acceptodds
Under review as a conference paper at ICLR 2027

PepMap: MultiModal Flow Maps for Few-Step All-Atom Peptide Binder Design

Abstract

Designing peptide binders for a target protein requires generating amino-acid sequences and bound conformations that are compatible with its binding site. Recent diffusion and flow-matching methods can jointly generate peptide sequences and structures, but their iterative sampling makes candidate generation costly. Few-step sampling is challenging because residue identities and backbone geometry must remain coordinated across large time intervals. We introduce , a pocket-conditioned multimodal Flow Map for few-step peptide sequence–structure co-design. PepMap learns finite-time transitions that jointly update continuous sequence states, residue positions, and backbone orientations. Flow-matching supervision guides local predictions, while multimodal semigroup consistency encourages agreement between direct and composed transitions. Geometric objectives regularize backbone geometry and peptide–pocket compatibility, and a conditional side-chain packing module produces full-atom peptide structures. On the LNR benchmark, PepMap improves structural recovery over iterative baselines at their standard sampling budgets while requiring substantially fewer network evaluations. These results support finite-time transport of coupled sequence and backbone states for efficient pocket-conditioned peptide design. Code is available at https://anonymous.4open.science/r/PepMap-anonymous-E018/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide Design, Flow Matching, Flow Maps, Few-step Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2msLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PepMap, a pocket-conditioned multimodal flow map for peptide co-design. By learning finite-time transitions over the joint state and combining the flow matching objectives with other consistency supervision, this method shows strong results on the benchmark for structural and binding metrics at low sampling steps.

Critical strengths
  1. Accelerating peptide generation is a meaningful research problem.
  2. The low-NFE results are promising.
Critical weaknesses
  1. The method seems to be a direct application of flow-map self-distillation. It's reasonable to say the peptide generation problem is a multimodal generation problem, and we need to ensure some joint consistency during training and sampling. But I'm still confused about the concrete difficulty when adapting flow-map to such a problem. Once the flow map has been defined over the joint state of different modalities, I think it's natural to use this formulation.
  2. The analysis of low-NFE generation performance is not sufficiently detailed to explain the method'pros and cons. I would like to see both finer-grained comparisons of final generation results and trajectory-level analysis, showing how sequence, backbone, and pocket compatibility evolve.
  3. Presentation issues: In Figure3, the triangular markers above the plotting area are not explained, and such overlapping makes it difficult to read/understand. In Figure 4, the second panel in the second row contains really long connecting lines crossing the protein pocket, and it's unclear what they mean.
Questions for authors
  1. What is the minimal difference between PepMap and directly applying an existing flow-map self-distillation framework to a peptide generation problem? What's the difficulty of the adaptation, and how do you solve these problems?
  2. Can you provide more fine-grained comparisons, e.g. 1,2,3,5,10 NFE? I wonder if improvements across different modalities or metrics occur at the same time/candidates?
  3. Can you fix the pocket, peptide length and initial noise, and show representative succesful/unsuccessful generation trajectories? I want to see the evolution of different modalities.
Reviewer_B9LgOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PepMap, a pocket-conditioned multimodal flow-map framework for few-step peptide sequence–structure co-design. The method jointly models continuous amino-acid representations, residue translations, and backbone rotations through finite-time transition operators. Training combines diagonal flow-matching supervision, semigroup self-distillation, and auxiliary geometric objectives. A separate conditional side-chain packing module reconstructs all-atom peptide structures.

The method is evaluated on the LNR benchmark of 93 protein–peptide complexes, using PepBench for training. The authors report competitive structural recovery with substantially fewer neural function evaluations than iterative diffusion and flow-matching baselines. Ablation experiments suggest that semigroup supervision contributes substantially to few-step performance.

However, the reported improvements primarily concern retrospective structural recovery. The evidence for improved peptide–receptor interface quality, all-atom physical plausibility, and practically useful binder generation is considerably weaker.

Critical strengths

1. Efficient few-step generation with promising structural recovery.

PepMap achieves competitive structural recovery with only one to ten neural function evaluations. In particular, its peptide-aligned L-RMSD is substantially lower than those of the evaluated baselines. The matched-budget comparisons in Appendix E.1 further support the effectiveness of the approach under restricted sampling budgets.

This is a meaningful computational advantage, although its practical significance depends on whether the resulting structures are physically plausible and can be selected without reference information.

2. Reasonably motivated multimodal formulation.

The combination of sequence, translation, and rotation within a shared finite-time transport framework is technically coherent. The semigroup ablation demonstrates a substantial performance degradation when consistency supervision is removed, suggesting that this component is useful for the proposed architecture.

Critical weaknesses

1. Limited methodological novelty.

The main contribution seems rather incremental. PepMap brings together existing ideas from flow-map self-distillation, Riemannian transport, and categorical generation, but it is not clear what is fundamentally new. In particular, the multimodal semigroup objective looks like a natural extension of existing methods to joint sequence–structure generation. A more direct comparison with related few-step methods would help justify the novelty.

2. Concerns about the evaluation protocol.

I am not fully convinced by the evaluation setup. The authors generate 100 candidates per target and independently select the best one for each metric, often using information from the native complex. This makes the reported results rather optimistic, as the best scores may come from different candidates. It remains unclear whether PepMap can reliably generate and select high-quality binders in a realistic setting where the native structure is unknown.

3. Insufficient validation of the all-atom generation pipeline.

The side-chain packing module raises some concerns. It is trained using native torsion targets even when conditioned on predicted sequence probabilities, but the paper does not clearly explain how differences between predicted and native residue identities are handled. The packer is also trained on diagonal endpoint predictions rather than actual few-step outputs. More importantly, the evaluation mainly reports Cα-level clashes, which is not sufficient to establish the physical validity of the generated all-atom structures.

4. Limited improvements in binding quality and lack of statistical analysis.

Despite the strong structural recovery results, the improvements in binding-related metrics are less convincing. At 10 NFEs, PepMap only slightly improves C-RMSD over UniMoMo (2.43 vs. 2.48 Å), while performing worse in DockQ (0.762 vs. 0.788), interface energy, and receptor–peptide clash rates (6.16% vs. 0.94%). The paper also lacks confidence intervals or results across multiple training seeds, making it difficult to judge whether the relatively small improvements are meaningful.

Questions for authors

Q1. The proposed multimodal semigroup objective appears closely related to applying existing flow-map consistency directly to a joint sequence–structure state. Could the authors clarify the essential algorithmic distinction and provide a controlled comparison against an existing few-step formulation using the same architecture, training data, and computational budget? In particular, does joint multimodal consistency outperform independently imposed modality-specific consistency?

Q2. Since different metrics independently select the best of 100 candidates, could the authors report average candidate quality and results under a single reference-free selection strategy? Does PepMap retain its advantages when all evaluation metrics are computed on the same selected candidate? This would help establish whether the method offers practical improvements beyond retrospective structural recovery.

Q3. Equations (38)–(39) indicate that the packer receives predicted sequence probabilities but is trained against native torsion targets and native valid-angle masks. When the predicted residue differs from the native residue, the number and chemical interpretation of valid torsions may differ. How are such mismatches handled during training? Additionally, why is the packer trained on diagonal endpoint predictions rather than the off-diagonal outputs encountered during one-step and few-step inference?

Q4. Please report peptide-bond lengths and angles, peptide-bond planarity, Ramachandran outliers, side-chain rotamer outliers, and all-atom intra-peptide and receptor–peptide clashes. Can the authors also quantify the contribution of the proposed packing module relative to Rosetta packing on identical generated backbones, before additional structural relaxation?

Q5. Could the authors provide paired target-level confidence intervals and results from multiple independent training seeds? In particular, is the 0.05 Å C-RMSD improvement over UniMoMo statistically meaningful? Additionally, could the authors report direct-versus-composed transition errors for translation, rotation, and sequence states on held-out targets to verify that the learned operator exhibits the intended semigroup behavior?

Reviewer_9d47OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PepMap, a few-step method for pocket-conditioned peptide design. It combines flow-matching supervision, semigroup consistency, and auxiliary geometric losses. The consistency objective encourages a direct transition to agree with two successive transitions through an intermediate state. On the LNR benchmark, PepMap achieves strong structural recovery with 1–10 generator evaluations. The paper’s main contribution is to bring few-step generation to joint peptide sequence and structure design. The reported results support gains in sampling efficiency and structural recovery under the chosen benchmark protocol.

Critical strengths
  1. Reducing sampling cost is a meaningful objective for peptide design.
  2. The reported structural recovery is encouraging. Matched-NFE comparisons and component ablations provide useful supporting evidence.
Critical weaknesses
  1. The introduction gives only a brief account of why peptide binder design matters. The Related Work section covers few-step generation, peptide design, and side-chain modeling, but does not clearly connect these three lines of work to the specific gap PepMap addresses.
  2. PepMap combines several established ideas. Its contribution beyond applying existing few-step methods to this task remains unclear, as does the practical value of faster generation when interface and side-chain quality remain mixed.
  3. The paper lacks a controlled comparison with an existing few-step method adapted to the same task.
Questions for authors

1.The introduction gives only a brief account of why peptide binder design matters. The Related Work section covers few-step generation, peptide design, and side-chain modeling, but does not clearly connect these three lines of work to the specific gap PepMap addresses. 2. PepMap combines several established ideas. Its contribution beyond applying existing few-step methods to this task remains unclear, as does the practical value of faster generation when interface and side-chain quality remain mixed. 3.The paper lacks a controlled comparison with an existing few-step method adapted to the same task.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.