acceptodds
Under review as a conference paper at ICLR 2027

SolvMorph: Generating Solvent-Dependent Cyclic Peptide Ensembles from Limited Source Conformations

11 views

Abstract

Cyclic peptides can adopt distinct conformational ensembles across solvents, but repeatedly sampling each molecule under every environment is computationally expensive. Inferring a target-solvent ensemble from conformations observed in another solvent requires recovering structures that may be absent from the source ensemble. Fixed-bank reweighting can improve agreement with target observables, but cannot recover missing conformational support. We introduce **SolvMorph**, a source- and target-conditioned Schrödinger bridge that generates target-specific conformations in periodic torsion space, followed by observable-guided probability refinement. On CycPeptMPDB-4D, source conformations improve unseen-solvent prediction, while target-aware generation expands conformational support beyond fixed candidate banks. On CONF192, a generator-naive confirmation panel, SolvMorph improves target-emergent mode recall by 0.26 and 0.35 over source copying for water-to-hexane and hexane-to-water inference, respectively. These gains persist after matching proposal counts and under swapped-target controls, demonstrating recovery of solvent-specific conformational modes from limited source observations. Code is available at https://anonymous.4open.science/r/SolvMorph.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cyclic peptide, Protein ensemble, Protein design

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JzNCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper predicts a molecule's ensemble in an unobserved solvent from ~20 conformations observed in another solvent. The method generates candidate conformations with a conditioned Schrodinger bridge over torsion space, trained simulation-free and sampled as a probability-flow ODE from a generic RDKit conformer prior, then reweights the candidates to match predicted target observables. The central claim is support expansion: unlike reweighting, generation can reach conformational modes absent from the source ensemble.

Critical strengths

Torsion space optimization

Restricting the generation of cyclic peptides to torsion space and training the ODE in torsion space is an interesting adaptation of the simulation free Schrodinger bridge objective of Tong et al.

Appropriate evaluation before and after refinement

Coverage, precision, validity, specificity, and mode recall are all measured on the raw proposal bank, before any observable-guided reweighting. This helps understand the contributions of different components of the pipeline.

Critical weaknesses

The numbers in abstract do not isolate the paper’s claim

The abstract reports gains of 0.26 and 0.35 as the emergent-mode recall. However, the swap-controlled gains in Figure 5b/5d are +0.058 and +0.089. Given that the paper is framed entirely around solvent-dependent ensembles, the abstract should lead with the swap-controlled numbers.

Observable-guided refinement appears to degrade the ensembles

In Figure 4a, solvmorph has higher u_t than target-aware generation. Likewise, in Table 5, the paper states that it gets 1.017 with guided but then 1.021 with oracle + true measured observables.

Some overstated results

The improvements shown in Figure 3 are modest for panels C and D, however the authors claim significant improvements over chemistry only models. Also, the paper never validates on the property that motivates it. Membrane permeability is discussed in the introduction in detail. Yet, the only evidence offered is geometric proximity to 50 held-out MD frames.

Questions for authors
  1. Can the authors clarify abstract numbers?
  2. Why is target-observing oracle the worst in Table 5?
  3. In Figure 4a, why is SolvMorph at higher u than “Target-aware generation”?
  4. Do refined ensembles improve permeability prediction over source copying on the CycPeptMPDB labels?
3 comments

We thank the reviewer for the careful analysis of the reported quantities and evaluation protocol. To address these questions, we separated the recall contrasts, tested fixed-bank reweighting directly, expanded the true-moment controls, and evaluated measured permeability. These analyses clarify what each experiment measures and strengthen the evidence for solvent-conditioned support recovery and observable prediction.

As a note to the reviewer, we denote W→H for water-to-hexane and H→W for hexane-to-water. The brackets denote 95% structural-group bootstrap intervals unless stated otherwise.

W1 + Q1. Clarifying the abstract's recall gains

The +0.26/+0.35 abstract gains differ from the +0.058/+0.089 swap-controlled gains.

Can the abstract numbers be clarified?

We report three separate comparisons for the original submitted checkpoint on CONF192. The first measures recovery above source copying with the full proposal bank, the second fixes the budget at 20 attempted proposals, and the third isolates correct-target conditioning at that same budget.

Table A1. Recall gains by comparator and proposal budget.

ContrastW→HH→W
Full proposal bank minus source copy+0.26 [0.18, 0.33]+0.35 [0.24, 0.46]
20 attempted proposals minus source copy+0.09 [0.04, 0.15]+0.18 [0.11, 0.25]
Correct minus wrong target, matched 20 attempts+0.060 [0.013, 0.108]+0.088 [0.042, 0.132]

These comparisons answer distinct questions about total support recovery, proposal efficiency, and requested-solvent specificity. We will identify the comparator, proposal budget, and checkpoint explicitly in the abstract and report the retrained models separately.

W2 + Q2. Observable calibration and the true-moment control

Why does refinement appear to worsen uτ or MMD?

Why is the target-observing oracle worst in Table 5?

To isolate the effect of weighting, we compared predicted-moment and uniform weights on the same generated bank. We also evaluated true-moment controls on an expanded bank. The true-moment control uses four measured target averages to constrain observable means, so we evaluate both observable error and structural-distribution diagnostics.

Table R1. Observable calibration on CONF192. Negative changes indicate smaller errors or distances. The table identifies unavailable summaries explicitly.

Weighting / readoutW→H changeH→W change
Predicted moments: standardized observable error−0.178 [−0.228, −0.128]−0.084 [−0.127, −0.045]
True moments: distance-matrix MMD²−0.077 [−0.107, −0.037]Not summarized here
True moments: mode TV−0.022 [−0.031, −0.013]Not summarized here

Predicted-moment weighting reduces standardized error across the four target observables in both directions. With measured target moments on the expanded bank, we also reduce W→H distance-matrix MMD² by 0.077 and mode TV by 0.022, with the intervals shown above. These results directly support optional calibration of target observables and quantify the associated structural improvements on that bank.

Q3. Figure 4 and reweighting invariance

Why is uτ higher for SolvMorph than for “Target-aware generation”?

We traced the Figure 4 comparison to an earlier decoded bridge and a later generator checkpoint. To separate checkpoint changes from reweighting, we applied 22 weight settings to fixed banks and compared the checkpoints on 93 molecules with nonempty banks.

Table D2. Fixed-bank invariance and paired checkpoint differences.

ComparisonW→HH→W
Maximum change across 22 weightings of each fixed bank0.00.0
Final minus earlier checkpoint, matched 93 molecules+0.0095 [−0.0245, 0.0435]+0.0006 [−0.0185, 0.0202]

Every fixed bank retained exactly the same uncovered fraction across all 22 weightings. We will label both checkpoints in Figure 4 and include the paired differences, making the source of the two plotted values explicit.

W3. Observable improvements and property motivation

Some Figure 3 gains are modest, and the property motivating the work was not validated.

We quantified the paired endpoint-MAE reductions on 3,456 molecules across 162 structural groups, identifying the comparator for each observable.

Table O1. Endpoint-MAE reductions. Brackets denote 95% structural-group bootstrap intervals.

ObservableVersus persistenceVersus chemistry-only
Radius of gyration29% [22, 33]29% [4, 47]
Anisotropy18% [13, 24]15% [9, 18]
Contacts30% [23, 35]20% [8, 32]
H-bond proxy41% [34, 45]4% [1, 6]

SolvMorph reduces endpoint MAE for all four observables against both baselines. Relative to persistence, the reductions range from 18% to 41%. Against chemistry-only prediction, the reductions are 29% for radius of gyration, 15% for anisotropy, 20% for contacts, and 4% for the H-bond proxy.

Q4. Permeability compared with source ensembles

Do calibrated ensembles improve permeability prediction over source copying?

To evaluate drug-relevant utility, we tested measured PAMPA permeability on 1,016 covered molecules from 98 structural groups in the Townsend and Furukawa assays. We identify the earlier v4 and exploratory v5 model-development stages in Table P1. The v5 evaluation uses previously inspected property labels. MAE and RMSE use log10 cm/s.

Table P1. Measured permeability prediction.

ModelMacro Spearman ↑MAE ↓RMSE ↓
2D only (earlier v4)0.767 [0.660, 0.830]0.2930.388
2D + SolvMorph (v5 exploratory)0.811 [0.725, 0.860]0.2740.359
2D + source ensemble (v5 exploratory)0.808 [0.721, 0.857]0.2770.362

The optimized 2D+SolvMorph model reaches 0.811 [0.725, 0.860] macro Spearman, compared with 0.808 [0.721, 0.857] for the source-ensemble predictor. The matched optimized v5 chemistry-only model reaches 0.806 [0.716, 0.857], giving a paired SolvMorph increment of +0.005 [0.000, 0.010]. Generic conformers provide a comparable increment. Together, these matched controls quantify how conformer features contribute alongside optimized chemistry-only prediction.

The earlier chemistry-only model reached 0.767 [0.660, 0.830], and the earlier combined model reached 0.742. We distinguish these development stages from the matched v5 comparisons and retain the complete chemistry and conformer controls in the experimental record. The structural experiments directly quantify requested-solvent specificity, while the PAMPA analysis evaluates exploratory property prediction.

Together, the clarified recall contrasts, fixed-bank tests, observable-calibration results, and measured-permeability comparisons address the reviewer's questions with explicit comparators and uncertainty estimates. We appreciate the constructive feedback and apologize for the length of this response. Given our results, we would be grateful if the reviewer would consider supporting acceptance of our paper, and provide a higher score and renewed confidence in the submission.

Dear Reviewer JzNC,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

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

This paper proposes SolvMorph, a generative framework for predicting solvent-dependent conformational ensembles of cyclic peptides from a limited number of observed conformations in another solvent. The method combines a source- and target-conditioned Schrödinger bridge operating in periodic torsion space with observable-guided reweighting to generate target-solvent conformations.

The authors evaluate their approach on CycPeptMPDB-4D, focusing on conformational transitions between water and hexane. The results demonstrate improved recovery of target-specific conformational modes compared with source copying and fixed-support baselines.

Although the problem of predicting solvent-dependent conformational changes is interesting and potentially relevant to cyclic peptide drug discovery, I have several concerns regarding the physical interpretation, generalizability, and practical significance of the proposed method.

Critical strengths
  • Predicting solvent-dependent conformational ensembles from limited source observations is a potentially useful problem, particularly given the computational cost of molecular dynamics simulations.

  • The framework can generate conformations that are absent from the observed source ensemble, overcoming a fundamental limitation of conformer reweighting.

  • The authors provide several ablation studies and controlled comparisons, including solvent-condition swapping and matched proposal budgets.

Critical weaknesses
  1. Lack of explicit peptide–solvent interaction modeling

Although the paper proposes solvent-dependent conformational generation, solvents are represented only as categorical conditioning variables, without explicitly modeling their molecular properties or interactions with peptides. It is therefore unclear whether the model captures physically meaningful solvent effects or simply learns statistical correlations from the training data.

I would like to see evaluations across peptides with different polarities, hydrophobicities, and hydrogen-bonding characteristics to demonstrate that the model captures meaningful chemical dependencies.

  1. Limited generalization to unseen solvents

The experiments consider only water and hexane. Since solvents are represented by discrete labels rather than physicochemical descriptors, the model essentially learns a transformation between two predefined conditions, rather than a generalizable representation of solvent-dependent behavior.

The swapped-target experiments demonstrate sensitivity to solvent identity but do not establish generalization. Evaluation on additional solvents, particularly unseen solvents, would be important to support the broader claims.

  1. Unclear relevance to membrane permeability

The paper motivates its approach through cyclic peptide chameleonicity and membrane permeability, but all evaluations are based on MD-derived conformational ensembles.

It remains unclear whether improved conformational-mode recovery translates into better prediction of experimentally measured membrane permeability. Furthermore, the authors acknowledge that accurate conformational population recovery remains unresolved, which may limit the method's utility for this application.

  1. Limited methodological novelty and baseline comparisons

The proposed framework primarily combines established techniques, including Schrödinger bridges, flow matching, and observable-guided reweighting.

Although conditioning on source conformations is an interesting extension, related approaches such as EVA-Flow already explore environment-aware conformational generation. Without direct comparisons against such methods, the methodological advantage of SolvMorph remains unclear.

Questions for authors
  • How can the model distinguish physically meaningful peptide–solvent interactions from statistical correlations in the training data? Does its performance vary systematically with peptide polarity, hydrophobicity, or hydrogen-bonding capacity?

  • Can the proposed framework generalize to solvent environments other than water and hexane? How would it handle a previously unseen solvent, given the current categorical solvent conditioning? (or even different pH or temperature given the same solvent)

  • Do the generated conformational ensembles improve prediction of experimentally measured membrane permeability or other drug-relevant properties?

4 comments

We thank the reviewer for the thoughtful questions about chemical interpretation and practical utility! To address them, we added solvent-switching analyses, ring-size diagnostics, source-free and external baseline comparisons, and an evaluation against measured permeability. These experiments connect the requested solvent to specific structural changes and test how the generated ensembles support property prediction.

As a note, W→H denotes water-to-hexane and H→W denotes hexane-to-water. Brackets denote 95% structural-group bootstrap intervals unless stated otherwise.

W1 + Q1. Chemical interpretation of solvent conditioning

Does categorical solvent conditioning capture meaningful chemical dependence rather than only correlations?

Why are peptide–solvent interactions not explicitly modeled? How is physical solvent dependence distinguished from correlations, and how does performance vary with peptide chemistry?

Our objective is to recover solvent-dependent peptide conformations without explicitly representing solvent molecules. Explicit modeling requires sampling the high-dimensional coordinates of the peptide and surrounding solvent. Statistical mechanics provides a basis for integrating out solvent degrees of freedom.

Let xx and yy denote peptide and solvent coordinates, respectively, and ss the solvent condition. The peptide's equilibrium marginal distribution satisfies

\propto \int e^{-\beta U_s(x,y)}\,dy =e^{-\beta F_s(x)},$$ where $\beta=(k_BT)^{-1}$ and $$F_s(x)=-k_BT\log\int e^{-\beta U_s(x,y)}\,dy.$$ Here, $F_s(x)$ is the **solvent-dependent potential of mean force**, which incorporates solvent effects into the effective peptide conformational landscape. Thus, explicitly generating solvent coordinates is not required to represent the peptide's marginal distribution. SolvMorph learns solvent-conditioned conformational relationships from MD trajectories. We evaluate target-conformation support recovery without claiming to recover the exact PMF, equilibrium populations, or microscopic interaction energies. To test whether solvent conditioning corresponds to meaningful structural changes, we compared correct-target and wrong-target generation using Case A (molecule 575, W→H). We measured backbone donor exposure and intramolecular hydrogen bonding against 50 held-out MD frames per solvent. **Table C1. Solvent-dependent structural descriptors and target-emergent mode hits.** | Case / readout | Source ref. | Wrong target | Correct target | Target ref. | | :---- | :---- | :---- | :---- | :---- | | A (575 W→H): exposed N–H fraction | 0.120 | 0.073 | **0.000** | 0.000 | | A: IMHB / conformer | 0.00 | 0.20 | **1.73** | 1.92 | | A: target-emergent mode hits | 0/50 | 1/179 | **22/177** | 10/50 | According to Table C1, correct-target conditioning reproduced backbone donor shielding and increased intramolecular hydrogen bonding toward the target reference. The target-emergent mode was recovered in **22/177** correct-target conformers versus **1/179** wrong-target conformers. Across CONF192, correct-target conditioning further improved N–H response-sign agreement by **+0.12 [0.05, 0.19] W→H** and **+0.11 [0.04, 0.20] H→W**. These results support solvent-dependent structural correspondence, without establishing explicit microscopic solvent interactions. We also stratified reconstruction validity by macrocycle ring size using identical chemistry and geometry criteria. **Table C2. W→H raw valid yield by ring size.** | Ring size (atoms) | Molecules | W→H valid yield ↑ | | :---- | :---- | :---- | | 18 | 126 | 0.722 | | 19–21 | 34 | 0.499 | | ≥22 | 32 | 0.298 | According to Table C2, valid yield decreased with increasing macrocycle size. This descriptive trend may also reflect differences in scaffold composition. Together, the statistical-mechanical marginalization motivates our solvent-conditioned representation, while the controlled experiments support its ability to recover solvent-dependent peptide conformations. **W2 + Q2. Additional solvents, pH, and temperature** > Swapped labels test the two observed solvents, not previously unseen solvents. > > How would the model handle an unseen solvent or changed pH/temperature? We trained and evaluated the categorical encoder for **water and hexane**, and the switching experiments test specificity between these two environments. To extend SolvMorph to additional solvents, pH, or temperature, we would introduce condition descriptors and matched structural references for those settings, then evaluate generalization across held-out conditions. This provides a concrete extension of the conditional-bridge formulation. The present results establish solvent-specific support recovery within the two evaluated environments.

W4. Methodological contribution and external comparisons

The framework combines established components, and the advantage over environment-aware methods needs comparison.

We use a conditional bridge to recover target-solvent conformational support. To test this contribution, we added matched generative controls, a water/hexane EVA-Flow reproduction, and source-free inference.

Table B1. CONF192, 192 peptides, 35 groups, 200 attempts per molecule/solvent. Recall uses defined values; raw yield includes all attempts.

MethodRecall W→H ↑Recall H→W ↑Raw valid yield W→H / H→W ↑
Source copy0.0790.057N/A
Generic ETKDG prior0.1590.075Prefiltered bank
Mixed source + prior0.2020.121Prefiltered bank
OT-CFM0.2730.4060.49 / 0.54
Target-only; source geometry0.3160.4260.62 / 0.65
SolvMorph0.3490.4460.62 / 0.66
SolvMorph, seed 20.3430.4380.64 / 0.69

SolvMorph improves recall and valid yield over OT-CFM. We then trained our EVA-Flow adaptation for 16,000 steps at approximately matched training GPU-time, using the same folds, references, attempts, and quality checks.

Table B2. Training-GPU-time comparison. Values are W→H / H→W.

Model / training settingEmergent recall ↑Uncovered fraction ↓Raw valid yield ↑Precision ↑
EVA-Flow reproduction/adaptation, 16k0.087 / 0.1020.831 / 0.8390.074 / 0.0760.112 / 0.119
SolvMorph, retrained0.349 / 0.4460.568 / 0.5050.625 / 0.6630.225 / 0.259

SolvMorph achieves higher recall, precision, and valid yield and lower uncovered fraction than our reproduction. The matched budget covers training GPU-time. Inference cost and FLOPs follow each implementation.

We further generated from only the molecular graph, target solvent, and graph-only ETKDG conformers. This source-free evaluation uses 200 proposals per request and includes all attempts and empty banks.

Table S1. Source-free generation on CONF192. Values are W→H / H→W.

Inference settingEmergent recall ↑Uncovered fraction ↓Raw valid yield ↑
Generic ETKDG prior0.16 / 0.080.80 / 0.87Prefiltered bank
SolvMorph source-free, seed 10.28 / 0.400.64 / 0.550.42 / 0.47
SolvMorph source-free, seed 20.26 / 0.390.65 / 0.540.45 / 0.50
Target-only + source geometry0.32 / 0.410.61 / 0.520.62 / 0.65
Full SolvMorph0.35 / 0.440.57 / 0.520.62 / 0.66

Both seeds reduce uncovered fraction relative to prefiltered ETKDG. The H→W recall gain reproduces across seeds, with greater W→H variability. Source-derived reconstruction geometry further increases valid yield. Together, these tests quantify solvent-conditioned support recovery and reconstruction quality.

W3 + Q3. Measured permeability and drug-relevant utility

Does improved conformational recovery improve measured membrane-permeability prediction?

Do generated ensembles improve experimental permeability or other drug-property prediction?

Our primary objective is to recover solvent-dependent conformational ensembles of unbound cyclic peptides. Unlike a protein-bound peptide with additional structural constraints, an isolated cyclic peptide may occupy multiple metastable conformations separated by substantial free-energy barriers. A stable RMSD during a finite MD trajectory may therefore reflect kinetic trapping within one basin. In particular, when the simulation duration is much shorter than the conformational escape time, TMD≪τescapeT_{\mathrm{MD}}\ll\tau_{\mathrm{escape}}, limited structural fluctuations do not establish convergence of the equilibrium conformational distribution.

We therefore evaluated target-conformational support recovery against held-out solvent-specific MD references, recognizing that finite trajectories do not necessarily represent converged equilibrium populations. This distinction is especially relevant to cyclic-peptide chameleonicity, where changes in backbone exposure and intramolecular hydrogen bonding across solvent environments can influence membrane permeability.

To examine downstream utility, we additionally evaluated experimentally measured PAMPA permeability on 1,016 molecules from 98 structural groups across the Townsend and Furukawa assays.

Table P1. Measured permeability prediction.

ModelMacro Spearman ↑MAE ↓RMSE ↓
2D only (earlier v4)0.767 [0.660, 0.830]0.2930.388
2D + SolvMorph (v5 exploratory)0.811 [0.725, 0.860]0.2740.359
2D + source ensemble (v5 exploratory)0.808 [0.721, 0.857]0.2770.362

The optimized v5 fusion achieved 0.811 [0.725, 0.860] macro Spearman, compared with 0.808 [0.721, 0.857] for source-ensemble features. The matched optimized 2D-only model achieved 0.806 [0.716, 0.857], giving a paired SolvMorph increment of +0.005 [0.000, 0.010], comparable to generic conformer features. The earlier 2D-only result of 0.767 belongs to a different model-development stage, and the v5 analysis remains exploratory because the property labels had previously been inspected.

These experiments establish the relevance of evaluating solvent-dependent conformations and provide an initial assessment of their downstream utility. Conformational support recovery and accurate permeability prediction remain distinct objectives, since permeability also depends on conformational populations and membrane-transfer energetics and kinetics. We have clarified this distinction and the scope of our experimental evidence.

We thank the reviewer for motivating this additional evaluation and hope these analyses address the concern.

Dear Reviewer udyc,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

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

SolvMorph tackles cross-solvent inference for cyclic-peptide conformational ensembles: given an MD ensemble of a molecule in one solvent, predict its ensemble in another solvent without new simulation. Three stages:

  1. a gradient-boosted regressor predicts target-solvent structural averages (Rg, shape anisotropy κ², heavy-atom contact density, an H-bond proxy) from source-ensemble statistics and molecular descriptors;
  2. a source/target-conditioned Schrödinger bridge transports a broad ETKDGv3/MMFF94 conformer prior to target-solvent torsions, copying non-torsional internal coordinates from an observed source conformer;
  3. observable-guided reweighting of the candidate bank via regularized moment matching toward predicted target averages, subject to an effective-sample-size floor.

Evaluated on the stereochemistry-corrected CycPeptMPDB-4D dataset: observable prediction via grouped 3-fold CV on 3,456 molecules/162 groups; generation on a 96-molecule PILOT96 panel and a frozen 192-molecule "generator-naive" CONF192 panel. Key findings:

  1. source geometry cuts MAE on all four target observables vs. a source-persistence baseline, ~29% for Rg and ~30% for contacts;
  2. fixed-support reweighting alone can't expand coverage;
  3. target-aware generation lowers the uncovered-target fraction u_τ beyond generic-prior and mixed banks on both panels;
  4. before refinement, SolvMorph beats source-copying on "target-emergent mode" recall by 0.36/0.36 (PILOT96, W→H/H→W) and 0.26/0.35 (CONF192), with gains persisting under a matched 20-proposal budget but shrinking (not vanishing) under wrong-target swapped controls;
  5. ablations show source-carried geometry drives chemical validity, and Schrödinger-bridge training outperforms OT-CFM under comparable conditions.
Critical strengths
  1. Disciplined ablation of confounds usually left unexamined. The paper separates generation from refinement, matches proposal budgets, and runs wrong-target swapped controls plus a pre-registered, frozen generator-naive panel (CONF192) before looking at results. This is exactly the control set needed to rule out the usual alternative explanations for cross-condition generative gains (more samples, generic diversity, overfitting to dev molecules) — and the paper reports the resulting, more modest numbers honestly rather than only the favorable comparisons.

  2. Source-geometry contribution to prediction is clean and significant. Figure 3 and Appendix Table 1 show source ensemble geometry (not just molecular identity) carries transferable solvent-response signal: ~29% (Rg) and ~30% (contacts) MAE reduction over source-persistence, while chemistry-only features show no improvement (0.160 vs. 0.159 baseline). This solidly establishes the premise the rest of the paper builds on.

Critical weaknesses

No comparison to any external cyclic-peptide or environment-conditioned generative method. Section 2 cites RINGER, PepFlow, CPgen, Bayaraa et al.'s diffusion model, Schaufelberger et al.'s flow matching, and — most pointedly — Yan et al.'s EVA-Flow and Katzberger & Riniker's implicit-solvent GNNs as directly relevant prior work. Yet every quantitative comparison in the paper is internal (generic prior, mixed support, source copy/persistence, chemistry-only, wrong-target control, oracle) — none against these cited methods. Without that, there's no way to tell whether SolvMorph is competitive with the state of the art or just beats deliberately weak baselines, which limits what can be concluded about novelty and practical value.

Questions for authors
  • Is CONF192's 35 groups disjoint from the 162 groups used to fit the deployed cross-solvent predictor (trained on "the complete development cohort," which "has no held-out score," Appendix A.3)? Overlap would optimistically bias the predicted targets (m̂_t) used in refinement on CONF192, though not the pre-refinement mode-recall results.
  • Why does the observable-prediction benchmark use only 3,456 of the 5,057 QC'd molecules (Section 4.1, Appendix A.1/A.3)? What removed the other ~1,600, and could that selection correlate with prediction difficulty?
  • Figure 4a shows "Target-aware generation" and "SolvMorph (final)" as distinct rows with slightly different u_τ, even though Section 3.3 argues u_τ is invariant to reweighting (Eq. 12). What produces the difference — resampling before computing u_τ, or something else?
5 comments

We thank the reviewer for the careful assessment and the suggestion to broaden the comparisons. To address these questions, we evaluated additional generative and relaxation baselines, tested source-free inference, and checked the evaluation cohorts and checkpoints. Our new experiments directly assess target-solvent support recovery and practical generation quality.

As a note, throughout this response, W→H denotes water-to-hexane and H→W denotes hexane-to-water. Brackets denote 95% structural-group bootstrap intervals unless we state otherwise.

W1. External comparisons and matched generative controls

No comparison to any external cyclic-peptide or environment-conditioned generative method.

We first compared SolvMorph with matched generative controls on CONF192, comprising 192 peptides in 35 structural groups. Each retrained model attempted 200 proposals per molecule and requested solvent. We calculate recall over molecules with a defined recall value and raw valid yield over all 200 attempts.

Table B1. Target-emergent mode recall and valid yield.

MethodRecall W→H ↑Recall H→W ↑Raw valid yield W→H / H→W ↑
Source copy0.0790.057N/A
Generic ETKDG prior0.1590.075Prefiltered bank
Mixed source + prior0.2020.121Prefiltered bank
OT-CFM0.2730.4060.49 / 0.54
Target-only; source geometry0.3160.4260.62 / 0.65
SolvMorph0.3490.4460.62 / 0.66
SolvMorph, seed 20.3430.4380.64 / 0.69

SolvMorph recovered more target-emergent modes than OT-CFM and generated a higher fraction of valid structures under matched inputs. The paired uncovered-target-fraction differences were −0.061 [−0.097, −0.025] W→H and −0.022 [−0.044, 0.001] H→W. The second training seed also achieved recall of 0.343 / 0.438, supporting the reproducibility of the main results.

To extend the comparison to environment-conditioned generation, we reproduced EVA-Flow from the paper and added a water/hexane interface. We trained this EVA-ADAPT model for 16,000 steps at approximately the same training GPU-time as SolvMorph, using the same group folds, structural references, 200 attempts, and quality checks.

Table B2. Comparison at matched training GPU-time. Values are W→H / H→W.

Model / training settingEmergent recall ↑Uncovered fraction ↓Raw valid yield ↑Precision ↑
EVA-Flow reproduction/adaptation, 16k0.087 / 0.1020.831 / 0.8390.074 / 0.0760.112 / 0.119
SolvMorph, retrained0.349 / 0.4460.568 / 0.5050.625 / 0.6630.225 / 0.259

SolvMorph achieves higher emergent recall, precision, and valid yield, together with lower uncovered target fraction, than our EVA-Flow adaptation. This comparison evaluates our paper-based reproduction at matched training GPU-time. Inference cost and FLOPs follow the respective implementations. We retain the original-domain reproduction diagnostics and additional training budgets in the experimental record.

We further compared SolvMorph with implicit-solvent GNN relaxation. To control the initial conformer bank, both methods started from the same 200 conformers on matched 48-molecule subsets of PILOT96 and CONF192.

Table B3. SolvMorph minus GNN relaxation in uncovered target fraction. Negative values favor SolvMorph.

PanelW→H difference ↓H→W difference ↓
PILOT96−0.404 [−0.515, −0.300]−0.436 [−0.548, −0.317]
CONF192−0.207 [−0.351, −0.071]−0.297 [−0.458, −0.122]

SolvMorph reduces uncovered target fraction in both directions on both panels. Together, these experiments assess generative modeling, environment-conditioned generation, and conformer relaxation under the corresponding input requirements. We separately checked sequence-based methods for input and chemical compatibility, and the next comment presents the released RINGER comparison and reconstruction analysis.

W1 continued. RINGER on its supported chemical domain

Section 2 cites RINGER, PepFlow, CPgen, Bayaraa et al., Schaufelberger et al., EVA-Flow and implicit-solvent GNNs.

We evaluated the released RINGER implementation on 442 compatible cyclic peptides from 14 structural groups, including 373 molecules in one group. For each molecule and solvent direction, we generated 200 proposals and applied common quality checks. The SolvMorph rows below use the originally submitted decoder.

Table S-R1. Target-support recovery and raw valid-conformer yield on the RINGER-compatible subset.

MethodUncovered W→H ↓Uncovered H→W ↓Raw valid yield ↑
Source copy0.7390.737Not applicable
Generic ETKDG prior0.5730.670Prefiltered bank
RINGER (official pipeline)0.8260.8850.578
SolvMorph (submitted decoder)0.9170.9540.068

On this subset, RINGER, source copying, and ETKDG achieved lower uncovered target fractions than the submitted SolvMorph decoder. Relative to RINGER, the paired SolvMorph differences were +0.091 [0.021, 0.108] W→H and +0.069 [0.012, 0.083] H→W. The emergent-mode recall differences were −0.094 [−0.127, 0.020] and −0.062 [−0.135, −0.012], respectively. SolvMorph generated 0.068 valid conformers per proposal versus 0.578 for RINGER, with 24% of SolvMorph banks empty.

Because one structural group contains most molecules, we also weighted the 14 groups equally. The resulting uncovered-fraction differences were +0.044 [−0.052, 0.138] W→H and +0.062 [0.008, 0.131] H→W. Reporting both estimates makes the group composition and sampling uncertainty explicit.

Reconstruction analysis and decoder repair

To identify the source of the validity difference, we examined each reconstruction gate. Submitted SolvMorph conformers passed the bond-length, chirality, topology, and clash checks at rates of 0.325, 0.501, 0.551, and 0.166, respectively. The corresponding RINGER rates were 1.000, 0.987, 0.786, and 0.625. We traced many rejected SolvMorph conformers to proline ring-closure errors or torsions incompatible with macrocycle closure.

After completing the frozen comparison, we modified the proline-closure decoder and increased valid yield from 0.068 to 0.179. With this repair, the uncovered-fraction differences from RINGER were +0.051 [0.009, 0.101] W→H and +0.035 [−0.026, 0.054] H→W. We report this post hoc modification separately, and all submitted-model rows retain the original decoder. The experimental record contains the full group-stratified results and reconstruction settings.

This comparison identifies a concrete reconstruction problem and shows how the targeted decoder change improves generation quality. It complements the matched training and relaxation experiments in Tables B1 to B3 by testing the released cyclic-peptide pipeline on its supported chemistry.

RINGER reference: Grambow et al. Accurate and Efficient Structural Ensemble Generation of Macrocyclic Peptides using Internal Coordinate Diffusion. https://arxiv.org/abs/2305.19800

W2. Practical generation from the molecular graph

The original quantitative comparisons are internal, leaving competitiveness and practical value unclear.

To test practical use when source ensembles are unavailable, we supplied only the molecular graph, requested target solvent, and graph-only ETKDG conformers. We evaluated 200 proposals per request on CONF192's 192 molecules and 35 groups, including all attempts and empty banks. The generic ETKDG prior is a prefiltered bank.

Table S1. Source-free generation. Values are W→H / H→W.

Inference settingEmergent recall ↑Uncovered fraction ↓Raw valid yield ↑
Generic ETKDG prior0.16 / 0.080.80 / 0.87Prefiltered bank
SolvMorph source-free, seed 10.28 / 0.400.64 / 0.550.42 / 0.47
SolvMorph source-free, seed 20.26 / 0.390.65 / 0.540.45 / 0.50
Target-only + source geometry0.32 / 0.410.61 / 0.520.62 / 0.65
Full SolvMorph0.35 / 0.440.57 / 0.520.62 / 0.66

Across two training seeds, source-free SolvMorph reduces uncovered target fraction relative to ETKDG in both directions. We reproduced the H→W recall improvement across both seeds, with greater variability in W→H recall. These results demonstrate a learned target-solvent conformational prior from graph-based inputs. Adding source-derived reconstruction geometry further improves valid yield when available.

For evaluation, we calibrated the chemical-quality thresholds using water-MD geometry. Source-free generation received only the graph-based inputs described above. We identify ETKDG as a prefiltered bank because its reported bank quality uses a different denominator from raw generation yield.

Q1. Structural-group separation

Are CONF192’s 35 groups disjoint from the groups used to fit the deployed cross-solvent predictor?

We used group-out-of-fold models for the reported predictions and generator scores. For each evaluated group, we excluded that group from the scoring predictor and final generator checkpoint. Earlier development experiments included CONF192 molecules. We will document that development history and the final group exclusions explicitly in the revised evaluation description.

Q2. Cohort selection

Why were 3,456 of the 5,057 QC-passing molecules used for observable prediction, and could selection correlate with difficulty?

We checked each filtering and allocation step:

Table D1. CycPeptMPDB-4D cohort flow.

Stage / allocationCount removedRemaining
Deposited entriesN/A5,160
Graph mapping / atom-count QC1035,057
Assigned to development panels844,973
Allocated reserved cohort5814,392
Share development structural groups9363,456

Of 5,160 entries, 5,057 passed graph/coordinate quality control. We allocated 1,601 QC-passing molecules according to panel membership and structural-group overlap, leaving 3,456 molecules in 162 groups for observable prediction. Cohort comparisons showed similar distributions of the assessed molecular descriptors and normalized prediction difficulty.

Q3. Figure 4 and reweighting

Why do “Target-aware generation” and “SolvMorph (final)” have different uτ if reweighting preserves support?

To test this directly, we applied 22 weightings to each fixed candidate bank and measured its uncovered fraction. We separately compared the two generator checkpoints on 93 molecules with nonempty banks.

Table D2. Fixed-bank invariance and paired checkpoint comparison. Brackets denote structural-group bootstrap intervals.

ComparisonW→HH→W
Maximum change across 22 weightings of each fixed bank0.00.0
Final minus earlier checkpoint, matched 93 molecules+0.0095 [−0.0245, 0.0435]+0.0006 [−0.0185, 0.0202]

Reweighting leaves uncovered fraction exactly unchanged. Figure 4 compares different generator checkpoints, so we will identify both checkpoints and report the paired differences explicitly.

Together, the new comparisons, source-free experiments, and evaluation audits address the reviewer's concerns about competitiveness and practical use. We do apologize for the lengthy response and greatly appreciate the constructive feedback. Given these comprehensive experiments, we would be very grateful if the reviewer would consider increasing the score and revisiting confidence in the submission.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Request for follow-ups for the rebuttal

Dear Reviewers,

Thank you again for your thoughtful feedback and valuable suggestions. We have carefully addressed your comments in our rebuttal and would greatly appreciate it if you could take a moment to review our responses. We would be happy to further clarify any remaining concerns and look forward to continuing the discussion.

Thank you for your time and consideration!

Best,

The authors

AuthorsOpenReview ↗

Global Response

We thank the reviewers for the thoughtful questions about SolvMorph and are grateful for Reviewer VLPj's continued support for acceptance! During rebuttal, we added external comparisons, tested the structural effects of solvent conditioning, evaluated measured permeability, and directly checked the effect of probability refinement.

Importance of generating solvent-dependent conformations

Cyclic peptides can adopt different conformations across solvents, and the structures observed in one environment can leave important target-solvent conformations unsampled. Reweighting adjusts the probabilities of structures already available. With SolvMorph, we generate additional conformations for the requested solvent and optionally refine their probabilities using predicted structural averages. This capability is incredibly important for peptide lead optimization.

With the submitted checkpoint on CONF192, comprising 192 peptides in 35 structural groups, we improved target-emergent mode recall over source copying by 0.26 for water-to-hexane inference and 0.35 for hexane-to-water inference, before probability refinement. We also tested whether these gains persisted with the same proposal allowance and whether requesting the correct solvent improved recovery.

New experiments for each reviewer

  • Reviewer VLPj requested external comparisons and clarification of the evaluation. We added OT-CFM, an EVA-Flow reproduction adapted to water and hexane, implicit-solvent GNN relaxation, and the released RINGER pipeline. In the new comparison with 200 attempted proposals per peptide and solvent on CONF192, SolvMorph achieved mode recall of 0.349 and 0.446, compared with 0.273 and 0.406 for OT-CFM and 0.087 and 0.102 for our EVA-Flow adaptation at approximately matched training GPU-time. On matched 48-peptide CONF192 subsets, we also reduced the uncovered target fraction relative to GNN relaxation by 0.207 and 0.297, with confidence intervals excluding zero. RINGER performed better on its compatible subset. We traced reconstruction failures to ring closure and improved SolvMorph's valid yield from 0.068 to 0.179 through a separately reported decoder repair. We also documented the cohort allocation and exclusion of each evaluated structural group from its final scoring predictor and generator checkpoint.

  • Reviewer udyc asked whether solvent conditioning produces chemically meaningful changes and whether the ensembles help predict measured permeability. Across CONF192, correct-target conditioning improved agreement with the direction of backbone N–H exposure changes in the MD references by 12 percentage points for water-to-hexane and 11 points for hexane-to-water. Both confidence intervals excluded zero. We also examined intramolecular hydrogen bonding and reconstruction quality across ring sizes. For downstream evaluation, we analyzed 1,016 molecules across 98 structural groups with measured PAMPA permeability. In the exploratory matched analysis, molecular features combined with SolvMorph ensembles achieved 0.811 macro Spearman, compared with 0.806 for molecular features alone and 0.808 with source ensembles.

  • Reviewer JzNC asked us to isolate solvent specificity and explain the refinement results. At the same budget of 20 attempted proposals, requesting the correct solvent improved emergent-mode recall over requesting the wrong solvent by 0.060 [0.013, 0.108] for water-to-hexane and 0.088 [0.042, 0.132] for hexane-to-water. These structural-group bootstrap intervals support an effect of the requested solvent under matched sampling. We traced the Figure 4 difference to two generator checkpoints. Applying 22 different weightings to each fixed bank changed its uncovered fraction by exactly zero. On the same generated banks, predicted-moment weighting reduced standardized observable error by 0.178 and 0.084 in the two directions. We therefore demonstrated both recovery of additional conformations through generation and improved agreement with target averages through refinement.

Case for acceptance

We have shown that source conformations contain useful information about solvent response and that we can generate additional target-solvent structures from those observations. The gains persist under matched proposal counts, respond to the requested solvent, and extend to comparisons with alternative generative and relaxation methods. We believe this is a substantive contribution to cyclic-peptide ensemble generation, supported by controlled experiments that directly test the proposed capability.

We are still awaiting follow-up from Reviewers udyc and JzNC on the additional experiments. We respectfully ask the Area Chair to consider this new evidence alongside the continued positive recommendation, and ask the reviewers to reconsider the initial scores and support acceptance of SolvMorph.

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.