acceptodds
Under review as a conference paper at ICLR 2027

MuCO2: Unlocking Multi-Stage Conformation Generation for Noncanonical Cyclic Peptides

Abstract

Cyclic peptides containing noncanonical amino acids (ncAAs) represent an indispensable modality in modern drug discovery, offering improved proteolytic stability and therapeutic profiles. However, efficiently modeling their conformational ensembles across complex energy landscapes remains challenging: existing methods like the HighFold series and MuCO mainly focus on canonical peptides and struggle with complicated structural perturbations. Building upon MuCO, we propose MuCO2, a multi-stage conformation generation framework for noncanonical cyclic peptides with α-backbone modifications, stereochemical transformations, and open-ended side-chain modifications. In particular, MuCO2 represents open-set ncAAs by aligning the atomic representations of Uni-Mol2 with the residue-level embeddings from ESM-2. Leveraging a curriculum learning paradigm driven by conformational energy and symmetries, MuCO2 learns a quaternion flow-based cyclic backbone generator coupled with a structure-guided Feynman-Kac steering mechanism, enabling efficient, generalizable cyclic peptide sampling conditioned on chirality parity and N-methylation. Extensive evaluations on canonical and noncanonical benchmarks show that MuCO2 achieves high-throughput cyclic peptide sampling with near-perfect stereochemical preservation, consistently outperforming state-of-the-art methods in physical stability, structural diversity, and potential-well coverage.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
noncononical amino acids, cyclic peptides, conformation generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_p8JDOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MuCO2, a framework for generating conformations of noncanonical cyclic peptides. It extends MuCO with aligned Uni-Mol2 and ESM-2 representations, modification-aware quaternion flow, and curriculum learning, together with a structure-guided Feynman–Kac steering strategy for backbone selection. Empirically, the authors report competitive performance across canonical and noncanonical cyclic-peptide benchmarks, with improvements in cyclization success and residue-level chirality preservation, particularly for chemically modified peptides.

Critical strengths

The paper addresses a meaningful and well-motivated problem. Noncanonical cyclic-peptide conformation generation is a relevant application area with genuine modeling challenges arising from ring-closure constraints, limited structural data for modified residues, and diverse stereochemical and side-chain modifications. The paper identifies a plausible and relevant tension between retaining canonical conformational priors and accommodating noncanonical chemistry. This framing is clear and helps motivate the architecture.

The method is conceptually coherent and reasonably well aligned with the problem structure. The decomposition into backbone generation, side-chain assembly, and physical refinement is intuitive. The chemical alignment mechanism is intended to accommodate residues outside the canonical vocabulary, while modification-aware conditioning introduces methylation and stereochemical information. The use of residue-level and pairwise representations as conditioning information is a sensible design for noncanonical cyclic-peptide generation. The overall architecture in Fig. 1 is understandable and the components form a coherent high-level pipeline.

The empirical section is comparatively broad. The paper evaluates on multiple benchmarks and reports several complementary metrics, including cyclization success, vacuum and implicit-solvent energy scores, backbone RMSD, and residue-level chirality preservation. This breadth is a strength. The ablation study is useful because it provides initial evidence about the contributions of backbone parameterization, curriculum training, and chemical representations. The external MeD and mixed-modification evaluations also support the claimed motivation beyond performance on canonical peptides.

Presentation of the overall framework is generally clear. The paper is organized around the main modules, the overview figure is informative, and the high-level generation process is easy to follow. The distinction between backbone generation, hybrid side-chain assembly, and force-field refinement is presented clearly. The method sections are organized around the core components and the experiments are reasonably well structured.

The contribution is potentially useful within a specialized but relevant area. Even if the scope is domain specific, extending cyclic-peptide models to accommodate chemical and stereochemical modifications can be practically meaningful. The paper also presents ideas, particularly chemical representation transfer and physics guided data augmentation, that may inform broader molecular conformation-modeling problems.

Critical weaknesses

The novelty is limited and not fully convincing as a major conceptual advance. While the full framework is reasonably well assembled, many of the ingredients are individually familiar, including hierarchical conformation generation, pretrained molecular representations, quaternion flow, template assembly, and force-field refinement. The paper’s main originality appears to lie in extending MuCO to noncanonical chemistry through the combination of chemical alignment, curriculum learning, and structure-guided selection. That can still be valuable, but the paper should do more to clarify precisely what is fundamentally new relative to prior cyclic-peptide generators and simpler adaptations of the existing framework.

The theoretical section is weaker than the empirical section. The analysis in Appendix C appears to provide a formal description and heuristic justification of FKS rather than a tightly algorithm specific demonstration of its claimed advantages. Based on the current presentation, it is not fully clear how geometric dispersion guarantees distinct physical-basin coverage or improved low-energy discovery. The DPP approximation and near-optimal dispersion claims also lack clearly defined approximation guarantees. I do not view the absence of a universal guarantee as a fatal flaw, but the theory currently reads as supportive intuition rather than central evidence.

The empirical comparisons are broad but still leave some important questions open. The paper compares against a useful set of learned baselines, but it is not fully clear whether relevant conventional conformer generation methods and other generative approaches are adequately represented under comparable computational budgets. Because the framework introduces several interacting components, I would like more detail on fairness of sampling and training budgets. In particular, main table inference uses (K=K_0), so FKS retains all proposed backbones, while its dedicated ablation increases the proposal pool from 10 to 100. Curriculum rounds also change data and training budget simultaneously, making the individual mechanisms difficult to isolate.

The practical significance of the contribution is somewhat uncertain. The paper presents the method as an efficient and generalizable framework for noncanonical cyclic-peptide sampling, but the experimental evidence is concentrated on structural and force-field-based metrics, with pseudo label generation, refinement, and energy evaluation all relying on CHARMM36. That is acceptable for an initial computational study, but it weakens the case for practical scientific significance unless the authors demonstrate a clear cost quality advantage over conventional tools, improved agreement with independent experimental data, or a useful downstream application. The absence of end-to-end runtime measurements particularly limits the claimed efficiency advantage.

Some claims are stronger than the available evidence strictly supports. The paper occasionally presents FKS as ensuring distinct basin exploration and the overall pipeline as providing near-perfect stereochemical preservation. However, cyclization success and residue level chirality averages do not directly establish complete molecular validity or accurate conformational distributions. Given the available evidence, more measured claims would be preferable. Several details also require clarification, including the whole peptide versus residue wise feature switch, the different side-chain partitions used in the main text and appendix, and the aggregation of dataset level energy results. These inconsistencies make the actual implementation and reported improvements harder to assess.

Questions for authors

What are the closest prior methods to MuCO2 in terms of noncanonical cyclic-peptide generation, and what is the minimal technical delta? A sharper comparison with MuCO and other compatible generators would help assess originality.

How fair are the comparisons in terms of sampling budget, training effort, and computational cost? Please provide end-to-end timing and compare FKS with random or simple diversity-based selection from the same candidate pool, alongside budget matched curriculum controls.

How directly does the analysis in Appendix C apply to the actual FKS sampling rule and subsequent force-field relaxation? Does it formally establish improved physical basin coverage or low-energy discovery relative to random selection, or mainly provide a conceptual justification?

How robust are the physical validity conclusions to the evaluation protocol? Since cyclization success and residue level chirality measure different aspects, please clarify complete molecule stereochemical validity, failure handling, and energy aggregation.

To what extent does MuCO2 provide practical benefits beyond improved scores under the shared CHARMM36 pipeline? Could the approach demonstrate a clear cost quality advantage over conventional conformer generation tools, improved experimental agreement, or a useful downstream application?

7 comments

Thank you for your thorough analysis, constructive criticism, and recognition of our model's coherence and breadth. Below is our detailed point-by-point response addressing each of your comments:

  • [Response to W1 & Q1] Conceptual Novelty and Technical Delta Over MuCO
  • [Response to W2 & Q3] Theoretical Framing vs. Empirical Role of Feynman-Kac Steering (FKS)
  • [Response to W3 & Q5] Comparison with Conventional Conformer Tools and Curriculum Rationale

and

  • [Response to W4 & Q4] Empirical Validation Beyond In-Sample Scores and Absolute Molecule Validity
  • [Response to W5] Implementation Clarifications: Whole-Peptide Feature Switch & Notations

Corresponding experiments and narrations are updated in revised paper.

[Response to W1 & Q1] Conceptual Novelty and Technical Delta Over MuCO

While MuCO (Wang et al., 2026; Page 14, Line 706) established a multi-stage paradigm, it is strictly restricted to canonical cyclic L-peptides and collapses on noncanonical chemistry. MuCO2 introduces three minimal, non-trivial technical contributions to overcome these fundamental physical barriers:

  1. Cross-Modal Chemical Alignment Adapter (gϕg_\phi, Eq. 2, Page 4): ESM-2 cannot encode ncAAs. We bridge frozen Uni-Mol2 atomic representations into the residue trunk. As shown in Table 14 (Page 38), removing this alignment and relying on sequence-only representations causes chirality preservation to plummet by 6.06% and degrades implicit water score by 56.47 kJ/mol on CycPeptMPDB-Mixed.
  2. Singularity-Free Quaternion Flow with Parity Conditioning (Eqs. 4 & 5, Pages 4–5): Standard Riemannian SE(3) flow matching on SO(3)\mathrm{SO}(3) suffers from numerical instabilities and chiral collapse under inversion. We formulate the flow on S3\mathbb{S}^3 via SLERP, gated by directional peptide-bond indicators (PijmjP_{ij}m_j) and relational parity (cicjc_ic_j). On CycPeptMPDB-MeD, this boosts Success@1 from 42.13% to 67.90% and chirality preservation from 78.37% to 89.21% (Table 10, Page 36).
  3. Symmetry-Driven Physics Curriculum (Algorithm 1, Page 6; Algorithm 2, Page 23): Solves the extreme data scarcity of multi-modified macrocycles. Without this curriculum, Success@10 on CPSea-MeD drops from 84.10% to 28.06% (Table 12, Page 37).

[Response to W2 & Q3] Theoretical Framing vs. Empirical Role of Feynman-Kac Steering (FKS)

We appreciate the reviewer's insight regarding theoretical guarantees versus physical realities. We agree that geometric dispersion does not unconditionally guarantee entry into distinct force-field local minima. We have explicitly clarified in Appendix C.3 (Page 27, Lines 1412–1414) that FKS provides an exploratory inductive bias motivated by extreme-value statistics (Eq. 25): by preventing sample degeneracy at the continuous flow stage, it enlarges the union volume of explored basins (Vol(⋃Ω(Bi))\text{Vol}(\bigcup \Omega(B_i)), Eq. 26), elevating KeffK_{\text{eff}} closer to the nominal budget K0K_0.

Empirically, under an identical evaluation budget of 100 full-atom candidates on CPSea-PDB (Figure 4e, Page 10), FKS drives down vacuum energy by 43.71 kJ/mol and implicit water score by 40.41 kJ/mol, while lowering Cα\text{C}_\alpha-RMSD from 2.98 A˚2.98\text{ Å} to 2.87 A˚2.87\text{ Å} over unsteered naive tree expansion. Figure 7 (Page 38) further confirms that FKS introduces negligible time overhead relative to downstream relaxation.


[Response to W3 & Q5] Comparison with Conventional Conformer Tools and Curriculum Rationale

  • Failure Modes of Conventional Distance Geometry: Standard distance geometry tools (such as RDKit ETKDG) fail to close macrocyclic rings under dense backbone constraints. In our tests, ETKDG achieved near-zero valid cyclization for macrocycles longer than 10 residues.
  • Prohibitive Cost of Physics-Based Search Protocols: Comprehensive sampling tools like CREST are computationally prohibitive for high-throughput discovery: CREMP reports an average runtime of 13.9 hours on 14 CPU cores (195 CPU-hours) per single cyclic hexapeptide, totaling 3.9 million CPU-hours across their dataset (Grambow et al., 2024; Page 12, Line 628). Moreover, incorporating novel ncAAs into Rosetta/MD incurs tedious parameterization bottlenecks for charges and bonded terms (Wijesiri & Brown, 2025; Page 14, Line 713). MuCO2 provides an amortized deep generative alternative that evaluates in seconds.
  • Curriculum Training Fairness (Table 12, Page 37): In our experiments, extending Round-1 training on single-site modifications did not improve generalization to multi-site peptides; instead, it led to severe overfitting on isolated modifications. Progressive expansion across Rounds 1–3 is physically necessary: earlier rounds generate viable geometric starting points that allow CHARMM36 to relax multi-modified structures with high success rates, creating valid training distributions for subsequent rounds.

[Response to W4 & Q4] Empirical Validation Beyond In-Sample Scores and Absolute Molecule Validity

Our validation is not confined to in-sample CHARMM36 scores:

  1. Independent Explicit-Solvent MD Distributions (Table 3, Page 9): We evaluated generated ensembles against 50 ns explicit-solvent MD in both water and hexane.
  2. Candidate Validity Metric (Valid@K, Eq. 42, Page 34): To directly address the reviewer's concern that micro-averaged chirality might mask partial stereochemical errors, we evaluate Candidate Validity (Eq. 42, Page 34), which strictly requires 100% correct chiral centers across all residues simultaneously and valid ring closure (δ≤0.1 A˚\delta \le 0.1\text{ Å}).

As shown in Table 3 (Page 9), MuCO2 achieves near-perfect whole-molecule validity, whereas AlphaFold3-based HighFold3 breaks down:

Dataset / BenchmarkMethodWater W1W_1@10 ↓\downarrowWater Cov.@10 ↓\downarrowDiv. ↑\uparrowValid@10 ↑\uparrow (%)
CycPeptMPDB-MeDHighFold-MeD21.519 Å1.121 Å1.153 Å62.7%
(Page 9, Table 3)HighFold31.500 Å1.233 Å0.852 Å41.5%
MuCO2 (Ours)1.354 Å1.076 Å1.177 Å100.0%
CycPeptMPDB-MixedHighFold31.275 Å0.798 Å1.193 Å31.4%
(Page 9, Table 3)MuCO2 (Ours)1.151 Å0.710 Å1.241 Å100.0%
  1. Energy Aggregation Protocol: As formalized in Appendix E.1 (Eq. 33, Page 32), total potential energy scales extensively with peptide length and net charge. To ensure unbiased evaluation, energy comparisons are computed strictly in a pairwise manner across identical target sequences, and then averaged across the dataset.

[Response to W5] Implementation Clarifications: Whole-Peptide Feature Switch & Notations

We have carefully tightened and aligned our descriptions across the main text and appendix:

  • Whole-Peptide Feature Switch (Appendix B.1, Page 19, Lines 1011–1015; Eq. 3, Page 4): We confirm that we adopt a whole-peptide feature switch. If a cyclic peptide contains any noncanonical side chain, the representation switches entirely to the projected Uni-Mol2 embeddings (gϕ⋆(ZR)g_{\phi^\star}(Z_R)) across all residue positions and pairs. Mixing ESM-2 and Uni-Mol2 representations within the same peptide creates latent geometric domain mismatch. As verified in Table 14 (Page 38), this unified Uni-Mol2 representation provides superior chemical and stereochemical grounding:
Dataset / BenchmarkEmbeddingsSuccess@10 ↑\uparrow (%)EwaterE_{\text{water}}@10 ↓\downarrow (kJ/mol)Cα\text{C}_\alpha-RMSD@10 ↓\downarrow (Å)Chirality ↑\uparrow (%)
HighFold2-CompatibleESM-2100.00%-3318.903.17 Å94.19%
(Page 38, Table 14)Uni-Mol2100.00%-3424.843.14 Å98.81%
CycPeptMPDB-MixedESM-2100.00%504.390.64 Å92.35%
(Page 38, Table 14)Uni-Mol2100.00%447.920.56 Å98.41%
  • Side-Chain Partition Notation (Ic/InI_c / I_n vs. Iknown/IopenI_{known} / I_{open}): In the main text (Page 3, Line 149), IcI_c and InI_n denote standard canonical versus noncanonical amino acid types. In Appendix B.1 (Page 19, Lines 1017–1025), IknownI_{known} generalizes IcI_c to include noncanonical residues that share canonical torsional degrees of freedom parameterized by exocyclic dihedrals (allowing EquiformerV2 neural packing), while IopenI_{open} designates open-set residues bypassing neural torsions and assembled via template attachment. We have unified this notation in the revised manuscript.
Reviewer_6SNuOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces MuCO2, a multi-stage conformation generation framework for cyclic peptides containing noncanonical amino acids (NCAAs). The framework aligns Uni-Mol2 features with ESM-2 representations to encode noncanonical chemistry, uses a chemically conditioned quaternion-flow model to generate backbones with noncanonical modifications, and applies physics-guided curriculum learning to address the scarcity of training structures. Additionally, structure-guided Feynman–Kac steering is used to improve sampling diversity. Results on canonical and noncanonical benchmarks demonstrate improvements in cyclization success, stereochemical correctness, and conformational quality.

Critical strengths
  1. The paper considers the valuable problem of extending cyclic peptide conformation generation to noncanonical amino acids. Supporting N-methylation, mixed chirality, and noncanonical side chains makes the framework more relevant to practical peptide design. To address these problems, the paper extends the MuCO framework to several practically relevant noncanonical modifications, improving the flexibility and controllability of the conformation generation pipeline.

  2. The experiments cover both canonical and noncanonical peptides, including test sets derived from CycPeptMPDB. The component ablations and improvements across training rounds provide evidence for the effectiveness of the design, while the external evaluations support its generalization beyond the training data.

Critical weaknesses
  1. The paper claims improvements in structural diversity and potential-well coverage in both the abstract (Line 027-028) and introduction (Line 087-088), but the evaluation does not directly establish these improvements. The metrics in Section 5.1 measure cyclization success, energy, reference RMSD, and stereochemical correctness. These are useful measures of candidate quality, but they do not show whether the generated conformations cover a broader range of structures or distinct low-energy basins. In particular, a lower RMSD to a reference structure supports better recovery of that structure, but does not by itself demonstrate the stronger multi-basin exploration discussed in Section 5.2. Since diversity and coverage are among the paper’s main claimed advantages, a quantitative analysis of the existing generated ensembles would more sufficiently support these claims. The reviewer would raise the score if this concern is adequately addressed.
Questions for authors
  1. Section 3.2 describes neural side-chain packing for canonical residues and template attachment for noncanonical residues, while Appendix B.1 extends neural packing to “known natural proxies.” Which noncanonical residues fall into each category? Is the use of templates motivated by limitations in the current torsion parameterization? Moreover, is a single template conformation used for each residue type?

  2. Section 5.1 describes @10 as “best-of-10,” while Algorithm 3 ranks valid candidates by effective energy. Are the reported best energy and RMSD values selected independently from the candidates, or are both evaluated on the candidate selected by energy?

  3. Appendix D.3 includes head-to-tail cyclic peptides with internal disulfide bonds. Out of curiosity, could MuCO2 also accommodate scaffold-linked bicyclic peptides, such as those formed by crosslinking three cysteine residues with TBMB? If not, what extensions would be needed to represent the shared scaffold and enforce its connections to multiple residues?

7 comments

Dear Reviewer 6SNu,

We sincerely appreciate your positive evaluation of our work's practical relevance and experimental breadth. In response to your comments, we have incorporated direct quantitative evaluations of ensemble diversity, potential-well coverage, and rotamer states, structured as follows, and corresponting experiments and narrations are updated in paper (in blue).

  • [Response to W1] Direct Quantification of Diversity and Potential-Well Coverage
  • [Response to Q1] Division of Scope: Neural Rotamer Packing vs. Template Assembly
  • [Response to Q2] Candidate Selection Protocols for Metric Reporting vs. Practical Inference
  • [Response to Q3] Feasibility and Extension Pipeline for TBMB-Crosslinked Bicyclic Peptides

[Response to W1] Direct Quantification of Diversity and Potential-Well Coverage

To provide direct quantitative evidence that MuCO2 covers multi-basin energy landscapes rather than collapsing into single modes, we benchmarked 10-conformer ensembles against 50 ns explicit-solvent MD distributions across both water and hexane environments (Page 9, Table 3), and highlighted landscape coverage on challenging individual cases (Page 9, Figure 3).

Evaluation BenchmarkMethodWater W1W_1@10 ↓\downarrowWater Cov.@10 ↓\downarrowHexane W1W_1@10 ↓\downarrowHexane Cov.@10 ↓\downarrowDiv. ↑\uparrowValid@10 ↑\uparrow (%)
CycPeptMPDB-MeDHighFold-MeD21.519 Å1.121 Å1.613 Å1.176 Å1.153 Å62.7%
(Page 9, Table 3)HighFold31.500 Å1.233 Å1.598 Å1.303 Å0.852 Å41.5%
MuCO2 (Ours)1.354 Å1.076 Å1.509 Å1.171 Å1.177 Å100.0%
CycPeptMPDB-MixedHighFold31.275 Å0.798 Å1.338 Å0.819 Å1.193 Å31.4%
(Page 9, Table 3)MuCO2 (Ours)1.151 Å0.710 Å1.209 Å0.706 Å1.241 Å100.0%
  • Global Ensemble Matching (Table 3, Page 9): On CycPeptMPDB-Mixed in water, MuCO2 achieves the lowest Wasserstein-1 transport distance (1.151 A˚1.151\text{ Å} vs. 1.275 A˚1.275\text{ Å}) and coverage error (0.710 A˚0.710\text{ Å} vs. 0.798 A˚0.798\text{ Å}), alongside higher pairwise diversity (1.241 A˚1.241\text{ Å} vs. 1.193 A˚1.193\text{ Å}) and 100.0% target validity.
  • Direct Multi-Well Landscape Visualizations (Figure 3, Page 9): As showcased for cyclic hexapeptide CP1682 (Page 9, Figure 3), MuCO2 simultaneously populates both independent MD potential wells (Energy Well 1 & Well 2) in PCA projection (Figure 3a), achieves markedly smaller nearest-MD distances on the cumulative distribution function (Figure 3b), and samples distinct backbone clusters (Figure 3c), reducing W1W_1 from 1.280 A˚1.280\text{ Å} to 0.552 A˚0.552\text{ Å} and coverage error from 0.859 A˚0.859\text{ Å} to 0.361 A˚0.361\text{ Å} relative to HighFold3.
  • Side-Chain Rotamer Diversity (Table 4, Page 10): By combining template placement with DPP rotamer exploration (Appendix C.4), pre-relaxation ncAA diversity increases from 1.255 A˚1.255\text{ Å} to 2.708 A˚2.708\text{ Å} and coarse χ\chi-states expand from 1.6531.653 to 5.2085.208, proving that multi-basin coverage spans both backbone and side-chain manifolds.

[Response to Q1] Division of Scope: Neural Rotamer Packing vs. Template Assembly

We have clarified the exact boundary between neural packing and template assembly in Section 3.2 (Page 5, Lines 235–247) and Appendix B.1 (Page 19, Lines 1017–1025):

  • Neural Packing (IcI_c / IknownI_{known}): Conditional EquiformerV2 operates on all 20 canonical amino acids and supported natural proxies that share canonical exocyclic dihedral connectivity (χ1\chi_1–χ4\chi_4). The reason is that flow matching over T4\mathbb{T}^4 requires fixed torsional degrees of freedom.
  • Template Attachment (InI_n / IopenI_{open}): Applied to open-set noncanonical residues (e.g., bulky rings, unnatural heterocycles, and peptoid-like NN-alkylations) that fall outside the T4\mathbb{T}^4 parameterization.
  • Multi-Rotamer Exploration via DPP: Rather than relying on a single static template, open-set residues are explored using partition-constrained DPP sampling (Appendix C.4, Page 27). For each surviving backbone branch, we generate an overcomplete pool of rotamers by perturbing exocyclic dihedrals, penalize local steric clashes via kernel qiq_i (Eq. 28), and select the most diverse, clash-free conformers prior to CHARMM36 relaxation.

[Response to Q2] Candidate Selection Protocols for Metric Reporting vs. Practical Inference

We have clarified the distinction between benchmark evaluation metrics and inference ranking in Section 5.1 (Page 7, Lines 341–354), Appendix B.1 (Page 21, Lines 1115–1118), and Appendix E (Pages 32–34):

  • Metric Reporting in Benchmark Tables (Tables 2 & 9): To assess the absolute physical and structural discovery limits of the generated distribution, metrics are reported independently:
    1. Evac@10E_{\text{vac}}@10 and Ewater@10E_{\text{water}}@10 report the ensemble potential minimum (min⁡1≤k≤10E(xk)\min_{1 \le k \le 10} E(x_k), Eqs. 34 & 37), measuring the deepest energy well colonized by the model.
    2. Cα−RMSD@10C_\alpha{-RMSD}@10 evaluates the structural oracle (min⁡1≤k≤10RMSD(xk,Y)\min_{1 \le k \le 10} \text{RMSD}(x_k, Y), Eq. 39), reflecting the closest geometric approach to the reference fold without being confounded by implicit-solvent ranking inaccuracies.
  • Practical Inference Deployment (Algorithm 3, Page 24, Lines 1278–1280): When selecting candidates in a blind setting without ground truth, valid candidates are ranked and selected strictly by their total effective physical energy Eeff(X⋆)=Evac(X⋆)+ΔGsolv(X⋆)E_{\text{eff}}(X^\star) = E_{\text{vac}}(X^\star) + \Delta G_{\text{solv}}(X^\star) (Line 1279).
  • Distribution Matching (Table 3, Page 9): Evaluates the entire 10-conformer distribution against the 100-frame MD reference ensemble using balanced optimal transport (W1@10W_1@10, Eq. 43).

[Response to Q3] Feasibility and Extension Pipeline for TBMB-Crosslinked Bicyclic Peptides

Accommodating scaffold-linked bicyclic peptides (such as those formed by crosslinking three Cysteine residues via 1,3,5-tris(bromomethyl)benzene, TBMB) is implicitly chemically compatible with MuCO2’s multi-stage architecture (we may integrate into our future work):

  1. Backbone Conditioning: Following composable geometric constraints (Jiang et al., ICML 2025; Page 13, Line 648), pairwise spatial distance priors (dCys-S…Cys-S≈6.0 A˚d_{\text{Cys-S}\dots\text{Cys-S}} \approx 6.0\text{ Å}) can be injected directly into the pair embedding matrix h~ij\tilde{h}_{ij} (Eq. 4, Page 4) during Stage-1 backbone flow generation.
  2. Scaffold Coordinate Fitting: In Stage-2 side-chain assembly, the rigid 3D mesitylene core of TBMB is placed by tripartite Kabsch alignment onto the three synthesized Cysteine sulfur coordinates, and three covalent C–S\text{C–S} thioether bonds are formed.
  3. Joint Physics Refinement: The entire crosslinked bicyclic system is relaxed via CHARMM36 using standard parameter stream files for thioether linkages. We have added this prospective discussion in Section 6 (Page 10, Lines 530–534).
Reviewer_8212OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MUCO2, a new model for structure prediction of cyclic peptides containing non-canonical amino acids (ncAA). Building on a previous model, MUCO, several technical contributions are combined in MUCO2: representation of ncAA as canonical priors with perturbations embedded using ESM-2 for residue level and Uni-Mol2 for atomic level; a Quaternion Flow model replacing the Riemannian SE(3)-Flow model for better support of ncAA and cycling sequences; curriculum learning-based training combining pretraining on canonical peptides, training for generalisation to single-site modifications using symmetry augmentation, and training for generalisation to multi-site modifications using simulated noncanonical conformations from the CHARMM36 force field; and improved diversity of the sampled structures by guiding the backbone selection using a particle crowding potential. The model is evaluated on canonical, backbone-mutated, and open-set noncanonical cyclic peptides and compared with several baseline models, including the previous model MUCO, and using a wide range of metrics, which I will comment on below.

Critical strengths

None

Critical weaknesses
  1. The evaluation does not sufficiently support the claims of accurate ncAA conformational ensemble generation.

Ring closure, stereochemical preservation, and low force-field energies establish useful aspects of structural validity, but do not demonstrate recovery of relevant conformational states or their relative populations. Cα-only RMSD against a single reference frame is particularly limited: it neither evaluates ensemble coverage nor directly tests the ncAA side-chain coordinates produced by the proposed packing procedure. Heavy-atom coverage and matching metrics against reference ensembles, under matched sampling budgets, together with torsion and cis/trans-state analyses, would more directly assess the claimed capabilities. Agreement with ensemble-averaged NMR observables would provide additional validation where suitable experimental data are available.

  1. Sampling of ncAA side-chain conformations is insufficiently demonstrated.

Canonical side-chain torsions are explicitly sampled, whereas ncAA side chains are attached through rigid template placement followed by restrained minimization. The procedure does not clearly establish how alternative ncAA rotamers or internal conformational states are explored. Direct evaluation of ncAA side-chain accuracy and diversity is therefore needed, together with clarification of force-field parameterization for the supported noncanonical chemistries.

Questions for authors

In addition to the questions raised by the two points above, I am also wondering about the accelerated sampling using the Quaternion Flow compared to the Riemannian SE(3) flow. Have the authors measured the runtimes under matched conditions? Reporting both backbone generation time and total time per valid conformation, including packing, relaxation, and failed attempts, would clarify whether QFlow delivers a meaningful improvement in sampling throughput.

5 comments

Dear Reviewer 8212,

Thank you for your constructive and insightful feedback, we have updated our paper (in blue). Below, we will address each of your concerns point-by-point, organizing into:

  • [Response to W1] Conformational Solution Ensemble Evaluation and Experimental NMR Validation.
  • [Response to W2] Noncanonical Side-Chain Sampling (MuCO2+DPP) and Force-Field Parameterization
  • [Response to Q1] Sampling Throughput and End-to-End Runtime Analysis

[Response to W1] Conformational Ensemble Evaluation and Experimental Solution NMR Validation

To directly address your concern regarding ensemble coverage and population fidelity beyond single-frame RMSD, we benchmarked 10-conformer ensembles against 50 ns explicit-solvent MD distributions across both polar (water) and apolar (hexane) environments (Page 9, Table 3), as well as on experimental solution NMR structures (Page 7, Section 5.1; Page 8, Table 2).

As shown in the excerpt below from Table 3 (Page 9), MuCO2 consistently captures physical conformational distributions superior to state-of-the-art baselines:

Benchmark / DatasetMethodWater W1W_1@10 ↓\downarrowWater Cov.@10 ↓\downarrowDiv. ↑\uparrowValid@10 ↑\uparrow (%)
CycPeptMPDB-MeDHighFold-MeD21.519 Å1.121 Å1.153 Å62.7%
(Page 9, Table 3)HighFold31.500 Å1.233 Å0.852 Å41.5%
MuCO2 (Ours)1.354 Å1.076 Å1.177 Å100.0%
CycPeptMPDB-MixedHighFold31.275 Å0.798 Å1.193 Å31.4%
(Page 9, Table 3)MuCO2 (Ours)1.151 Å0.710 Å1.241 Å100.0%
  • Ensemble Distribution Matching: On CycPeptMPDB-Mixed in water, MuCO2 reduces Wasserstein-1 distance (W1W_1, defined via sequence-ordered optimal transport in Eq. 43, Page 34) from 1.275 A˚1.275\text{ Å} to 1.151 A˚1.151\text{ Å} and coverage error (Eq. 45, Page 34) from 0.798 A˚0.798\text{ Å} to 0.710 A˚0.710\text{ Å} over HighFold3, while enhancing pairwise diversity (1.241 A˚1.241\text{ Å} vs. 1.193 A˚1.193\text{ Å}) and attaining 100.0% target validity.
  • Experimental Solution NMR Validation: On the HighFold2-Compatible experimental NMR benchmark (Page 30, Appendix D.3; 9 solution NMR entries), MuCO2 attains reference fidelity comparable to specialized predictors on deterministic @1 inference: 3.248 A˚3.248\text{ Å} vs. 3.282 A˚3.282\text{ Å} for HighFold2 on macrocycles <20 aa<20\text{ aa} (7 cases), and 3.782 A˚3.782\text{ Å} vs. 3.639 A˚3.639\text{ Å} on ≥20 aa\ge 20\text{ aa} (2 cases). Crucially, these NMR structures are deposited in the PDB and were already present in the AlphaFold pretraining data utilized by HighFold2/3, whereas MuCO2 achieves this recovery zero-shot without prior structural exposure.

[Response to W2] Noncanonical Side-Chain Sampling (MuCO2+DPP) and Force-Field Parameterization

To actively explore alternative ncAA rotameric states beyond a single rigid template, we introduced quality-weighted Determinantal Point Process sampling (MuCO2+DPP, detailed in Appendix C.4, Pages 27–28). MuCO2+DPP stochastically perturbs exocyclic dihedrals, penalizes local steric overlap via a clash severity kernel (qiq_i, Eq. 28), and selects diverse, clash-relieved rotamers before relaxation.

The quantitative impact on standard and noncanonical side chains is reported in Table 4 (Page 10):

StageMethodSide-Chain Diversity (Noncan.) ↑\uparrowSide-Chain W1W_1 (Noncan.) ↓\downarrowCoarse χ\chi-States (Noncan.) ↑\uparrowOverall W1W_1 ↓\downarrow
Before RelaxationMuCO2 (Template)1.255 Å2.587 Å1.6531.490 Å
(Page 10, Table 4)MuCO2+DPP2.708 Å1.928 Å5.2081.365 Å
After RelaxationMuCO2 (Template)2.098 Å2.354 Å3.6041.373 Å
(Page 10, Table 4)MuCO2+DPP2.791 Å1.840 Å4.3471.274 Å

As demonstrated above, pre-relaxation ncAA diversity surges from 1.255 A˚1.255\text{ Å} to 2.708 A˚2.708\text{ Å} and coarse χ\chi-states expand from 1.6531.653 to 5.2085.208, reducing post-relaxation ncAA W1W_1 against water MD from 2.354 A˚2.354\text{ Å} to 1.840 A˚1.840\text{ Å} (measured across heavy atoms, Eq. 47, Page 34).

Force-Field Parameterization: All supported chemistries follow CHARMM36-compatible protocols: N-methylations use our charge-neutral NME patch derived from official CHARMM parameters, D-residues employ dedicated CHARMM36 topologies, and noncanonical side chains are parameterized using QM-refined CHARMM36/CGenFF-compatible parameters (Croitoru et al., 2021).

[Response to Q1] Sampling Throughput and End-to-End Runtime Analysis

Quaternion Flow (QFlow) delivers substantial speedups at both the backbone generator level and the full-atom pipeline level (Page 9, Section 5.3; Page 10, Figure 4a; Page 38, Figure 7 & Appendix H):

  • Backbone Step Speedup: By parameterizing rotations directly on S3\mathbb{S}^3 and integrating via SLERP, QFlow eliminates iterative geodesic projections and SO(3)\mathrm{SO}(3) logarithmic maps. It accelerates backbone training from 212212 to 77 s/epoch77\text{ s/epoch} (2.75×2.75\times speedup) and batch inference from 6.726.72 to 5.02 s/batch5.02\text{ s/batch} at batch size 100 (1.34×1.34\times speedup) (Page 10, Figure 4a).
  • End-to-End Throughput per Valid Conformation: In end-to-end conformer generation, downstream force-field relaxation accounts for ∼50%\sim 50\% of wall-clock runtime (Figure 7, Page 38). Crucially, QFlow drastically improves cyclization and physical validity rates. On CycPeptMPDB-MeD, single-attempt validity (Success@1) surges from 42.13%42.13\% (SE(3)-Flow) to 67.90%67.90\% (QFlow) (Page 36, Table 10). Because uncyclized or clashing candidates from SE(3) lead to discarded relaxation time and require costly repeated sampling attempts, QFlow significantly reduces the net amortized runtime required to yield a verified, physically valid cyclic peptide conformation.

Comments on the paper

1 comment

General Response to Area Chair and Reviewers

We sincerely thank the Area Chair and all reviewers (8212, 6SNu, and p8JD) for their constructive feedback and active engagement throughout the discussion phase. We are grateful that all reviewers confirmed our revisions addressed their concerns and updated or reaffirmed their positive recommendations (6SNu: 4; 8212: 3; p8JD: 3, noting 3.5 for the AC).

Below is a brief summary of the major updates incorporated into the revised manuscript (highlighted in blue):

  • MD ensemble evaluation and physical validity (8212, 6SNu, p8JD). Added 50 ns explicit-solvent MD distribution-matching benchmarks (water and hexane), multi-basin landscape visualizations, full-molecule validity (Valid@K), and zero-shot validation on experimental solution NMR structures. See Section 5.2, Table 3, Figure 3, and Appendices D.3 and E.4–E.5.
  • Noncanonical side-chain exploration (MuCO2+DPP) (8212, 6SNu). Introduced quality-weighted, partition-constrained DPP rotamer sampling to actively explore open-set ncAA conformations beyond single templates, and clarified the neural vs. template packing scope as well as CHARMM36/CGenFF parameterization. See Sections 3.2 and 5.3, Table 4, and Appendices B.1, C.4, and D.2.
  • End-to-end runtime and budget-matched controls (8212, p8JD). Added component-wise and amortized end-to-end runtime analyses across peptide lengths, along with budget-matched unsteered controls confirming the gains of FKS diversity selection. See Section 5.3, Figure 4(a, e), and Appendix H (Figure 7).
  • Methodological, theoretical, and structural clarifications (p8JD, 6SNu). Sharpened the technical delta over MuCO, refined the theoretical framing of FKS as an exploratory inductive bias, unified implementation notations (whole-peptide feature switch, Ic/InI_c/I_n, pairwise energy aggregation, and oracle vs. EeffE_{eff} ranking), and added a prospective discussion on scaffold-linked bicyclic peptides (e.g., TBMB). See Sections 5.1 and 6, and Appendices B.1 and C.3.

All additional clarifications raised during the discussion will be cleanly integrated into the final version alongside the public release of our codebase and parameter files. Thank you again for your time and support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.