- 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
- 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.
- 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
- 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.
- 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:
- 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.
- 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
[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.
| Method | 4-mer coupling | 4-mer JS error ↓ | 5-mer coupling | 5-mer JS error ↓ | 6-mer coupling | 6-mer JS error ↓ |
|---|---|---|---|---|---|---|
| RINGER | 0.695 ± 0.002 | 0.099 ± 0.005 | 0.589 ± 0.021 | 0.126 ± 0.006 | 0.492 ± 0.049 | 0.191 ± 0.001 |
| ET-Flow-SO(3) | 0.948 ± 0.049 | 0.137 ± 0.003 | 0.612 ± 0.028 | 0.187 ± 0.002 | 0.358 ± 0.006 | 0.251 ± 0.002 |
| AvgFlow$_{\mathrm{DiT}}$ | 0.933 ± 0.046 | 0.159 ± 0.006 | 0.639 ± 0.004 | 0.199 ± 0.004 | 0.450 ± 0.031 | 0.220 ± 0.003 |
| CycloFlow, stage 1 only | 0.759 ± 0.028 | 0.092 ± 0.001 | 0.736 ± 0.016 | 0.119 ± 0.002 | 0.572 ± 0.071 | 0.195 ± 0.011 |
| CycloFlow | 0.663 ± 0.030 | 0.096 ± 0.003 | 0.651 ± 0.016 | 0.123 ± 0.003 | 0.530 ± 0.008 | 0.195 ± 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.
[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.
| Method | Ring | All-atom | ||||
|---|---|---|---|---|---|---|
| n | COV-F1 ↑ | MAT (Å) ↓ | n | COV-F1 ↑ | MAT (Å) ↓ | |
| Alpha only, 21 ring atoms | ||||||
| RINGER | 121 | 24.7 ± 7.0 | 0.606 ± 0.042 | 56 | 20.2 ± 3.1 | 1.273 ± 0.086 |
| ET-Flow-SO(3) | 121 | 0.7 ± 0.3 | 0.754 ± 0.005 | 121 | 4.4 ± 0.4 | 1.721 ± 0.022 |
| AvgFlow$_{\mathrm{DiT}}$ | 121 | 3.6 ± 0.6 | 0.466 ± 0.013 | 121 | 36.1 ± 0.9 | 1.155 ± 0.029 |
| CycloFlow | 121 | 31.1 ± 3.7 | 0.398 ± 0.016 | 121 | 58.8 ± 1.3 | 0.950 ± 0.027 |
| 22, 23 and 26 ring atoms | ||||||
| RINGER | 0 | – | – | 0 | – | – |
| ET-Flow-SO(3) | 58 | 0.1 ± 0.2 | 0.929 ± 0.011 | 58 | 2.2 ± 0.5 | 2.036 ± 0.019 |
| AvgFlow$_{\mathrm{DiT}}$ | 58 | 0.5 ± 0.0 | 0.645 ± 0.008 | 58 | 14.7 ± 0.6 | 1.549 ± 0.013 |
| CycloFlow | 58 | 10.5 ± 1.7 | 0.675 ± 0.025 | 58 | 35.2 ± 3.0 | 1.497 ± 0.039 |
Table 5. CREMP-CycPeptMPDB 7-mers, group split (n = 103). Mean ± sd over 3 seeds. n as in Table 4.
| Method | Ring | All-atom | ||||
|---|---|---|---|---|---|---|
| n | COV-F1 ↑ | MAT (Å) ↓ | n | COV-F1 ↑ | MAT (Å) ↓ | |
| Alpha only, 21 ring atoms | ||||||
| RINGER | 66 | 24.5 ± 0.5 | 0.594 ± 0.017 | 3 | 3.2 ± 0.2 | 0.947 ± 0.036 |
| ET-Flow-SO(3) | 66 | 0.6 ± 0.1 | 0.727 ± 0.010 | 66 | 3.6 ± 0.2 | 1.773 ± 0.017 |
| AvgFlow$_{\mathrm{DiT}}$ | 66 | 1.8 ± 0.4 | 0.449 ± 0.010 | 66 | 25.4 ± 2.0 | 1.212 ± 0.035 |
| CycloFlow | 66 | 28.2 ± 2.6 | 0.386 ± 0.009 | 66 | 52.4 ± 2.5 | 1.028 ± 0.032 |
| With beta, 22 ring atoms | ||||||
| RINGER | 0 | – | – | 0 | – | – |
| ET-Flow-SO(3) | 37 | 0.1 ± 0.1 | 0.834 ± 0.016 | 37 | 2.3 ± 0.6 | 1.952 ± 0.034 |
| AvgFlow$_{\mathrm{DiT}}$ | 37 | 0.3 ± 0.2 | 0.590 ± 0.014 | 37 | 14.2 ± 2.4 | 1.513 ± 0.043 |
| CycloFlow | 37 | 16.4 ± 1.2 | 0.546 ± 0.008 | 37 | 40.8 ± 3.2 | 1.368 ± 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.
[Reviewer3 R1]. Torsion prior and Cartesian control (W1, Q1)
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.
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.
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.