acceptodds
Under review as a conference paper at ICLR 2027

Generalized Peptide Design with Physically Inspired Diffusion

Abstract

Peptides, as short amino acid polymers, have attracted increasing attention due to their favorable pharmacological properties. Recent advances in deep learning have greatly facilitated the design and discovery of peptides. However, most existing generative approaches rely primarily on static structural representations, and the generated peptides generally adhere closely to patterns in the training data, which inherently limits the exploration of functionally diverse conformations and weakens generalization capability to unseen scenarios. Inspired by the dynamic physical process of molecular docking, where hotspot regions form key interactions and flexible regions enable adaptive fit, we propose two complementary strategies for peptide generation: (i) **Flexibility-Conditioned Out-of-Distribution Exploration**, which encourages the model to dynamically explore the conformational landscape based on structural flexibility, thereby promoting high novelty; and (ii) **Hotspot-Driven Hierarchical Affinity Guidance**, which steers the exploration process toward high-affinity binding modes by prioritizing key interactions at hotspot regions. Integrating these strategies, we introduce **PhyDiff**, a physically informed diffusion-based framework that adaptively explores the chemical space while simultaneously discovering high-affinity binding patterns, thereby producing peptides that exhibit both strong binding affinity and novelty, and ultimately enhancing the model's generalization capability in real-world drug design scenarios. Extensive experiments across diverse peptide design benchmarks consistently demonstrate the effectiveness and generalization capability of our approach.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion model, peptide design, AI for science

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Z4FZOpenReview ↗
Rating
4(+2)
Confidence
2(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces two techniques for peptides design which are supposed to bind to a given protein target. One of them is Flexibility-conditioned Out-of-Distribution Exploration (FlexOOD), which is supposed to enable wider exploration for peptides which are more flexible. Second technique is Hotspot-Driven Hierarchical Affinity Guidance (HHAG), which guides the generation towards high-affinity peptides and more probable hotspot residues. The authors validated their methods with computational experiments on de novo peptide design. In all cases their method is better than beslines or comparable. Ablation studies shows that FlexOOD improves diversity and HHAG improves quality of generated peptides.

Critical strengths
  1. The Authors propose novel property-conditioned Out Of Distribution exploration. The authors as conditioning property use flexibility, but the formulation is general, so any property can condition Out Of Distribution exploration.
  2. The Authors adapt classifier guidance technique to guide the model toward higher overall binding affinity.
Critical weaknesses
  1. The Authors don’t measure the impact of “Hotspot-level local guidance” and “Peptide-level global affinity guidance” separately. It seems that “Hotspot-level local guidance” term is always 0 – see Question 2).

  2. The Authors propose mechanism FlexOOD, which leverages structural flexibility to explore beyond the training distribution, but they don’t argue why exactly flexibility is the property that should steer beyond the training distribution. It seems arbitrary and unjustified.

In the introduction The Authors say: “Most mainstream approaches overlook peptide flexibility and the dynamic nature of protein-peptide recognition, both of which are recognized as essential for accurate binding. [...] This dynamic interplay is a hallmark of high-affinity binding.” However, FlexOOD does not incorporate flexibility into affinity binding prediction or high-affinity guidance. Instead, flexibility drives OOD exploration specifically, which again appears insufficiently justified.

The authors didn’t sufficiently analyze the importance and behavior of conditioning OOD exploration specifically on flexibility property - see Questions 1)

Questions for authors
  1. Why is it beneficial to have an adaptive OOD weight conditioned on flexibility specifically? Can you compare it with constant η\eta, which equals for example the average flexibility?
  • How does flexibility prediction vary along the generation trajectory for a given target? Does it increase/decrease/is constant? What is the variance of the predictions at the beginning of trajectory or at the end?

  • There is also a drawback that when OOD weight is enlarged then the flexibility predictor also gets OOD samples so its predictions are unreliable. Is it possible to somehow measure such behaviour?

  1. In the description of equation (11) there is writen that the predicted hotspot residue set H\mathcal{H} is obtained based on the EG\mathcal{E}_G prediction. The input of EG\mathcal{E}_G is (Grec,G^tpep)(\mathcal{G}^{rec}, \hat{\mathcal{G}}^{pep}_t), so H\mathcal{H} is a subset of indices {1,…,∣G^tpep∣}\{1, …, |\hat{\mathcal{G}}^{pep}_t| \}, so G^hot\hat{\mathcal{G}}^{hot} is a subset of G^tpep\hat{\mathcal{G}}^{pep}_t. And because of that xihot=x^ix_i^{hot} = \hat{x}_i and sihot=s^is_i^{hot} = \hat{s}_i, so the value of (11) is always zero. Is it true?
  • If it is not always zero, then why? What is the effect of “Hotspot-level local guidance” and “Peptide-level global affinity guidance” to the overall performance of PhyDiff? It would be good to see this effect in ablation studies, not necessarily setting γ=γ1=γ2\gamma = \gamma_1 = \gamma_2
  1. The Authors measure Sequence Validity (SeqValid) as “the proportion of generated sequences containing only valid amino acid identities (no unknown or invalid tokens)”. How is it possible that Decoder outputs invalid amino acid identities?
  • Additionally, The Authors report that they follow PepTri (Nguyen et al., 2026) evaluation setup, however in PepTri they have a different definition of Sequence Validity – “fraction of generated sequences passing biochemical criteria”. Why did The Authors change the definition of this metric?
  1. Are the values in Table 3 reported with or without relaxation? Without such information it is difficult to compare it to Table 1.

  2. Is it correct that the experimental setup in Section 4.2, “Generalization to Unseen Targets,” differs from that in Section 4.1, “Experiments on De Novo Peptide Design,” only in the test dataset used? If so, it would be beneficial to report all metrics from Section 4.1 on the PepPro-96 test dataset as well.

Reviewer_5o5sOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PhyDiff, a physically informed diffusion framework for peptide design. It incorporates two strategies: Flexibility-Conditioned Out-of-Distribution Exploration to promote structural novelty and Hotspot-Driven Hierarchical Affinity Guidance to improve binding affinity. Together, these strategies aim to enhance peptide design performance and generalizability.

Critical strengths
  1. Incorporation of Molecular Dynamics Information: Leveraging molecular dynamics (MD) data to train a flexibility predictor and subsequently using it to guide peptide generation is an interesting idea that introduces physically motivated information into the diffusion process.
Critical weaknesses
  1. Limited Methodological Novelty: Although PhyDiff incorporates molecular flexibility and hotspot information into peptide generation, both diffusion guidance and hotspot-based design have been extensively explored. Moreover, MD information has already been utilized in previous peptide generation methods. The proposed approach primarily introduces additional guidance signals into an existing diffusion framework. The authors should better clarify the methodological novelty compared with prior work.
  2. Misleading Discussion of Related Work: The "Hotspot-Driven Methods" section discusses AlphaProteo and PepHAR, where AlphaProteo's hotspots are defined on the target protein, whereas PhyDiff and PepHAR define hotspots on the peptide side. These formulations have different conditioning mechanisms and objectives. The authors should clarify this distinction.
Questions for authors
  1. Comparison with NeoPep: Could the authors compare PhyDiff with NeoPep [1], which also incorporates MD information into peptide generation through MD-derived training data? A discussion of the methodological differences would help establish the advantages and novelty of the proposed approach, and it is better if an experimental comparison could be established.

  2. Effectiveness of Flexibility-Conditioned Exploration: Could the authors compare the proposed flexibility-conditioned guidance with a simple time-dependent schedule, such as (λ(t)=λbaseα(t)\lambda(t)=\lambda_{\mathrm{base}}\alpha(t))? Since the flexibility predictor operates on estimated denoised structures, its predictions may be unreliable at early diffusion timesteps. A simple annealing schedule might achieve a similar exploration effect without requiring an additional predictor. Such an ablation would help establish whether flexibility information provides benefits beyond standard time-dependent guidance.

  3. Definition of Hotspot Residues: If the hotspot is to promote specific interactions, would it not be more appropriate to condition on the amino acid identities and spatial coordinates of peptide hotspots rather than only their positions? Could the authors clarify the motivation for this design choice? In pepHAR, they got the hotspots first, and then design the scaffold to match the hotspots.

  4. Reliability of the Affinity Predictor: Does the affinity predictor use estimated denoised structures as input during reverse diffusion? The Appendix only specifies the input formulation for the flexibility predictor. If so, since the affinity predictor is trained on ground-truth structures, its predictions may be unreliable on intermediate denoised structures, particularly at early timesteps. Could the authors report the Spearman correlation between predicted affinity and Rosetta-computed energy for estimated denoised structures at different diffusion timesteps to validate the reliability of the guidance signal?

  5. Quantitative Validation of Hotspot Guidance: The authors claim that HHAG improves binding by focusing on key interactions at hotspot regions. Could the authors provide quantitative evidence supporting this mechanism, such as comparing hotspot-specific intermolecular contacts or interaction energies with and without HHAG?

References

[1] Li, M., Liu, Y., Zhong, J., Shi, X., Zheng, J., Wang, K., ... & Zhang, J. (2026). Accurate de novo design of peptides from programming biophysical landscape. bioRxiv, 2026-06.

Reviewer_op4eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhyDiff, a physically inspired diffusion framework for protein-conditioned peptide sequence–structure co-design. Motivated by the roles of conformational flexibility and hotspot-mediated interactions in protein–peptide recognition, the authors introduce two guidance strategies, FlexOOD and HHAG. The authors evaluate PhyDiff on de novo peptide design, scaffold-constrained generation, and cross-domain generalization tasks. Compared with existing methods, including PepGLAD and PepTri, PhyDiff achieves improvements in predicted affinity scores, structural quality, and generation diversity across the evaluated benchmarks.

Critical strengths
  • The paper addresses a relevant and relatively underexplored direction in peptide generation. Peptide flexibility and hotspot-mediated interactions are important considerations in practical peptide drug discovery, yet their explicit incorporation into generative models remains relatively underexplored.

  • The proposed method demonstrates improvements across multiple peptide design benchmarks. PhyDiff achieves competitive or improved performance compared with existing methods.

Critical weaknesses
  1. Insufficient justification and validation of the proposed guidance strategies. While peptide flexibility and hotspot-mediated interactions are important in practical drug discovery, I am not fully convinced that incorporating these concepts through the proposed guidance mechanisms is sufficiently justified. In particular, it is unclear why greater peptide flexibility should necessarily lead to stronger out-of-distribution exploration, or whether the hotspot-based and affinity-based guidance terms provide complementary benefits. The current ablation studies demonstrate improvements from adding these modules, but do not sufficiently establish the necessity of the specific physical signals or their formulations.

  2. Potential evaluation bias and overstated claims regarding binding affinity and practical applicability. I find claims such as “high-affinity peptide binders” and “improved practical applicability in drug design” too strong without experimental validation. The affinity predictor used in HHAG is trained on Rosetta-derived binding scores, while two primary evaluation metrics, ΔG and Success Rate, are also based on Rosetta. This raises concerns about evaluation independence, as the observed improvements may primarily reflect optimization toward the same scoring framework rather than genuinely improved binding affinity. Although the computational results are encouraging, they do not directly establish that the generated peptides exhibit stronger binding affinity or greater practical value in drug discovery. I therefore believe the authors should moderate these claims or provide additional independent validation to support them.

Questions for authors
  • The paper introduces peptide flexibility and hotspot-mediated interactions as its two central motivations. However, the current evaluation primarily focuses on conventional metrics such as predicted binding scores, structural quality, and generation diversity, without directly assessing whether the generated peptides exhibit improved flexibility-related properties or hotspot interactions. Could the authors provide dedicated evaluations to demonstrate that FlexOOD and HHAG actually improve the corresponding physical properties of the generated peptides, rather than merely improving overall benchmark performance?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.