acceptodds
Under review as a conference paper at ICLR 2027

HyBridge: Receptor-Guided Cyclic-Peptide Generation on Compiled Covalent Support

Abstract

Cyclic peptides offer attractive scaffolds for protein-targeted therapeutics by coupling rich amino-acid chemistry with a constrained conformational landscape. However, existing generative methods using geometric conditioning or cyclic encodings leave ring constraints unenforced during the sampling process. These geometric inconsistencies can compromise both the physical validity and receptor compatibility of their generated designs. Here, we propose HyBridge, a hybrid sampler that preserves compiled covalent constraints throughout joint sequence–structure generation. HyBridge leverages a receptor-free reference model to capture intrinsic peptide diversity. Following geometric placement, a shared receptor potential guides backbone motion and residue substitutions, with accepted updates preserving the compiled constraints. To learn pairing preferences, we incorporate crossed peptide–receptor panels alongside chemical supervision, canceling additive single-chain scoring offsets. Evaluated on held-out receptor targets under a head-to-tail (H2T) benchmark, HyBridge achieves 61.2% structural delivery under a common closure, contact, and severe-overlap criterion. Specifically, suffix guidance improves delivery by 15.1 percentage points over execution without the learned potential and by 5.6 percentage points over terminal-only use. Experiments demonstrate receptor-responsive cyclic-peptide generation with persistent ring constraints and improved structural delivery, complemented by construction examples spanning disulfide, lactam, and bicyclic topologies.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cyclic peptides, Molecular generative modeling, Receptor-conditioned generation, Sequence–structure co-design, Geometric constraints

Reviews

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

This paper proposes HYBRIDGE, a hybrid sampler for receptor-guided joint sequence–structure generation of cyclic peptides. A receptor-free reference model supplies peptide variation, while a shared receptor potential guides backbone motion and residue substitutions under compiled covalent constraints. The potential is trained using crossed peptide–receptor panels and chemical supervision. On 103 held-out receptor components, HYBRIDGE achieves 61.2% H2T structural delivery, gaining 15.1 percentage points over execution without the learned potential and 5.6 over terminal-only use.

Critical strengths
  1. Constraint preservation throughout generation. HYBRIDGE explicitly preserves compiled covalent constraints during both backbone updates and residue substitutions, allowing receptor guidance to act on cyclic candidates throughout sampling.
  2. Controlled evidence for guidance during sampling. With the model, initialization, placement, seeds, and output slots fixed, guided generation improves structural delivery by 5.6 percentage points over terminal-only scoring and 15.1 points over disabling the learned potential. These comparisons directly support the benefit of adapting candidates before selection.
  3. Pair-specific receptor modeling. Crossed peptide-receptor comparisons cancel additive partner-specific score offsets. Chemistry counterfactuals and receptor-swap controls provide complementary evidence that the learned potential captures pairing preferences and drives receptor-dependent changes.
  4. Detailed and transparent evaluation. The paper reports component-level held-out evaluation, confidence intervals, failure-stage accounting, and path-validity controls. Including missing and invalid outputs in the denominator makes the delivery metric informative about practical generation reliability.
Critical weaknesses
  1. Cross-method comparisons are difficult to interpret. The methods use different native sampling budgets, and much of HYBRIDGE’s delivery advantage arises before interface evaluation, from successful backbone construction and certification. This supports pipeline reliability, but provides less direct evidence of superior receptor-specific design under comparable computational resources.
  2. Limited validation of binding quality. Structural delivery measures geometric validity, contact, and severe-overlap avoidance rather than affinity or specificity. The independent AF3 analysis covers only seven reciprocal receptor pairs, and the improvement over terminal-only scoring has a confidence interval that includes zero. Short molecular dynamics trajectories provide limited additional evidence.
  3. Unclear exploration beyond source scaffolds. Generation starts from perturbed training-bank backbones and preserves source-derived geometric targets along each path. The paper does not sufficiently quantify how this dependence affects conformational diversity, scaffold novelty, or coverage of structures unavailable in the source bank.
Questions for authors
  1. How do runtime and computational cost compare across methods? Could you report results under matched budgets and show whether the advantage persists when receptor-interface quality is evaluated conditional on successful backbone construction?
  2. How similar are generated peptides to their source scaffolds and nearest training examples, in both sequence and cyclic-aligned structure? How sensitive are delivery and diversity to source-bank size and composition?
  3. Could you report sensitivity to the potential scale β, guidance activation time, and sampling budget to assess the robustness of the structural delivery gains?
6 comments
AuthorsOpenReview ↗

[1/3] Response to Critical Weakness 1: Computational Cost, Conditional Interfaces, and Guidance Attribution

Thank you for recognizing constraint-preserving generation, the controlled guidance evidence, pairing-specific receptor modeling, and the evaluation’s transparency. Below we address Critical Weakness 1 with complete pipeline accounting, comparisons after valid-backbone construction, and matched guidance schedules; new material is marked in blue.

1. Complete Pipeline Cost and Resource-Normalized Output

MetricHYBRIDGECP-ComposerRFpeptides
Complete resource time (h) ↓35.0189.2352.20
Qualified designs per resource-hour (estimated) ↑43.191.401.66
Resource-minutes per qualified design (estimated) ↓1.3942.9036.25

The timing corpus contains 2,472 attempts per method. CP-Composer follows its official protocol, including 2.97 GPU-h plus 86.26 additional CPU-h for OpenMM cyclization; resource-hours sum GPU-hours and additional cyclization CPU-hours. Qualified yield applies each method’s delivery fraction from the common registered panel to its measured attempted throughput. HYBRIDGE’s complete cost includes its source previews and terminal continuations. These estimates quantify delivered-design output per unit of measured computation (Appendix I.3, p. 42, lines 2226–2255; Table 24, p. 42, lines 2247–2255).

On GPU consumption alone, HYBRIDGE uses 35.01 versus 52.20 GPU-h for RFpeptides, with 61.2% versus 3.5% delivery. At d_min ≥ 2 Å, estimated qualifying yields are 23.72 versus 12.93 designs/GPU-h for HYBRIDGE and CP-Composer, excluding CP-Composer’s additional CPU work. Across 24 outputs, topology-valid fractions are 90.3% versus 23.0% (Table 18, p. 33, lines 1729–1742).

2. Interface Formation After Valid-Backbone Construction

MetricHYBRIDGECP-ComposerRFpeptides
Delivery among valid Canonical-L backbones ↑67.7%19.8%13.3%
Median buried surface area, delivered (Ų) ↑808.3800.3299.6
Median peptide contact coverage, delivered ↑83.3%80.0%23.2%
CHARMM36 target Rank@1, delivered designs ↑78.7%69.2%44.4%

Conditional delivery uses each method’s valid Canonical-L backbones; the remaining rows summarize delivered complexes. HYBRIDGE retains higher interface delivery after backbone failures have been removed from the comparison. The broader interface engagement and higher intended-receptor energy ranking provide additional evidence at the interface stage (Table 3, p. 8, lines 379–398; Tables 22–23, pp. 41–42, lines 2183–2224).

3. Learned Guidance Within the Same Execution Framework

HYBRIDGE guidance scheduleStructural delivery ↑CHARMM36 target Rank@1, all registered slots ↑
No learned potential46.0%55.5%
Terminal-only scoring55.5%60.0%
Guidance during generation61.2%63.7%

The reference checkpoint, initialization, placement, covalent support, seeds, and registered slots are shared. Each schedule uses the same source-preview and terminal-branching settings. Generation-time guidance yields 29 additional delivered designs, a 5.6-percentage-point gain over terminal-only scoring, with receptor-component-paired bootstrap 95% CI [+1.2, +10.1] percentage points. Intended-receptor Rank@1 over all registered slots also increases from 60.0% to 63.7% (Table 1, p. 6, lines 297–304; Table 29, p. 45, lines 2419–2429).

To address your readability feedback, Section 5 and Appendix I.8 now explain in plain language why receptor adaptation and ring preservation must be coupled, how the updates work, and which controls support them (p. 46, lines 2430–2467).

Together, the resource accounting measures computational efficiency, the conditional comparison examines performance after valid-backbone construction, and the matched HYBRIDGE schedules test receptor adaptation with the surrounding execution machinery held common.

AuthorsOpenReview ↗

[2/3] Response to Critical Weaknesses 2 and 3: Binding-Related Quality, AF3 Coverage, and Backbone Diversity

Critical Weakness 2: Binding-Related Quality, AF3 Coverage, and Geometric Persistence

1. External Predicted Affinity, Stricter Steric Screening, and Pocket Specificity

Following your request for stronger validation, we added official PRODIGY v2.4.0/FreeSASA v2.2.1 predictions on intended complexes, stricter clash screening, and 11-pocket geometric controls under common preprocessing (all 515 slots; best-target medians select the lowest-ΔG delivered design per covered target; geometric specificity requires intended-pocket compatibility with 0/10 compatible controls).

MetricHYBRIDGECP-ComposerRFpeptides
Delivery at d_min ≥ 2.00 Å ↑33.6%1.6%2.7%
Delivered + Kd ≤ 1 μM ↑24.7%2.5%0.0%
Delivered + Kd ≤ 100 nM ↑11.3%1.2%0.0%
Best-target median ΔG (kcal/mol) ↓−8.78−8.29−6.21
Best-target median Kd (nM) ↓365.7830.527,740.0
Geometric specificity (%) ↑43.3%3.7%0.6%

HYBRIDGE achieves the highest affinity-qualified yields and strongest best-target ΔG and Kd medians, complementing the buried-area, contact-coverage, and delivered CHARMM36 results under Critical Weakness 1 (Table 3, p. 8, lines 379–398; Tables 22–23, pp. 41–42, lines 2183–2224).

2. Common Independent AF3 Evaluation and Geometric Persistence

We ran common AF3 prediction across all 103 targets with designed peptide coordinates withheld. Valid cyclic designs pass the Canonical-L check. Chain-pair ipTM (AF3 peptide–receptor interface confidence) and RMSDs use the median across five AF3 predictions per design; reported percentages use all 515 registered slots. Backbone RMSD fits peptide N/Cα/C/O atoms, and pose RMSD fits receptor Cα atoms.

MetricHYBRIDGECP-ComposerRFpeptides
Canonical-L validity ↑90.3%25.4%26.2%
Valid cyclic, ipTM ≥ 0.60 ↑61.6%17.7%17.1%
Valid cyclic, ipTM ≥ 0.65 ↑47.2%13.2%12.6%
Valid cyclic, ipTM ≥ 0.70 ↑31.7%8.7%8.2%
Delivered + ipTM ≥ 0.65 ↑33.6%2.3%1.6%
Backbone RMSD ≤ 2 Å ↑28.5%7.6%6.2%
Pose RMSD ≤ 10 Å ↑27.4%9.5%0.0%

HYBRIDGE leads at every AF3 confidence cutoff, reaching 14.4× and 21.6× baseline yields on delivered + ipTM ≥ 0.65 alongside stronger backbone and pose agreement (Appendix I.4, pp. 42–43, lines 2257–2294; Table 25, p. 43, lines 2269–2282). Within HYBRIDGE, seven-pair AF3 estimates a +0.0130 guided-minus-terminal margin (95% CI [−0.0030, +0.0270], Table 21, p. 40, lines 2133–2145), paired with +5.6 points delivery gain across 515 slots (CI [+1.2, +10.1], Table 1, p. 6, lines 297–304). Topology trajectories measure covalent and geometric persistence over the simulated interval (Figure 6, p. 10, lines 486–505; Table 16, p. 31, lines 1621–1627).

Critical Weakness 3: Backbone and Sequence Diversity Beyond Source Scaffolds

Compiled support retains local bond lengths, modeled bond angles, the peptide-bond branch, and H2T closure while backbone torsions and residue identities adapt within those constraints.

MeasurementHYBRIDGE
Median cyclic sequence identity vs. bank ↔45.5%
Median cyclic Cα RMSD vs. bank ↔0.93 Å
Exact cyclic training-sequence repeats ↓0
Mean distinct valid sequences / 24 slots ↑21.7

Bank comparisons span 2,232 outputs, and diversity counts distinct ring-valid sequences across 24 slots per target (Appendix I.5, Table 26, p. 43, lines 2303–2317; Figure S15, p. 44, lines 2322–2347). Receptor swapping shifts guided backbones by median 2.08 Å vs. ~0.13 Å under controls (Figure 4f, p. 9, lines 432–459). In the 60-slot study, shrinking the source pool while keeping 64 donors/seed holds H2T validity and receptor support at 83.3% (0.99× median runtime, Table 27, p. 44, lines 2365–2375).

AuthorsOpenReview ↗

[3/3] Response to Questions 1, 2, and 3: Computational Cost, Training-Bank Comparison, and Sensitivity

Question 1: Runtime, Computational Cost, and Conditional Interface Quality

As detailed under Critical Weakness 1, complete resource totals are 35.01, 89.23, and 52.20 h for HYBRIDGE, CP-Composer, and RFpeptides. Estimated qualified yields are 43.19, 1.40, and 1.66 designs per resource-hour. All totals include required pipeline stages; CP-Composer follows its official OpenMM cyclization protocol. HYBRIDGE’s accounting includes 64 source previews shared across eight lineages per seed and eight terminal continuations per lineage. The timing corpus uses 2,472 attempts per method; resource-hours sum GPU-hours and additional cyclization CPU-hours (Appendix I.3, p. 42, lines 2226–2255; Table 24, p. 42, lines 2247–2255).

After successful Canonical-L construction, delivery is 67.7%, 19.8%, and 13.3%. Among delivered complexes, median BSA is 808.3, 800.3, and 299.6 Ų, contact coverage is 83.3%, 80.0%, and 23.2%, and CHARMM36 target Rank@1 is 78.7%, 69.2%, and 44.4%. These directly quantify the interface stage conditional on backbone construction and subsequent delivery (Table 3, p. 8, lines 379–398; Tables 22–23, pp. 41–42, lines 2183–2224). Resource-normalized output quantifies cross-method efficiency, while the paired HYBRIDGE guidance schedules test the learned contribution with common execution settings (Table 1, p. 6, lines 297–304).

Question 2: Source Similarity, Training-Bank Comparison, and Source-Bank Sensitivity

Donors match length and H2T topology, exclude the target receptor component, and include at most one template per donor component. Each seed samples 64 donors without replacement, followed by cyclic recutting, constrained coordinate perturbation, and sequence corruption. Local covalent targets are retained while torsions and residue identities adapt (Appendix I.5, pp. 43–44, lines 2295–2347).

As detailed under Critical Weakness 3, the 2,232-output training-bank comparison gives 45.5% median cyclic sequence identity, 0.93 Å median cyclic-aligned Cα RMSD, and zero exact cyclic sequence repeats. Output diversity averages 21.7 distinct ring-valid cyclic sequences per 24 registered outputs per receptor component (Appendix I.5 and Table 26, p. 43, lines 2303–2317; Figure S15, p. 44, lines 2322–2347).

The source-pool intervention retains approximately half of eligible components by component-ID hash, with a minimum of 64, and still samples 64 donors per seed. The length-6 pool changes from 94 to 64 components. Full and reduced pools both give 83.3% validity and receptor support, at relative runtimes 1.00× and 0.99×, on the paired 60-slot development panel (Table 27, p. 44, lines 2365–2375). This tests eligible-pool dependence with the sampled-donor count fixed.

Question 3: Sensitivity to Score Scale β, Guidance Activation Time, and Sampling Budget

We added a one-factor-at-a-time study on the same four development targets of lengths 6, 10, 13, and 15, three paired seeds, and first five registered slots per target–seed combination: 60 slots per setting. The baseline uses β = 7.1727, onset 0.50, eight particles, and 50 reference steps.

ConfigurationH2T validity ↑Receptor support ↑Relative median runtime ↓
Baseline83.3%83.3%1.00×
Lower β (3.5863)83.3%83.3%1.03×
Later onset (0.76)83.3%83.3%0.92×
Reduced source bank83.3%83.3%0.99×
More particles (16)83.3%83.3%2.08×
More reference steps (100)91.7%91.7%1.66×

HYBRIDGE retains both pass rates under the tested scale, onset, and source-pool changes. Increasing reference steps from 50 to 100 raises validity and receptor support by 8.3 percentage points at 1.66× median runtime. These results quantify robustness and the response to additional computation (Appendix I.6, p. 44, lines 2349–2375; Table 27, p. 44, lines 2365–2375).

Thank you for your questions; they helped us make computational accounting, conditional interface quality, source dependence, and sensitivity explicit.

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

The paper introduces HyBridge, a receptor-guided framework for joint cyclic-peptide sequence and structure generation. A receptor-free reference model proposes backbone and residue updates, while a learned receptor potential guides both within covalent constraints. The potential is trained using crossed peptide-receptor comparisons and chemical counterfactuals. On 103 held-out receptor components, the method achieves a structural delivery rate of 61.2%, improving by 5.6 percentage points over terminal-only selection.

Overall, the paper addresses a relevant problem: receptor-guided updates can disrupt cyclic geometry, and subsequent closure can alter the intended interface. Maintaining constraints throughout sampling is useful, and the matched guidance-schedule comparison provides evidence for adapting candidates before terminal selection.

However, the performance comparison does not clearly separate the benefits of learned generation from internal screening, geometry correction, and sampling budget. Beyond structural delivery, the evidence for improved binder design quality is also limited.

Critical strengths
  1. The method explicitly preserves the modeled covalent constraints during coordinate and sequence updates. The hard-constraint versus soft-penalty path controls distinguish constraint satisfaction throughout generation from endpoint closure, providing a useful diagnostic for cyclic-peptide generation.

  2. Guidance during generation improves structural delivery by 5.6 percentage points over terminal-only selection, indicating the value of receptor guidance during generation. Crossed-pair and pocket-chemistry interventions further show that the learned potential responds to pairing.

Critical weaknesses
  1. The substantial improvement in structural delivery is difficult to attribute to the learned model itself. HyBridge screens 64 source candidates per seed, searches for supported placements, corrects or rejects coordinate updates, and selects one output from eight terminal continuations per lineage (Algorithm S1; Table 11). Although non-emitted designs remain in the denominator, the reported rates do not account for internal search costs. Computational budgets are not matched across baselines.

  2. The evaluation establishes successful structural delivery more clearly than improved binding quality. The secondary receptor-ranking screen holds peptide poses fixed and compares them against rigidly transported pockets; the resulting rankings may be affected by severe steric overlap introduced by this transport (Appendix D.1). AF3 provides an independent check, but only compares internal guidance schedules on seven reciprocal receptor pairs, where guidance during generation shows no clear advantage over terminal-only selection. There is also no corresponding independent comparison against CP-Composer and RFpeptides. The 1 ns topology simulations support short-time geometric persistence, but add little evidence about sustained binding. These limitations leave the practical benefit for binder design uncertain.

Questions for authors
  1. What are the runtime and internal candidate costs per returned design and per successfully delivered design for each method? Does the delivery advantage persist under a matched computational budget?

  2. Can the authors provide a common, independent sequence-to-complex evaluation for all three methods, reporting interface confidence and agreement between the repredicted structures and the designed backbones and poses?

  3. What is the principal algorithmic advance over existing constraint-conditioned cyclic-peptide generators? Please clarify this through a more explicit comparison with prior work.

5 comments
AuthorsOpenReview ↗

[1/3] Response to Critical Weaknesses 1 and 2: Guidance Attribution, Pipeline Accounting, and Interface Quality

Thank you for recognizing constraint preservation, the path controls, and the evidence for generation-time guidance and pairing sensitivity. Below we report the added pipeline accounting, interface measurements, affinity predictions, and common AF3 evaluations (blue in the revised PDF).

Critical Weakness 1: Attribution and Internal Candidate Accounting

To isolate learned guidance within identical execution machinery, the matched schedules hold the reference checkpoint, source and placement procedure, covalent support, seeds, and registered slots fixed, varying only when the same learned potential acts.

Guidance scheduleDelivery ↑CHARMM36 Rank@1, all slots ↑
No learned potential46.0%55.5%
Terminal-only scoring55.5%60.0%
Guidance during generation61.2%63.7%

Within HYBRIDGE, generation-time guidance adds 29 delivered designs (+5.6 percentage points, 10.1% more designs over terminal-only selection; target-paired bootstrap 95% CI [+1.2, +10.1] points). With construction and placement shared, this gain isolates candidate adaptation to receptor preferences before terminal selection (Table 1, p. 6, lines 297–304; Table 29, p. 45, lines 2419–2429).

Pipeline accounting explicitly counts 64 source previews shared across eight lineages per seed (admitted in frozen order with target-score ranking disabled) and eight terminal continuations per lineage sharing the earlier trajectory, all included in the end-to-end costs in Appendix I.3 (p. 42, lines 2226–2255) and Table 24 (p. 42, lines 2247–2255).

Critical Weakness 2: Interface Quality and Transported-Pocket Controls

1. Stricter Steric Screening, Conditional Delivery, and Predicted Affinity

To test whether higher delivery translates into binding-related quality, we added stricter clash screening, interface measurements, and official PRODIGY v2.4.0/FreeSASA v2.2.1 predictions on intended complexes under common preprocessing.

MetricHYBRIDGECP-ComposerRFpeptides
Delivery at d_min ≥ 2.00 Å ↑33.6%1.6%2.7%
Delivery (valid Canonical-L) ↑67.7%19.8%13.3%
Median buried area, delivered (Ų) ↑808.3800.3299.6
Median contact coverage, delivered ↑83.3%80.0%23.2%
Delivered + Kd ≤ 1 μM ↑24.7%2.5%0.0%
Delivered + Kd ≤ 100 nM ↑11.3%1.2%0.0%
Best-target median ΔG (kcal/mol) ↓−8.78−8.29−6.21
Best-target median Kd (nM) ↓365.7830.527,740.0

Conditional delivery conditions on valid Canonical-L backbones; area and coverage summarize delivered complexes, and best-target medians select the lowest-ΔG delivered design per covered target. HYBRIDGE achieves the highest affinity-qualified yields and strongest best-target ΔG and Kd medians (Table 3, p. 8, lines 379–398; Table 22, p. 41, lines 2183–2199).

2. Eleven-Pocket Geometric and Energetic Specificity

MetricHYBRIDGECP-ComposerRFpeptides
Target geometry pass rate, evaluable ↑67.7%19.8%13.3%
Decoy geometry pass rate, evaluable ↓3.23%2.82%7.85%
Target-to-decoy pass-rate ratio ↑21.0×7.0×1.7×
CHARMM36 Rank@1, delivered ↑78.7%69.2%44.4%

Geometry pass rates measure spatial compatibility and CHARMM36 measures fixed-coordinate energy ranking, providing complementary evidence of intended-receptor preference (Appendix I.2, pp. 41–42, lines 2200–2224; Table 23, p. 42, lines 2215–2224).

Common 103-target AF3 gives 47.2%, 13.2%, and 12.6% valid cyclic designs at median ipTM ≥ 0.65 (Table 25, p. 43, lines 2269–2282). Within HYBRIDGE, seven-pair AF3 estimates a +0.0130 guided-minus-terminal margin (95% CI [−0.0030, +0.0270], Table 21, p. 40, lines 2133–2145), alongside +5.6 points across 515 slots (Table 1, p. 6, lines 297–304). Question 2 details AF3 protocol and RMSDs.

AuthorsOpenReview ↗

[2/3] Response to Questions 1 and 2: Pipeline Computational Cost and Independent AF3 Evaluation

Question 1: Runtime, Internal Candidate Costs, and Sampling Budgets

The timing corpus contains 2,472 attempts per method, including HYBRIDGE’s source previews and terminal continuations.

MetricHYBRIDGECP-ComposerRFpeptides
Complete resource time (h) ↓35.0189.2352.20
Attempted designs per resource-hour ↑70.6127.7047.36
Qualified designs per resource-hour (estimated) ↑43.191.401.66
Resource-minutes per qualifying design (estimated) ↓1.3942.9036.25

CP-Composer follows its official protocol: 2.97 GPU-h for generation plus 86.26 CPU-h for OpenMM cyclization. Resource-hours are their arithmetic accounting sum. Qualified yield applies the common panel’s delivery fractions to measured attempted throughput (Appendix I.3, p. 42, lines 2226–2255; Table 24, p. 42, lines 2247–2255).

On GPU consumption alone, HYBRIDGE uses 35.01 versus 52.20 GPU-h for RFpeptides, with 61.2% versus 3.5% delivery. At d_min ≥ 2 Å, estimated yields are 23.72 versus 12.93 qualifying designs/GPU-h for HYBRIDGE and CP-Composer, excluding CP-Composer’s CPU work. Across 24 outputs, topology-valid fractions are 90.3% versus 23.0% (Table 18, p. 33, lines 1729–1742).

Paired HYBRIDGE sampling results use the same development targets, seeds, and 60 slots:

Sampling configurationH2T validity ↑Receptor support ↑Relative median runtime ↓
8 particles, 50 reference steps83.3%83.3%1.00×
16 particles, 50 reference steps83.3%83.3%2.08×
8 particles, 100 reference steps91.7%91.7%1.66×

Increasing reference steps raises validity and receptor support by 8.3 points (Table 27, p. 44, lines 2365–2375). The matched guidance comparison gives 61.2% versus 55.5% delivery with the remaining execution settings fixed (Table 1, p. 6, lines 297–304).

Question 2: Common Independent Sequence-to-Complex AF3 Evaluation

We added AF3 across the full 103-component benchmark, withholding designed peptide coordinates. Official AF3 v3.0.1 uses identical receptor inputs, query-only MSAs, receptor-backbone-only templates, H2T bonds, seed, and recycle settings across methods. ipTM is AF3’s confidence score for the predicted peptide–receptor interface; higher scores mean greater confidence. Intended-receptor chain-pair ipTM and RMSDs use the median across five AF3 predictions per design.

MetricHYBRIDGECP-ComposerRFpeptides
Canonical-L validity ↑90.3%25.4%26.2%
Valid cyclic designs, ipTM ≥ 0.60 ↑61.6%17.7%17.1%
Valid cyclic designs, ipTM ≥ 0.65 ↑47.2%13.2%12.6%
Valid cyclic designs, ipTM ≥ 0.70 ↑31.7%8.7%8.2%
Delivered + ipTM ≥ 0.65 ↑33.6%2.3%1.6%
Delivered + ipTM ≥ 0.65, within Canonical-L ↑37.2%9.2%5.9%
Median pose RMSD, delivered (Å) ↓16.5324.6327.53
Slots with backbone RMSD ≤ 2 Å ↑28.5%7.6%6.2%
Slots with pose RMSD ≤ 10 Å ↑27.4%9.5%0.0%

Slot rates use all 515 registered slots; the within-Canonical-L row uses valid backbones. Backbone RMSD aligns peptide N/Cα/C/O atoms; pose RMSD aligns receptor Cα atoms. RMSD medians use available reference coordinates; the delivered-pose row uses originally delivered designs.

HYBRIDGE yields more originally delivered designs meeting AF3 confidence and more designs meeting either agreement threshold. At ipTM ≥ 0.65, original-delivery confidence yields are 14.4× and 21.6× the baselines. Appendix I.4 and Table 25 report the common protocol, interface confidence, backbone agreement, and receptor-aligned pose agreement (pp. 42–43, lines 2257–2294; Table 25, p. 43, lines 2269–2282).

AuthorsOpenReview ↗

[3/3] Response to Question 3: Principal Algorithmic Advance and Comparison with Prior Methods

Your question asks what HYBRIDGE adds to receptor-conditioned cyclic-peptide generation. Its principal advance is a sampler that adapts sequence and backbone together to a receptor while preserving the modeled covalent constraints at every accepted step. The covalent graph defines which changes are allowed; one learned receptor potential guides their effect on the evolving interface. We have clarified this mechanism in Section 5 (p. 10, lines 526–527) and expanded its comparison with prior methods and supporting controls in Appendix I.8 (p. 46, lines 2430–2463).

1. Why Geometry and Receptor Adaptation Must Be Coupled

In a closed peptide chain, backbone motion is constrained by the ring: moving residues toward a receptor can disturb closure geometry, while subsequent closure repair can move those residues away from the contacts being optimized. Preserving the cyclic molecule and improving its interface are therefore coupled requirements. Receptor-directed changes need to remain compatible with the covalent geometry throughout generation.

2. How HYBRIDGE Implements This Coupling

HYBRIDGE compiles the declared covalent graph into allowed molecular states and updates. It projects receptor-driven backbone motion onto constraint-compatible directions and retracts proposals to the constrained geometry. Sequence changes preserve the graph and its attachment requirements. Every accepted state satisfies the modeled covalent constraints, while backbone torsions and residue identities remain free to adapt.

One learned peptide–receptor potential coordinates the two updates. Its coordinate gradients guide backbone motion, while its score differences between allowed mutations adjust residue-substitution rates. Sequence preferences are evaluated on the updated backbone, so the amino-acid choices respond to the interface being formed. A receptor-free reference model supplies peptide variation without requiring paired receptor data; crossed-pair and chemistry supervision train the shared receptor-dependent signal.

The key difference is how receptor adaptation is kept compatible with the ring. CP-Composer encodes residue-type and distance requirements as conditions for joint sequence–structure diffusion. These conditions guide samples toward closure, but do not themselves require every intermediate update to preserve the modeled covalent geometry. RFpeptides generates backbones with cyclic positional encoding and then uses ProteinMPNN for sequence design, separating backbone generation from sequence design. HYBRIDGE addresses the coupling described above directly: every accepted update preserves the modeled covalent constraints, while one receptor-dependent potential guides backbone motion and residue changes together. Sequence choices can therefore respond to the evolving interface as the backbone adapts, keeping ring preservation and interface optimization coupled throughout generation.

3. What the Controlled Results Establish

The path control tests geometry preservation directly: all 480 hard-support paths remain valid, while every soft-penalty path leaves the constraint support (Figure 4a, p. 9, lines 432–459). The matched guidance comparison then tests adaptation within the same execution framework. Holding the reference, initialization, placement, seeds, and registered slots fixed, guidance during generation yields 29 additional delivered designs, a 5.6-percentage-point gain over terminal-only scoring (Table 1, p. 6, lines 297–304). Together, these controls support both persistent ring validity and the benefit of receptor-directed adaptation before selection.

The shared signal is trained through crossed peptide–receptor comparisons that cancel additive peptide-only and receptor-only score offsets, making partner compatibility the learning target. Held-out pairing results are:

Held-out pairing measurementResult
Quartet accuracy ↑92.8%
Factual-pair preference ↑85.9%

During generation, this signal produces a frozen-readout crossed-response gain of +0.069 over terminal-only scoring and +0.093 over the geometry/support control across eleven reciprocal receptor pairs (Figure S16, p. 45, lines 2384–2403). These results connect the learned pairing preferences to receptor-dependent changes in generated candidates.

Thank you for your question; it helped us explain why covalent validity and receptor adaptation need to be coupled, how HYBRIDGE implements that coupling, and which controls support it.

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

The paper introduces HYBRIDGE, a generative framework for receptor-conditioned cyclic peptide design that preserves covalent constraints throughout sequence and backbone generation. The approach combines a receptor-free generative model with a learned receptor potential trained using crossed peptide–receptor comparisons and chemical supervision. This potential jointly guides backbone dynamics and residue substitutions, while projection and retraction operations maintain the prescribed cyclic topology. On 103 held-out receptor targets, HYBRIDGE achieves 61.2% structural delivery, substantially outperforming CP-Composer (5.0%) and RFpeptides (3.5%) under their native pipelines. Ablation studies demonstrate improvements from applying receptor guidance during generation rather than only at the final selection stage, while additional experiments assess constraint preservation and applicability to different cyclic topologies.

Critical strengths

The paper addresses an important limitation of generative approaches for cyclic peptides: preserving covalent constraints throughout sampling rather than imposing them through post-processing. The combination of constrained backbone dynamics, discrete sequence updates, and a shared receptor-dependent potential provides a coherent framework for joint sequence–structure generation. The ability to specify different covalent topologies through a common representation is particularly appealing.

The paper provides relatively extensive ablation studies and controls to investigate individual methodological components. In particular, the comparison between full guidance, terminal-only scoring, and unguided generation provides evidence that receptor information is beneficial when incorporated directly into the sampling process. The experiments examining constraint preservation along trajectories and receptor-dependent conformational changes further support the proposed framework.

Critical weaknesses

The principal performance metric, structural delivery, only requires successful cyclic backbone construction, receptor contact, and the absence of severe steric overlaps (interchain heavy-atom distances below 1.5 Å). Although useful for assessing geometric feasibility, these criteria provide limited evidence that the generated peptides form physically plausible, stable, or biologically meaningful binding interfaces. In particular, receptor contact alone does not demonstrate binding specificity or affinity, and the relatively permissive clash criterion could classify unrealistic complexes as successful designs. While the paper includes receptor-response controls and mentions AF3 re-prediction analyses, a more systematic evaluation using independent structure prediction, interface quality, energetic metrics, or experimentally characterised binders would considerably strengthen its conclusions.

The reported improvement over CP-Composer and RFpeptides is striking, but much of this difference arises from backbone construction and materialisation failures in the baselines rather than demonstrated improvements in receptor-specific binding. Moreover, the methods are evaluated using their native sampling budgets, making it difficult to establish how much of the performance gap reflects the proposed generative methodology rather than differences in constraint enforcement and implementation. The authors demonstrate a more modest 5.6 percentage-point improvement over terminal-only scoring within HYBRIDGE, while the specific contribution of crossed-pair supervision to downstream design performance remains insufficiently isolated. Finally, despite the emphasis on general covalent topologies, quantitative benchmarking focuses almost exclusively on head-to-tail cyclic peptides, with other topologies supported primarily by a few illustrative examples and short molecular dynamics trajectories.

Questions for authors

None

10 comments
AuthorsOpenReview ↗

[1/3] Response to Critical Weakness 1: Interface Plausibility, Predicted Affinity, and Independent AF3 Evaluation

Thank you for recognizing constraint-preserving generation, controlled guidance, and pairing-specific receptor modeling. Below we report the new steric, interface, predicted-affinity, receptor-control, and common AF3 evaluations (marked in blue in the revised PDF).

1. Stricter Clash Screening

We raised the interchain minimum distance from 1.50 to 2.00 Å while holding preprocessing, contact rules, outputs, and denominators fixed.

↑ higher is better; ↓ lower is better; ↔ descriptive.

MetricHYBRIDGECP-ComposerRFpeptides
Delivery at d_min ≥ 1.50 Å ↑61.2%5.0%3.5%
Delivery at d_min ≥ 2.00 Å ↑33.6%1.6%2.7%
Covered targets at 2.00 Å ↑84/1038/1038/103

HYBRIDGE retains higher delivery and target coverage at 2.00 Å, addressing sensitivity to the clash cutoff (Table 3, p. 8, lines 379–398; Appendix I.1, p. 41, lines 2162–2199).

2. Interface Extent

To assess multi-residue engagement beyond single-atom contact, we measured buried surface area and peptide-residue contact coverage within 4 Å among delivered complexes.

MetricHYBRIDGECP-ComposerRFpeptides
Median buried area (Ų) ↑808.3800.3299.6
Median contact coverage ↑83.3%80.0%23.2%
Delivered / covered targets ↑315 / 9626 / 2118 / 11

HYBRIDGE engages most peptide residues across 96/103 targets, matching CP-Composer in buried area and exceeding RFpeptides across broader receptor coverage (Table 3, Panel D, p. 8, lines 379–398; Table 22, p. 41, lines 2183–2199).

3. External Predicted Affinity and Receptor Preference

We added official PRODIGY v2.4.0/FreeSASA v2.2.1 predictions on intended complexes and a fixed-pose 11-pocket screen under common preprocessing (all 515 registered slots unless noted; best-target medians select the lowest-ΔG delivered design per covered target; geometric specificity requires intended-pocket compatibility with 0/10 compatible controls).

MetricHYBRIDGECP-ComposerRFpeptides
Delivered + Kd ≤ 1 μM ↑24.7%2.5%0.0%
Delivered + Kd ≤ 100 nM ↑11.3%1.2%0.0%
Best-target median ΔG (kcal/mol) ↓−8.78−8.29−6.21
Best-target median Kd (nM) ↓365.7830.527,740.0
CHARMM36 Rank@1, all slots ↑63.7%8.2%5.0%
CHARMM36 Rank@1, delivered ↑78.7%69.2%44.4%
Geometric specificity (%) ↑43.3%3.7%0.6%

HYBRIDGE yields substantially more micromolar and nanomolar candidates, stronger best-target ΔG and Kd medians, and higher intended-receptor preference (Table 3, p. 8, lines 379–398; Table 22, Panel B, p. 41, lines 2183–2199; Table 23, p. 42, lines 2215–2224).

4. Common Independent AF3 Evaluation

Following your suggestion, we ran common AF3 prediction across the 103-component benchmark with designed peptide coordinates withheld. Qualifying designs pass the shared cyclic-backbone check. Chain-pair ipTM (AF3 peptide–receptor interface confidence) and RMSDs use the median across five AF3 predictions per design; reported percentages use all 515 registered slots. Backbone RMSD fits peptide backbones, and pose RMSD fits receptor Cα atoms.

MetricHYBRIDGECP-ComposerRFpeptides
Canonical-L validity ↑90.3%25.4%26.2%
Valid cyclic, ipTM ≥ 0.60 ↑61.6%17.7%17.1%
Valid cyclic, ipTM ≥ 0.65 ↑47.2%13.2%12.6%
Valid cyclic, ipTM ≥ 0.70 ↑31.7%8.7%8.2%
Delivered + ipTM ≥ 0.65 ↑33.6%2.3%1.6%
Backbone RMSD ≤ 2 Å ↑28.5%7.6%6.2%
Pose RMSD ≤ 10 Å ↑27.4%9.5%0.0%

HYBRIDGE leads at every AF3 confidence cutoff, reaching 14.4× and 21.6× baseline yields on delivered + ipTM ≥ 0.65 alongside stronger backbone and pose agreement (Appendix I.4, pp. 42–43, lines 2257–2294; Table 25, p. 43, lines 2269–2282).

AuthorsOpenReview ↗

[2/3] Response to Critical Weakness 2: Conditional Interface Quality, Computational Cost, and Learned Guidance

1. Interface Formation After Valid-Backbone Construction

To examine the interface stage once backbone construction has succeeded, we compare delivery conditional on valid Canonical-L backbones and then measure the delivered interfaces.

MetricHYBRIDGECP-ComposerRFpeptides
Delivery among valid Canonical-L backbones ↑67.7%19.8%13.3%
Median buried surface area, delivered (Ų) ↑808.3800.3299.6
Median peptide contact coverage, delivered ↑83.3%80.0%23.2%

The higher conditional delivery shows that HYBRIDGE also improves the conversion of valid cyclic backbones into receptor-contacting, sterically acceptable complexes. Area and coverage quantify the resulting engagement. Thus, the comparison addresses interface formation within the successfully constructed subset (Appendix I.1, p. 41, lines 2162–2199; Table 22, p. 41, lines 2183–2199; Table 3, p. 8, lines 379–398).

2. Complete Computational Accounting

Your budget question calls for counting the internal work needed to produce an output. We therefore added full native-pipeline resource accounting, including source screening, terminal continuations, and prescribed cyclization. Each HYBRIDGE seed uses 64 source previews shared across eight lineages, with eight terminal continuations per lineage sharing the preceding trajectory. These operations are included in the totals below.

MetricHYBRIDGECP-ComposerRFpeptides
Complete resource time (h) ↓35.0189.2352.20
Attempted designs per resource-hour ↑70.6127.7047.36
Qualified designs per resource-hour (estimated) ↑43.191.401.66
Resource-minutes per qualifying design (estimated) ↓1.3942.9036.25

CP-Composer follows its official protocol, including 2.97 GPU-h for generation and 86.26 additional CPU-h for OpenMM cyclization. Resource-hours sum GPU-hours and additional cyclization CPU-hours. Qualified yield is estimated by applying the delivery fractions from the common registered panel to the attempted throughput in the timing corpus. This accounting relates complete pipeline cost to qualifying output, and the matched guidance comparison evaluates the learned potential within a common execution framework (Appendix I.3, p. 42, lines 2226–2255; Table 24, p. 42, lines 2247–2255).

On GPU consumption alone, HYBRIDGE uses 35.01 versus 52.20 GPU-h for RFpeptides, with 61.2% versus 3.5% delivery. At d_min ≥ 2 Å, estimated qualifying yields are 23.72 versus 12.93 designs/GPU-h for HYBRIDGE and CP-Composer, excluding CP-Composer’s additional CPU work. Across 24 outputs, topology-valid fractions are 90.3% versus 23.0% (Table 18, p. 33, lines 1729–1742).

3. Benefit of Guidance During Generation

To test the contribution of learned guidance within the same pipeline, the matched schedules hold the reference checkpoint, source and placement procedure, covalent support, seeds, and registered slots fixed. They change when the same potential acts: during generation or at terminal selection.

HYBRIDGE guidance scheduleStructural delivery ↑CHARMM36 target Rank@1, all registered slots ↑
No learned potential46.0%55.5%
Terminal-only scoring55.5%60.0%
Guidance during generation61.2%63.7%

Within HYBRIDGE, generation-time guidance provides 29 additional delivered designs, a 5.6-percentage-point increase and 10.1% more delivered designs over terminal-only scoring. The receptor-component-paired bootstrap 95% interval is +1.2 to +10.1 percentage points. HYBRIDGE’s CHARMM36 target Rank@1 over all registered slots also rises from 60.0% with terminal-only scoring to 63.7% with generation-time guidance, connecting the delivery gain to intended-receptor preference. Because the surrounding execution framework is shared, these results support receptor-guided adaptation while sequence and conformation are developing (Table 1, p. 6, lines 297–304; Table 29, p. 45, lines 2419–2429).

AuthorsOpenReview ↗

[3/3] Response to Critical Weakness 2: Crossed-Pair Supervision and Covalent-Topology Support

1. Role of Crossed-Pair Supervision

Your question concerns how crossed-pair supervision contributes to receptor-specific generation. Crossed comparisons exchange the same peptide and receptor partners, canceling additive partner-specific score offsets and directing learning toward their joint compatibility. We examine this pairing signal through held-out recognition and receptor-dependent generation controls.

EvaluationResult
Held-out quartet accuracy ↑92.8%
Held-out factual-pair preference ↑85.9%
Frozen-readout crossed-response gain over terminal-only ↑+0.069
Frozen-readout crossed-response gain over geometry/support control ↑+0.093
Median receptor-swap cyclic Cα RMSD, guided ↔2.08 Å
Median receptor-swap cyclic Cα RMSD, either control ↔0.13 Å

The held-out results show that the potential recognizes compatible partner assignments. Across eleven reciprocal receptor pairs, guidance using this signal improves the frozen-readout crossed response over both terminal-only scoring and the geometry/support control (Figure S16, p. 45, lines 2384–2403). Under matched receptor exchange, guided backbones change by a median 2.08 Å, compared with 0.13 Å under either control, showing that generation responds to receptor identity (Figure 4f, p. 9, lines 432–459).

Appendix B.5 shows how crossed contrasts cancel partner-specific offsets; Proposition 3 explains how peptide-only score terms affect sampling (p. 17, lines 870–891). Complementary retraining gives pocket-chemistry preference of 60.0% for retrieval only, 61.1% without pocket-counterfactual loss, and 67.2% for the full objective (Tables 4 and 28; Table 28, p. 45, lines 2409–2415). Together, recognition and generation controls show that the learned pairing signal guides receptor-dependent adaptation.

2. General Covalent Support and Benchmarking Scope

Your comment asks how the breadth of the framework is supported by its evaluation. We agree that the rationale should be clear. Our experimental design follows two requirements: compare methods on a common chemical task, and test the shared constraint machinery across distinct covalent graphs.

  • Shared topology for interpretable comparison: Covalent topology determines which bonds must remain closed, which residue attachments are permitted, and which geometric constraints each design must satisfy. These requirements define the molecular generation task. We therefore chose H2T as the common setting for all three pipelines on the same 103-component receptor panel, with identical backbone-validity, contact, and steric evaluation criteria. Holding the requested chemistry consistent gives the delivery and interface comparisons a common basis.
  • Multiple topologies for testing general covalent support: The framework compiles a declared graph, geometric constraints, and residue attachment requirements into a shared constraint-preserving execution procedure. Disulfide, lactam, and bicyclic programs exercise different bond patterns, attachment requirements, and cycle counts. Their execution directly tests the graph compilation and constraint enforcement that motivate general covalent support. The H2T benchmark and these topology tests thus address complementary requirements of the proposed framework.

Figure S6 shows products from six compiled programs. Its target-free terminal-lactam path passes the topology certificate at all eight steps, with final amide closure 1.330 Å (p. 30, lines 1602–1606). The disulfide and H2T-plus-disulfide trajectories additionally assess closure, chirality, and receptor-contact persistence over 1 ns (Figure 6, p. 10, lines 486–505; Tables 16–17, p. 31, lines 1621–1654).

We have clarified this rationale in Section 4.4 (p. 9, lines 476–479) and Appendix I.8 (p. 46, lines 2430–2467): the common H2T task supports quantitative cross-method performance conclusions, and the additional covalent graphs provide concrete evidence for the breadth of the shared execution mechanism. Section 5 and Appendix I.8 also explain ring preservation and joint receptor adaptation in plain language, connecting the intuition to the formal update rules.

Thank you for your questions on pairing supervision and topology support; they helped us clarify how the mechanism, controlled results, and topology demonstrations fit together.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Manuscript Revisions and Additional Evaluations

Dear Area Chair and Reviewers,

Thank you for the time, care, and thoughtful feedback in your reviews. We truly appreciate that across your reviews, you recognized keeping the peptide ring closed throughout generation rather than fixing it afterward, learning true partner-specific pocket compatibility rather than single-partner shortcuts, and backing each claim with transparent, controlled ablations. Those points go straight to a real physical catch in cyclic-peptide design. In a closed chain, pulling residues toward a pocket can disturb ring closure, while repairing closure afterward pulls those residues back away from the contacts being formed, and choosing sequence only after fixing the backbone keeps amino-acid choices and ring shape from adapting together. HYBRIDGE resolves this directly during generation. Every step projects and retracts backbone motion onto compiled covalent constraints so torsions adapt while the ring stays closed, while one pairing-trained potential guides coordinate updates and residue changes together on the updated backbone so sequence and ring shape co-adapt to the interface (Section 5, Appendix I.8).

Your reviews also raised fair, important questions on predicted binding affinity, independent folding, cyclization and compute budgets, and backbone diversity. Taking every point to heart, we added six pages in Appendix I (pp. 41–46, Tables 22–29, Figures S15–S16) plus blue updates in Table 3 and Sections 4.4 and 5.

1. Independent AlphaFold 3 validation across all 103 targets. Following your suggestion, we ran AF3 v3.0.1 across the full benchmark (515 slots per method, five predictions per available design) with designed peptide coordinates withheld. HYBRIDGE reaches 47.2% valid cyclic designs at median ipTM ≥ 0.65 vs. 13.2% (CP-Composer) and 12.6% (RFpeptides), and 33.6% vs. 2.3% and 1.6% with original delivery, matching designs at 28.5% vs. 7.6% and 6.2% within 2 Å backbone RMSD and 27.4% vs. 9.5% and 0.0% within 10 Å pose RMSD (Appendix I.4, Table 25).

2. Predicted binding affinity, stricter clash checks, and valid-ring interfaces. Re-evaluating all 1,545 slots under identical preprocessing to test affinity and rule out cyclization or clash artifacts, PRODIGY and FreeSASA give 24.7% HYBRIDGE delivered designs at predicted Kd ≤ 1 μM vs. 2.5% and 0.0%, and 11.3% vs. 1.2% and 0.0% at nanomolar Kd ≤ 100 nM, with best-per-target median Kd 365.7 nM (ΔG −8.78 kcal/mol) vs. 830.5 nM (−8.29) and 27,740.0 nM (−6.21). Within valid Canonical-L backbones alone, HYBRIDGE still delivers 67.7% vs. 19.8% and 13.3% (808.3 Ų median buried area across 83.3% of residues). At a 2.00 Å clash cutoff, delivery stays 33.6% across 84 of 103 targets vs. 1.6% and 2.7%, with higher 11-pocket specificity (43.3% vs. 3.7% and 0.6%) and delivered CHARMM36 Rank@1 (78.7% vs. 69.2% and 44.4%, Table 3, Appendix I.1–I.2, Tables 22–23).

3. End-to-end compute accounting and guidance ablations. Timing 2,472 attempts per method (7,416 total), including 64 previews per seed, 8 continuations per lineage, and CP-Composer’s OpenMM cyclization, gives 35.01, 89.23, and 52.20 resource-h (43.19 vs. 1.40 and 1.66 qualifying designs per resource-hour). On GPU time alone, HYBRIDGE uses less compute than RFpeptides (35.01 vs. 52.20 GPU-h) and yields more 2.00 Å designs per GPU-hour than CP-Composer (23.72 vs. 12.93, without its 86.26 CPU-h). With HYBRIDGE fixed, guiding during generation adds 29 delivered designs (+5.6 points, 95% CI [+1.2, +10.1]) over terminal-only scoring and lifts all-slot CHARMM36 Rank@1 (60.0% → 63.7%, Table 1, Appendix I.3, Tables 24 and 29).

4. Backbone diversity and sensitivity checks. Comparing 2,232 outputs to the training bank gives 45.5% median cyclic sequence identity (zero exact repeats), 0.93 Å median cyclic Cα RMSD, and 21.7 mean distinct ring-valid sequences per 24 slots per target. Across six settings (60 paired slots each varying β, onset, source bank, particles, and steps), validity and support hold at 83.3% and reach 91.7% at 100 steps (Appendix I.5–I.6, Tables 26–27, Figure S15).

5. Pairing supervision, clearer writing, and topology scope. We expanded pairing and chemistry checks (92.8% quartet accuracy, +0.069 crossed-response gain, 2.08 Å vs. 0.13 Å receptor-swap adaptation), rewrote Section 5 and Appendix I.8 in plain English, reported all CIs openly, and clarified how H2T benchmarking complements disulfide, lactam, and bicyclic tests (Figure 4f, Sections 4.4–5, Appendix I.7–I.8, Figure S16, Tables 28–29).

Detailed replies appear under each of your reviews, and we welcome any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.