acceptodds
Under review as a conference paper at ICLR 2027

DC-Flow: Target-Conditioned Cyclic Peptide Design via Delayed-Coupling Flow Matching

Abstract

Target-conditioned cyclic peptide design requires coordinating interfacial side chains with macrocyclic closure. Existing end-to-end generators concurrently update geometric and chemical variables from noise, assigning residues to disordered coordinates. Meanwhile, predictor-assisted workflows often get stuck in local traps because they lack feedback in the propose-and-filter stage and ignore 3D structural proposals during sequence search. To overcome the drawbacks of these two paradigms, we propose DC-Flow, a staged generative framework built around Delayed-Coupling Flow Matching. In principle, DC-Flow postpones sequence and side-chain activation until a target-conditioned backbone takes shape, resolving structural ambiguity before chemical assignment. Then, it feeds delayed activations back into the backbone, achieving co-evolution of backbone geometry and emerging residue identities by coupling their flows. This mechanism enables more reliable macrocyclic closure and favorable interface packing. For predictor-assisted workflows, DC-Flow supplies 3D proposals with in-trajectory guidance, enabling cascaded test-time scaling to prune redundant backbones and terminate unpromising refinement early. Extensive experiments show that DC-Flow improves cyclization fidelity, diversity, and interface energetics, substantially increasing the yield of confident cyclic peptide binders under a fixed final-candidate budget.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cyclic Peptide Design, Predictor-Guided Optimization, Protein-Protein Interactions

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_177UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes DC-Flow, a target-conditioned cyclic peptide design framework based on delayed-coupling flow matching. The method first generates a receptor-conditioned backbone and progressively activates sequence and side-chain generation while allowing the backbone and sequence to continuously co-adapt. It further incorporates a frozen ProteinMPNN as a structural sequence prior and introduces cascaded test-time scaling to improve candidate exploration and allocate computational resources to promising designs.

Critical strengths
  1. The proposed delayed-coupling strategy is intuitive. Delaying sequence and side-chain generation until the backbone has developed, while still allowing the modalities to co-evolve afterward, is a reasonable approach to cyclic peptide design.
  2. The paper is generally well written and organized, with clear explanations and high-quality figures that help illustrate the motivation and overall framework.
Critical weaknesses
  1. The performance gains of DC-Flow on the cofolding benchmark appear to depend heavily on predictor guidance. Without generation-time guidance, DC-Flow achieves the lowest strict-hit rate among the compared methods after AfCycDesign optimization (24.2%, compared with 33.3% for HFGuidedDesign and 29.2% for RFPeptides). Its advantage only emerges when guidance is enabled, raising concerns about how much of the improvement can actually be attributed to the proposed delayed-coupling mechanism. Moreover, since generation-time guidance and final evaluation rely on the same AlphaFold-Multimer model family, it is difficult to rule out predictor-specific optimization or reward hacking. Although HFGuidedDesign provides a relevant guided baseline, the comparisons remain limited, and it is unclear whether the methods use comparable computational budgets. The authors should demonstrate that DC-Flow outperforms existing approaches under matched structure-predictor inference budgets and provide stronger evidence isolating the contribution of delayed coupling.
  2. DC-Flow involves substantial computational overhead, requiring repeated structure-predictor evaluations during generation-time guidance, followed by expensive AfCycDesign optimization and potentially additional cascaded search. However, the paper does not provide a comprehensive comparison of inference costs, such as the number of predictor calls, GPU-hours, or wall-clock time. Another major concern is that the final evaluation still relies on AlphaFold-Multimer, the same model family used for guidance and optimization. Evaluating the generated candidates with an independent predictor, such as AlphaFold 3, would provide more convincing evidence that the improvements generalize beyond the optimization oracle. Additionally, the motivation for generating the backbone before the sequence is not fully convincing. The authors should further justify this design choice and investigate alternatives, such as generating the sequence first and then refining the backbone, to establish whether the proposed generation order is indeed preferable.
Questions for authors
  1. Could the authors report the number of structure-predictor evaluations, optimization steps, and total inference costs (e.g., GPU-hours) for DC-Flow and the baselines? Although HFGuidedDesign also uses predictor guidance, it is unclear whether the methods receive comparable amounts of guidance and optimization compute.
  2. Since generation-time guidance, post-generation optimization, and final evaluation all rely on the AlphaFold-Multimer model family, could the authors evaluate the generated candidates using an independent structure predictor, such as AlphaFold 3 or another suitable model?
Reviewer_FvCcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The key novelty of DC-FLOW lies in its delayed-coupling flow matching framework, which postpones sequence and side-chain generation until the target-conditioned backbone has begun to form, while maintaining continuous backbone–sequence co-adaptation throughout the remaining trajectory. It further integrates a frozen ProteinMPNN prior into the generation process to improve sequence–structure compatibility, and introduces cascaded test-time scaling to adaptively allocate computational resources through diverse backbone selection, guided sequence branching, and early pruning of unpromising candidates. Experiments on direct full-atom generation and predictor-assisted cofolding demonstrate improvements in cyclization success, structural and sequence diversity, interface energetics, and the yield of high-confidence cyclic peptide binders.

Critical strengths
  • The paper introduces three technical improvements for cyclic peptide design: delayed-coupling flow matching to coordinate backbone and sequence generation, ProteinMPNN prior integration to improve sequence–structure compatibility, and cascaded test-time scaling to allocate computational resources more efficiently during candidate generation and refinement. These components are well motivated and integrated into a coherent framework, providing a clear technical pathway for addressing structural ambiguity, sequence–backbone compatibility, and search efficiency.

  • The experimental evaluation is comprehensive, covering direct full-atom generation on the LNR benchmark and predictor-assisted cofolding on 12 published targets. The authors compare against six full-atom generation baselines and two major proposal-generation methods, evaluate cyclization success, sequence and structural diversity, interface energetics, and prediction confidence, and conduct ablation studies on activation timing, backbone coupling, ProteinMPNN integration, and test-time search.

Critical weaknesses

The evaluation primarily focuses on cyclization success, structural diversity, interface energetics, and prediction confidence. However, other important properties for drug development, such as solubility, proteolytic stability, membrane permeability, selectivity, and synthetic feasibility, are not assessed, limiting the understanding of the practical applicability of the generated candidates.

Questions for authors
  • The ablation study evaluates only a limited number of sequence activation thresholds ((τs\tau_s = 0, 0.2,) and (0.5)), leaving the overall relationship between activation timing and model performance insufficiently characterized. Could the authors systematically investigate a broader range of (\tau_s) values and analyze the resulting performance curves across cyclization success, sequence and structural diversity, and interface energetics? Such an analysis would help clarify how the timing of sequence activation affects the trade-off between exploration and structural compatibility, and whether the default (τs\tau_s = 0.5) represents a robust operating point or a task-dependent optimum.
  • How could the evaluation framework be extended to incorporate additional drug-relevant metrics, such as solubility, proteolytic stability, membrane permeability, selectivity, and synthetic feasibility, to provide a more comprehensive assessment of the generalizability of DC-FLOW to real-world cyclic peptide drug design?
Reviewer_2Z5fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DC-Flow generates target-conditioned cyclic peptides by staggered flow matching: the backbone runs first, sequence and side chains turn on at later times, and the backbone keeps moving so geometry and sequence co-adapt. A frozen ProteinMPNN prior enters as a residual, and a cascade prunes backbones, branches sequences under cofolding guidance, and spends full AfCycDesign refinement on a shortlist. Evaluation is direct generation on 93 LNR complexes and cofolding on 12 published targets.

Critical strengths
  1. Retraining PepFlow, PepGLAD, DiffPepBuilder, APM, Genie 3, and APCyc on the same filtered corpus makes the LNR comparison fairer than usual. The in-distribution / out-of-distribution split by receptor sequence identity (37 receptors below 50%, none above 70%) is the right check for this data regime.

  2. The coupling ablation is decisive. Freezing the backbone at sequence activation collapses out-of-distribution head-to-tail Rosetta interface energy from −28.37 to −4.54 REU and turns stability from −940 to +113. Continued co-adaptation is doing real work.

  3. Head-to-tail DC-Flow is the best LNR entry on out-of-distribution interface energy (−28.37) and sequence diversity (0.940), with cyclization at 97.5–100%. The disulfide variant instead leads in structural diversity (0.991 / 0.943).

  4. On the 12-target cofolding set, mean ipTM, peptide pLDDT, iPAE, composite score, and strict-hit rate all edge RFPeptides and HFGuidedDesign (strict hits 36.7% versus 33.3% versus 29.2%), and the Rosetta interface panel moves in the same direction.

Critical weaknesses
  1. Cyclization is already saturated (most baselines are above 92%), so the headline closure claim has little room. Head-to-tail in-distribution structural diversity falls to 0.459, behind PepGLAD (0.817) and Genie 3 (0.796). PepFlow still has the best out-of-distribution stability (−961 versus −940).

  2. Delay is a smaller effect than coupling. Setting τ_s = 0 worsens out-of-distribution head-to-tail interface energy from −28.37 to −20.40; removing the ProteinMPNN prior only moves it to −27.56. The emphasis on postponing chemical assignment is only partly what the ablations isolate. Earlier activation (τ_s = 0.2) also improves diversity, so the default schedule is one point on a tradeoff.

  3. The search result, strict hits from 32.6% to 75.5% at 32 final candidates, is the largest number in the paper and the least controlled. The cascade launches K backbones, branches M sequences, queries a cofolding oracle along the trajectory, then fully refines a shortlist. Baselines are matched on final candidates, not on oracle calls. A fixed output budget can hide a much larger search budget.

Questions for authors
  1. Please report strict-hit rate against number of AfCycDesign evaluations, with HFGuidedDesign and RFPeptides given the same evaluation count the cascade actually uses.

  2. Why does freezing the backbone destroy energy while sequence and structural diversity stay almost unchanged? Is the frozen scaffold geometrically open, or does the sequence trunk fail without further backbone motion?

  3. What fraction of LNR successes remain if cyclization is judged before Rosetta FastRelax, and if the receptor is required to be a real complex rather than a CPSea-style pseudo-complex?

Reviewer_15W4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper tackles target-conditioned cyclic peptide binder design. The model must jointly solve receptor docking, macrocyclic closure, sequence design, and side-chain packing, all under scarce complex-structure supervision. The authors argue that both dominant paradigms have a structural flaw. End-to-end full-atom generators (PepFlow, PepGLAD, DiffPepBuilder, APCyc) update geometry and chemistry simultaneously from noise, assigning residue identities over disordered backbones. Predictor-assisted pipelines (RFPeptides, AfCycDesign, HFGuidedDesign) achieve high self-consistency but pay prohibitive cofolding cost and lack structure-sequence feedback during search.

DC-FLOW proposes a staged flow-matching framework with three contributions. (i) Delayed-coupling flow matching: per-modality clocks activate backbone, sequence, and side-chain flows at thresholds 0 < τ_s < τ_c < 1 (default 0, 0.5, 0.8). Before τ_s the sequence inputs are mask tokens. After activation all modalities co-evolve in a shared geometric trunk, rather than freezing the backbone. (ii) A frozen ProteinMPNN residual prior, fused into the sequence logits through a trainable scalar gate g with a stop-gradient on the coordinate input, meant to compensate for cyclic-complex data scarcity. This yields a product-of-experts distribution at inference and places the prior inside the trajectory rather than as a post-hoc step. (iii) Cascaded test-time scaling: K backbone trajectories are pruned to K₀ diverse roots at τ_s, each branching into M oracle-guided sequence continuations. After a 5-step AfCycDesign warm-up, candidates are probabilistically screened by an interface-quality score Q, with early stopping when scores plateau.

All validation is in silico, along two lines. As a direct full-atom generator on a 93-receptor LNR benchmark (37 OOD / 56 ID receptors by MMseqs2 identity), DC-FLOW(HT/SS) is compared against six retrained baselines on cyclization success, Rosetta stability, sequence/structure diversity, and interface dG. As a cofolding proposal generator on the 12 HFGuidedDesign targets, it is compared with RFPeptides and HFGuidedDesign on AlphaFold-Multimer-v3 confidence metrics, a composite score, strict-hit rate, and Rosetta interface quality. Ablations cover the activation schedule (τ_s ∈ {0, 0.2, 0.5}), backbone freezing, and removal of the MPNN prior. A final experiment reports that cascaded search raises the strict-hit rate from 32.6% to 75.5%.

Critical strengths
  1. The delayed-coupling mechanism is a well-motivated design that the ablations show is load-bearing, not decorative.

The biophysical premise (backbone scaffold dictates side-chain presentation at PPI interfaces) gives a sensible reason to stage generation. More importantly, the paper does not stop at "stage it." The distinction between delaying and decoupling is the real contribution. The w/o coupling ablation isolates exactly this: freezing the backbone at τ_s collapses OOD HT interface energy from −28.37 to −4.54 REU and total stability from −940.05 to +112.60 REU, while leaving both diversity metrics essentially unchanged (Table 3). Whatever the confound (see Weakness 2), a swing of this size cannot be explained by diversity and supports the paper's claim that continued co-adaptation, not merely ordering, is what matters. This is consequential to my assessment because it is the one place where the paper's conceptual thesis is tested against a sharply defined alternative and produces a clear answer.

  1. Unusually thorough and self-critical experimental documentation.

Appendices D and E specify stereochemical filtering criteria, the closure-restraint functional form and its parameters, the SS-to-HT fallback policy for baselines, the exclusion of closure restraints from reported energies, and the guidance schedule down to specific sampling steps. All six full-atom baselines are retrained on the identical corpus and split. Several statements work against the authors' own interests: equal final-candidate counts do not imply equal oracle cost (App. E.2), the main and search comparisons use different seed sets (App. E.5), Rosetta and confidence scores are design criteria rather than affinity measurements (App. E.5), and CPSea-derived supervision remains largely synthetic (Sec. 5). This matters to my assessment because it makes the paper's weaknesses diagnosable and fixable rather than opaque, and it is the main reason I regard the concerns below as addressable.

Critical weaknesses
  1. The headline cofolding comparison (Table 2) is not statistically resolved, and the paper's own appendix supplies the counter-evidence.

The claimed advantage over HFGuidedDesign is ipTM 0.683 vs 0.674, interface iPAE 0.326 vs 0.334, composite S 1.982 vs 1.957, and strict-hit rate 36.7% vs 33.3%, the last corresponding to roughly 13 extra candidates out of 384. All of these are reported as single point estimates, with nothing said about their stability. Appendix E.5 then states that the main and search comparisons use separate seed sets and that "their baseline values need not coincide." Indeed, the same DC-FLOW configuration (guidance + AFCyc optimization, 32 candidates per target) reports 36.7% in Table 2 and 32.6% in the search experiment. That 4.1-point seed-induced swing is larger than the 3.4-point gap to HFGuidedDesign that the paper's main claim rests on. As written, the evidence does not distinguish the proposed method from the strongest baseline on the predictor-assisted task. This is consequential because the cofolding result, not the LNR table, is where the paper argues it bridges the two paradigms.

  1. The two results that would otherwise carry the paper are each confounded by a design choice.

(a) The w/o coupling ablation does not implement the pipeline it claims to mimic. It freezes the backbone at B_{τ_s}, a state at t = 0.5 of a 200-step flow, i.e. a partially denoised backbone, and describes this as "mimicking conventional decoupled two-stage pipelines (i.e., fixed-scaffold inverse folding)." Real two-stage pipelines perform inverse folding on a completed backbone. The +112.60 REU stability therefore plausibly measures the cost of packing onto an unfinished scaffold rather than the cost of decoupling. The ablation conflates "coupling helps" with "half-denoised backbones are bad." The honest comparison, running the backbone flow to t = 1 and only then freezing for design, is missing.

(b) The cascaded-search result is not compute-matched. The 32.6% to 75.5% jump is the paper's largest effect, but the search configuration uses K = 16 initial trajectories, N = 64 continuations, and per-event oracle calls, while the baselines produce 32 candidates. Appendix E.2 concedes that equal final-candidate counts do not imply equal oracle cost. Since the paper frames its argument explicitly around efficiency under a fixed compute budget, a search result reported without oracle-call or wall-clock accounting, and without baselines given a comparable budget (e.g. RFPeptides/HFGuidedDesign with 64 proposals and matched oracle calls, or simple best-of-N selection over 64 DC-FLOW samples), cannot establish that the cascade is the source of the gain rather than the extra compute. Together with Weakness 1, these two items leave each of the paper's three contributions supported by evidence that is either statistically unresolved or confounded.

  1. (Contributing, not decisive) Evaluation is partly circular and the LNR picture is mixed.

AfCycDesign/AlphaFold-Multimer is used both as the in-trajectory guidance oracle and as the source of the ipTM/iPAE/strict-hit metrics. Guided methods are effectively optimized against their own evaluator. The authors acknowledge this risk in Section 5 but do not mitigate it with an independent held-out scorer (e.g. HighFold, or Boltz/Chai-class predictors). On LNR, DC-FLOW does not dominate: HT has the lowest ID structural diversity in Table 1 (0.459, against 0.670 to 0.817 for baselines), SS has the lowest ID sequence diversity (0.205), and neither variant leads on Stability for ID receptors. The text reads this as a "complementary topological profile," which is fair, but it means the direct-generation claim rests on cyclization success and OOD interface energy alone.

Questions for authors
  1. Table 2 reports 36.7% strict hits for DC-FLOW (guide + AFCyc optim.), while the search experiment reports 32.6% for what appears to be the same configuration, attributed in Appendix E.5 to different seed sets. The gap between those two numbers (4.1 points) exceeds the gap to HFGuidedDesign in Table 2 (3.4 points). Two questions: (a) do the two runs differ in anything beyond the random seeds, such as oracle-call budget, guidance settings, or filtering pipeline? (b) Given that the range of values the same configuration takes across seeds already spans the between-method gap, in what sense does the Table 2 advantage hold, and on what basis? Please make that case explicitly.

  2. For the w/o coupling ablation: what happens if the backbone flow is run to completion (t = 1) and only then frozen for sequence and side-chain design? This is the pipeline the ablation claims to mimic, and it would separate "coupling helps" from "partially denoised backbones are poor scaffolds." Do you have this number, or an estimate of how much of the +112.60 REU is attributable to scaffold incompleteness?

  3. For cascaded search, can you report total oracle calls (AlphaFold forward passes) and wall-clock time per target for DC-FLOW+Search and for each baseline, and add at least one compute-matched comparison, e.g. baselines generating 64 proposals with matched oracle budget, and/or naive best-of-64 selection over unguided DC-FLOW samples under the same budget as the cascade?

  4. Has any subset of designs been evaluated with a predictor not used for guidance (HighFold, Boltz-2, Chai-1, or similar)? Even a small held-out comparison would indicate how much of the strict-hit advantage survives outside the AF2 family that produced it.

  5. Table 1 shows DC-FLOW(HT) with the lowest ID structural diversity (0.459) among all methods. Is this an intended consequence of the ProteinMPNN prior and delayed activation concentrating probability mass, and does it indicate reduced coverage of alternative binding modes on targets close to the training distribution?

  6. The gate g is initialized to 0.1. What value does it converge to after training, and how sensitive are the LNR and cofolding results to it? A near-zero or very large learned gate would change how the prior's contribution should be interpreted.

  7. Appendix E.3 states that failed baseline outputs may fall back from SS to HT closure while DC-FLOW is scored against its requested topology, with linkage changes counted as failures. Could the authors confirm that this adapter does not systematically advantage DC-FLOW's cyclization-success numbers, ideally by reporting baseline Cyc. under both policies?

Suggestions for future work (not affecting my score): reporting results under a fixed wall-clock budget rather than a fixed candidate count would make the efficiency argument much more direct. A small wet-lab validation, as the authors propose in Section 5, would address the synthetic-supervision concern more convincingly than any additional in-silico metric.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.