acceptodds
Under review as a conference paper at ICLR 2027

CycloFlow: Internal Coordinate Flow Matching and Geometric Ring Closure for Cyclic Peptide Conformer Generation

Abstract

A cyclic peptide is a short peptide whose backbone is closed into a ring. Cyclic peptides can engage broad protein interfaces while remaining cell membrane permeable. Both properties depend on a conformational ensemble rather than on a single structure, so the conformational ensemble is what has to be predicted. A generative model of that ensemble has to choose the coordinates it works in. Internal coordinates suit a cyclic peptide. Ring closure ties the torsion angles to one another, and the representation encodes that dependency for free where a Cartesian one has to learn it. We present CycloFlow, which transports a bond length, a bond angle and a torsion angle at every heavy atom by flow matching, each from a prior chosen for that internal coordinate, and closes the ring by solving six of these internal coordinates. On CREMP tetra-, penta- and hexapeptides, CycloFlow leads the strongest baseline by 4.8 and 10.4 points of ring and all-atom coverage, reduces the Wasserstein distance to the reference ensembles, and clashes in only 1.4% of its conformers. It follows the reference distribution of each internal coordinate closely, and places pairs of ring torsion angles, holding more of their coupling. The code is available at https://anonymous.4open.science/r/cycloflow.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
macrocyclic peptides, conformer generation, flow matching, internal coordinates, ring closure, generative models for molecules

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

This paper proposes Cycloflow, a framework for generating cyclic peptide conformers through internal coordinate flow matching, geometric ring closure, and steric refinement, instead of directly generating the conformation in the Cartesian space. Experiments show that Cycloflow achieves the best agreement with reference ensembles compared with other baselines. Besides, it achieves the best internal coordinate distribution and local geometry quality. Ablations further show the importance of steric refinement and Gauss–Newton ring closure for improving the geometry.

Critical strengths
  1. The motivation for the method proposed is well-supported due to the nature of cyclic peptides. Two-stage flow matching training with different objectives and post-refinement of geometry are carefully designed to address the limitations of the cyclic peptide generation problem.
  2. The authors have done extensive experiments to evaluate the generated cyclic peptide ensembles, and ablations show the effectiveness of geometric closure and refinement in improving the properties, which also provides evidence for the method's soundness.
Critical weaknesses
  1. The provided ablations are mostly for the post-refinement stage, but they lack evidence to show the effectiveness of some components in the flow matching stage that the authors claim as important, e.g. the different priors for the bond length, bond angles, and torsion angles, and pure Cartesian space generation versus internal coordinates.
  2. As the work also points out, there are some limitations to the usefulness of this framework. For example, the results in Figure 3c show that Cycloflow has the worst coupling strength with the reference for 4-mers compared with other baselines. Besides, the experiments are on cyclic peptides with 4-6 residues, limiting its wider usage potential.
Questions for authors

Clarification:

  1. I am curious whether the authors have ablation results on the effect of different prior options on model performance. The paper says a mode-specific prior is important but doesn't show enough data on how that changes learning performance.
  2. From table 1, it seems like Gauss–Newton ring closure is one of the most important steps for performance compared with RINGER. Have the authors considered trying RINGER + Gauss–Newton ring closure to see what the performance is like? This can also help isolate the effect of the generative model.
6 comments
AuthorsOpenReview ↗

[Reviewer3 R1]. Torsion prior and Cartesian control (W1, Q1)

Weakness 1. The provided ablations are mostly for the post-refinement stage, but they lack evidence to show the effectiveness of some components in the flow matching stage that the authors claim as important, e.g. the different priors for the bond length, bond angles, and torsion angles, and pure Cartesian space generation versus internal coordinates.

Question 1. I am curious whether the authors have ablation results on the effect of different prior options on model performance. The paper says a mode-specific prior is important but doesn't show enough data on how that changes learning performance.

Review point. The ablations in the paper concern the refinement stage. The reviewer asks for evidence on the flow matching stage, namely the per coordinate priors and generation in Cartesian space versus internal coordinates.


Experimental setup

Torsion prior. CycloFlow gives each coordinate type its own prior. Bond lengths and bond angles start from a Gaussian around the training mean. Torsion angles, which are periodic and multimodal, start from a uniform prior on [−π, π) with a path that wraps at ±π. We retrained with the torsion prior replaced by a Gaussian, N(0, π²/3), which has the same variance as the uniform prior and a straight path, as for the other coordinates. Everything else is unchanged.

Cartesian control. We also trained a control that differs from CycloFlow only in the space in which the flow is defined. Its state is heavy atom coordinates, its prior is an isotropic Gaussian and its velocity is read out equivariantly. The trunk, features, training setup, sampling budget and steric refinement are shared, and no closure step is needed. Its read out is symmetric under reflection, so about half of its conformers are mirror images. We therefore also report it with the post hoc parity switch of ET-Flow, which reflects a conformer whose stereocenters disagree with the input. We also fine tuned it on the CREMP-CycPeptMPDB random split in the same way as CycloFlow (Response 3).


Results

Table 1. Flow matching ablations on the 1,000 CREMP test molecules. Mean ± sd over 3 seeds.

MethodCOV-F1 ↑MAT (Å) ↓W1 (Å) ↓Clash ↓
RingAll-atomRingAll-atomRingAll-atom
Gaussian torsion prior64.5 ± 1.071.8 ± 0.40.182 ± 0.0030.622 ± 0.0090.240 ± 0.0060.961 ± 0.0241.99 ± 0.25
Cartesian control48.5 ± 0.843.5 ± 0.70.355 ± 0.0021.057 ± 0.0070.702 ± 0.0041.799 ± 0.0190.21 ± 0.11
Cartesian control, parity switch72.5 ± 1.169.7 ± 1.00.164 ± 0.0040.599 ± 0.0130.252 ± 0.0111.017 ± 0.0190.21 ± 0.11
CycloFlow72.4 ± 1.375.2 ± 0.30.164 ± 0.0050.555 ± 0.0070.198 ± 0.0120.841 ± 0.0061.35 ± 0.39

Table 2. COV-F1 on the CREMP-CycPeptMPDB 7-mers, random split (n = 179), by ring size. Both models are fine tuned on its training set. Mean ± sd over 3 seeds.

MethodAlpha onlyLarger rings
21 atoms (n = 121)22, 23 and 26 atoms (n = 58)
RingAll-atomRingAll-atom
Cartesian control9.9 ± 10.927.1 ± 6.94.3 ± 6.513.4 ± 9.1
Cartesian control, parity switch17.1 ± 17.944.7 ± 9.77.7 ± 11.423.5 ± 14.6
CycloFlow31.1 ± 3.758.8 ± 1.310.5 ± 1.735.2 ± 3.0

Torsion prior. The mode specific prior matters. With the Gaussian prior, ring COV-F1 falls from 72.4 to 64.5 and all atom COV-F1 from 75.2 to 71.8, and every other metric degrades as well. A straight path cannot cross the periodic boundary at ±π, so the flow places torsion angles in the wrong well more often. The drop is larger on larger rings. Each coordinate type therefore needs a prior that matches its geometry.

Cartesian control. Without the parity switch the Cartesian control has a ring COV-F1 of only 48.5, because half of its conformers are mirror images. Even with the parity switch it is worse than CycloFlow on every all atom metric and on ring W1. Only the ring COV-F1 and ring MAT are similar on CREMP, 72.5 against 72.4 and 0.164 for both. On the larger 7-mers of CREMP-CycPeptMPDB the difference becomes larger (Table 2). With the parity switch the Cartesian control has a ring COV-F1 of 17.1 against 31.1 for CycloFlow on the alpha only rings and 7.7 against 10.5 on the larger rings of 22, 23 and 26 atoms, and its standard deviation across seeds is larger. The advantage of internal coordinates grows with the number of degrees of freedom.

AuthorsOpenReview ↗

[Reviewer3 R2]. Coupling strength in Figure 3c (W2)

Weakness 2. As the work also points out, there are some limitations to the usefulness of this framework. For example, the results in Figure 3c show that Cycloflow has the worst coupling strength with the reference for 4-mers compared with other baselines. Besides, the experiments are on cyclic peptides with 4-6 residues, limiting its wider usage potential.

Review point. Figure 3c shows that CycloFlow has the weakest coupling strength for 4-mers among the compared methods.


Experimental setup

We reran the Figure 3c analysis on all 1,000 test molecules and 3 seeds, and added our stage 1 checkpoint, taken before the fine tuning on heavy atom distances.


Results

Table 3. Coupling strength and combination error, the JS divergence, by ring size on the 1,000 CREMP test molecules. Mean ± sd over 3 seeds.

Method4-mer coupling4-mer JS error ↓5-mer coupling5-mer JS error ↓6-mer coupling6-mer JS error ↓
RINGER0.695 ± 0.0020.099 ± 0.0050.589 ± 0.0210.126 ± 0.0060.492 ± 0.0490.191 ± 0.001
ET-Flow-SO(3)0.948 ± 0.0490.137 ± 0.0030.612 ± 0.0280.187 ± 0.0020.358 ± 0.0060.251 ± 0.002
AvgFlow$_{\mathrm{DiT}}$0.933 ± 0.0460.159 ± 0.0060.639 ± 0.0040.199 ± 0.0040.450 ± 0.0310.220 ± 0.003
CycloFlow, stage 1 only0.759 ± 0.0280.092 ± 0.0010.736 ± 0.0160.119 ± 0.0020.572 ± 0.0710.195 ± 0.011
CycloFlow0.663 ± 0.0300.096 ± 0.0030.651 ± 0.0160.123 ± 0.0030.530 ± 0.0080.195 ± 0.009

Coupling strength is the generated mutual information of a ring torsion pair over the reference's, so it measures how tightly the pair moves together. Combination error is the JS divergence between the generated and reference combination shares, so it measures whether the pair uses the right combinations in the right proportions. CycloFlow has the lowest or near lowest combination error at every ring size, and on 5-mers and 6-mers it also holds more coupling than every baseline. The weak coupling is limited to 4-mers.

The weaker 4-mer coupling comes from the second training stage, the fine tuning on heavy atom distances. Stage 1 alone holds 0.759, and fine tuning lowers it to 0.663. Combination error stays almost the same, from 0.092 to 0.096, and all atom COV-F1 rises from 72.5 to 75.2. Fine tuning loosens how tightly ring torsion pairs move together, but it keeps the combinations the model generates and improves the whole conformer.

Higher coupling alone does not give better conformers. ET-Flow-SO(3) and AvgFlow$_{\mathrm{DiT}}$ hold 0.948 and 0.933 on 4-mers, but they pair the wrong wells tightly. Their 4-mer combination error is 1.4 to 1.7 times ours, and over all 1,000 test molecules their ring COV-F1 is 35.7 and 45.7 against our 72.4. Ring accuracy follows combination error, and CycloFlow has the lowest 4-mer combination error of the compared methods. The low 4-mer coupling in Figure 3c does not lower conformer quality.

AuthorsOpenReview ↗

[Reviewer3 R3]. Longer cyclic peptides (W2)

Weakness 2. As the work also points out, there are some limitations to the usefulness of this framework. For example, the results in Figure 3c show that Cycloflow has the worst coupling strength with the reference for 4-mers compared with other baselines. Besides, the experiments are on cyclic peptides with 4-6 residues, limiting its wider usage potential.

Review point. The experiments cover only cyclic peptides of 4 to 6 residues.


Experimental setup

We evaluated on CREMP-CycPeptMPDB, Zenodo 13334335. It was computed with the same CREST and GFN2-xTB protocol as CREMP and shares no molecule with it. It contains only head to tail cyclic peptides of 6, 7 and 10 residues, so 8 and 9 residue rings cannot be evaluated. Its 6-mers fall in the size range of CREMP, so we use its 1,732 7-mers and 15 10-mers. The 10-mers are too few to evaluate, so we report the 7-mers. The 7-mers differ from CREMP in three ways.

  • Ring size. A 7 residue ring has 21 backbone atoms, against 12 to 18 for the 4 to 6 residue rings of CREMP. Five 7-mers have 24 or 26 ring atoms.
  • Beta amino acids. A beta amino acid contributes four backbone atoms to the ring instead of three. 552 of the 1,732 7-mers (31.9%) contain beta homophenylalanine. Of these 552 7-mers, 490 contain one beta homophenylalanine and have 22 ring atoms, and 62 contain two beta homophenylalanines and have 23 ring atoms.
  • Unseen residues. Of the 30 residue types in the 7-mers, 12 residue types never occur in CREMP. They account for 17.0% of all residue occurrences. 1,327 7-mers (76.6%) contain at least one of them.

We built two train/validation/test splits. The random split (1,366/187/179) is an 80/10/10 shuffle. The group split (1,449/180/103 molecules, 39/12/9 patterns) shares no pattern between the test set and the other two partitions. A pattern is the sequence with stereochemistry and N-methylation removed.


Results

Table 4. CREMP-CycPeptMPDB 7-mers, random split (n = 179). Mean ± sd over 3 seeds. n is the number of molecules in each MAT, and RINGER has no Cα template for Aib and N-alkyl glycines. COV-F1 counts a molecule a method cannot build as zero.

MethodRingAll-atom
nCOV-F1 ↑MAT (Å) ↓nCOV-F1 ↑MAT (Å) ↓
Alpha only, 21 ring atoms
RINGER12124.7 ± 7.00.606 ± 0.0425620.2 ± 3.11.273 ± 0.086
ET-Flow-SO(3)1210.7 ± 0.30.754 ± 0.0051214.4 ± 0.41.721 ± 0.022
AvgFlow$_{\mathrm{DiT}}$1213.6 ± 0.60.466 ± 0.01312136.1 ± 0.91.155 ± 0.029
CycloFlow12131.1 ± 3.70.398 ± 0.01612158.8 ± 1.30.950 ± 0.027
22, 23 and 26 ring atoms
RINGER0––0––
ET-Flow-SO(3)580.1 ± 0.20.929 ± 0.011582.2 ± 0.52.036 ± 0.019
AvgFlow$_{\mathrm{DiT}}$580.5 ± 0.00.645 ± 0.0085814.7 ± 0.61.549 ± 0.013
CycloFlow5810.5 ± 1.70.675 ± 0.0255835.2 ± 3.01.497 ± 0.039

Table 5. CREMP-CycPeptMPDB 7-mers, group split (n = 103). Mean ± sd over 3 seeds. n as in Table 4.

MethodRingAll-atom
nCOV-F1 ↑MAT (Å) ↓nCOV-F1 ↑MAT (Å) ↓
Alpha only, 21 ring atoms
RINGER6624.5 ± 0.50.594 ± 0.01733.2 ± 0.20.947 ± 0.036
ET-Flow-SO(3)660.6 ± 0.10.727 ± 0.010663.6 ± 0.21.773 ± 0.017
AvgFlow$_{\mathrm{DiT}}$661.8 ± 0.40.449 ± 0.0106625.4 ± 2.01.212 ± 0.035
CycloFlow6628.2 ± 2.60.386 ± 0.0096652.4 ± 2.51.028 ± 0.032
With beta, 22 ring atoms
RINGER0––0––
ET-Flow-SO(3)370.1 ± 0.10.834 ± 0.016372.3 ± 0.61.952 ± 0.034
AvgFlow$_{\mathrm{DiT}}$370.3 ± 0.20.590 ± 0.0143714.2 ± 2.41.513 ± 0.043
CycloFlow3716.4 ± 1.20.546 ± 0.0083740.8 ± 3.21.368 ± 0.022

CycloFlow remains the best method on the 7-mers. It has the highest ring and all-atom COV-F1 in every group of both splits and the lowest MAT in almost every column. The difference from the baselines is largest on the rings with a beta residue. RINGER cannot generate conformers for these rings (–), and the other baselines have a low ring COV-F1 on them. CycloFlow has a ring COV-F1 of 10.5 in the random split and 16.4 in the group split. RINGER assigns ring atoms to residues using a three atom backbone and a fixed residue library, so it cannot handle beta residues. CycloFlow works at the atom level and needs neither.

Reviewer_B7SHOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposed CycloFlow, an all-atom flow matching model for cyclic peptide ensemble generation. This method denotes heavy atoms as bond lengths, bond angles and torsional angles, then introduces differential NeRF distance loss in training process to suppress error accumulations. In inference, a closed loop can be achieved by performing GN iterations. The method outperforms RINGER baselines in ensemble and physical metrics in CREMP benchmark.

Critical strengths
  1. Geometric closure is a constraint in model. Combining learning with constraint solving is more applicable than relying solely on the network to learn closure on its own.
  2. Proposed a new SE(3) invariant all atom formulations for cyclic peptide, overcoming complex representation and modeling of cartesian coordinates in Vilya-1 and information loss of internal coordinates in RINGER.
Critical weaknesses
  1. Chirality is input as soft condition but not hard constraint. Chirality preservation or accuracy is not discussed in paper.
  2. Performance plummets with increasing ring size. And experiment truncated the reference ensemble to top conformations may mask model's abilities of exploring rare metastable states.
Questions for authors
  1. What is the accuracy rate of chirality? How would the model generalize on larger rings like MeD cases in CycPeptMPDB?
  2. How much of the advantage comes from the generative model and how much from post processing? How would it will be if we replace post processing method with forcefield like CHARMM or Rosetta? Can this method outperform general end-to-end model like HighFold3 (AF3) along with fast relax?
9 comments
AuthorsOpenReview ↗

[Reviewer2 R1]. Chirality preservation (W1, Q1)

Weakness 1. Chirality is input as soft condition but not hard constraint. Chirality preservation or accuracy is not discussed in paper.

Question 1. What is the accuracy rate of chirality? How would the model generalize on larger rings like MeD cases in CycPeptMPDB?

Review point. Chirality enters the model as a soft condition rather than a hard constraint, and the paper does not report how often it is preserved.


Experimental setup

For every stereocenter we recompute its chirality label, R or S under the Cahn–Ingold–Prelog (CIP) rules, from the generated 3D coordinates with RDKit and compare it with the label of the input molecule. We do this for all 1,000 test molecules and 3 seeds, using the ensembles from the main results table of the paper. Applied to the reference conformers, the same procedure returns 100 percent (%), so detection error does not enter the measurement. A conformer counts as preserved only if every one of its stereocenters matches the input.


Results

Table 1. Chirality preservation on the 1,000 CREMP test molecules, in percent. The 4-mer, 5-mer and 6-mer columns are at the conformer level.

MethodRepresentationConformer level ↑Stereocenter level ↑4-mer ↑5-mer ↑6-mer ↑
ETKDGv3RDKit embedding100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00
+MMFF94RDKit embedding100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00
Torsional Diffusiontorsion angles only100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00
CycloFlowinternal coordinates98.69 ± 0.3199.59 ± 0.1297.71 ± 0.5999.87 ± 0.0399.60 ± 0.13
RINGERinternal coordinates97.99 ± 0.9899.56 ± 0.2098.33 ± 0.8297.52 ± 1.1297.83 ± 1.28
AvgFlow$_{\mathrm{DiT}}$Cartesian68.80 ± 1.2372.94 ± 1.4369.58 ± 2.3569.73 ± 0.9062.64 ± 1.25
ET-Flow-SO(3)Cartesian40.10 ± 0.7677.84 ± 0.3849.31 ± 1.3431.31 ± 0.1124.60 ± 0.24
MCF-SCartesian37.59 ± 0.7576.35 ± 0.3446.41 ± 1.0028.87 ± 0.5323.69 ± 0.32
AGDiffCartesian4.62 ± 0.0949.89 ± 0.066.45 ± 0.142.93 ± 0.091.39 ± 0.03

98.69% of CycloFlow conformers preserve every stereocenter, and 99.59% of individual stereocenters are correct. ETKDGv3, ETKDGv3+MMFF94 and Torsional Diffusion preserve every stereocenter by construction, since they fix the configuration during RDKit embedding and change only torsion angles. Among the methods that generate the geometry itself, CycloFlow has the highest rate, and RINGER is within run-to-run variation of it. CycloFlow stays at 97.7% or higher at every ring size, since each stereocenter is almost always right.

The concern that a soft condition loses chirality holds for Cartesian generation. AGDiff is correct at 49.89% of stereocenters, the rate expected by chance, because a representation built on distances cannot tell a molecule from its mirror image. ET-Flow-SO(3) and MCF-S are correct at 76 to 78% of stereocenters, but their errors are spread across conformers. With an error at about one stereocenter in four, the chance that every stereocenter is right falls as their number grows with ring size, so only about a quarter of their 6-mer conformers are correct. AvgFlow$_{\mathrm{DiT}}$ shows a different failure. Its conformer level and stereocenter level rates are nearly equal, 68.8 and 72.9, which means about 30% of its conformers are whole molecule mirror images rather than conformers with scattered errors.

AuthorsOpenReview ↗

[Reviewer2 R2]. Longer cyclic peptides and the MeD molecule (Q1)

Question 1. What is the accuracy rate of chirality? How would the model generalize on larger rings like MeD cases in CycPeptMPDB?

Review point. Generalization to larger rings such as the MeD cases in CycPeptMPDB.


Experimental setup

We used CREMP-CycPeptMPDB, computed with the CREMP protocol and sharing no molecule with it. meD, N-methyl-L-aspartate, occurs in one molecule, the 6-mer L.dP.meY.L.Me_dL.meD.

For larger rings we fine-tuned on its 7-mers and 10-mers and report the 7-mers, since the 10-mers are too few.

  • Ring size. A 7 residue ring has 21 backbone atoms, against 12 to 18 for the 4 to 6 residue rings of CREMP. Five 7-mers have 24 or 26 ring atoms.
  • Beta amino acids. A beta amino acid contributes four backbone atoms to the ring instead of three. 552 of the 1,732 7-mers (31.9%) contain beta homophenylalanine. Of these 552 7-mers, 490 contain one beta homophenylalanine and have 22 ring atoms, and 62 contain two beta homophenylalanines and have 23 ring atoms.
  • Unseen residues. Of the 30 residue types in the 7-mers, 12 residue types never occur in CREMP. They account for 17.0% of all residue occurrences. 1,327 7-mers (76.6%) contain at least one of them.

We built two train/validation/test splits. The random split (1,366/187/179) is an 80/10/10 shuffle. The group split (1,449/180/103 molecules, 39/12/9 patterns) shares no pattern between the test set and the other two partitions. A pattern is the sequence with stereochemistry and N-methylation removed.


Results

Table 2. CycloFlow inference on the MeD molecule with the CREMP checkpoint, 80 conformers against the top 40. Mean ± sd over 3 seeds.

MoleculeCOV-F1 ↑MAT (Å) ↓W1 (Å) ↓
RingAll-atomRingAll-atomRingAll-atom
L.dP.meY.L.Me_dL.meD87.0 ± 16.197.9 ± 1.60.065 ± 0.0090.339 ± 0.0290.172 ± 0.0160.636 ± 0.032

meD never occurs in CREMP, which has only unmethylated Asp, so the model has seen Asp and N-methylation but never together.


Table 3. CREMP-CycPeptMPDB 7-mers, random split (n = 179). Mean ± sd over 3 seeds. n is the number of molecules in each MAT, and RINGER has no Cα template for Aib and N-alkyl glycines. COV-F1 counts a molecule a method cannot build as zero.

MethodRingAll-atom
nCOV-F1 ↑MAT (Å) ↓nCOV-F1 ↑MAT (Å) ↓
Alpha only, 21 ring atoms
RINGER12124.7 ± 7.00.606 ± 0.0425620.2 ± 3.11.273 ± 0.086
ET-Flow-SO(3)1210.7 ± 0.30.754 ± 0.0051214.4 ± 0.41.721 ± 0.022
AvgFlow$_{\mathrm{DiT}}$1213.6 ± 0.60.466 ± 0.01312136.1 ± 0.91.155 ± 0.029
CycloFlow12131.1 ± 3.70.398 ± 0.01612158.8 ± 1.30.950 ± 0.027
22, 23 and 26 ring atoms
RINGER0––0––
ET-Flow-SO(3)580.1 ± 0.20.929 ± 0.011582.2 ± 0.52.036 ± 0.019
AvgFlow$_{\mathrm{DiT}}$580.5 ± 0.00.645 ± 0.0085814.7 ± 0.61.549 ± 0.013
CycloFlow5810.5 ± 1.70.675 ± 0.0255835.2 ± 3.01.497 ± 0.039

Table 4. CREMP-CycPeptMPDB 7-mers, group split (n = 103). Mean ± sd over 3 seeds. n as in Table 3.

MethodRingAll-atom
nCOV-F1 ↑MAT (Å) ↓nCOV-F1 ↑MAT (Å) ↓
Alpha only, 21 ring atoms
RINGER6624.5 ± 0.50.594 ± 0.01733.2 ± 0.20.947 ± 0.036
ET-Flow-SO(3)660.6 ± 0.10.727 ± 0.010663.6 ± 0.21.773 ± 0.017
AvgFlow$_{\mathrm{DiT}}$661.8 ± 0.40.449 ± 0.0106625.4 ± 2.01.212 ± 0.035
CycloFlow6628.2 ± 2.60.386 ± 0.0096652.4 ± 2.51.028 ± 0.032
With beta, 22 ring atoms
RINGER0––0––
ET-Flow-SO(3)370.1 ± 0.10.834 ± 0.016372.3 ± 0.61.952 ± 0.034
AvgFlow$_{\mathrm{DiT}}$370.3 ± 0.20.590 ± 0.0143714.2 ± 2.41.513 ± 0.043
CycloFlow3716.4 ± 1.20.546 ± 0.0083740.8 ± 3.21.368 ± 0.022

CycloFlow remains the best method on the 7-mers. It has the highest ring and all-atom COV-F1 in every group of both splits and the lowest MAT in almost every column. The difference from the baselines is largest on the rings with a beta residue. RINGER cannot generate conformers for these rings (–), and the other baselines have a low ring COV-F1 on them. CycloFlow has a ring COV-F1 of 10.5 in the random split and 16.4 in the group split. RINGER assigns ring atoms to residues using a three atom backbone and a fixed residue library, so it cannot handle beta residues. CycloFlow works at the atom level and needs neither.

AuthorsOpenReview ↗

[Reviewer2 R3]. COV-P against the full ensemble (W2)

Weakness 2. Performance plummets with increasing ring size. And experiment truncated the reference ensemble to top conformations may mask model's abilities of exploring rare metastable states.

Review point. Performance drops with ring size, and scoring against only the top 40 conformers may hide the model's ability to find rare metastable states.


Experimental setup

We did not retrain any model on the full ensemble. We kept every generated ensemble fixed and replaced only the test reference, from the top 40 conformers by Boltzmann weight to the full CREMP ensemble, which holds tens to 5,000 unique conformers per molecule. A generated conformer that matches no top 40 conformer but does match a conformer in the full ensemble is a real conformer that the top 40 reference counted as a false positive.


Results

Table 5. COV-P in percent (%) against the top 40 and against the full ensemble on the 1,000 CREMP test molecules. Mean over 3 seeds.

MethodRing COV-P ↑All-atom COV-P ↑Recovered among missed ↑
top 40 → fulltop 40 → full
All6-merAll6-merRingAll-atom
ETKDGv30.0 → 0.00.0 → 0.00.4 → 2.60.0 → 0.00.0%2.2%
+MMFF945.5 → 12.30.0 → 0.14.7 → 15.20.1 → 0.97.2%11.1%
Torsional Diffusion0.0 → 0.00.0 → 0.00.2 → 1.00.0 → 0.00.0%0.8%
RINGER57.7 → 65.230.8 → 46.256.2 → 66.033.0 → 49.218.0%22.6%
MCF-S20.4 → 25.80.0 → 0.122.7 → 32.82.6 → 6.66.9%13.1%
ET-Flow-SO(3)25.6 → 35.40.5 → 1.923.1 → 34.71.7 → 5.413.3%15.1%
AGDiff2.7 → 5.00.0 → 0.00.9 → 2.20.0 → 0.02.4%1.2%
AvgFlow$_{\mathrm{DiT}}$35.3 → 47.13.3 → 9.433.3 → 47.47.7 → 18.818.4%21.2%
CycloFlow64.6 → 75.439.4 → 60.070.4 → 81.543.9 → 63.630.7%37.6%

Table 5 rescores the same generated conformers against the full ensemble instead of the top 40, the lowest energy conformers. CycloFlow's ring COV-P then rises from 64.6 to 75.4, and from 39.4 to 60.0 on 6-mers. Of the conformers that the top 40 counts as wrong, 30.7% on the ring and 37.6% on all atoms match a conformer in the full ensemble, the highest of any method. The full ensemble conformers they match are higher in energy than the top 40. Relative to the lowest energy conformer of each molecule, the 40th conformer lies at a median 2.3 kcal/mol, while these matched conformers lie at 4.3 kcal/mol for the ring and 3.3 kcal/mol for all atoms. A large part of what the top 40 reference counts as wrong is CycloFlow sampling real metastable states.

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

This paper proposes CycloFlow, a cyclic peptide conformer generation method based on internal-coordinate Flow Matching, which predicts bond lengths, bond angles, and torsion angles. The method uses a six-coordinate geometric optimization for ring closure, followed by steric refinement to reduce atomic clashes. The model is trained and evaluated on 4–6 residue cyclic peptides from the CREMP dataset. Experimental results show that CycloFlow outperforms RINGER and other baselines in conformer coverage, matching, and Wasserstein distance.

Critical strengths
  1. The paper presents a reasonable framework combining internal-coordinate Flow Matching with geometric ring closure. It generates both backbone and side-chain conformations and uses geometric optimization to close the ring.

  2. Under the given CREMP benchmark setting, the experiments are relatively thorough, including comparisons with multiple baselines, ablation studies, and conformer distribution analysis. The results show clear improvements in conformer coverage and Wasserstein distance.

Critical weaknesses
  1. The experiments are limited to CREMP, which only contains short cyclic peptides with 4–6 residues. The performance already drops significantly on 6-mers. It is unclear whether the method can generalize to longer and more flexible cyclic peptides.

  2. The baseline comparison is not comprehensive enough. Most baselines are conventional small-molecule conformer generation methods or RINGER. The paper does not compare with recent protein structure prediction or all-atom models, such as AF3, Boltz-2, and Vilya-1. This makes it difficult to assess the advantages of CycloFlow over these methods.

  3. The methodological novelty is relatively limited. Internal-coordinate generation, Flow Matching, and geometric ring closure have all been explored in previous work. The paper mainly combines these existing techniques. Although the ablation shows that the proposed closure solver is effective, it is still unclear how much of the improvement comes from the generative model rather than the closure algorithm.

Questions for authors
  1. Can the authors evaluate CycloFlow on longer cyclic peptides, such as 7–10 residues, to demonstrate its generalization ability? Given the significant performance drop on 6-mers, this would help clarify the applicability of the method.

  2. Can the authors include comparisons with recent structure prediction or all-atom models, such as AF3, Boltz-2, and Vilya-1? If these models cannot be directly applied to the current ensemble generation setting, it would be helpful to explain why.

  3. Can the authors further separate the contributions of the generative model and the ring closure solver? For example, applying the proposed six-coordinate closure solver to RINGER would allow a comparison under the same closure method. This would help clarify whether the improvement mainly comes from Flow Matching or the new closure algorithm.

6 comments
AuthorsOpenReview ↗

[Reviewer1 R1]. Longer cyclic peptides (W1, Q1)

Weakness 1. The experiments are limited to CREMP, which only contains short cyclic peptides with 4–6 residues. The performance already drops significantly on 6-mers. It is unclear whether the method can generalize to longer and more flexible cyclic peptides.

Question 1. Can the authors evaluate CycloFlow on longer cyclic peptides, such as 7–10 residues, to demonstrate its generalization ability? Given the significant performance drop on 6-mers, this would help clarify the applicability of the method.

Review point. Generalization to longer cyclic peptides. CREMP covers only 4 to 6 residues.


Experimental setup

We evaluated on CREMP-CycPeptMPDB, Zenodo 13334335. It was computed with the same CREST and GFN2-xTB protocol as CREMP and shares no molecule with it. It contains only head to tail cyclic peptides of 6, 7 and 10 residues, so 8 and 9 residue rings cannot be evaluated. Its 6-mers fall in the size range of CREMP, so we use its 1,732 7-mers and 15 10-mers. The 10-mers are too few to evaluate, so we report the 7-mers. The 7-mers differ from CREMP in three ways.

  • Ring size. A 7 residue ring has 21 backbone atoms, against 12 to 18 for the 4 to 6 residue rings of CREMP. Five 7-mers have 24 or 26 ring atoms.
  • Beta amino acids. A beta amino acid contributes four backbone atoms to the ring instead of three. 552 of the 1,732 7-mers (31.9%) contain beta homophenylalanine. Of these 552 7-mers, 490 contain one beta homophenylalanine and have 22 ring atoms, and 62 contain two beta homophenylalanines and have 23 ring atoms.
  • Unseen residues. Of the 30 residue types in the 7-mers, 12 residue types never occur in CREMP. They account for 17.0% of all residue occurrences. 1,327 7-mers (76.6%) contain at least one of them.

We built two train/validation/test splits. The random split (1,366/187/179) is an 80/10/10 shuffle. The group split (1,449/180/103 molecules, 39/12/9 patterns) shares no pattern between the test set and the other two partitions. A pattern is the sequence with stereochemistry and N-methylation removed.


Results

Table 1. CREMP-CycPeptMPDB 7-mers, random split (n = 179). Mean ± sd over 3 seeds. n is the number of molecules in each MAT, and RINGER has no Cα template for Aib and N-alkyl glycines. COV-F1 counts a molecule a method cannot build as zero.

MethodRingAll-atom
nCOV-F1 ↑MAT (Å) ↓nCOV-F1 ↑MAT (Å) ↓
Alpha only, 21 ring atoms
RINGER12124.7 ± 7.00.606 ± 0.0425620.2 ± 3.11.273 ± 0.086
ET-Flow-SO(3)1210.7 ± 0.30.754 ± 0.0051214.4 ± 0.41.721 ± 0.022
AvgFlow$_{\mathrm{DiT}}$1213.6 ± 0.60.466 ± 0.01312136.1 ± 0.91.155 ± 0.029
CycloFlow12131.1 ± 3.70.398 ± 0.01612158.8 ± 1.30.950 ± 0.027
22, 23 and 26 ring atoms
RINGER0––0––
ET-Flow-SO(3)580.1 ± 0.20.929 ± 0.011582.2 ± 0.52.036 ± 0.019
AvgFlow$_{\mathrm{DiT}}$580.5 ± 0.00.645 ± 0.0085814.7 ± 0.61.549 ± 0.013
CycloFlow5810.5 ± 1.70.675 ± 0.0255835.2 ± 3.01.497 ± 0.039

Table 2. CREMP-CycPeptMPDB 7-mers, group split (n = 103). Mean ± sd over 3 seeds. n as in Table 1.

MethodRingAll-atom
nCOV-F1 ↑MAT (Å) ↓nCOV-F1 ↑MAT (Å) ↓
Alpha only, 21 ring atoms
RINGER6624.5 ± 0.50.594 ± 0.01733.2 ± 0.20.947 ± 0.036
ET-Flow-SO(3)660.6 ± 0.10.727 ± 0.010663.6 ± 0.21.773 ± 0.017
AvgFlow$_{\mathrm{DiT}}$661.8 ± 0.40.449 ± 0.0106625.4 ± 2.01.212 ± 0.035
CycloFlow6628.2 ± 2.60.386 ± 0.0096652.4 ± 2.51.028 ± 0.032
With beta, 22 ring atoms
RINGER0––0––
ET-Flow-SO(3)370.1 ± 0.10.834 ± 0.016372.3 ± 0.61.952 ± 0.034
AvgFlow$_{\mathrm{DiT}}$370.3 ± 0.20.590 ± 0.0143714.2 ± 2.41.513 ± 0.043
CycloFlow3716.4 ± 1.20.546 ± 0.0083740.8 ± 3.21.368 ± 0.022

CycloFlow remains the best method on the 7-mers. It has the highest ring and all-atom COV-F1 in every group of both splits and the lowest MAT in almost every column. The difference from the baselines is largest on the rings with a beta residue. RINGER cannot generate conformers for these rings (–), and the other baselines have a low ring COV-F1 on them. CycloFlow has a ring COV-F1 of 10.5 in the random split and 16.4 in the group split. RINGER assigns ring atoms to residues using a three atom backbone and a fixed residue library, so it cannot handle beta residues. CycloFlow works at the atom level and needs neither.

AuthorsOpenReview ↗

[Reviewer1 R2]. Structure prediction models (W2, Q2)

Weakness 2. The baseline comparison is not comprehensive enough. Most baselines are conventional small-molecule conformer generation methods or RINGER. The paper does not compare with recent protein structure prediction or all-atom models, such as AF3, Boltz-2, and Vilya-1. This makes it difficult to assess the advantages of CycloFlow over these methods.

Question 2. Can the authors include comparisons with recent structure prediction or all-atom models, such as AF3, Boltz-2, and Vilya-1? If these models cannot be directly applied to the current ensemble generation setting, it would be helpful to explain why.

Review point. Comparison with recent structure prediction and all atom models, namely AF3, Boltz-2 and Vilya-1.


Experimental setup

Vilya-1 is not included because no implementation or weights are public.

AF3 has no cyclic peptide support. We give it the sequence with N-methyl residues as CCD modifications, and it is not told that the peptide is cyclic. Boltz-2 is given the cyclic flag and a bond entry on every peptide bond, with its steering potentials on. HighFold4, the successor of HighFold3, changes only the encoding of AF3. At its own defaults it does not close the ring, so we run it with the cyclic position offset on and polymer bonds off.

We run each model three times, as for every method. In each run we collect twice as many structures as reference conformers, up to 80, from several model seeds and several diffusion samples per seed. Before relaxation we apply the same validity filter to all three models. A structure passes only if every stereocenter has the input chirality, no N-methyl carbon has collapsed, its heavy atom connectivity matches the molecular graph and every 1-3 heavy atom pair is at least 2.0 Å apart, since relaxing another stereoisomer or an open ring does not relax the requested molecule. Per seed, 1.8, 53.6 and 54.2 percent (%) of the AF3, Boltz-2 and HighFold4 structures pass. We then relax them with GFN2-xTB, the method that produced the references, and remove duplicates with CREGEN, the most favorable setting for these models.


Results

Table 3. Structure predictors on the 1,000 CREMP test molecules. COV-F1 counts a molecule with no remaining structure as zero. MAT and clash are averaged over the n covered molecules and W1 is their median.

MethodCOV-F1 ↑MAT (Å) ↓W1 (Å) ↓Clash ↓n covered
RingAll-atomRingAll-atomRingAll-atom
AF30.08 ± 0.012.86 ± 0.040.602 ± 0.0011.790 ± 0.0030.835 ± 0.0012.267 ± 0.0090.07 ± 0.001000
+ validity filter0.09 ± 0.010.60 ± 0.050.566 ± 0.0141.777 ± 0.0150.637 ± 0.0232.085 ± 0.0130.00 ± 0.00170.7 ± 10.0
+ GFN2-xTB + CREGEN4.22 ± 0.512.04 ± 0.240.492 ± 0.0291.632 ± 0.0230.580 ± 0.0381.989 ± 0.0180.00 ± 0.00169.0 ± 10.5
Boltz-20.00 ± 0.000.86 ± 0.040.638 ± 0.0021.690 ± 0.0020.794 ± 0.0022.104 ± 0.00429.44 ± 0.041000
+ validity filter0.00 ± 0.000.90 ± 0.040.653 ± 0.0021.733 ± 0.0010.793 ± 0.0022.109 ± 0.0035.36 ± 0.13977.3 ± 2.5
+ GFN2-xTB + CREGEN16.19 ± 0.459.27 ± 0.130.523 ± 0.0021.554 ± 0.0030.706 ± 0.0041.983 ± 0.0020.00 ± 0.00965.3 ± 2.5
HighFold40.01 ± 0.011.41 ± 0.010.637 ± 0.0011.714 ± 0.0010.797 ± 0.0022.114 ± 0.00614.04 ± 0.151000
+ validity filter0.01 ± 0.011.38 ± 0.050.637 ± 0.0001.722 ± 0.0020.796 ± 0.0012.092 ± 0.0031.80 ± 0.17856.3 ± 4.7
+ GFN2-xTB + CREGEN15.23 ± 0.278.33 ± 0.190.516 ± 0.0031.560 ± 0.0040.713 ± 0.0011.990 ± 0.0220.00 ± 0.00840.3 ± 2.5
CycloFlow72.4 ± 1.375.2 ± 0.30.164 ± 0.0050.555 ± 0.0070.198 ± 0.0120.841 ± 0.0061.35 ± 0.391000
+ validity filter73.6 ± 0.975.9 ± 0.20.158 ± 0.0030.545 ± 0.0050.183 ± 0.0090.812 ± 0.0090.01 ± 0.011000

All three models have a ring COV-F1 near zero because their rings differ from every reference conformer by more than the 0.1 Å threshold. Their ring MAT is 0.60 to 0.64 Å, six times the threshold and nearly four times that of CycloFlow, so more samples would not raise ring COV-F1. Relaxation with GFN2-xTB and CREGEN lowers ring MAT only to 0.49 to 0.52 Å and raises ring COV-F1 to at most 16.2, still far below the 72.4 of CycloFlow. The main training target of these models is a single structure rather than an ensemble, so a comparison in the main table would not be fair. We will add these results to the appendix for reference. Under the same validity filter CycloFlow keeps 95.7 ± 1.0% of its conformers and all 1,000 molecules, and its ring COV-F1 rises from 72.4 to 73.6 and its all atom COV-F1 from 75.2 to 75.9. We report CycloFlow without the filter, the conservative choice.

AuthorsOpenReview ↗

[Reviewer1 R3]. Generative model versus closure solver (W3, Q3)

Weakness 3. The methodological novelty is relatively limited. Internal-coordinate generation, Flow Matching, and geometric ring closure have all been explored in previous work. The paper mainly combines these existing techniques. Although the ablation shows that the proposed closure solver is effective, it is still unclear how much of the improvement comes from the generative model rather than the closure algorithm.

Question 3. Can the authors further separate the contributions of the generative model and the ring closure solver? For example, applying the proposed six-coordinate closure solver to RINGER would allow a comparison under the same closure method. This would help clarify whether the improvement mainly comes from Flow Matching or the new closure algorithm.

Review point. Separating the contribution of the generative model from that of the closure solver.


Experimental setup

We applied our six coordinate closure solver, Gauss Newton (ours), to RINGER's generated samples and changed nothing else. For CycloFlow we swap the closure solver in the same way, removing it or replacing it with SLSQP (RINGER).


Results

Table 4. Ring metrics on the seam and the Non-Seam atoms. Both are superposed on the Non-Seam atoms.

ModelClosure solverRing
SeamNon-Seam
COV-F1 ↑MAT (Å) ↓W1 (Å) ↓COV-F1 ↑MAT (Å) ↓W1 (Å) ↓
RINGERnone43.8 ± 20.00.399 ± 0.0570.627 ± 0.10380.6 ± 0.40.107 ± 0.0030.146 ± 0.005
SLSQP61.2 ± 1.00.542 ± 0.0040.888 ± 0.01271.2 ± 0.70.187 ± 0.0010.301 ± 0.010
Gauss Newton65.3 ± 2.50.313 ± 0.0200.515 ± 0.02480.6 ± 0.40.107 ± 0.0030.147 ± 0.005
CycloFlownone13.0 ± 0.30.443 ± 0.0140.652 ± 0.03782.6 ± 0.30.101 ± 0.0010.109 ± 0.002
SLSQP24.3 ± 0.20.680 ± 0.0111.279 ± 0.02866.3 ± 0.60.222 ± 0.0030.388 ± 0.005
Gauss Newton50.8 ± 1.80.298 ± 0.0100.403 ± 0.02082.0 ± 0.40.104 ± 0.0020.115 ± 0.004

1. Ring, generator and closure solver. Table 4 splits the ring by our Gauss Newton closure solver. The seam is the ring atoms it moves to close the ring, and the Non-Seam atoms are the rest, which it never moves. Both generators are scored on the same atoms. On the Non-Seam atoms CycloFlow's generator is better than RINGER's even without a closure solver, with ring COV-F1 82.6 against 80.6 and ring W1 0.109 against 0.146. The closure solver then closes the seam. With the default closure solvers, Gauss Newton for CycloFlow and SLSQP for RINGER, CycloFlow has a ring COV-F1 of 72.4 against 67.6 (Table 5).


Table 5. Closure solver swapped between RINGER and CycloFlow on the 1,000 CREMP test molecules, without steric refinement. Mean ± sd over 3 seeds.

ModelClosure solverRingAll-atomClash ↓
COV-F1 ↑MAT (Å) ↓W1 (Å) ↓COV-F1 ↑MAT (Å) ↓W1 (Å) ↓
RINGERnone71.3 ± 3.40.198 ± 0.0220.286 ± 0.04167.3 ± 0.80.691 ± 0.0171.057 ± 0.03014.38 ± 2.21
SLSQP67.6 ± 1.00.293 ± 0.0030.458 ± 0.02164.8 ± 0.20.810 ± 0.0051.265 ± 0.00921.28 ± 1.37
Gauss Newton74.8 ± 1.20.168 ± 0.0080.250 ± 0.01767.6 ± 0.40.678 ± 0.0091.041 ± 0.01915.01 ± 1.90
CycloFlownone34.8 ± 0.50.230 ± 0.0060.310 ± 0.01974.0 ± 0.40.604 ± 0.0080.891 ± 0.0129.79 ± 0.64
SLSQP47.4 ± 0.50.367 ± 0.0050.649 ± 0.01265.3 ± 0.50.822 ± 0.0101.343 ± 0.01925.64 ± 0.99
Gauss Newton72.4 ± 1.30.165 ± 0.0050.200 ± 0.01375.1 ± 0.30.556 ± 0.0080.845 ± 0.0074.56 ± 0.79

2. All atom. With the same closure solver the two generators score similarly on the ring, but CycloFlow is better on every all atom metric. With Gauss Newton its all atom COV-F1 is 75.1 against 67.6 for RINGER and its all atom MAT 0.556 against 0.678, and it also leads without a closure solver. The all atom metrics add the side chains, so CycloFlow also places the side chains more accurately.


CycloFlow's design has two parts, generation and closure, and Tables 4 and 5 separate their contributions. The generator is better than RINGER's on the Non-Seam atoms (Table 4) and on every all atom metric (Table 5). The closure solver then sets the six internal coordinates that closure determines. Gauss Newton does this without moving the rest of the ring, while SLSQP moves every ring atom and lowers the Non-Seam scores of both generators. CycloFlow's performance comes from combining the two, flow matching with separate priors for bond lengths, bond angles and torsion angles, followed by our Gauss Newton closure solver.

Comments on the paper

2 comments
AuthorsOpenReview ↗

[All reviewers]. Summary of new experiments and analyses

We thank all reviewers for their careful and constructive reviews. Their comments raised four main questions. Does CycloFlow generalize beyond its training data? How does it compare with structure prediction models? Which of its components drive its performance? Does it preserve chirality? To answer them we ran new experiments and reanalyzed our existing results. We tested longer cyclic peptides from CREMP-CycPeptMPDB and compared CycloFlow with AF3, Boltz-2 and HighFold4. We swapped the closure solver between RINGER and CycloFlow, ablated the mode specific prior and the coordinate space, and replaced our post processing with general force fields. We also measured chirality, COV-P against the full ensemble and coupling strength on the existing ensembles. We incorporated most of the results below into the revised paper. The MeD molecule is a single case and is left out, and Figure 3c needed no new experiment.

R = Response. Each location is given as the reviewer and the response number.


New results

1. New experiments

1.1 Ablation

#What is addedLocationRevised paper
AB1Closure solver swap (SLSQP and Gauss Newton)Reviewer1 R3 · Reviewer3 R4Table 1, Appendix U
AB2Gaussian vs uniform torsion priorReviewer3 R1Table 1, Appendix C.4
AB3Cartesian vs internal coordinatesReviewer3 R1Table 1, Appendix C.4
AB4Post processing with general force fieldsReviewer2 R4Appendix P

1.2 Main dataset (CREMP)

#What is addedLocationRevised paper
CR1AF3, Boltz-2 and HighFold4 on the CREMP test setReviewer1 R2 · Reviewer2 R5Appendix O

1.3 Additional dataset (CREMP-CycPeptMPDB)

#What is addedLocationRevised paper
CP1Longer cyclic peptides (CREMP-CycPeptMPDB 7-mers)Reviewer1 R1 · Reviewer2 R2 · Reviewer3 R3Section 4.3, Table 2, Appendix Q
CP2Zero shot evaluation of the MeD moleculeReviewer2 R2Not included (single case)

2. Analysis of existing results

#What is addedLocationRevised paper
A1Chirality preservationReviewer2 R1Appendix M
A2COV-P against the full ensembleReviewer2 R3Appendix N
A3Figure 3c coupling analysis on all 1,000 moleculesReviewer3 R2No new experiment

Responses by reviewer

[Reviewer1]

ResponseTitleReview item
R1Longer cyclic peptidesW1, Q1
R2Structure prediction modelsW2, Q2
R3Generative model versus closure solverW3, Q3

[Reviewer2]

ResponseTitleReview item
R1Chirality preservationW1, Q1 (chirality)
R2Longer cyclic peptides and the MeD moleculeQ1 (larger rings, MeD)
R3COV-P against the full ensembleW2
R4Post processing and force fieldsQ2
R5Structure prediction modelsQ2

[Reviewer3]

ResponseTitleReview item
R1Torsion prior and Cartesian controlW1, Q1
R2Coupling strength in Figure 3cW2 (Figure 3c)
R3Longer cyclic peptidesW2 (ring size)
R4Generative model versus closure solverQ2
AuthorsOpenReview ↗

[All reviewers]. Revised paper uploaded

We have uploaded a revised version of the paper. Every new or changed passage, table and caption is in blue. The revision incorporates the experiments and analyses from our responses. We are grateful to the reviewers, whose feedback helped us strengthen our contributions and gave us the opportunity to add experiments from several different angles.


#What is addedRevised paper
AB1Closure solver swap (SLSQP and Gauss Newton)Table 1, Appendix U
AB2Gaussian vs uniform torsion priorTable 1, Appendix C.4
AB3Cartesian vs internal coordinates, with and without chirality correctionTable 1, Appendix C.4
AB4Post processing with general force fieldsAppendix P
CR1AF3, Boltz-2 and HighFold4 on the CREMP test setAppendix O
CP1Longer cyclic peptides (CREMP-CycPeptMPDB 7-mers), zero-shot and fine-tunedSection 4.3, Table 2, Appendix Q
A1Chirality preservationAppendix M
A2COV-P against the full ensembleAppendix N

Adding the longer cyclic peptides as Section 4.3 moves the former Section 4.3 to Section 5.1.


We sincerely thank the reviewers once again for their thoughtful and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.