sequence–structure design in Table 2 (4.68% to 40.78% AAR).

A: We ran ProteinMPNN [1] and LigandMPNN [2] on two sets of EditFlow backbones under the exact Table 2 protocol (64 candidates, Table 2, Appendix G.4, Table 12):

Backbone + Sequence DesignBB-RMSD (Å) ↓AAR (%) [95% CI] ↑
Sequence-blind backbone + ProteinMPNN / LigandMPNN3.4312.60 [11.37, 13.88] / 11.45 [10.21, 12.73]
EditFlow joint bb. + ProteinMPNN / LigandMPNN2.5526.66 [23.79, 29.61] / 25.41 [22.52, 28.36]
EditFlow joint output2.5540.78 [36.18, 45.51]

Moving from 3.43 Å sequence-blind backbones to 2.55 Å joint backbones doubles MPNN AAR, confirming the role of backbone accuracy. Yet on the exact same 2.55 Å backbones, EditFlow still beats ProteinMPNN and LigandMPNN by +14.12 and +15.36 pp in paired AAR, showing that the gain comes from both better backbones and joint sequence–structure design.

W3: Complex-level paired analysis for optimization to control inter-complex variance.

A: Mean paired differences across complexes ΔHV\Delta\mathrm{HV} (EditFlow minus PepFlow) under matched 200-step integration (64 proposals) control inter-complex variation and are strictly positive across all four policies (Section 3.3, Appendix I.2, Table 24):

Search PolicyPaired Mean ΔHV\Delta\mathrm{HV} ↑95% Bootstrap CIPaired Wilcoxon pp
BoN+0.0510[0.0342, 0.0690]8.8×10−98.8 \times 10^{-9}
SPIE+0.0261[0.0123, 0.0428]9.5×10−59.5 \times 10^{-5}
Beam+0.0256[0.0116, 0.0405]3.7×10−33.7 \times 10^{-3}
MCTS+0.0315[0.0201, 0.0439]3.9×10−73.9 \times 10^{-7}

Window refinement vs. 200-step PepFlow likewise has paired 95% CIs strictly above zero across all four policies (Appendix I.2).

AuthorsOpenReview ↗

To Reviewer GhZr Part II

Q1: What is the percentage of test receptors that have a homolog in the training set?

A: Using MMseqs2 [3] on receptor-chain sequences available in the structures (≥40%\ge 40\%% sequence identity, ≥80%\ge 80\% %test-chain coverage, E≤10−3E \le 10^{-3}, matching PepFlow's clustering cutoff), 46.52% of test complexes have a training-set receptor homolog (45.99% under bidirectional ≥80%\ge 80\%% coverage, Appendix G.2).

Q2: Why is PepFlow missing from Table 2?

A: We added locally masked PepFlow [4] (200 steps), adapted PepHAR [5], and shared-backbone MPNN controls to Table 2 under matched segments, context, and minimization (64 candidates). EditFlow (50 steps: 40.78% AAR, 2.55 Å BB-RMSD, 65.10% Contact F1, 99.25% closure) beats PepFlow (30.09%, 2.98 Å, 60.12%, 98.47%) by +10.69 pp AAR, −0.43 Å BB-RMSD, and +4.98 pp Contact F1 (Appendix H.3, Tables 10–12).

Q3: What is the error of the Caco-2 predictor, given the small numerical differences?

A: PeptiVerse's [6] Caco-2 head (log⁡Papp\log P_{\mathrm{app}}) has validation RMSE 1.30 and Spearman ρ=0.80\rho = 0.80. Because canonical hydrophilic peptides cluster tightly near −5.85 (SD 0.011), small raw shifts reflect consistent rank-order gains (Appendix I.1). On a held-out PepLand [7] nc-CPP permeability predictor (zero overlap, RMSE 0.9426, ρ=0.6236\rho = 0.6236), saved Caco-2 (SPIE) endpoints gain +0.2817 [0.0944, 0.4715] without re-selection (Appendix I.4, Table 26).

Q4: Can you run ProteinMPNN on your own backbones on the Table 2 regions to check if the gap stays?

A: Yes (see W2 above and Appendix Table 12). On the exact 2.55 Å joint EditFlow backbones, ProteinMPNN reaches 26.66% AAR (63.96% Contact F1, 99.25% closure) vs. 40.78% AAR (65.10% Contact F1) for EditFlow, leaving a +14.12 pp paired AAR lead (+1.14 pp Contact F1) on identical coordinates.

References:
[1] Dauparas et al., Robust deep learning–based protein sequence design using ProteinMPNN, Science 2022.
[2] Dauparas et al., Atomic context-conditioned protein sequence design using LigandMPNN, Nat. Methods 2025.
[3] Steinegger & Söding, MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets, Nat. Biotechnol. 2017.
[4] Li et al., Full-Atom Peptide Design based on Multi-modal Flow Matching, ICML 2024.
[5] Li et al., Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension, ICLR 2025.
[6] Zhang et al., PeptiVerse: A unified platform for therapeutic peptide property prediction, Nat. Commun. 2026.
[7] Zhang et al., PepLand: a large-scale pre-trained peptide representation model for a comprehensive landscape of both canonical and non-canonical amino acids, Brief. Bioinform. 2025.

Thank you for the additional experiments and clarifications. My main concerns have been addressed. I have updated my scores accordingly.

AuthorsOpenReview ↗

Thank you for your Comment

Thank you very much for your time and for raising your scores. Your constructive feedback helped us strengthen both the soundness and the presentation of the paper, and we will make sure these improvements are reflected in the final version.

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

This paper proposes EditFlow, a flow-matching model for peptide design and optimization. It allows users to selectively modify peptide sequences and structures while keeping other parts unchanged. The model is combined with search algorithms to optimize multiple peptide properties, including binding energy, toxicity, permeability, and half-life. Experiments show promising results in peptide generation, redesign, and property optimization.

Critical strengths

The paper addresses an important and practical problem in peptide design and optimization.

The proposed framework is flexible, supporting different types of sequence and structure editing within a single model.

The method shows promising improvements over existing baselines across several computational benchmarks.

Critical weaknesses

Preservation of original properties. While the method preserves unedited residues and structures, it would be helpful to further evaluate whether important original properties, such as binding affinity, are maintained during optimization.

Validation of property improvements. The optimization results are based on computational predictors. Additional validation, such as independent predictors or wet-lab experiments, would help strengthen the conclusions. Since the same predictors are used for optimization and evaluation, it would be useful to assess how well the improvements generalize to independent evaluation methods.

The joint redesign evaluation primarily compares against relatively weak baselines. Stronger comparisons with modern conditional generative models would better demonstrate the benefits of the proposed framework.

Questions for authors

Please refer to the weaknesses

5 comments
AuthorsOpenReview ↗

To Reviewer qUGA

We thank you for recognizing the practical value of selective local peptide editing and the flexibility of EditFlow, and for your constructive comments that helped us improve the manuscript. In the revision, we (1) evaluated binding affinity preservation during optimization via pre-/post-search interface energy, dual-objective property–energy search, and budget scaling (W1); (2) assessed how property improvements generalize to independent evaluators (W2); and (3) added comparisons with modern conditional generative models (PepFlow and PepHAR) and shared-backbone MPNN controls for joint redesign (W3). We address your three concerns as follows.

W1: Evaluating whether original binding affinity is maintained during property optimization.

A: Evaluating single-objective EditFlow + SPIE endpoints with Rosetta REF2015 [1] shows that optimizing properties alone degrades interface energy (ΔE>0\Delta E > 0). Thanks to the flexibility of EditFlow in decoupling generation from scoring, we can perform dual-objective search with interface energy without retraining, optimizing target properties while maintaining interface energy. Scaling the budget from 64 to 256 proposals improves both hypervolume and ΔE\Delta E further (Section 3.3, Appendix I.5, Tables 27–28):

Property (EditFlow + SPIE)Single-Obj. ΔE\Delta E (REU)Dual-Obj. ΔE\Delta E: 64 → 256Dual-Obj. ΔHV\Delta\mathrm{HV}: 64 → 256
Non-toxicity+13.2097−0.3905 → −0.45260.1443 → 0.2270
Cell penetration+15.6881−0.1857 → −0.45200.1113 → 0.2053
Caco-2 permeability+15.9170−0.2671 → −0.51820.1065 → 0.2126
Half-life+14.0581−0.4147 → −0.61100.2666 → 0.6896

Note: ΔE\Delta E is endpoint minus start (negative is better). Dual-objective endpoints pick the highest-property archive candidate with search-score ΔE≤0\Delta E \le 0, keeping the start otherwise.

Beyond the search scorer, five-objective optimization also preserves receptor binding in additional all-atom Rosetta FastRelax evaluations (−28.29 vs. −22.69 REU for unconstrained single-objective Caco-2, p=2.4×10−8p = 2.4 \times 10^{-8}) and in template-free AlphaFold2-Multimer v3 co-folding from sequence alone (0.595–0.649 ipTM across policies vs. 0.563–0.578 for single-property SPIE, Appendix I.6, Table 29).

W2: Generalization of property improvements to independent evaluators excluded from search.

A: We re-evaluated our saved starting and endpoint peptides without candidate re-selection using external predictors excluded from search: the official pretrained ToxinPred 3.0 [2] Hybrid model for non-toxicity, external predictors built on frozen PepLand [3] and ESM-2 [4] encoders for cell penetration and half-life, and a permeability predictor trained on frozen PepLand representations and public nc-CPP data for Caco-2 candidates, with zero exact overlap between evaluated candidates and predictor training splits.

All four properties show positive shifts under SPIE:

Optimized ObjectiveHeld-Out External EvaluatorMean Shift [95% CI] (SPIE)
Non-toxicityToxinPred 3.0 Hybrid (AUC 0.900)+0.204 [0.160, 0.250]
Cell penetrationPepLand + ESM-2 classifier (AUC 0.938)+0.393 [0.343, 0.443]
Caco-2 permeabilityPepLand nc-CPP predictor (RMSE 0.9426, ρ=0.6236\rho = 0.6236)+0.2817 [0.0944, 0.4715]
Half-lifePepLand + ESM-2 SVR (5-fold ρ=0.473\rho = 0.473)+0.398 [0.302, 0.496] h

Under SPIE, the predicted non-toxic fraction rises from 62.0% to 93.6%, supporting cross-predictor transfer. Multi-round SPIE also exceeds single-round BoN across all four external evaluators in Table 26, indicating that iterative local edits accumulate transferable property gains rather than overfitting single-round scorer rankings. Full experimental protocols (data sources, splits, validation metrics, and overlap controls) and results across all four search policies are detailed in Section 3.3 and Appendix I.4 (Table 26).

AuthorsOpenReview ↗

To Reviewer qUGA Part II

W3: Comparisons with modern conditional generative models in joint redesign.

A: We added locally masked PepFlow [5] and PepHAR [6], alongside shared-backbone ProteinMPNN [7] and LigandMPNN [8] controls, evaluated on 64 candidates per complex under identical test complexes, editable segments, and retained context:

MethodAAR (%) ↑BB-RMSD (Å) ↓Contact F1 (%) ↑Closure (%) ↑
Sequence-blind backbone + ProteinMPNN / LigandMPNN12.60 / 11.453.43 / 3.4351.19 / 51.7499.42 / 99.42
EditFlow backbone + ProteinMPNN / LigandMPNN26.66 / 25.412.55 / 2.5563.96 / 64.5799.25 / 99.25
PepHAR14.644.7146.1992.09
Locally masked PepFlow30.092.9860.1298.47
EditFlow40.782.5565.1099.25

Because EditFlow explicitly samples residue and variable masks during training rather than applying masks only at inference time like PepFlow, it improves AAR by +10.69 pp, lowers BB-RMSD from 2.98 to 2.55 Å, and raises Contact F1 from 60.12% to 65.10%. Even on the exact same 2.55 Å EditFlow backbones, joint EditFlow outperforms ProteinMPNN and LigandMPNN, confirming the value of joint sequence–structure editing over two-stage design. Detailed experimental protocols and full results are provided in Table 2, Appendices G.4 and H.3, and Tables 10–12.

References:
[1] Alford et al., The Rosetta All-Atom Energy Function for Macromolecular Modeling and Design, JCTC 2017.
[2] Rathore et al., ToxinPred 3.0: An improved method for predicting the toxicity of peptides, Comput. Biol. Med. 2024.
[3] Zhang et al., PepLand: a large-scale pre-trained peptide representation model for a comprehensive landscape of both canonical and non-canonical amino acids, Brief. Bioinform. 2025.
[4] Lin et al., Evolutionary-scale prediction of atomic-level protein structure with a language model, Science 2023.
[5] Li et al., Full-Atom Peptide Design based on Multi-modal Flow Matching, ICML 2024.
[6] Li et al., Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension, ICLR 2025.
[7] Dauparas et al., Robust deep learning–based protein sequence design using ProteinMPNN, Science 2022.
[8] Dauparas et al., Atomic context-conditioned protein sequence design using LigandMPNN, Nat. Methods 2025.

AuthorsOpenReview ↗

Follow up on our response

Dear Reviewer qUGA,

As the discussion period is nearing its end, we would like to briefly follow up on our response above, which addressed your three points: binding-affinity preservation during optimization (W1, Tables 27-29), generalization of the property gains to independent evaluators excluded from the search (W2, Table 26), and stronger baselines for joint redesign (W3, Tables 10-16).

If any of these remains unclear or unconvincing, we would be grateful if you could let us know which one, and we will do our best to address it with additional analysis before the deadline.

Thank you again for your time.

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

This paper introduces EditFlow, a flow-matching framework for selectively editing receptor-bound peptides. Residue–variable masks specify which amino-acid identities, backbone frames, and torsion angles may change, while preserving the receptor and unselected peptide variables. A shared model is trained on full generation and four local editing tasks: side-chain packing, sequence redesign, backbone remodeling, and joint sequence–structure redesign. The frozen editor is subsequently combined with external property evaluators and search strategies for single- and multi-objective optimization. Experiments on protein–peptide complexes evaluate pocket-conditioned generation, local reconstruction, sparse-anchor completion, and property optimization. The authors report improved reconstruction metrics in several settings and higher five-objective hypervolume than PepFlow under matched search configurations, supporting the use of a unified editing model for computational peptide design and lead optimization.

Critical strengths
  1. The paper addresses selective modification of receptor-bound peptides through explicit residue–variable masks. This formulation provides control over both the locations and molecular variables being redesigned, while preserving the unselected context.
  2. A unified model supports side-chain packing, sequence redesign, backbone remodeling, joint redesign, and full generation. Its integration with different external objectives and search policies without retraining demonstrates the flexibility of the editing interface.
Critical weaknesses
  1. The joint-redesign baselines are insufficient to establish the claimed performance advantage. In Table 2, independent priors and sequential Rosetta achieve amino-acid recovery rates of 6.14% and 4.68%, respectively, close to uniform random assignment. The sequential baseline first samples a backbone from a geometric prior and then performs fixed-backbone sequence design. Its poor performance may therefore primarily reflect backbone quality rather than an inherent limitation of sequential design. These comparisons demonstrate improvements over the selected baselines, but do not sufficiently establish state-of-the-art performance without stronger learned alternatives under matched editing conditions.
  2. The contributions of the model, sampling procedures, and geometric refinement are not sufficiently separated. The reported 99.27% closure rate is obtained after Cartesian minimization, leaving unclear whether EditFlow directly generates well-connected peptide structures or relies substantially on subsequent repair. Window refinement also benefits different search policies inconsistently, and equal proposal counts do not necessarily imply equal computational costs. Although the matched comparison with PepFlow provides useful evidence, additional pre-refinement results and compute-matched sampling ablations are needed to identify the sources of improvement.
Questions for authors
  1. How does EditFlow compare with stronger learned baselines on joint redesign? In particular, locally masked PepFlow and a learned backbone-reconstruction pipeline followed by ProteinMPNN or LigandMPNN should be evaluated under the same editable regions, retained context, sampling budget, and post-processing.
  2. How much of the reported structural quality is achieved before Cartesian minimization? Please report backbone RMSD, interface contact F1, and closure rates before and after refinement. Comparing joint integration and window refinement under matched network-evaluation budgets would also help distinguish model improvements from additional sampling computation.
  3. To what extent are the results in Table 1 comparable across evaluation protocols? Please clarify the differences in baseline settings and discuss the lower amino-acid recovery and diversity alongside the geometric improvements. A matched reevaluation of key baselines, particularly PepFlow, would strengthen the conclusions about full-generation performance.
2 comments
AuthorsOpenReview ↗

To Reviewer bjiV

We thank you for recognizing our formalization of residue and variable masks with dependency closure and the practical design of combining a unified local editor with different search policies, and for your constructive comments that helped us improve the manuscript. In the revision, we (1) added learned baselines (PepFlow, PepHAR, and PocketXMol) and shared backbone MPNN controls for joint redesign (W1, Q1); (2) separated model generation, sampling budget, and local minimization via pre and post refinement ablations and optimization with matched integration steps (W2, Q2); and (3) clarified Table 1 evaluation protocols and reevaluated PepFlow under matched full generation settings (Q3). We address your concerns and questions as follows.

W1: Comparisons with stronger baselines for joint redesign.

A: Beyond geometric priors and sequential Rosetta [1] (6.14% and 4.68% AAR), we added locally masked PepFlow [2], adapted PepHAR [3], shared-backbone ProteinMPNN [4] and LigandMPNN [5], and PocketXMol [6] under matched segments, retained context, and local minimization (64 candidates). Protocols and results are included in the revised PDF (Section 3.1, Table 2, Appendices G.4 and H.3, Tables 10–16; paired tests in Q1):

  • Selected Table 2 baselines: | Method | AAR (%) ↑ | BB-RMSD (Å) ↓ | Contact F1 (%) ↑ | |---|---:|---:|---:| | PepHAR | 14.64 | 4.71 | 46.19 | | PepFlow (200 steps) | 30.09 | 2.98 | 60.12 | | EditFlow (50 steps) | 40.78 | 2.55 | 65.10 |

EditFlow achieves 99.25% closure and samples in 0.18 s per candidate before minimization, 3.9× faster than PepFlow (0.70 s).

  • PocketXMol (Table 13): Because 48.13% of test complexes appear in PocketXMol's PepBDB splits (52.49% valid rate, resampled to 64 candidates), we stratify by overlap. On the non-overlapping subset (51.87%), EditFlow leads all three metrics at 37.43% vs. 14.59% AAR, 2.98 vs. 3.29 Å BB-RMSD, and 59.43% vs. 56.33% Contact F1 (full set: 17.47% AAR, 2.48 Å, 67.23%; raw BB-RMSD 2.43 vs. 2.65 Å).
  • Co-folding (Tables 15–16): Under five-model AlphaFold2-Multimer v3 [7] from sequence alone, EditFlow ranks first among all ten redesign methods (0.655 ipTM, 2.42 Å scRMSD, 35.83% pass rate), with all paired pass-rate differences having 95% CIs above zero.

W2: Separating flow generation, sampling budget, and local Cartesian minimization.

A: We separated these three factors via pre-minimization ablations, stereochemical checks, and search with matched integration steps (Appendices G.7, H.3, and I.2; tables in Q2):

  • Flow generation vs. minimization: Before minimization, EditFlow already achieves the lowest raw BB-RMSD (2.43 Å) and 64.09% Contact F1, with 74.12% of edited C–N bonds ≤1.40\le 1.40 Å (95.13% Cα clash-free). Because AlphaFold2-style [8] SE(3)\mathrm{SE}(3) rigid frames do not hard-constrain inter-residue C–N junctions during ODE integration, strict whole-segment closure (1.33±0.201.33 \pm 0.20 Å on every bond) is 11.33% pre-minimization; minimization leaves AAR unchanged, shifts BB-RMSD to 2.55 Å (+0.13 Å), and raises closure to 99.25%. Across all seven checks from ProteinBench [9], AlphaFold [8], and MolProbity [10] (Table 11), EditFlow has lower violation rates than PepFlow both before and after minimization.
  • Integration steps vs. window refinement: With both models using 200 integration steps per proposal (134.8 vs. 135.3 s per complex), EditFlow outperforms PepFlow across all four search policies (p≤3.7×10−3p \le 3.7 \times 10^{-3}). Window refinement benefits broad parallel search (BoN, Beam), whereas single-pass 200-step integration suffices for iterative small-window search (SPIE, MCTS).

Q1: Paired comparisons against locally masked PepFlow and shared-backbone MPNN pipelines.

A: Paired complex-level differences under the matched Table 2 settings (Appendix Table 12) are:

Baseline Compared with 50-Step EditFlowBaseline AAR (%)Paired Δ\DeltaAAR (pp) ↑Paired Δ\DeltaBB-RMSD (Å) ↓Paired Δ\DeltaContact F1 (pp) ↑
Locally masked PepFlow (200 steps)30.09+10.69 [7.67, 13.65]−0.43 [−0.52, −0.33]+4.98 [3.70, 6.26]
Locally masked PepFlow (matched 50 steps)30.03+10.74 [7.73, 13.73]−0.45 [−0.54, −0.35]+5.06 [3.84, 6.28]
Seq.-blind EditFlow bb. (3.43 Å) + ProteinMPNN / LigandMPNN12.60 / 11.45+28.17 / +29.33−0.87 / −0.87+13.91 / +13.36
Joint EditFlow bb. (2.55 Å) + ProteinMPNN / LigandMPNN26.66 / 25.41+14.12 / +15.360.00 / 0.00+1.14 / +0.53

PepFlow stays near 30.0% AAR at both 50 and 200 steps, confirming that EditFlow's gain stems from mask-based training rather than step count. Upgrading from 3.43 Å sequence-blind backbones to 2.55 Å joint backbones doubles MPNN AAR, yet a +14.12 and +15.36 pp paired gap remains on identical 2.55 Å backbones because fixed-backbone design cannot adjust ~2.5 Å coordinate deviations.

AuthorsOpenReview ↗

To Reviewer bjiV Part II

Q2: Pre/post-minimization structural metrics and optimization comparison with matched integration steps.

A: Pre- vs. post-minimization metrics (Appendix Tables 10 and 13) and five-objective hypervolumes under matched 200-step joint integration and window refinement (Figure 4g, Appendix Table 24) are summarized below:

Generative ModelBB-RMSD (Å, Pre → Post) ↓Contact F1 (%, Pre → Post) ↑Closure (%, Pre → Post) ↑
PepHAR (adapted)6.39 → 4.7142.12 → 46.1918.40 → 92.09
Locally masked PepFlow (200 / 50 steps)4.09 → 2.98 / 4.11 → 3.0054.17 → 60.12 / 54.11 → 60.048.51 → 98.47 / 7.53 → 98.49
EditFlow (50 steps, unfiltered)2.43 → 2.5564.09 → 65.1011.33 → 99.25
PocketXMol (100 steps, full set, resampled)†^\dagger2.65 → 2.4867.38 → 67.2398.49 → 100.00

Note: †^\daggerEvaluated after resampling 47.51% invalid peptide graphs across the full test set (on the non-overlapping subset (51.87% of test complexes), EditFlow outperforms PocketXMol at 2.98 vs. 3.29 Å and 59.43% vs. 56.33% Contact F1).

Model and Sampler (64 Proposals)Runtime / Complex (s)BoN HVSPIE HVBeam HVMCTS HV
Locally masked PepFlow (200-step joint)135.30.11980.07110.08790.0698
EditFlow (200-step joint)134.80.17080.09730.11350.1013
EditFlow (window refinement)168.80.18220.09110.13380.0939

With both models using 200 integration steps, paired ΔHV\Delta\mathrm{HV} (EditFlow minus PepFlow) is strictly positive on all four policies: BoN +0.0510 [0.0342, 0.0690] (p=8.8×10−9p = 8.8 \times 10^{-9}), SPIE +0.0261 [0.0123, 0.0428] (p=9.5×10−5p = 9.5 \times 10^{-5}), Beam +0.0256 [0.0116, 0.0405] (p=3.7×10−3p = 3.7 \times 10^{-3}), and MCTS +0.0315 [0.0201, 0.0439] (p=3.9×10−7p = 3.9 \times 10^{-7}). On the subset where window refinement requires no more function evaluations than the joint-integration control, its SPIE HV (0.0929) still beats 200-step PepFlow (0.0871).

Q3: Protocol comparability in Table 1 and matched re-evaluation of PepFlow.

A: Table 1 compiles published numbers from differing protocols: PepHAR conditions on K=3K = 3 ground-truth hotspot anchors (matched anchor comparisons are in Table 3), while high TM-align [11] diversity in ProteinGenerator [12] (54.00%) and Diffusion (57.00%) reflects low pocket recovery (BSR 24.62% and 49.83%, RMSD 4.35 and 3.28 Å). To remove relaxation and reference-pipeline differences between PepFlow and EditFlow, we ran a matched anchor-free re-evaluation on native crystal complexes (protocols and results in Section 3.1, Appendices G.3 and G.8). Under this shared protocol, EditFlow and PepFlow reach 42.62% and 44.19% AAR (−1.58 pp), 1.80 and 1.83 Å peptide Cα RMSD (4.38 and 4.59 Å complex RMSD), 76.38% and 76.26% SSR, and 87.98% and 87.96% BSR (78.24% and 77.61% under strict 6 Å Cβ), while EditFlow achieves higher structural diversity at 38.41% compared with 36.48% (+1.93 pp [+1.23, +2.64]).

References:
[1] Leaver-Fay et al., ROSETTA3: An object-oriented software suite for the simulation and design of macromolecules, Methods Enzymol. 2011.
[2] Li et al., Full-Atom Peptide Design based on Multi-modal Flow Matching, ICML 2024.
[3] Li et al., Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension, ICLR 2025.
[4] Dauparas et al., Robust deep learning–based protein sequence design using ProteinMPNN, Science 2022.
[5] Dauparas et al., Atomic context-conditioned protein sequence design using LigandMPNN, Nat. Methods 2025.
[6] Peng et al., Unified modeling of 3D molecular generation via atomic interactions with PocketXMol, Cell 2026.
[7] Evans et al., Protein complex prediction with AlphaFold-Multimer, bioRxiv 2021.
[8] Jumper et al., Highly accurate protein structure prediction with AlphaFold, Nature 2021.
[9] Ye et al., ProteinBench: A Holistic Evaluation of Protein Foundation Models, ICLR 2025.
[10] Chen et al., MolProbity: all-atom structure validation for macromolecular crystallography, Acta Crystallogr. D 2010.
[11] Zhang & Skolnick, TM-align: a protein structure alignment algorithm based on the TM-score, NAR 2005.
[12] Lisanza et al., Multistate and functional protein design using RoseTTAFold sequence space diffusion, Nat. Biotechnol. 2025.

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

This paper proposed EditFlow, a flow matching model for pocket conditioned all-atom peptide editing and lead optimization. By combining residue-variable masks with dependency closure, the method controls local modifications across sequence, backbone frames, and side chains while strictly preserving unselected atoms. The frozen model is coupled with external scoring and discrete search algorithms to iteratively optimize molecular properties without model retraining.

Critical strengths
  1. Unified sequence redesign, backbone remodeling, side-chain repacking, and joint co-design into a single flow-matching model, allowing explicit control over edit regions and physical variables while keeping the unedited context intact.
  2. Decouples candidate generation from external property evaluation and search. A single set of frozen model weights serves diverse search policies (BoN, Beam Search, MCTS, SPIE) across multiple property objectives.
Critical weaknesses
  1. Lacks sequence-conformation consistency validation metrics by co-folding. Metrics like RMSD, SSR and BSR only assess geometric similarity to native structures, local relaxation and energy evaluation starting from generated coordinates cannot prove designed sequences will autonomously adopt the bound pose.
  2. Baselines are insufficient to establish practical advantages in joint redesign, the Rosetta baseline (random backbone sampling + fixed backbone sequence design) is an weak control that fail to isolate the benefit of joint modeling. Furthermore, the BSR definition diverges from cited literature (PepFlow, I mean 10\AA is a rough threshold), and the paper lacks direct comparisons with methods capable of motif-scaffolding and local conditioning such as RFD3 (even recent RFD4-Proteina).
Questions for authors
  1. Fixing unedited coordinates is an algorithmic constraint, not evidence of preserved binding. Without independent co-folding checks, how do you distinguish genuine lead optimization from merely optimizing proxy scores on an artificially rigidified pose? or you can just provide some cofolding results.
  2. In the joint redesign experiment, how do you decouple the contribution of joint sequence-structure modeling from backbone generation quality? Can you provide some more powerful baseline with matched prior information and sampling budget to verify that EditFlow is more useful?
5 comments
AuthorsOpenReview ↗

To Reviewer szBc

We thank you for recognizing EditFlow's unified editing formulation and its decoupling of candidate generation from external evaluation and search, and for your constructive comments that helped us improve the manuscript. In the revision, we (1) added sequence only AlphaFold2 Multimer v3 cofolding and constrained relaxation across redesign methods and optimization archives (W1, Q1); (2) added conditional generative baselines (PepFlow, PepHAR, PocketXMol, and RFdiffusion3) and shared backbone MPNN controls, and clarified the BSR definition (W2, Q2); and (3) evaluated interface energy preservation under dual and five objective search (Q1). We address your concerns and questions as follows.

W1. Sequence–conformation consistency validation via co-folding from sequence alone.

A: In template-free AlphaFold2-Multimer v3 [1, 2] co-folding from sequence alone, EditFlow ranks first across all ten redesign methods on every metric (protocols and results in Section 3.1, Appendices G.8 and H.3, Tables 15–16). We folded the first candidate per complex from unpaired receptor MSAs and single-sequence peptide inputs without templates or coordinates, followed by constrained Rosetta FastRelax [3]:

MethodipTM ↑scRMSD (Å) ↓Pass Rate (%) ↑Design Contact F1 (%) ↑Constrained ΔG\Delta G (REU) ↓
Reference peptide sequence0.7631.5959.3662.83−34.11
Sequential Rosetta redesign0.5083.5011.2320.09−2.40
EditFlow bb. + ProteinMPNN0.6342.5327.2737.28−20.90
EditFlow bb. + LigandMPNN0.6442.4729.4139.94−20.89
PepHAR0.5993.0622.9930.63−17.07
PocketXMol0.6182.6025.1340.64−21.33
Locally masked PepFlow0.6462.7227.8139.25−22.41
EditFlow0.6552.4235.8343.55−23.49

Note: Pass Rate requires ipTM>0.6\mathrm{ipTM} > 0.6 and scRMSD<2\mathrm{scRMSD} < 2 Å. Design Contact F1 measures 5 Å heavy-atom contacts after receptor-pocket Cα alignment; Native Contact F1 is 45.81% for EditFlow, compared with 40.93% for PepFlow and 39.72% for PocketXMol.

In paired bootstrap tests, EditFlow's 35.83% pass rate has 95% CIs strictly above zero against all nine baselines, beating PepFlow by +8.02 pp, PocketXMol by +10.70 pp, and shared-backbone ProteinMPNN and LigandMPNN by +8.56 and +6.42 pp (−2.59 and −2.60 REU ΔG\Delta G). See Q1 for optimization co-folding.

W2. Stronger joint-redesign baselines and the 10 Å BSR contact threshold.

A: We added three sets of baseline and metric controls:

  • Conditional baselines in Table 2. Across 64 candidates per complex under matched segments, retained context, and local minimization, EditFlow achieves the best AAR (40.78%) and Contact F1 (65.10%) and tied-lowest BB-RMSD (2.55 Å, with 99.25% closure) in Table 2, with 0.18 s pre-minimization sampling per candidate. It outperforms locally masked PepFlow [4] at 30.09% AAR, 2.98 Å BB-RMSD, and 60.12% Contact F1, PepHAR [5] at 14.64% AAR and 4.71 Å BB-RMSD, and two-stage ProteinMPNN [6] and LigandMPNN [7] controls discussed in Q2, while achieving lower violation rates than PepFlow on all seven checks from ProteinBench [8] and MolProbity [9] (protocols and results in Section 3.1, Appendices G.4 and H.3, Tables 10–14).
  • PocketXMol and RFdiffusion3. We stratify PocketXMol [10] results by overlap with its PepBDB splits, which contain 48.13% of our test complexes (Appendices G.4 and H.3, Table 13). Following your suggestion, we evaluated RFdiffusion3 [11]; RFD4-Proteina was released too recently for us to include in this comparison. RFD3's PDB training data extend through December 2024, and all our test structures predate this cutoff. Potential training overlap limits the fairness of this comparison: | Cohort and Method | AAR (%) ↑ | BB-RMSD (Å) ↓ | Contact F1 (%) ↑ | |---|---:|---:|---:| | Non-overlap subset: PocketXMol | 14.59 | 3.29 | 56.33 | | Non-overlap subset: EditFlow | 37.43 | 2.98 | 59.43 | | Full set: RFdiffusion3 | 21.11 | 2.40 | 67.39 | | Full set: EditFlow | 40.78 | 2.55 | 65.10 |

EditFlow leads all three metrics on the non-overlapping PocketXMol subset. On the full set, with invalid PocketXMol reconstructions resampled, EditFlow retains a +23.31 pp AAR lead, lower pre-minimization BB-RMSD (2.43 vs. 2.65 Å), and a higher co-folding pass rate (W1). RFD3 achieves lower BB-RMSD and higher Contact F1, while EditFlow achieves higher AAR.

Stricter 6 Å Cβ BSR. Tables 1 and 3 follow the 10 Å Cα cutoff in the released PepFlow and PepHAR scripts for comparability, where EditFlow and PepFlow reach 87.98% and 87.96% BSR in matched full-peptide generation. Under a stricter 6 Å Cβ cutoff, EditFlow reaches 78.24% BSR compared with 77.61% for PepFlow (Appendix G.8), and Table 2 evaluates atomic 5 Å heavy-atom Contact F1.

AuthorsOpenReview ↗

To Reviewer szBc Part II

Q1. Distinguishing preserved binding from fixed-pose proxy optimization and co-folding checks on optimized peptides.

A: Incorporating interface energy into multi-objective editing mitigates the degradation in relaxed interface energy and sequence-only co-folding metrics seen in unconstrained single-property search. When hotspot priors are available, residue masks can lock key binding anchors directly while generating the rest of the peptide freely (2.18 Å peptide Cα RMSD from two anchors, Table 3). When the editable window shifts across rounds without predefined hotspot anchors, two independent evaluations assess binding and structural consistency.

  • Multi-objective energy retention. Decoupling generation from scoring allows interface energy to join any property search without retraining. Under the search scorer, dual-objective SPIE achieves positive 2D ΔHV\Delta\mathrm{HV} across all four properties (+0.1065 to +0.2666 at 64 proposals and +0.2053 to +0.6896 at 256) with negative mean ΔE\Delta E (−0.4147 to −0.1857 REU at 64 and −0.6110 to −0.4520 REU at 256, Section 3.3, Appendix I.5, Tables 27–28). In additional all-atom Rosetta FastRelax evaluations, unconstrained single-property searches shift relaxed ΔG\Delta G upward by +6.51 to +15.92 REU, whereas five-objective EditFlow retains strong binding energy at −28.29 REU (−28.99 REU best-energy), substantially outperforming unconstrained single-property optimization (−22.69 REU, p=2.4×10−8p = 2.4 \times 10^{-8}, with 4.94 vs. 3.79 interface H-bonds).
  • Sequence-only co-folding on optimization archives. Five-model AlphaFold2-Multimer v3 co-folding across single-objective endpoints and all 14,249 feasible five-objective Pareto candidates provides independent evidence of sequence–conformation consistency (Section 3.3, Appendix I.6, Table 29). Single-objective interface-energy search maintains native-like binding (0.739–0.757 ipTM vs. 0.763 start, 58.21%–62.25% Native Contact F1 vs. 62.83% start), whereas unconstrained single-property SPIE drops to 0.563–0.578 ipTM and 29.66%–33.47% Native Contact F1. Five-objective archives (0.595–0.649 ipTM, constrained ΔG\Delta G −19.03 to −23.58 REU) improve over the four-property average of single-property searches under every policy. Under matched 200-step five-objective SPIE, EditFlow significantly outperforms PepFlow in ipTM (0.612 vs. 0.574, p=1.7×10−3p = 1.7 \times 10^{-3}), peptide pLDDT (52.06 vs. 48.22, p=1.8×10−3p = 1.8 \times 10^{-3}), and scRMSD (3.25 vs. 3.65 Å, p=2.1×10−4p = 2.1 \times 10^{-4}), while achieving a higher pass rate (24.06% vs. 16.04%).

Q2. Decoupling joint sequence–structure modeling from backbone quality under matched budgets.

A: Shared-backbone inverse-folding controls and a matched 50-step generative comparison separately evaluate backbone quality and joint sequence–structure co-design (Section 3.1, Table 2, Appendix G.4, Table 12).

  • Sequence-blind backbones (3.43 Å BB-RMSD). ProteinMPNN and LigandMPNN reach 12.60% and 11.45% AAR.
  • Shared joint EditFlow backbones (2.55 Å BB-RMSD). AAR rises to 26.66% and 25.41%, isolating the gain from improved backbone geometry.
  • EditFlow joint output (same 2.55 Å backbones). AAR reaches 40.78%, exceeding ProteinMPNN and LigandMPNN by +14.12 and +15.36 pp on identical coordinates.

Fixed-backbone inverse folding cannot adjust ~2.5 Å coordinate deviations in generated backbones, whereas joint flow co-adapts sequence and structure simultaneously. Separately, under this matched setting, locally masked PepFlow reaches 30.03% AAR and 3.00 Å BB-RMSD, confirming EditFlow's lead.

We hope these co-folding, baseline, and energy-relaxation results address your concerns, and we welcome any further questions.

References:
[1] Evans et al., Protein complex prediction with AlphaFold-Multimer, bioRxiv 2021.
[2] Mirdita et al., ColabFold: making protein folding accessible to all, Nat. Methods 2022.
[3] Alford et al., The Rosetta All-Atom Energy Function for Macromolecular Modeling and Design, JCTC 2017.
[4] Li et al., Full-Atom Peptide Design based on Multi-modal Flow Matching, ICML 2024.
[5] Li et al., Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension, ICLR 2025.
[6] Dauparas et al., Robust deep learning–based protein sequence design using ProteinMPNN, Science 2022.
[7] Dauparas et al., Atomic context-conditioned protein sequence design using LigandMPNN, Nat. Methods 2025.
[8] Ye et al., ProteinBench: A Holistic Evaluation of Protein Foundation Models, ICLR 2025.
[9] Chen et al., MolProbity: all-atom structure validation for macromolecular crystallography, Acta Crystallogr. D 2010.
[10] Peng et al., Unified modeling of 3D molecular generation via atomic interactions with PocketXMol, Cell 2026.
[11] Butcher et al., De novo design of all-atom biomolecular interactions with RFdiffusion3, bioRxiv 2025.

Thank you for the thorough and honest rebuttal. The new experiments directly address my main concerns:

  1. Co-folding and binding retention: The sequence-only AlphaFold2-Multimer v3 evaluations across the redesign baselines and the five-objective Pareto candidates provide solid evidence of sequence–conformation consistency. I also appreciate the transparent acknowledgment that unconstrained single-objective optimization degrades binding affinity, and the demonstration that incorporating interface energy into multi-objective search—together with independent external predictor validation—mitigates proxy-score overfitting.
  2. Stronger baselines and decoupling joint modeling: Evaluating ProteinMPNN and LigandMPNN on the exact 2.55 Å EditFlow backbones cleanly separates backbone quality from joint sequence–structure co-design (+14.12 to +15.36 pp AAR gain on identical coordinates). Adding locally masked PepFlow, PepHAR, stratified PocketXMol, RFdiffusion3, and the stricter 6 Å Cβ BSR metric makes the comparison much more rigorous and convincing.

While minor limitations remain—such as the reliance on local Cartesian minimization for strict whole-segment closure and a moderate ipTM gap between optimized multi-objective peptides and native references—my primary soundness concerns have been satisfactorily resolved.

Reviewer_6D5qOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EditFlow, a flow-matching model for structured full-atom editing of bound peptides. Residue–variable masks allow a shared model to perform different editing operations while preserving the retained molecular state. The model is further combined with external property evaluation and search for peptide lead optimization.

Critical strengths
  1. Flexible and unified editing framework. EditFlow supports side-chain packing, sequence redesign, backbone remodeling, and joint redesign within a single model.

  2. Clear presentation. The paper is well organized and easy to follow. The figures clearly illustrate the proposed editing framework and different editing operations.

Critical weaknesses
  1. Limited validation of peptide lead optimization. The property optimization experiments rely on four PeptiVerse predictions and Rosetta interface energy (Sec. 3.3). The same property scores are used for candidate selection and evaluation. Although EditFlow improves the reported scores and hypervolume, the paper does not provide independent property evaluation or experimental validation of the edited peptides. It remains unclear whether these improvements reflect actual gains in peptide properties and target engagement, rather than optimization toward the chosen predictors.

  2. Insufficient learned baselines for joint sequence–structure redesign. Table 2 compares EditFlow only with independent priors and sequential Rosetta redesign. e.g., PocketXMol [1] supports constrained peptide design with fixed residue positions and types, while PepFlow [2] supports joint full-atom generation and has been adapted to preserve selected residues during sampling in PepHAR [3].

[1]. Unified modeling of 3D molecular generation via atomic interactions with PocketXMol

[2]. Full-Atom Peptide Design based on Multi-modal Flow Matching

[3]. Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension

Questions for authors

See Weaknesses

4 comments
AuthorsOpenReview ↗

To Reviewer 6D5q

We thank you for recognizing EditFlow's flexible and unified editing framework and the clear presentation of our paper, and for your constructive comments that helped us improve the manuscript. In the revision, we (1) validated optimized peptides via unseen external predictors, interface energy preservation, and sequence only cofolding (W1); and (2) added PocketXMol, PepFlow, PepHAR, and shared backbone MPNN baselines for joint redesign (W2). We address your two concerns as follows.

W1. Independent validation of property improvements and receptor binding retention during lead optimization.

A: Evaluating our saved starting and endpoint peptides directly without candidate re-selection across external property predictors excluded from search, all-atom Rosetta FastRelax, and template-free AlphaFold2-Multimer v3 co-folding confirms both cross-evaluator property gains and receptor binding retention (Section 3.3, Appendices I.1 and I.4–I.6, Tables 26–29).

  • External property evaluators (Appendix I.4, Table 26). With zero exact sequence or molecular-structure overlap between evaluated candidates and external training splits, we re-evaluated saved endpoints with official pretrained ToxinPred 3.0 [1] and external heads built on frozen PepLand [2] and ESM-2 [3] representations. All four optimized properties improve across all four search policies, with all primary paired 95% CIs strictly above zero.
Optimized ObjectiveExternal Evaluator Excluded from SearchBoNSPIEBeamMCTS
Non-toxicityToxinPred 3.0 Hybrid (AUC 0.900)+0.166+0.204+0.206+0.199
PenetrationPepLand + ESM-2 classifier (AUC 0.938)+0.308+0.393+0.357+0.350
Caco-2PepLand nc-CPP permeability (ρ=0.6236\rho = 0.6236)+0.173+0.282+0.422+0.284
Half-lifePepLand + ESM-2 SVR (5-fold ρ=0.473\rho = 0.473)+0.328 h+0.398 h+0.385 h+0.301 h

Under SPIE, the predicted non-toxic fraction on ToxinPred 3.0 rises from 62.0% to 93.6%. Notably, multi-round iterative editing (SPIE) surpasses single-round sampling (BoN) across all four external evaluators in Table 26, and supplementary checks on unoptimized aqueous solubility also shift positively (+0.098 on penetration and +0.079 on Caco-2 endpoints), indicating that iterative local edits accumulate transferable physicochemical improvements rather than overfitting single-round scorer rankings. Appendices I.1 and I.4 report validation metrics for the external evaluators and PeptiVerse Caco-2 (RMSE=1.30\mathrm{RMSE} = 1.30, ρ=0.80\rho = 0.80, starting peptide SD 0.011 near −5.85).

  • Receptor binding-energy retention. When hotspot priors are available, residue masks can lock key binding anchors directly while generating the rest of the peptide freely (2.18 Å Cα RMSD and 86.80% BSR from two anchors, Table 3). Without locked peptide residues, optimizing sequence-level properties in isolation ignores pocket complementarity and shifts all-atom Rosetta REF2015 [4] FastRelax ΔG\Delta G upward by +6.51 to +15.92 REU. Because EditFlow decouples pocket-conditioned proposal generation from external scoring (Section 2), incorporating interface energy turns lead optimization into a retraining-free Pareto search that filters out pocket-disruptive edits. Under the search scorer, dual-objective SPIE yields positive 2D ΔHV\Delta\mathrm{HV} (+0.1065 to +0.2666 at 64 proposals and +0.2053 to +0.6896 at 256) with negative mean ΔE\Delta E (−0.4147 to −0.1857 REU at 64 and −0.6110 to −0.4520 REU at 256, Section 3.3, Appendix I.5, Tables 27–28), and five-objective EditFlow retains strong relaxed binding energy at −28.29 REU (−28.99 REU best-energy) vs. −22.69 REU for unconstrained single-objective Caco-2 (p=2.4×10−8p = 2.4 \times 10^{-8}, with 4.94 vs. 3.79 interface H-bonds).
  • Template-free co-folding (Appendix I.6, Table 29). Sequence-only five-model AlphaFold2-Multimer v3 [5] co-folding across all feasible five-objective Pareto candidates (0.595–0.649 ipTM, constrained ΔG\Delta G −19.03 to −23.58 REU) improves over the four-property average of unconstrained single-property endpoints under every policy. Under matched 200-step five-objective SPIE, EditFlow also significantly outperforms PepFlow across ipTM (0.612 vs. 0.574, p=1.7×10−3p = 1.7 \times 10^{-3}), peptide pLDDT (52.06 vs. 48.22, p=1.8×10−3p = 1.8 \times 10^{-3}), scRMSD (3.25 vs. 3.65 Å, p=2.1×10−4p = 2.1 \times 10^{-4}), and pass rate (24.06% vs. 16.04%).
AuthorsOpenReview ↗

To Reviewer 6D5q Part II

W2. Comparisons with learned conditional baselines (PocketXMol, PepFlow, PepHAR) for joint redesign in Table 2.

A: Across 64 valid candidates per complex under identical editable segments, retained context, and local Cartesian minimization, 50-step EditFlow ranks first in Table 2 across AAR (40.78%), BB-RMSD (2.55 Å), and Contact F1 (65.10%), surpassing locally masked PepFlow [6], adapted PepHAR [7], shared-backbone ProteinMPNN [8] and LigandMPNN [9] controls, and PocketXMol [10] on the non-overlapping subset (51.87% of test complexes) (Section 3.1, Table 2, Appendices G.4 and H.3, Tables 10–16).

Method (Table 2)AAR (%) ↑BB-RMSD (Å) ↓Contact F1 (%) ↑Closure (%) ↑
Independent priors / Sequential Rosetta6.14 / 4.687.52 / 7.2231.56 / 32.5575.97 / 83.13
Sequence-blind backbone + ProteinMPNN / LigandMPNN12.60 / 11.453.43 / 3.4351.19 / 51.7499.42 / 99.42
EditFlow bb. + ProteinMPNN / LigandMPNN26.66 / 25.412.55 / 2.5563.96 / 64.5799.25 / 99.25
PepHAR (adapted)14.644.7146.1992.09
Locally masked PepFlow (200 steps)30.092.9860.1298.47
EditFlow (50 steps)40.782.5565.1099.25
  • Mask-aware training vs. inference-time masking and two-stage design. Increasing locally masked PepFlow from 50 to 200 steps barely changes performance (30.03% to 30.09% AAR and 3.00 to 2.98 Å BB-RMSD, Table 12), demonstrating that EditFlow's +10.69 pp AAR lead (−0.43 Å BB-RMSD, +4.98 pp Contact F1), 3.9× faster sampling (0.18 vs. 0.70 s per candidate, Table 14), and lower violation rates across all seven stereochemical checks (Table 11) stem from training directly under conditional residue–variable masks (Eqs. 2–5) rather than clamping unedited channels only at inference time. Likewise, decoupling sequence and structure hurts both directions (Section 3.1, Table 9). Holding sequence fixed yields 3.51 Å BB-RMSD vs. 2.55 Å under joint redesign, while running ProteinMPNN or LigandMPNN on the shared 2.55 Å EditFlow backbones trails joint EditFlow by 14.12 and 15.36 pp AAR (Table 12) because fixed-backbone inverse folding cannot adjust ~2.5 Å coordinate deviations, whereas joint flow co-adapts identity, frames, and torsions simultaneously (Eq. 6).

  • PocketXMol (Appendix G.4, Table 13). Because 48.13% of test complexes match PocketXMol's PepBDB splits (100 steps, reconstructing valid peptides on 52.49% of initial draws and resampled to 64 accepted candidates per complex), Table 13 stratifies by split overlap. On the non-overlapping subset (51.87% of test complexes), EditFlow leads all three metrics: | Method | AAR (%) ↑ | BB-RMSD (Å) ↓ | Contact F1 (%) ↑ | |---|---:|---:|---:| | PocketXMol | 14.59 | 3.29 | 56.33 | | EditFlow | 37.43 | 2.98 | 59.43 |

  • Sequence-only co-folding (Section 3.1, Appendix H.3, Tables 15–16). Under five-model AlphaFold2-Multimer v3 from sequence alone, 50-step EditFlow ranks first across all ten redesign methods at 0.655 ipTM, 2.42 Å scRMSD, and a 35.83% pass rate, ahead of LigandMPNN (29.41%), PepFlow (27.81%), ProteinMPNN (27.27%), PocketXMol (25.13%), and PepHAR (22.99%), with all paired pass-rate 95% CIs strictly above zero.

We hope these external validation, co-folding, and baseline comparisons address your concerns, and we welcome any further questions.

References:
[1] Rathore et al., ToxinPred 3.0: An improved method for predicting the toxicity of peptides, Comput. Biol. Med. 2024.
[2] Zhang et al., PepLand: a large-scale pre-trained peptide representation model for a comprehensive landscape of both canonical and non-canonical amino acids, Brief. Bioinform. 2025.
[3] Lin et al., Evolutionary-scale prediction of atomic-level protein structure with a language model, Science 2023.
[4] Alford et al., The Rosetta All-Atom Energy Function for Macromolecular Modeling and Design, JCTC 2017.
[5] Evans et al., Protein complex prediction with AlphaFold-Multimer, bioRxiv 2021.
[6] Li et al., Full-Atom Peptide Design based on Multi-modal Flow Matching, ICML 2024.
[7] Li et al., Hotspot-Driven Peptide Design via Multi-Fragment Autoregressive Extension, ICLR 2025.
[8] Dauparas et al., Robust deep learning–based protein sequence design using ProteinMPNN, Science 2022.
[9] Dauparas et al., Atomic context-conditioned protein sequence design using LigandMPNN, Nat. Methods 2025.
[10] Peng et al., Unified modeling of 3D molecular generation via atomic interactions with PocketXMol, Cell 2026.

Thanks to the authors' response. My main concerns are addressed, so I will raise my score.

AuthorsOpenReview ↗

Thank you for your Comment

Thank you very much for your time and for raising your score. We truly appreciate your constructive feedback on independent property validation and comparisons with strong baselines, which has helped us strengthen the paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the Area Chair and all five reviewers for their thoughtful feedback. We are encouraged that the reviewers recognized EditFlow's unified formulation of four local peptide-editing operations via residue–variable masks and its retraining-free integration with external search policies. Below we summarize the five sets of controlled experiments added in the revised manuscript.

  • Conditional joint-redesign and shared-backbone baselines (All Reviewers). Under matched editable segments, retained context, and local Cartesian minimization, 50-step EditFlow ranks first in Table 2 across AAR (40.78%), BB-RMSD (2.55 Å), and Contact F1 (65.10%), while sampling in 0.18 s per candidate (Appendix Table 14). It outperforms 200-step locally masked PepFlow, adapted PepHAR, PocketXMol on the non-overlapping subset (51.87% of test complexes), and ProteinMPNN and LigandMPNN run on the shared 2.55 Å EditFlow backbones (Section 3.1, Table 2, Appendices G.4 and H.3, Tables 10–14).
  • Template-free AlphaFold2-Multimer v3 co-folding (Reviewers szBc, 6D5q, qUGA, bjiV). When refolded from sequence alone without templates or input peptide coordinates, 50-step EditFlow ranks first across all ten evaluated redesign methods at 0.655 ipTM, 2.42 Å scRMSD, and a 35.83% pass rate (Section 3.1, Appendices G.8 and H.3, Tables 15–16). Across all five-objective Pareto candidates, incorporating interface energy improves co-folding confidence and contact retention over unconstrained single-property searches, and matched 200-step EditFlow outperforms PepFlow in ipTM, peptide pLDDT, and scRMSD (Section 3.3, Appendix I.6, Table 29).
  • Binding-energy retention via multi-objective search (Reviewers szBc, 6D5q, qUGA). Without retraining, incorporating interface energy into dual-objective (64 and 256 proposals) and five-objective searches improves target properties while keeping dual-objective search-score ΔE\Delta E non-positive (Section 3.3, Appendix I.5, Tables 27–28). In additional Rosetta FastRelax evaluations, five-objective EditFlow substantially curbs the binding-energy drift of unconstrained single-property search (−28.29 vs. −22.69 REU, p=2.4×10−8p = 2.4 \times 10^{-8}).
  • Cross-predictor property validation (Reviewers 6D5q, qUGA, GhZr). When saved endpoints are evaluated without candidate re-selection on official pretrained ToxinPred 3.0 and external PepLand and ESM-2 predictors, all four optimized tasks show consistent positive paired shifts across all four search policies (Section 3.3, Appendices I.1 and I.4, Table 26).
  • Pre-minimization geometry, search with the same step budget, and full-generation controls (Reviewers szBc, bjiV, GhZr). We added pre- versus post-minimization metrics and seven stereochemical checks, five-objective comparisons using the same 200 integration steps, and a full-peptide generation comparison (Sections 3.1 and 3.3, Figure 4g, Appendices G.2–G.3, G.7–G.8, H.3, and I.2, Tables 6, 10–11, and 24).

We have uploaded a revised manuscript PDF with all changes highlighted in blue, and provide detailed point-by-point responses to each reviewer in our official comments below.

AuthorsOpenReview ↗

Update on the Discussion Period

Dear Area Chair and Reviewers,

We thank all five reviewers for their thoughtful feedback and engagement. Following our revision and individual responses, Reviewers szBc, 6D5q, and GhZr have confirmed that their main concerns are resolved.

As summarized in our General Response above, these resolved points directly overlap with the core concerns raised across all five reviews, including stronger baselines for joint redesign, binding affinity preservation and co-folding validation, independent property evaluation, and minimization and compute ablations. All of these are addressed in our responses to each reviewer and in the revised manuscript.

We remain glad to provide any further clarification during the remainder of the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.