acceptodds
Withdrawn from ICLR 2027

Full-Atom Peptide Design with Dual-Flow Contrastive Guidance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Peptides can bind protein surfaces with high specificity and engage interfaces that are typically hard to drug, making them a highly promising class of therapeutics. Recent advances in structure-based peptide design leverage diffusion and flow-matching models to simultaneously generate binding peptide sequences and structures. While existing methods can optimize binding affinities through property guidance or preference alignment, they require extensive model retraining or auxiliary classifiers, and still struggle with poor side-chain modeling and severe steric clashes. To address these limitations, we introduce a lightweight and transferable approach that concurrently optimizes binding energy and side-chain conformations. Specifically, we propose PepContra, a dual-flow contrastive guidance method for full-atom peptide co-design. This contrastive guidance integrates dual flows during sampling, leveraging their difference to drive the generative process toward high-quality regions and away from suboptimal spaces. We further optimize the joint sampler by synchronizing discrete amino-acid types with continuous side-chain torsions at each Euler step. Experimental results demonstrate that PepContra outperforms existing approaches across co-design benchmarks, achieving superior binding affinities and highly competitive stereochemistry.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide Design, Drug Design, Flow Matching, Classifier Free Guidance

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gLXnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes PepContra, a guidance framework that trains two model variants for contrastive vector fields. The dual-flow guidance uses a "low-quality" vector field and a "high-quality" one to guide translation, rotation, torsion, and type. The results show meaningful improvements across different spans of in silico metrics.

Critical strengths
  • The paper deals with a relevant, concrete modeling problem in the AI4Science domain and provides a meaningful approach for guiding existing generative models towards desirable properties.
  • The paper's dual flow formulation extends to Riemannian components necessary for modeling peptide structures.
  • The paper shows improvements over the base unguided model and provides concrete evidence of the previous generative models' failure modes corrected by energy guidance.
Critical weaknesses
  • The theoretical justification contains two substantive gaps in terms of distributional claims. These substantially weaken the current evidence.
    • Sec 3.2 claims that the guided distribution pˉ∝pH1+γ/pLγ\bar p\propto p_H^{1+\gamma}/p_L^\gamma assigns more probability than the true distribution qq to desirable samples, and claims it is "straightforward to prove". However, this is incorrect. Consider the following example, with the first two in CH\mathcal{C}_H and the last in CL\mathcal{C}_L:
p_H=(0.3,0.4,0.3),\quad p_L=(0.2,0.1,0.7).$$ This follows the authors' definition of high-quality and low-quality distributions. However, with $\gamma=1$, direct calculation gives: $$\bar p\approx(0.207,0.734,0.059).$$ The first good state loses probability relative to $q$. This means that **there is no guarantee that every good point can be promoted** compared to the original distribution, because the normalizing constant is now different. - The claim in Sec 3.3 that the combined velocity field generates the proposed marginal distributions is also incorrect, as the normalizing constant also changes. The author may claim this as an intuitive justification: linearly combining such vector fields may provide guidance (e.g., CFG), but it does not rigorously guarantee the marginals. - The idea of using a good and bad model for contrastive guidance is **already explored in existing works**. Specifically, **autoguidance** (cited, but not compared) explicitly uses a bad version of the generative model, with the additional benefit of being training-free. More directly, **Diffusion-NPO** [1] explicitly trains a negative model representing dispreferred outputs and combines it with the preferred outputs during sampling. The proposed method therefore seems more like an application of these models specifically to the peptide domain. - The paper claims itself as "without large-scale retraining or modification". While this is true compared to retraining a new model, the authors **do not compare with training-free approaches like autoguidance**, which does not require fine-tuning at all. To support such a dual-flow framework, 10k samples must first be generated and scored by the designated scorer, and two models need to be fine-tuned. Even more, each different score/reward requires a different fine-tuned model pair. This pipeline **limits scalability to other rewards**, especially those with non-negligible cost, such as AlphaFold2 confidence. - Similarly, the authors did not compare existing guidance methods like autoguidance and Diffusion-NPO as baselines to better support the superiority of the proposed approach. Other domain-specific guided generative models for peptides are cited (e.g., PepTri, PepMimic), but are never directly compared. - The paper optimizes Rosetta energy, which **does not yet establish improved binding affinity**. This supports successful optimization of that proxy, but offers limited independent evidence of binding quality. Indeed, the results in Tab 1 & 2 show a **tradeoff, rather than broad superiority**. Guidance considerably reduces sequence and structural diversity. AAR drops, C-RMSD increases, and both dihedral-distribution distances worsen. Together with the clear energy gains, this should be better interpreted as a trade-off. [1] Wang, F. Y., Shui, Y., Piao, J., Sun, K., & Li, H. (2025, May). Diffusion-npo: Negative preference optimization for better preference aligned generation of diffusion models. In International Conference on Learning Representations (Vol. 2025, pp. 71440-71462).
Questions for authors

Please refer to the above weaknesses.

Reviewer_9TfUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents PepContra for designing peptide sequences and full-atom structures for a given target protein. It fine-tunes two copies of a base model on high and low quality generated peptides, then uses their predictions to guide sampling.

Critical strengths
  1. The paper addresses a useful and challenging problem: designing peptide sequences and full-atom structures together.
  2. PepContra improves computational interface energy and reduces peptide-internal clashes relative to its base model.
Critical weaknesses
  1. Main contribution is hard to identify. The section of introduction does not clearly link each problem to the proposed method. The dual-model guidance in Figure 1 is close to prior work.
  2. Fine-tuning is unclear. Which parts of the model were updated? It also lacks a control with the same amount of extra training but without quality filters.
  3. I am concerned that the test set was used to choose the base checkpoint, the number of fine-tuning epochs, and the guidance settings. Final results should be checked on a separate set of targets.
  4. Energy scores and internal clashes improve, but diversity and some structure measures get worse. The authors should be cautious about claiming better binding without more direct evidence.
  5. The code is not available, and some training and sampling details are missing. This makes the results hard to reproduce and check.
Questions for authors
  1. Which parts of the base model are updated during fine-tuning?
  2. Were the base checkpoint, fine-tuning length, and guidance settings chosen using the reported test set?
  3. How much of the improvement comes from the high-quality model alone?
  4. Will the authors provide the missing training and sampling details so the results can be reproduced? 5. If they also address the questions above, I would consider raising my score.
Reviewer_ry6cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents PepContra, a full-atom peptide sequence–structure co-design method that uses models trained on high- and low-quality peptides to guide generation toward better peptides. It also introduces synchronized sampling to maintain side-chain torsion states as residue identities change. On 93 LNR complexes, PepContra reports improved Rosetta interface energy and IMP and fewer internal clashes than its base model, alongside declines in several recovery and diversity metrics.

Critical strengths
  1. Synchronized sampling addresses a real difficulty in full-atom peptide generation. When the predicted amino-acid identity changes during sampling, its side-chain torsion state must also change consistently. The proposed sampler maintains type-specific torsion states, and the reported comparison shows fewer torsion and amino-acid collapse events and reduced peptide-internal clashes.

  2. The HQ/LQ guidance improves some important design outcomes. On the 93-complex LNR benchmark, PepContra reports the strongest Rosetta interface energy and IMP among the compared methods. In the ablation, adding LQ guidance to HQ-only reduces internal clashes from 0.283 to 0.149, although the corresponding gains in energy and IMP are modest and several recovery and diversity metrics decline.

Critical weaknesses
  1. The claimed distributional justification for dual-flow guidance does not follow from the stated assumptions. In Equation (2), the conditions pH(xH)>pL(xH)p_{H}(x_{H}) > p_{L}(x_{H}) and pH(xL)<pL(xL)p_{H}(x_{L}) < p_{L}(x_{L}) do not imply pˉ(xH)>q(xH)\bar{p}(x_{H}) > q(x_{H}) or pˉ(xL)<q(xL)\bar{p}(x_{L}) < q(x_{L}), because the normalized density also depends on the partition function and on how pHp_{H} and pLp_{L} relate to qq.

  2. The evaluation is not independent, and the specific contribution of the LQ flow is not sufficiently isolated. Appendix B.2 and Appendix G use LNR to select the base checkpoint, HQ/LQ fine-tuning epochs, and all three guidance coefficients, while the main results are reported on the same benchmark. Thus, LNR is functioning as both validation and test data, and the final numbers may be affected by benchmark-informed selection. In addition, HQ-only already obtains ΔG=−28.61\Delta G = -28.61 and IMP=26.9\mathrm{IMP} = 26.9\\%, compared with ΔG=−29.51\Delta G = -29.51 and IMP=28.0\mathrm{IMP} = 28.0\\% for full PepContra in the 40-sample ablation. The dedicated LQ branch clearly helps internal clashes, but its additional energy and IMP gains are modest and are reported without uncertainty. Since the LQ set also combines two qualitatively different subsets and has a different length distribution from HQ, the experiment does not yet establish which signal the negative flow contributes. An untouched test set and a stronger reference such as HQ–Base guidance are needed to substantiate the central empirical claim.

Questions for authors
  1. What additional assumptions are needed for Equation (2) to guarantee the claimed changes in high- and low-quality sample probabilities relative to the data distribution after normalization? Can the authors provide a proof under those assumptions? If this result is intended only as motivation for a heuristic, would the authors state that more explicitly?

  2. Was any validation set separate from LNR used for checkpoint, fine-tuning-epoch, or guidance-scale selection? If not, can the authors select these quantities without LNR and report results on an untouched test set?

  3. Can the authors report a controlled HQ–Base guidance baseline, vH+γ(vH−vB)v_{H}+\gamma(v_{H}-v_{B}), and fuller statistics for the final HQ, LQ-part-1, and LQ-part-2 datasets? This would clarify whether a deliberately trained LQ flow is necessary and whether its effect is confounded by peptide length or the heterogeneous LQ construction.

  4. Can the authors repeat Tables 1–2 across multiple independent sampling seeds and report means and standard deviations or confidence intervals using the same number of candidates per target?

  5. How do the authors reconcile the statement that PepContra is best or second-best on almost every metric with the degradation from PepContra-base in AAR, C-RMSD, L-RMSD, DockQ, JSDbb, JSDsc, and both diversity metrics? Please also report these metrics for HQ-only and the other ablations so that the source of the trade-off can be identified. Compared with UniMoMo, PepContra improves Rosetta energy, IMP, DockQ, and JSDbb, whereas UniMoMo performs substantially better on clashes, JSDsc, and diversity; which operating regime do the authors consider preferable?

  6. How does this work differ methodologically from Dual Flows with Contrastive Guidance for Generating Highly Designable Proteins, which already uses HQ/LQ multimodal flows and contrastive guidance over protein sequence and structure? Please cite and position this closely related work and clarify whether the main methodological novelty here is the peptide/full-atom extension, synchronized sampling, or another component.

  7. Please report wall-clock duration and GPU-hours for base-model training and for HQ and LQ fine-tuning separately. This information is relevant to the characterization of the method as lightweight.

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

Related papers

Discussion (0)

Sign in to comment.