acceptodds
Under review as a conference paper at ICLR 2027

StapleBridge: Finite-Support Intervention Control for Stapled Peptide Optimization

Abstract

Hydrocarbon stapling covalently links two peptide side chains to constrain peptide conformation and can improve proteolytic stability and cellular uptake. Recent computational cyclic-peptide design has largely focused on de novo generation, yet redesigning a functional peptide from scratch can perturb sequence and structural context important for its function. Constrained local editing requires a framework to select chemically and geometrically valid modifications despite scarce paired linear-to-stapled supervision. To address this, we introduce **StapleBridge**, a chemistry-aware discrete Schr\"odinger-control framework for optimizing existing peptides through hydrocarbon stapling. For each peptide, StapleBridge constructs a finite set of executable interventions and maps each to a canonical minimal-edit product, making the optimal intervention distribution exactly computable from terminal utility. Distilling this distribution into an amortized controller provides plan-level supervision without paired examples and enables efficient intervention selection on unseen peptides. On test leads supported by the fixed, property-independent intervention catalog, StapleBridge improves predicted permeability for 81.13% of leads, with a mean permeability gain of 0.169 compared with 0.140 for the strongest matched plan-level baseline. Cross-predictor permeability evaluation and an orthogonal, post-selection 3D evaluation support our observed predicted gains and structural compatibility, while computational experiments on N-methylation and disulfide cyclization support extension to additional chemistries. Overall, StapleBridge enables property-guided refinement of existing peptide leads by learning which feasible chemical edits to apply without paired linear-to-stapled supervision. Code is available at https://anonymous.4open.science/r/Staplebridge-A05A/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
hydrocarbon-stapled peptides, peptide lead optimization, Schrödinger control, stochastic control, constrained molecular optimization, peptide design

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_uvieOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of optimizing existing peptides to make them have constrained, localized chemical modifications, with the primary focus on hydrocarbon stapling. For each peptide, this method constructs a finite intervention space and maps each modification to a canonical product. The experimental results show that the proposed method can actually make accurate modifications to the peptide leads.

Critical strengths
  1. I find that improving original peptides by adding chemical modifications is a really important application question.
  2. The evaluation and experimental results can support the claims, and the comparison also helps find the benefits of each part of the method.
Critical weaknesses
  1. The number of evaluated peptides is relatively small, only 77 test peptides, which limits the strength of the method.
  2. Also, there are only a few baselines; for example, I would like to see more comparisons with a broader set of peptide optimization or constrained selection methods.
Questions for authors
  1. How diverse are the test peptides in terms of sequence, length, and receptor structures? Are the improvements distributed across different samples, or are the concentrations in a small subset of peptides that are similar to each other?
  2. Can the authors explain the motivation for including current baselines? I think you could include more peptide optimization or constrained selection methods.
3 comments

Dear Reviewer uvie,

Thank you for your positive assessment of StapleBridge, particularly your recognition of the importance of peptide lead optimization and the comprehensiveness of our evaluation. We appreciate your constructive suggestions and have conducted additional analyses and baseline comparisons to address both concerns.

Response to Q1: Test-Set Diversity and Distribution of Improvements

To assess test-set diversity and whether improvements are broadly distributed, we conducted additional analyses of peptide sequences, receptor structures, and per-peptide performance.

1. Sequence and receptor diversity.

The 77 test peptides comprise 77 unique sequences spanning 5–21 residues (median length: 11). The fixed 53-peptide hard-supported subset forms 44 sequence clusters at 30% global identity and 53 clusters at 50% identity, indicating substantial sequence diversity. The test set also contains 77 distinct receptor sequences, with no receptor pairs exceeding 30% whole-chain sequence identity under our alignment protocol. These receptors span helix-rich, sheet-rich, and mixed secondary-structure classes.

Furthermore, as detailed in Appendix A, our benchmark construction incorporates receptor-homology-aware splitting at 40% sequence identity, excluding development examples sharing receptor clusters with the LNR test set.

2. Distribution and robustness of improvements.

StapleBridge improves predicted permeability for 43/53 supported peptides (81.1%), with improvements distributed across peptide lengths: 9/14 (5–9 residues), 16/19 (10–14 residues), and 18/20 (15–21 residues). Consistent improvements are also observed across receptor structural classes: 19/22 helix-rich, 16/20 mixed, and 8/11 sheet-rich receptors.

To assess whether the average improvement is dominated by a few exceptional cases, we excluded the ten peptides with the largest positive gains. The remaining 43 peptides still achieve a mean predicted permeability improvement of +0.0953 over their linear parents.

These results indicate that StapleBridge's improvements extend across diverse peptide and receptor contexts rather than being concentrated in a small subset of closely related examples. We have also added a test-set diversity analysis and corresponding visualizations to Appendix A, including peptide-length distributions, pairwise sequence identities, and receptor sequence diversity.

Response to Q2: Baseline Motivation and Extended Comparisons

We appreciate the suggestion to broaden our baseline comparisons. Our original baselines were designed to isolate distinct methodological components under identical chemical, geometric, and editing constraints. We have additionally evaluated an established cyclic-peptide design framework.

1. Motivation for the existing baselines.

Each baseline addresses a specific methodological question:

  • Random Feasible: Tests whether uniform selection within the chemically feasible intervention space is sufficient.
  • Empirical Reference: Tests whether population-level staple preferences alone can guide effective intervention selection.
  • Generic-SB: Evaluates bridge-based optimization without explicit intervention-level control.
  • Direct Plan Ranker (DPR): Isolates the contribution of exact distribution-level supervision beyond matched plan representations and input features.

Together, these baselines distinguish the contributions of feasible intervention selection, empirical reference preferences, structured control, and distribution-level supervision.

2. Additional comparison with CP-Composer.

To extend our evaluation beyond matched intervention-selection baselines, we additionally compared StapleBridge with CP-Composer (Jiang et al., 2025), an established zero-shot cyclic-peptide design framework that uses composable geometric constraints for cyclization.

Since CP-Composer's native stapling chemistry uses K–D/E lactam linkages rather than S5/R8 hydrocarbon staples, we introduced a task-compatible HC Adapter while retaining its released pretrained model without retraining. The adapter translates CP-Composer's structural proposals into hydrocarbon-stapling interventions, which are constructed and evaluated under the same chemical feasibility constraints and property-evaluation protocol as StapleBridge. We evaluated CP-Composer + HC Adapter on all 53 hard-supported test peptides across three seeds.

Table R1. Comparison with CP-Composer under the common hydrocarbon-stapling protocol.

MethodPermeability Improved ↑Mean ΔPermeability ↑Plan Regret ↓
CP-Composer + HC Adapter73.58 ± 3.77%0.1283 ± 0.00460.8467 ± 0.0389
StapleBridge81.13 ± 0.00%0.1690 ± 0.00440.2092 ± 0.0062

StapleBridge achieves a 7.55-percentage-point higher improvement rate and a 31.7% larger mean predicted permeability gain, together with 75.3% lower Plan Regret. These results support the effectiveness of lead-specific intervention optimization compared with an adapted zero-shot geometry-guided cyclic-peptide design baseline under a common hydrocarbon-stapling evaluation protocol.

Overall, the additional diversity analyses and broader baseline comparisons strengthen the evidence for StapleBridge's generalization across peptide contexts and the practical value of lead-specific intervention control.

We thank the reviewer again for the constructive feedback, which has helped us improve the evaluation and presentation of our work.

I appreciate the authors’ response and their careful answers to my questions.

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

The paper describes a discrete diffusion approach to peptide stabling that is introducing chemical bridges between amino acids to stabilise the structure. This is a very relevant application in for example drug design where a stable structure is often desired. Discrete diffusion can be used because the problem is formulated as a sequence of discrete (local) edits of the initial peptide sequence.

The paper contributes new methodology - amortised inference for the control problem of predicting the edit sequence. It is argued that editing is an attractive to the usual de novo design approach. This is seems likely.

A relatively small dataset is used: "A fixed preprocessing pipeline produced 4,020 training, 111 validation, and 77 test leads of at most 25 residues." The validation is despite this quite extensive in that other baseline methods are investigated and predicted molecules are investigated with computational methods.

Critical strengths

Novel approach to relevant very specialised problem. Methodology can be transferred to other domain. Extensive validation on limited dataset.

Critical weaknesses

The dataset used is relatively small and I cannot see anywhere that homology of the systems have been taken in consideration. So there might be overlap between training and test set.

Questions for authors

For final version of the paper: Give a better introduction to the problem. More details on the model used. Could not find it. Perhaps I hav overlooked.

2 comments

Dear Reviewer upYX,

Thank you for your positive assessment of StapleBridge and for recognizing its methodological novelty, practical relevance, and extensive validation. We appreciate your suggestions regarding dataset homology and model presentation, and address them below.

1. Dataset homology and test-set diversity.

We clarify that receptor-homology-aware splitting was already incorporated into our benchmark construction, as described in Appendix A. Receptor sequences were clustered at 40% sequence identity, and all development examples sharing a receptor cluster with the LNR test set were excluded from training and validation. This removed 1,834 potentially overlapping complexes, leaving 4,157 training and 114 validation complexes before preprocessing.

We additionally analyzed test-set diversity. The 77 test peptides contain 77 unique sequences spanning 5–21 residues. The 53 hard-supported peptides form 44 sequence clusters at 30% global identity and 53 clusters at 50% identity. Moreover, all 77 test receptors form distinct clusters at 30% global sequence identity. StapleBridge improves predicted permeability for 43/53 supported peptides (81.1%), with improvements distributed across peptide lengths: 9/14 (5–9 residues), 16/19 (10–14 residues), and 18/20 (15–21 residues).

These results demonstrate substantial sequence and receptor diversity within the held-out benchmark, while the homology-aware splitting protocol limits receptor-level overlap between development and test sets. We have also added a brief description of the test-set diversity analysis and a corresponding figure 3 in Appendix A to further illustrate the distributions of peptide lengths, sequence identities, and receptor diversity.

2. Clarification of problem formulation and model architecture.

StapleBridge addresses lead-specific peptide optimization rather than de novo sequence generation. It first constructs a finite set of chemically and geometrically feasible stapling interventions, each uniquely specifying a canonical minimal-edit product. This structure enables exact computation of the optimal intervention distribution, which is then distilled into an amortized controller.

The controller combines frozen ESM2-650M sequence representations with staple chemistry, anchor-position, and geometric features through a trainable plan-scoring network (hidden dimension 128). It is trained using KL distillation against the exact intervention distribution. A separate plan-conditioned execution policy learns to realize the selected intervention through valid local edits. At inference, the controller selects an intervention without requiring terminal-property evaluations.

We will improve the introduction to clarify the lead-optimization setting and make the architecture and training procedure more explicit in the method and implementation sections.

Thank you again for your constructive suggestions, which will help improve the clarity and accessibility of our manuscript.

Reviewer_upYXOpenReview ↗

Happy with answer to my concerns, but point about exhaustive search raised by other reviewers lower my score

The authors' willingness to answer the reviewers' questions is much appreciated.

Two reviewers raise the major concern that the problem is so constrained that exhaustive search can be applied as well.

Taken together with other minor concerns: small test set and a bit hard to follow presentation, I will change my recommendation to reject with encouragement to extend the methodology to a problem that requires machine learning.

Reviewer_KiLhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed StapleBridge for lead-specific hydrocarbon stapling of existing linear peptides. The method computes an exact target distribution over a predefined finite feasible intervention space based on surrogate property utilities and constraints, and distills it into an amortized controller to eliminate test-time property queries. The selected interventions are evaluated via surrogate predictors and post-hoc 3D analyses.

Critical strengths
  1. Explicitly incorporates chemical rules, geometric feasibility, and protected positions into a finite candidate space, establishing a clear mapping between stapling plans and canonical products.
  2. Compared to a matched plan-level ranker (DPR) sharing identical features and candidate spaces, complemented by cross-predictor validations and post-hoc 3D analyses. it has comprehensive experimental setup.
Critical weaknesses
  1. Methodology contribution is limited, in Eq 11 control optimization only adjusts plans and conditional execution is unchanged. The method reduces to a finite-candidate distillation, and its advantage against DPR may arise form soft-target supervision but not control framework. Full distribution matching is also insufficient justified when deployment uses only argmax, and its value over direct enumeration remains unclear.
  2. Without validation against measured permeability, its unclear whether the model has actual improvements or just learns to exploit the surrogate's inherent preferences (such as for hydrophobic modifications). Low Cα-RMSD under restrained relaxation establishes do NOT indicate conformational stability or receptor binding affinity. So there's insufficient validation of biological utility.
Questions for authors
  1. Would a listwise ranker trained on identical prior-adjusted soft targets, followed by deterministic product construction, match StapleBridge? Relative to direct enumeration, what benefit justifies training and distillation, and how much utility is lost through approximation versus the reference prior changing the selected plan?
  2. Among candidates with matched staple types and comparable hydrophobicity, does the method still identify better modification sites? Can this ranking be validated against measured permeability changes, alongside evidence of conformational stability without backbone restraints and retained receptor binding? Fully resolve my questions may raise my rating.
4 comments

Dear Reviewer KiLh,

Thank you for your careful evaluation and constructive feedback. To address your concerns, we conducted additional exhaustive optimization, experimental site-ranking, and structural validation experiments. Our responses are provided below.

Response to Q1: Methodological Contribution, Listwise Equivalence, and Amortization

We thank the reviewer for this insightful question. We clarify the relationship between StapleBridge and listwise ranking, and provide additional experiments to quantify the effects of the reference prior, amortization, and distribution-level supervision. 

1. Listwise equivalence and methodological contribution.

We agree that a listwise ranker trained with identical prior-adjusted soft targets and parameterization could recover the same intervention distribution. This equivalence is consistent with our formulation, as our contribution lies in deriving the exact supervision target rather than introducing a new ranking objective.

StapleBridge formulates chemically constrained peptide editing as a stochastic-control problem. The canonical plan–product correspondence enables exact computation of the optimal intervention distribution while separating intervention selection from chemical execution.

This formulation offers three advantages: (i) exact chemistry-aware supervision without paired linear-to-stapled examples; (ii) distribution-level preferences over feasible interventions rather than a single optimal-plan label; and (iii) amortized intervention selection across unseen peptide leads without test-time property evaluations.  Thus, StapleBridge transforms chemically constrained peptide optimization into a tractable supervision problem and a reusable intervention policy.

2. Exhaustive comparison, reference-prior effects, and amortization.

To quantify these effects, we conducted exhaustive comparisons under the same feasible intervention space: Exhaustive PV-Max selects the intervention with the highest PeptiVerse-predicted permeability; Exact Energy MAP (no prior) minimizes terminal energy without reference-prior regularization; and Exact Control MAP selects the optimal intervention under the full reference-regularized control objective.

Table R1. Exhaustive optimization and amortization.

MethodMean ΔPermeability ↑Plan Regret ↓Property-scored interventions ↓
Exhaustive PV-Max0.1860.3921,012
Exact Energy MAP (no prior)0.1660.2671,012
Exact Control MAP0.17201,012
StapleBridge0.1690.2090

The exhaustive comparisons highlight the complementary benefits of chemistry-aware reference regularization and amortized control. Incorporating the reference prior improves mean predicted permeability gain from 0.166 (Exact Energy MAP) to 0.172 (Exact Control MAP), demonstrating its value in guiding intervention selection. StapleBridge achieves a closely matched gain of 0.169, retaining 98.3% of the exact control solution's predicted permeability improvement while generalizing intervention preferences to unseen peptide leads.

The key practical advantage is efficient, reusable intervention optimization. Compared with Exhaustive PV-Max, StapleBridge retains 90.9% of the maximum predicted permeability gain while achieving 4.57× faster inference (7.82 vs. 35.72 s across 77 leads) and eliminating 1,012 test-time intervention-level property evaluations. As the number of peptide leads increases, these computational savings accumulate, yielding larger absolute time savings in large-scale screening. This combination of near-exhaustive performance and efficient amortized inference is valuable for large-scale peptide lead optimization.

3. Value of distribution-level supervision under argmax selection.

Although deployment uses deterministic argmax selection, the full soft target provides relative preference information across feasible interventions rather than supervising only the highest-ranked plan.

Our matched ablation confirms this benefit: replacing soft distributional supervision with hard top-1 supervision reduces mean ΔPermeability from 0.169 to 0.160, increases Plan Regret from 0.209 to 0.247, and decreases Spearman agreement with the exact target from 0.661 to 0.548. These results demonstrate that distribution-level supervision improves intervention ranking and final selection even under deterministic deployment.

Overall, StapleBridge derives exact, chemistry-aware supervision from constrained stochastic control and amortizes the resulting intervention preferences into a reusable policy. The additional experiments quantify the distinct effects of reference regularization and approximation, while demonstrating near-exhaustive predicted permeability gains with substantially lower inference costs.

Response to Q2: Site-Specific Stapling Preferences and Structural Compatibility

We thank the reviewer for these constructive suggestions. To further establish the biological relevance of StapleBridge, we conducted additional hydrophobicity-controlled cross-predictor evaluations, retrospective experimental validation, and restraint-free structural analyses.

1. Site-specific selection beyond staple chemistry and hydrophobicity.

To examine whether StapleBridge identifies favorable modification sites beyond chemical and hydrophobicity preferences, we compared different stapling positions within the same parent peptide, controlling for staple chemistry (S5–S5 or R8–S5) and product cLogP. Specifically, δ denotes the maximum permitted absolute difference in cLogP between matched candidate products, with δ = 0.25 representing stricter matching than δ = 0.50.

At δ = 0.50, the evaluation includes 35 held-out peptides, 130 matched intervention groups, and 446 candidate pairs. We evaluated frozen intervention selections using four CPMP permeability endpoints (PAMPA, Caco-2, RRCK, and MDCK) and a complementary PeptiVerse PAMPA predictor. None of these prediction heads guided intervention selection. We additionally evaluated PAMPA predictions under the stricter δ = 0.25 threshold, covering 28 peptides.

Table R2. Hydrophobicity-controlled stapling-site selection across permeability predictors.

EvaluatorcLogP δLeadsSB−DPR [95% CI] ↑SB−Hydrophobicity [95% CI] ↑
CPMP PAMPA0.5035+0.0564 [+0.0109, +0.1124]+0.0551 [+0.0103, +0.1283]
CPMP Caco-20.5035+0.116 [+0.035, +0.191]+0.046 [+0.003, +0.093]
CPMP RRCK0.5035+0.016 [−0.004, +0.036]+0.029 [+0.005, +0.054]
CPMP MDCK0.5035+0.080 [+0.017, +0.143]+0.045 [+0.008, +0.081]
PeptiVerse PAMPA0.5035+0.0748 [−0.0042, +0.2070]+0.0860 [+0.0078, +0.1918]
CPMP PAMPA0.2528+0.0438 [+0.0109, +0.0791]+0.0548 [+0.0073, +0.1261]
PeptiVerse PAMPA0.2528+0.1519 [+0.0344, +0.2869]+0.1588 [+0.0471, +0.2869]

Candidate comparisons are restricted to the same parent peptide and staple chemistry, with absolute product cLogP differences bounded by δ.

StapleBridge achieves positive predicted permeability advantages over DPR and hydrophobicity-only selection across all four CPMP endpoints, with seven of eight confidence intervals excluding zero. The complementary PeptiVerse PAMPA evaluation shows consistent positive trends, with both comparisons excluding zero under the stricter δ = 0.25 threshold. Overall, 12 of 14 confidence intervals exclude zero.

Importantly, the reference prior assigns equal probabilities to different anchor positions within the same peptide and staple mode. The observed advantages therefore support lead-specific site selection beyond fixed staple preferences and overall lipophilicity. Their consistency across permeability prediction tasks also indicates that the benefits are not confined to the original optimization surrogate.

2. Retrospective validation against experimental cellular-uptake rankings.

We further evaluated StapleBridge against published experimental hydrocarbon-stapling positional scans from Bird et al. (2016), which measured cellular uptake across different staple placements within the BIM and SOS1 peptide families. Of the 17 experimentally characterized BIM variants, 16 satisfy our predefined geometric feasibility constraints and are included alongside nine SOS1 variants.

Table R3. Spearman correlation with experimentally measured cellular-uptake rankings on feasible stapling interventions.

MethodBIM (n=16) ↑SOS1 (n=9) ↑
cLogP0.520.80
Direct Plan Ranker0.490.68
StapleBridge0.670.85

StapleBridge achieves Spearman correlations of 0.67 and 0.85 on BIM and SOS1, respectively, outperforming DPR on both families (+0.18/+0.17) and recovering the experimentally highest-uptake site within each evaluated feasible set. These results provide direct experimental-ranking evidence that StapleBridge captures peptide-specific stapling preferences without test-time property evaluations.

Both CPMP PAMPA and PeptiVerse PAMPA are trained on experimentally measured PAMPA permeability data, providing complementary, experimentally grounded assessments of the selected interventions. StapleBridge's consistent predicted advantages across both evaluators (Table R2) suggest that its learned intervention preferences capture permeability-relevant properties that generalize beyond the original optimization surrogate. Together with the direct experimental cellular-uptake rankings, these results provide converging evidence for the biological relevance and reliability of StapleBridge's site-specific intervention selection.

3. Restraint-free conformational and receptor-binding validation.

To further evaluate structural compatibility, we performed restraint-free GFN-FF/ALPB-water conformational ensemble analyses of linear parents and StapleBridge products under matched protocols. We additionally assessed receptor docking and native-interface preservation.

Table R4. Conformational and receptor-binding compatibility.

MetricResult [95% CI]
A. Restraint-free conformational analysis
ΔCα-RMSD vs. Linear (Å) ↓−0.945 [−1.118, −0.782]
ΔHelical fraction ↑+0.0342 [+0.0269, +0.0416]
ΔEnsemble dispersion (Å) ↓−0.605 [−0.746, −0.460]
ΔPolar SASA fraction ↓−0.0382 [−0.0460, −0.0308]
B. Receptor docking
Mean DockQ ↑0.5633 [0.4791, 0.6477]
Acceptable-or-better DockQ ↑87.2%
C. Native-interface preservation (n=31)
Contact retention: SB / DPR ↑0.795 / 0.730
iRMSD: SB / DPR (Å) ↓0.883 / 1.280

Section A reports paired changes relative to linear parents without backbone restraints. Section B evaluates frozen, QC-passing stapled products without DockQ-based filtering. Section C uses the common 31-peptide interface subset. Brackets denote bootstrap 95% confidence intervals where available.

Without backbone restraints, StapleBridge products exhibit lower backbone deviation, increased helicity, reduced conformational dispersion, and decreased exposed polar surface fraction, with all four confidence intervals excluding zero. These results support favorable conformational preorganization beyond the original restrained-relaxation analysis.

For receptor compatibility, the docked complexes achieve a mean DockQ of 0.5633, with 87.2% reaching acceptable-or-better docking quality. StapleBridge also retains more native contacts (0.795 vs. 0.730) and achieves lower iRMSD (0.883 vs. 1.280 Å) than DPR, supporting preservation of native-like receptor-binding geometry.

Overall, these additional evaluations provide complementary evidence across three dimensions: favorable stapling-site selection beyond hydrophobicity, agreement with experimentally measured cellular-uptake rankings, and improved conformational and receptor-interface compatibility. Together, they strengthen the biological relevance of StapleBridge beyond surrogate-score optimization.

We hope these additional results and clarifications address your concerns. We sincerely appreciate your constructive feedback and the opportunity to strengthen our manuscript.

Thank you for the detailed rebuttal and the additional experiments. After carefully reviewing your responses to my questions as well as the discussions with other reviewers, I appreciate the effort put into the retrospective analyses and additional evaluations. However, my core concerns regarding the methodological necessity and biological validation remain unresolved, so I maintain my recommendation for rejection.

  1. Methodological Contribution and Necessity of Amortization vs. Exhaustive Search

    • Equivalence to listwise ranking: In your response to Q1, you acknowledge that a listwise ranker trained on the same prior-adjusted soft targets would recover the same intervention distribution. Because each discrete plan p=(b,a,Δx)p = (b, a, \Delta x) deterministically maps to a canonical product zˉ(x,p)=G(x,p)\bar{z}(x, p) = G(x, p) (Eq. 1) and the optimal path law factorizes with an unchanged execution process (Eq. 11), the Schrödinger-control framing and learned execution policy Pϕ(τ∣x,p)P_\phi(\tau \mid x, p) appear unnecessary: the formulation effectively reduces to distilling a prior-weighted Boltzmann softmax over a small finite candidate set.
    • Unfavorable cost–accuracy trade-off relative to enumeration: Your new Table R1 highlights that exhaustive scoring across all 1,012 feasible candidates for the 77 test leads takes only 35.72 s on a single T4 GPU (~0.46 s per peptide). In a lead-optimization setting—where candidate evaluation is followed by 3D conformational relaxation, docking, or wet-lab synthesis—saving ~0.36 s per peptide at deployment is negligible. Meanwhile, distillation incurs substantial approximation error: as shown in Table 2, the amortized controller achieves only 51.57% Top-1 agreement with its exact target q⋆q^\star, dropping mean Δ\DeltaPermeability from 0.186 (Exhaustive PV-Max) / 0.172 (Exact Control MAP) to 0.169. Furthermore, training still requires evaluating the oracle across all feasible candidates for the 4,020 training leads. As also noted by other reviewers, this constrained finite-support setting does not convincingly justify amortized RL/control over direct enumeration.
  2. Experimental and Structural Validation (Q2)

    • Limited retrospective experimental evidence: While the Bird et al. (2016) evaluation is a welcome addition, it covers only two helical peptide families (BIM n=16n=16, SOS1 n=9n=9) measuring cellular uptake rather than direct membrane permeability. On SOS1 (n=9n=9), simple hydrophobicity (cLogP) already reaches a Spearman correlation of 0.80 compared to 0.85 for StapleBridge (well within statistical noise for n=9n=9), and on BIM, StapleBridge (0.67) underperforms the underlying PeptiVerse teacher (0.72). Meanwhile, the cLogP-controlled CPMP analysis (Table R2) still evaluates surrogate-against-surrogate agreement and lacks a comparison to exact oracle selection.
    • Missing baselines in restraint-free structural & docking checks (Table R4): In Table R4, Sections A (restraint-free conformational ensemble) and B (receptor docking) report metrics only for StapleBridge, without any baselines (Random Feasible, DPR, or Original Linear). Any covalent staple restricts conformational dispersion and tends to stabilize helicity relative to a linear peptide; without matched baselines, these results do not demonstrate that StapleBridge's site selection preserves native conformation or docking quality better than random or ranker baselines. Additionally, Section C (0.795 / 0.730 contact retention and 0.883 / 1.280 iRMSD) appears to directly reproduce the unrefined rigid-frame numbers from Table 4B in the main text rather than providing a new restraint-free or affinity-based receptor-binding comparison.

Taken together, while the problem of peptide lead editing is practically relevant, the finite intervention space makes exhaustive search trivial and accurate, and the current validation does not sufficiently establish methodological necessity or biological utility beyond the surrogate. I encourage the authors to extend this framework to a setting with a genuinely intractable combinatorial edit space where amortized control is strictly needed.

Reviewer_SUG2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces StapleBridge, a chemistry-aware discrete Schrödinger-control framework for improving the properties of existing linear peptides through hydrocarbon stapling. Given a peptide lead, the method first enumerates chemically and geometrically feasible stapling interventions, including i,i+4 and i,i+7 hydrocarbon staples. It then uses a property predictor to construct an optimal intervention distribution through KL-regularized control and distills this distribution into a learned controller, enabling intervention selection without paired linear-to-stapled training examples. The authors evaluate the method primarily through computationally predicted permeability improvements. StapleBridge improves predicted permeability for 81.13% of supported test leads, with an average gain of 0.169 compared with 0.140 for the strongest matched plan-level baseline. Additional cross-predictor and 3D structural evaluations are provided, along with computational experiments extending the framework to other peptide modifications.

Critical strengths
  • A practically relevant peptide optimization problem: The paper addresses a practical setting: optimizing existing peptide leads rather than generating entirely new sequences. This is potentially valuable because preserving the original peptide sequence and its functional properties is often desirable during lead optimization.
Critical weaknesses
  1. The necessity and added value of the proposed Schrödinger-control framework are insufficiently established: The intervention space considered in this work is relatively small. For a peptide of length L, the two supported single-stapling modes (i,i+4 and i,i+7) yield at most 2L−11 candidate interventions. For peptides of length at most 25 residues, this corresponds to no more than 39 candidates, even before applying feasibility constraints. Consequently, a straightforward approach would be to enumerate all feasible stapling positions, evaluate each candidate using the same property predictor, and select the highest-scoring intervention. It is therefore unclear what substantial advantage the learned controller provides over direct exhaustive enumeration and oracle-based ranking. Such an advantage might arise from reduced inference cost, but this needs to be demonstrated through explicit computational cost comparisons. So, this make it difficult to assess whether the proposed control framework provides a meaningful methodological improvement over a simple method of exhaustion.
  2. The main evaluation relies heavily on a property predictor whose reliability for hydrocarbon-stapled peptides is unclear. The primary optimization objective and reported improvements rely on the PeptiVerse permeability-penetrance predictor. However, this predictor was developed using cell-penetrating peptide classification data, rather than experimentally measured permeability changes across different hydrocarbon stapling positions. Its predictive performance on the original classification benchmark does not necessarily establish its ability to rank different stapled variants of the same peptide.
Questions for authors
  1. Given the small number of feasible stapling interventions per peptide, how does StapleBridge compare against simply enumerating all feasible candidates and selecting the one with the highest predicted permeability using the same oracle? Please report the achieved property scores, inference time, oracle evaluation costs, and the performance gap relative to exhaustive oracle selection.
  2. Has the PeptiVerse permeability-penetrance predictor been evaluated on experimentally characterized hydrocarbon-stapled peptides? In particular, is there evidence that it can accurately rank different stapling positions for the same parent peptide? Could the authors evaluate their approach using published experimental measurements of stapled-peptide cellular uptake or permeability?
2 comments

Dear Reviewer SUG2,

We sincerely thank you for your careful evaluation of our manuscript and your thoughtful, constructive questions. We greatly appreciate the opportunity to clarify the methodological contributions of StapleBridge and strengthen its experimental validation. In response to your concerns, we conducted additional exhaustive-search comparisons and retrospective site-specific ranking evaluations using published experimental hydrocarbon-stapling data. We are grateful for your valuable feedback and address each of your questions in detail below.


Response to Q1: Exhaustive Comparison and Amortized Intervention Control

We thank the reviewer for this important question. The finite intervention space is central to StapleBridge, enabling exact control supervision during training and efficient amortized optimization across unseen peptide leads.

1. Exhaustive comparison and inference efficiency.

We compared StapleBridge against Exhaustive PV-Max, which selects the highest PeptiVerse-predicted permeability candidate, and Exact Control MAP, which optimizes the same reference-regularized objective as StapleBridge.

Table R1. Exhaustive optimization and inference efficiency.

MetricExhaustive PV-MaxExact Control MAPStapleBridge
Mean ΔPermeability ↑0.1860.1720.169
Inference time (77 leads) ↓35.72 s—7.82 s
Property-scored interventions ↓1,0121,0120

Property improvements are evaluated on the fixed 53 hard-supported leads, while inference times cover all 77 test leads on the same NVIDIA T4 GPU with preloaded models, excluding initialization and training. Exact Control MAP uses the same exhaustive candidate pool.

StapleBridge achieves a mean ΔPermeability of 0.169, retaining 90.9% of the Exhaustive PV-Max gain (gap: 0.017) and 98.3% of the Exact Control MAP gain (gap: 0.003). Meanwhile, it achieves 4.57× faster inference (7.82 vs. 35.72 s) while avoiding 1,012 test-time intervention-level property evaluations. This advantage becomes increasingly valuable for large-scale peptide screening, where exhaustive property-scoring costs accumulate across thousands of leads.

2. Methodological contribution: exact control supervision and amortization.

Beyond computational efficiency, StapleBridge provides a principled formulation for learning chemically constrained intervention policies. Each feasible intervention uniquely determines a canonical minimal-edit product, making terminal energy invariant across its valid execution trajectories. This structure enables the exact factorization

Px⋆(p,τ)=q⋆(p∣x)Rcan(τ∣x,p),P_x^\star(p,\tau)=q^\star(p\mid x)R_{\mathrm{can}}(\tau\mid x,p),

with the optimal intervention distribution

\exp[-\beta E_T(G(x,p);x)].$$ Crucially, this converts the chemically constrained control problem into an exactly solvable supervision problem without paired linear-to-stapled examples. Distillation then transfers these intervention preferences into a reusable lead-conditioned controller, enabling efficient selection on unseen peptides without test-time terminal-property evaluations. Overall, StapleBridge exploits finite chemical intervention spaces to derive exact control supervision and amortize intervention optimization across peptide leads, combining near-exhaustive predicted permeability gains with substantially lower deployment-time oracle costs.

Response to Q2: Experimental Validation of Site-Specific Stapling Preferences

To directly assess the reliability of our optimization signal and the biological relevance of StapleBridge, we conducted additional retrospective evaluations using published experimental hydrocarbon-stapling data, complemented by hydrophobicity-controlled computational analyses.

1. Reliability of the PeptiVerse penetrance predictor.

PeptiVerse has been benchmarked on experimentally measured permeability datasets comprising 6,869 PAMPA and 606 Caco-2 measurements, providing empirical support for its permeability prediction capabilities. To directly assess its ability to rank different stapling positions, we evaluated the PeptiVerse penetrance head against the experimental hydrocarbon-stapling positional scans reported by Bird et al. (2016). These experiments systematically varied staple placements within the BIM and SOS1 peptide families and quantified cellular uptake using fluorescence-based measurements, providing experimental rankings of site-specific stapling preferences.

Our evaluation focuses on chemically compatible, geometry-supported stapling interventions under the predefined StapleBridge feasibility constraints. Of the 17 experimentally characterized BIM variants, 16 satisfy the geometric feasibility constraints and are included in the analysis. The remaining variant is excluded because it falls outside the predefined feasible intervention space. We additionally evaluate nine SOS1 stapling variants.

The PeptiVerse penetrance predictor achieves Spearman correlations of 0.72 (BIM, n=16) and 0.81 (SOS1, n=9) with experimentally measured cellular-uptake rankings. These results provide direct evidence that the optimization signal captures experimentally relevant differences among stapling positions within the same peptide family.

2. Experimental validation of StapleBridge's site selection.

Using the same experimental measurements and feasible intervention sets, we compared StapleBridge against Direct Plan Ranker (DPR), the PeptiVerse penetrance predictor, and a hydrophobicity-based cLogP baseline.

Table R2. Retrospective experimental cellular-uptake ranking on chemically compatible, geometry-supported stapling interventions. Values report Spearman correlations with experimentally measured cellular uptake.

MethodBIM (n=16) ↑SOS1 (n=9) ↑
cLogP0.520.80
Direct Plan Ranker0.490.68
PeptiVerse penetrance0.720.81
StapleBridge0.670.85

StapleBridge achieves Spearman correlations of 0.67 and 0.85 on BIM and SOS1, respectively, outperforming DPR on both families (+0.18/+0.17) and identifying the experimentally highest-uptake stapling site in each. These results demonstrate that the amortized controller captures experimentally relevant site-specific preferences without requiring test-time property evaluations.

Overall, these evaluations provide direct experimental support for the site-specific ranking capability of the PeptiVerse penetrance predictor and the ability of StapleBridge to select interventions consistent with measured cellular-uptake preferences.

We hope these additional results and clarifications address your concerns. We sincerely appreciate your constructive feedback, which has helped us strengthen the manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all four reviewers for the thoughtful discussion of StapleBridge. Across the reviews, the reviewers recognized the practical importance of optimizing an existing peptide lead, the care we took in defining chemically feasible staples, and the strength of the comparisons. The main disagreement concerned whether learning to select a staple is worthwhile when we can score every feasible candidate. Here, we agree that exhaustive scoring is a good option for an individual peptide. We also believe our results make a strong case for learning which staple to select, and we are grateful that Reviewer uvie’s recommendation remains that of acceptance.

Importance of the stapling formulation

Given a peptide lead, we want to choose where to place a hydrocarbon staple while preserving protected residues and satisfying chemical and geometric constraints. Each feasible choice specifies a unique stapled product. This lets us score every option during training, calculate the exact distribution we want the controller to learn, and train it without paired examples of linear and stapled peptides. For a new lead, we can then select a staple without running the property predictor on every candidate.

Because we can enumerate the candidates, we can also measure exactly how much we give up by using the controller. On 53 supported test leads, exhaustive selection gave a mean predicted permeability gain of 0.186. Exact selection under our reference-regularized objective gave 0.172, and the learned controller achieved 0.169. We therefore retained 98.3% of the gain under the objective we trained for and 90.9% of the gain from exhaustive permeability scoring.

New experiments for each reviewer

  • Reviewer SUG2 asked us to compare directly with enumeration. Across all 77 test leads on the same T4 GPU, exhaustive scoring took 35.72 seconds for the full set, which is about 0.46 seconds per lead. StapleBridge took 7.82 seconds for the full set and required zero test-time property evaluations, compared with 1,012 for enumeration. Both are fast at this scale. The practical distinction is that we can train the controller once and reuse it across new leads while retaining nearly all of the predicted gain under its training objective.

  • Reviewer KiLh asked whether we learned which sites work for a particular peptide. We tested this directly against the published cellular-uptake measurements of Bird et al. (2016), who varied staple placement within the same peptide. StapleBridge selected the experimentally highest-uptake site in both BIM and SOS1 among the feasible sites tested. Across all 16 BIM and nine SOS1 variants, our rankings correlated with measured uptake at 0.67 and 0.85, compared with 0.49 and 0.68 for the matched direct plan ranker. We also compared sites within the same parent peptide and staple chemistry at similar hydrophobicity. StapleBridge selected sites with higher predicted permeability across all four CPMP endpoints.

  • Reviewer upYX asked about homology and the diversity of the test set. We explained how we excluded 1,834 development complexes whose receptors shared a sequence cluster with a test receptor. We then checked diversity among the test peptides themselves. The 53 supported peptides spanned 44 clusters at 30% sequence identity, and we improved predicted permeability for 43 of 53 leads, with gains in every peptide-length group. Reviewer upYX expressed satisfaction with our response to the original concerns.

  • Reviewer uvie asked for a broader baseline and a check that a few peptides did not account for the gains. We adapted CP-Composer to the same hydrocarbon-stapling task and evaluated both methods under the same constraints across three seeds. StapleBridge improved predicted permeability for 81.1% of leads versus 73.6% for adapted CP-Composer, with a 31.7% larger mean gain. Even after removing our ten largest gains, the remaining 43 peptides still improved by 0.0953 on average.

Case for acceptance

While we understand that enumeration is practical for the current stapling space, we believe the formulation and results provide a strong case for acceptance. We derived exact supervision for chemically feasible interventions and trained a controller that improved predicted permeability for 43 of 53 held-out peptides, retaining 98.3% of the gain from exact optimization under the same objective. We further demonstrated experimentally relevant site selection by identifying the highest-uptake staple placement in both BIM and SOS1 from published measurements. Together with the broader baseline comparisons, these results support the effectiveness of our approach for peptide stapling and its value for learning modification preferences across leads. We respectfully ask the Area Chair and reviewers to support acceptance of StapleBridge.

Thank you,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.