acceptodds
Under review as a conference paper at ICLR 2027

CycPepFlow: Reconciling Ring Coupling and Stereochemical Fidelity in Cyclic Peptide Ensemble Generation

Abstract

Cyclic peptides are promising therapeutic scaffolds, but generating their conformational ensembles requires capturing long-range ring coupling while preserving local chemical geometry and stereochemistry. We introduce CycPepFlow, an all-atom Cartesian flow model that generates conformers directly from molecular graphs. CycPepFlow combines SO(3)-equivariant local interactions with gated global attention, while atomic chirality conditioning and signed-volume supervision promote tetrahedral stereochemical fidelity. An optional all-pair graph-geodesic attention bias (APG) incorporates pairwise topology into global communication. Evaluated on 1,000 held-out CREMP cyclic peptides containing four to six residues, CycPepFlow-L achieves the highest coverage F1 among the compared methods (76.13%), whereas CycPepFlow APG-L achieves the highest generated-set coverage precision (79.78%) and lowest generated-to-reference matching distance (0.419 Å). On the six-residue subset, APG-L achieves the best coverage F1 (61.61%) together with 99.71% conformer-level tetrahedral accuracy. CycPepFlow thus combines broad reference-ensemble coverage with high generated-set precision and stereochemical fidelity, with its six-mer performance highlighting the potential for extension to larger macrocycles.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cyclic peptides, conformer generation, flow matching, equivariant neural networks, molecular generative models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_21uiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CycPepFlow, an all-atom Cartesian flow-matching framework for cyclic peptide conformer ensemble generation. The method combines SO(3)-equivariant local message passing, gated global attention, Attention Residuals, and stereochemical supervision. An optional all-pair graph-geodesic attention bias is also introduced to incorporate molecular topology. Experiments on the CREMP benchmark demonstrate improvements in ensemble generation quality and stereochemical accuracy compared with several existing conformer generation approaches.

Critical strengths
  1. Important research problem. Cyclic peptide conformer generation is an important and challenging problem with potential applications in molecular modeling and drug discovery.

  2. Clear motivation, and the paper is easy to follow in general. The paper clearly explains the challenges associated with cyclic peptide generation, particularly the coupling between local geometry and global ring conformation, the diversity of noncanonical amino acids, and the importance of preserving stereochemistry.

Critical weaknesses
  1. Limited methodological novelty. My main concern is the lack of a clearly identifiable technical contribution. Most components of CycPepFlow, including Cartesian flow matching, graph-harmonic priors, SO(3)-equivariant message passing, gated attention, Attention Residuals, graph-geodesic attention biases, and chirality supervision, are adapted from existing methods. The paper primarily combines these techniques without introducing a fundamentally new modeling mechanism or providing substantial new insights into cyclic peptide generation. While combining existing approaches can be valuable, I would expect stronger methodological insights or more compelling empirical evidence to justify the contribution at a top-tier machine learning conference.

  2. Insufficient experimental results. The proposed method achieves improvements over existing baselines, but the overall gains in ensemble coverage are relatively modest. It remains unclear whether the method generalizes to larger macrocycles or more diverse chemical structures. Furthermore, the paper lacks sufficient analysis of computational efficiency and physical validity.

  3. Ablation studies do not fully justify the component choices. The authors provide several component-removal ablations, which are appreciated. However, these experiments do not sufficiently explain why the selected techniques are preferable to reasonable alternatives. For example, the paper does not isolate the benefit of gated attention over standard attention, and removing Attention Residuals does not reduce the overall coverage F1 score.

  4. The presentation could be more focused. The manuscript devotes too much space to background, related work, and figures, particularly the extensive discussion of existing protein-derived and molecular conformer generation methods.

Questions for authors

See weakness

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer 21ui

Weakness 1.

Limited methodological novelty. My main concern is the lack of a clearly identifiable technical contribution. Most components of CycPepFlow, including Cartesian flow matching, graph-harmonic priors, SO(3)-equivariant message passing, gated attention, Attention Residuals, graph-geodesic attention biases, and chirality supervision, are adapted from existing methods. The paper primarily combines these techniques without introducing a fundamentally new modeling mechanism or providing substantial new insights into cyclic peptide generation. While combining existing approaches can be valuable, I would expect stronger methodological insights or more compelling empirical evidence to justify the contribution at a top-tier machine learning conference.

Response: We thank the reviewer for this comment. Similar to recent advances in molecular modeling such as DiTMC and RFpeptides, CycPepFlow builds on established generative foundations while introducing targeted architectural innovations to jointly resolve local geometry, global ring closure, and stereochemical fidelity. Specifically, we introduce two key designs tailored to macrocycles: (i) atomic chirality conditioning with signed-volume supervision, which achieves 99.71% conformer-level tetrahedral accuracy on 6-mer macrocycles without post-hoc corrections; and (ii) an all-pair graph-geodesic attention bias that embeds cyclic topology directly into global attention to capture long-range ring constraints. Crucially, our experiments demonstrate that CycPepFlow outperforms the strongest baselines on the evaluated 7-mer and NCAA-containing 6-mer panels, underscoring the practical value of this unified design for cyclic peptide drug discovery.

Weakness 2.

Insufficient experimental results. The proposed method achieves improvements over existing baselines, but the overall gains in ensemble coverage are relatively modest. It remains unclear whether the method generalizes to larger macrocycles or more diverse chemical structures. Furthermore, the paper lacks sufficient analysis of computational efficiency and physical validity.

Response: We thank the reviewer for these constructive suggestions. Improvements on established in-distribution benchmarks are often modest given strong baselines (e.g., on GEOM-DRUGS, ETFlow improves by ~2% over MCF and DiTMC by ~1% over ETFlow). However, our method demonstrates substantial gains in out-of-distribution generalization. In our new experiments (Figure 5), CycPepFlow APG-L achieves a COV-F1 of 56.13% on larger 7-mer macrocycles (vs. 40.09% for DiTMC rPE-L), and CycPepFlow-L achieves 34.38% on chemically diverse NCAA-containing 6-mers (vs. 18.93% for LoQI), outperforming the strongest external baseline on each subset. Furthermore, our newly added discussion in Section 5.2 demonstrates both strong physical validity and competitive computational efficiency.

Weakness 3.

Ablation studies do not fully justify the component choices. The authors provide several component-removal ablations, which are appreciated. However, these experiments do not sufficiently explain why the selected techniques are preferable to reasonable alternatives. For example, the paper does not isolate the benefit of gated attention over standard attention, and removing Attention Residuals does not reduce the overall coverage F1 score.

Response: We thank the reviewer for this insightful comment. To isolate the effect of gated attention over standard attention, we added an ablation in Figure 10 confirming that gated attention outperforms standard attention. Regarding Attention Residuals, while removing them leaves overall coverage F1 unchanged, it reduces COV-F1 on the more challenging 6-mer subset by 1.46 percentage points—suggesting that Attention Residuals benefit performance on this subset and may improve generalization to larger macrocycles. We also highlighted this trend in the caption of Figure 10 to make it clearer to readers.

Weakness 4.

The presentation could be more focused. The manuscript devotes too much space to background, related work, and figures, particularly the extensive discussion of existing protein-derived and molecular conformer generation methods.

Response: We thank the reviewer for this helpful suggestion. In the revised manuscript, we condensed Section 2 to four short paragraphs (p. 2) and moved the detailed discussion of protein-derived and small-molecule conformer generators to Appendix G (pp. 39–40). The recovered space was used for the new generalization experiments on 7-mers and NCAA-containing 6-mers and for the physical-validity and efficiency analyses, so the main text is now weighted toward results rather than background.

Reviewer_du8vOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CycPepFlow, an all-atom Cartesian flow-matching framework for cyclic-peptide conformational ensemble generation. The method integrates SO(3)-equivariant local interactions with gated global attention to jointly capture local geometric constraints and long-range dependencies in cyclic structures. It further incorporates explicit chirality conditioning and signed-volume supervision to improve stereochemical fidelity, together with an optional all-pair graph-geodesic (APG) attention bias for encoding molecular topology. The method is evaluated on 1,000 held-out cyclic peptides from the CREMP benchmark, covering peptides with four to six residues, using bidirectional ensemble coverage and matching metrics as well as stereochemical quality measures. Experimental results show that CycPepFlow-L achieves the best overall COV-F1 among the compared methods, while APG-L provides stronger generated-set precision and particularly competitive performance on six-residue cyclic peptides.

Critical strengths
  1. The paper addresses an important and technically challenging problem. Cyclic peptide conformer generation requires jointly modeling long-range dependencies induced by the cyclic topology and local stereochemical constraints. The proposed combination of local equivariant geometric modeling, global information exchange, and explicit chirality supervision is well motivated for this setting.

  2. The experimental evaluation is relatively comprehensive. The authors evaluate both reference-set coverage and generated-set precision rather than relying on a single directional metric. On the full CREMP benchmark, CycPepFlow-L achieves the best COV-F1 (76.13%), while APG-L obtains the best COV-P (79.78%) and MAT-P (0.419 Å). The strong stereochemical accuracy observed on six-residue cyclic peptides further supports the effectiveness of the proposed approach.

Critical weaknesses
  1. The claimed generality is currently limited by the evaluation regime. The benchmark contains only four- to six-residue cyclic peptides, with only 120 six-mers. Therefore, the evidence does not yet establish that the global-attention/APG design will scale to substantially larger macrocycles. This is particularly relevant because global attention is quadratic in atom count and APG preprocessing is reported as cubic. The paper itself acknowledges that evaluation beyond six residues and broader noncanonical chemistry is needed.

  2. Some of the performance gain is difficult to attribute cleanly to the central methodological ideas. CycPepFlow combines several ingredients—local equivariant interactions, global attention, AttnRes, gated attention, Conditional ProductLite, SOAP optimization, chirality supervision, and APG. Although component ablations are provided, the resulting system is sufficiently composite that the source of the overall benchmark advantage is less clear than desirable. In addition, comparator training protocols are model-specific rather than strictly compute matched. The paper explicitly states that model families retain their own training procedures and budgets.

Questions for authors
  1. Can the authors provide results on cyclic peptides larger than six residues, even on a smaller controlled subset? Reporting generation time and peak memory across molecular sizes, including the overhead of APG, would also help assess the practical scalability of the approach.

  2. Does the advantage over the strongest retrained DiTMC baseline persist under approximately matched training compute and across multiple random seeds? Please report parameter counts and training costs for this comparison. A separately tuned AdamW control would also help clarify the contribution of SOAP relative to the architectural components.

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer du8v

Weakness 1.

The claimed generality is currently limited by the evaluation regime. The benchmark contains only four- to six-residue cyclic peptides, with only 120 six-mers. Therefore, the evidence does not yet establish that the global-attention/APG design will scale to substantially larger macrocycles. This is particularly relevant because global attention is quadratic in atom count and APG preprocessing is reported as cubic. The paper itself acknowledges that evaluation beyond six residues and broader noncanonical chemistry is needed.

Response: We agree that 4- to 6-residue rings alone do not establish broad generality, so we evaluated pretrained models without retraining on two additional panels (Section 5.3, Figure 5): on 100 7-mers with seen residues, CycPepFlow APG-L achieves 56.13% COV-F1 vs. 40.09% for the strongest baseline, DiTMC rPE-L; on 50 6-mers spanning 10 noncanonical residue types, CycPepFlow-L achieves 34.38% vs. 18.93% for LoQI. Peer baselines (DiTMC, LoQI, ETFlow) also incur quadratic complexity, whereas sub-quadratic MCF sacrifices substantial accuracy for throughput (Table 12). While quadratic complexity may constrain scaling to much larger macrocycles—an avenue for future work—it remains practical for the evaluated cyclic peptide sizes. The generation-time benchmark across molecular sizes (Table 14), showing that generation time increases modestly with the evaluated molecular size. The cubic APG shortest-path preprocessing executes once per fixed graph (independent of conformer or sampling step count), requiring just 31.4s on one CPU thread for all 31,678 training graphs.

Weakness 2.

Some of the performance gain is difficult to attribute cleanly to the central methodological ideas. CycPepFlow combines several ingredients—local equivariant interactions, global attention, AttnRes, gated attention, Conditional ProductLite, SOAP optimization, chirality supervision, and APG. Although component ablations are provided, the resulting system is sufficiently composite that the source of the overall benchmark advantage is less clear than desirable. In addition, comparator training protocols are model-specific rather than strictly compute matched. The paper explicitly states that model families retain their own training procedures and budgets.

Response: We thank the reviewer for raising both component attribution and baseline comparison protocols. On 6-mers—the most challenging macrocycle regime in Figure 10—full CycPepFlow-B reaches 57.94% COV-F1 compared to 56.48% without Attention Residuals (the strongest single ablation), while removing five components and switching from SOAP to AdamW drops performance to 33.44% (a 24.50 percentage-point gap), demonstrating substantial gains from the combined design. We have highlighted these 6-mer ablations in the caption of Figure 10. Regarding comparator training, we clarify that all models were trained for 200 epochs, and we preserved each comparator's original architecture and training protocol rather than imposing a uniform recipe that could mis-tune or alter external baselines; we have revised the text to state this accurately.

Question 1.

Can the authors provide results on cyclic peptides larger than six residues, even on a smaller controlled subset? Reporting generation time and peak memory across molecular sizes, including the overhead of APG, would also help assess the practical scalability of the approach.

Response: We added a generation time and peak memory benchmark across molecular sizes in Table 14. Both runtime and peak memory increase modestly over the evaluated size range, and APG adds minimal overhead. Unlike the throughput experiment in Table 12, this test generates one conformer per molecule at batch_size = 1; scaling to batch_size = 192 in practice substantially improves throughput while keeping APG overhead negligible.

Question 2.

Does the advantage over the strongest retrained DiTMC baseline persist under approximately matched training compute and across multiple random seeds? Please report parameter counts and training costs for this comparison. A separately tuned AdamW control would also help clarify the contribution of SOAP relative to the architectural components.

Response: We added the retrained DiTMC-L results in Table 13 and parameter counts in Table 12. With fewer parameters and lower training compute, CycPepFlow-B (18.30M; 61.09 GPU-hours) achieves 76.05% COV-F1, consistently outperforming the retrained DiTMC rPE-L (74.35%; 28.29M; 68.21 GPU-hours) and aPE-L (75.39%; 28.29M; 68.80 GPU-hours). The AdamW ablation in Figure 10 shows that SOAP improves COV-F1 by 2.06 percentage points, demenstrating its contribution beyond the architecture.

Reviewer_Br7gOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CycPepFlow generates 3D conformations of a given cyclic peptide, conditioned on its atoms, bonds, and stereochemistry. Training combines velocity regression with a signed-volume penalty for chirality. The method aims to capture coupled ring motions while preserving stereochemistry, without post-generation geometry relaxation.

Critical strengths
  1. The model combines nearby atom interactions with communication across the whole ring. This suits cyclic peptides, where moving one part can affect distant atoms and the ring closure.

  2. The paper retrains strong comparison methods and compares full and simplified models using the same optimizer. These checks support the architecture's contribution without geometry cleanup after generation.

Critical weaknesses
  1. The chirality loss can keep changing bond geometry after the handedness is correct. The paper lacks separate checks of bond angles, clashes, and strain to show that the structures remain physically reasonable.

  2. The input chirality labels depend on how neighboring atoms are ordered. Training handles this consistently, but the paper does not test whether different descriptions of the same molecule produce equally reliable stereochemistry.

  3. The method measures velocity error without squaring it, unlike standard flow matching. No comparison with squared error shows whether this change improves the range or quality of generated conformers.

Questions for authors
  1. How do the generated structures compare with other methods on ring closure, bond angles, strain, and handedness?

  2. Does adding the chirality loss improve handedness without worsening bond angles or other local geometry?

  3. Does the model preserve stereochemistry equally well when the same molecule is written with a different atom ordering?

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer Br7g

Weakness 1.

The chirality loss can keep changing bond geometry after the handedness is correct. The paper lacks separate checks of bond angles, clashes, and strain to show that the structures remain physically reasonable.

Response: We thank the reviewer for this helpful suggestion. We added PoseBusters physical-validity checks to Table 12 in Appendix F, which show that bond-length, bond-angle, and internal-clash failure rates decrease when chirality loss is included (0.0185% / 0.00342% / 0.147% vs. 0.0299% / 0.00541% / 0.164% without it). This suggests that the loss preserves—and modestly improves—local structural validity.

Weakness 2.

The input chirality labels depend on how neighboring atoms are ordered. Training handles this consistently, but the paper does not test whether different descriptions of the same molecule produce equally reliable stereochemistry.

Response: Thank you for pointing out this nuance. We canonicalize input SMILES prior to graph construction at inference, which ensures consistent stereochemical assignments across equivalent molecular descriptions. We removed two confusing statements from SI and clarified this procedure in Appendix A.3 (p. 19).

Weakness 3.

The method measures velocity error without squaring it, unlike standard flow matching. No comparison with squared error shows whether this change improves the range or quality of generated conformers.

Response: We thank the reviewer for suggesting a direct comparison. We added a squared velocity loss ablation for CycPepFlow-B in Figure 10, keeping the architecture and training hyperparameters fixed. On the 1,000-peptide test set, the squared loss reduced COV-F1 from 76.05% to 75.08% and conformer-level stereochemical accuracy from 99.63% to 99.03%. These results support retaining the unsquared loss under the tested training configuration.

Question 1.

How do the generated structures compare with other methods on ring closure, bond angles, strain, and handedness?

Response: We added PoseBusters physical-validity checks to Table 12 in Appendix F. All four CycPepFlow variants achieve lower bond-angle failure rates (0.00325–0.00421%) than competing methods alongside minimal bond-length failures (below 0.034%), implying intact ring closure, while low internal-clash rates (below 0.15%) indicate minimal steric strain. Furthermore, CycPepFlow leads in handedness (Figure 4), reaching 99.71% on 6-mers (APG-L) vs. 93.14% for the strongest comparator, DiTMC aPE-L.

Question 2.

Does adding the chirality loss improve handedness without worsening bond angles or other local geometry?

Response: We added PoseBusters physical-validity checks to Table 12 in Appendix F, which show that bond-length, bond-angle, and internal-clash failure rates decrease with chirality supervision (0.0185% / 0.00342% / 0.147% vs. 0.0299% / 0.00541% / 0.164% without the loss). This suggests that chirality supervision improves local geometry instead of worsening it.

Question 3.

Does the model preserve stereochemistry equally well when the same molecule is written with a different atom ordering?

Response: We canonicalize input SMILES prior to graph construction at inference which ensures identical stereochemistry regardless of atom ordering.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.