acceptodds
Under review as a conference paper at ICLR 2027

CovaTide: Covalent Topology-Conditioned All-Atom Flow Matching For Linear and Cyclic Peptide Design

Abstract

Peptides offer a versatile therapeutic modality for targeting protein surfaces, but conformational flexibility and susceptibility to proteolysis can limit their utility. Cyclization can mitigate these limitations, yet cyclic peptide design requires coordinating receptor-bound conformations with the local geometry of covalent closure. Existing methods often rely on post-generation filtering and endpoint mutation to obtain cyclic peptides, decoupling closure constraints from the generation of receptor-bound conformations. We present CovaTide, a receptor-conditioned all-atom flow model that jointly generates peptide sequences and structures under a prescribed covalent topology. CovaTide uses a deep topology adapter to repeatedly integrate residue-level connectivity and atom-level bond information, coupling global peptide conformations with local closure geometry throughout generation. This enables a single model to design linear peptides as well as head-to-tail, disulfide, and isopeptide cyclic peptides without post-generation endpoint mutation. To rank generated candidates, we further introduce a generation-aware confidence model that estimates peptide and interface errors. CovaTide outperforms the evaluated baselines in sequence and binding pose recovery for linear peptides and in cyclization, bond, and energy-based success rates for cyclic peptides. Together, these results support a unified formulation of linear and cyclic peptide binder design through multiscale covalent topology conditioning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI-driven peptide design, cyclic peptide, all-atom generation, flow matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PQCkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes CovaTide, a receptor-conditioned all-atom flow model that jointly generates peptide sequences and structures under a prescribed covalent topology. CovaTide uses a deep topology adapter to repeatedly integrate residue-level connectivity and atom-level bond information, coupling global peptide conformations with local closure geometry throughout generation. The main idea is to directly incorporate the prescribed covalent topology as a generation condition, rather than first generating linear peptides and then obtaining cyclic peptides through filtering, endpoint mutation, or post-processing.

Critical strengths
  1. A single model can generate linear peptides as well as head-to-tail, disulfide, and isopeptide cyclic peptides. This unified treatment of different peptide topologies is practically meaningful.
  2. The experimental design is relatively reasonable, and the proposed method is evaluated from multiple aspects.
Critical weaknesses
  1. The paper repeatedly emphasizes joint sequence–structure generation, but in practice, only the coordinates follow the flow trajectory, while the sequence is predicted by a readout after generation.
  2. For each target, 100 peptides are generated, and the best candidate is selected separately for AAR, C-RMSD, L-RMSD, energy, and other metrics, which may exaggerate the apparent performance advantage.
  3. The main result tables do not seem to demonstrate that the proposed method is broadly competitive across the evaluated metrics.
Questions for authors

The authors should clarify the main advantage of the proposed method. Although performance is not the only consideration for a method with clear methodological novelty, it is still important. CovaTide underperforms existing methods on several metrics, and even the ablation results do not consistently support the superiority of the full model. Could the authors provide a clear explanation? If only the coordinates follow the flow trajectory, the sequence is predicted at the end, and the sequence loss cannot affect the coordinate trunk due to stop-gradient, in what sense are sequence and structure “jointly generated” or “co-designed”? Why does the linear peptide evaluation use an independent best-of-100 selection for each metric?

Reviewer_yc9YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CovaTide is a receptor-conditioned all-atom flow matching model that generates peptide sequences and structures under a prescribed covalent topology, specified by a linkage type (linear, head-to-tail, disulfide or isopeptide) and two closure residue indices whose identities are fixed to Cys and Cys for disulfide and Lys and Asp for isopeptide. The goal is one model covering linear peptides and all three cyclic types without post-generation filtering or endpoint mutation.

The mechanism is a deep topology adapter at two resolutions. At the residue level, an endpoint role code and a shortest-path distance on the peptide graph augmented with the closure edge are embedded and injected with learned per-site scalar gates after each pairformer block and before each flow transformer block. At the atom level, the role embedding is propagated to atom slots and a typed closure-edge indicator is injected into the atom-pair context before each atom encoder and decoder block, with sparse-attention neighbourhoods augmented so the two closure partners stay in each other's neighbour lists. Coordinates follow a linear flow; residue identities are read out at the end from stop-gradient decoder features, so only coordinates follow the flow trajectory. Training adds a closure loss on the five atoms around the junction and a geometry-based chemical-consistency penalty, and a separately trained confidence model predicts pLDDT, PDE and iPDE for candidate ranking.

Linear training uses ProtFrag and PepBench; cyclic training uses a subset of CPBind, which is the subset of CPSea with Vina below -6 and Rosetta dG below -25, CPSea itself being built from AlphaFold Database predictions. Evaluation uses LNR, 93 targets for linear and a 57-target subset for cyclic. CovaTide leads on all linear recovery and energy metrics and on cyclization, bond and energy success for cyclic peptides, and additional results cover internal cyclization sites, molecular dynamics on two complexes, a rhodopsin case study and an adapter ablation.

Critical strengths
  • Linear-peptide results beat directly comparable baselines under the same 100-sample budget per target on recovery, pose, energy and both clash rates (L302-304).
  • None of that depends on the topology machinery, so the underlying generator is strong on its own terms. Internal cyclization at residue pairs that never appear in training still reaches 77.8 percent bond and 75.2 percent joint success over 540 candidates (L1399-1440).
  • That is a test of the conditioning mechanism rather than of the training distribution, and the most convincing result in the paper.
  • Implementation detail is sufficient to reimplement, down to injection sites, gate initialization and per-topology closure atom sets (L864-1025).
Critical weaknesses
  • Only CovaTide receives the cyclization type, the fixed endpoint identities and a closure loss; baselines cannot condition on type and are classified post hoc by endpoint identity (L324-328).
  • Removing the adapter entirely still leaves 0.581 cyclization and 0.753 closure range for disulfide, so head-to-tail is the only topology where it is decisive.
  • The shallow single-injection adapter beats the deep one on head-to-tail and is within 0.02 to 0.05 elsewhere, which does not establish the second stated contribution (L096-099).
  • No seeds or confidence intervals appear anywhere in the paper.
  • Training data is filtered by Vina below -6 and dG below -25 (L735-739) while cyclic success is scored as dG below 0 and Vina below -3 (L1176-1182), so the success rates largely restate the training filter.
  • No experimental cyclic complexes exist for the 57 test targets, so every cyclic metric is a self-consistency proxy and none speaks to binding.
  • The leakage filter removed only 14 of 234,513 samples (L759-774) and will miss domain-level homology between the AlphaFold-derived receptors and the test receptors.
Questions for authors
  • How many runs or seeds underlie the ablation table, and what is the run-to-run spread? With deep-versus-shallow margins of 0.02 to 0.05 and the shallow adapter winning head-to-tail, a single run cannot support the contribution stated at L096-099.
  • Can you run the cyclic baselines with their endpoint residues constrained to the same positions and identities you prescribe, so that the bond-success comparison isolates the conditioning mechanism rather than knowledge of which residues must bond (L324-328)?
  • Recovery and energy metrics are reported as the per-target best over 100 samples (L306-313) while CovaTide has the lowest sequence and structural diversity, 0.412 and 0.362 against 0.654 and 0.772 for PepGLAD. Can you report these as distributions and as best-of-k curves for k equal to 1, 10 and 100?
  • Side-chain coordinates are generated in generic slots before the residue type is known, with the type assigned afterwards from stop-gradient features (L1450-1456). What fraction of generated side-chain geometries are incompatible with the finally assigned residue type, and how is that resolved at output time?
  • What fraction of CPBind receptors share a Pfam or CATH domain, or a 30 percent-identity local alignment without the coverage requirement, with one of the 93 test receptors, and does a domain-level filter change the cyclic results (L759-774)?
  • The within-target correlations for the confidence model are described only as having smaller magnitudes (L382-385). Please report their median and interquartile range and the number of evaluable targets.
  • The molecular dynamics comparison uses two complexes with one trajectory each and contrasts a generated cyclic peptide with a native linear one (L391-400), confounding cyclization with design provenance. Would you run the generated linear design as the matched control, with replicates?
Reviewer_xZdzOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a new method for design of linear and cyclic peptides. The peptide sequence and structure are generated jointly. In the cyclic case, the joint coordinates are also generated. The paper does a number of evaluations and ablations that shows favourable performance to recent methods.

Critical strengths

A new approach for an important problem. Good benchmark performance. MD simulations for a few targets. Illustrative and simple ablations.

Critical weaknesses

Not critical per se. General criticism of the peptide generation field: The construction training data is done extraction substructures from proteins. It must have limitations. Specific criticism for the method: We are probably going to transition to fully "atomistic" design methods in a near future (after the ICLR deadline RFDiffusion4/Proteina has be released). So it is not clear that dedicated methods like this one is the right way to go.

Questions for authors

Why make design methods for specialised cases?

Comments on the paper

1 comment

Thank you for the clarifications and additional experiments. After considering the responses, I am maintaining my rating of 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.