acceptodds
Under review as a conference paper at ICLR 2027

PPDiff: Position-Priority-Aware Masked Diffusion for Controllable Peptide Generation

Abstract

Controllable peptide generation requires careful refinement of residues that determine the requested properties. Masked discrete diffusion offers flexible generation trajectories, but uniform training corruption allocates reconstruction practice independently of downstream property relevance. We introduce PPDiff, a controllable generation framework that uses an external property planner at both stages. A lightweight sequence random forest with in silico mutagenesis estimates residue criticality and uncertainty. The resulting prior reallocates a fixed training mask budget and, during sampling, is combined with denoiser confidence, margin, and temporal stability to decide when each residue is ready to commit. The selected pretrained masked-diffusion backbone is retained without architectural changes, and models used for evaluation are separate from the guide. Across permeability, hemolysis safety, and nonfouling, PPDiff improves the joint objective under three separate sequence evaluators while supporting target-specific candidate design in computational screens against seven proteins. A multi-seed factorial study and a shuffled-prior control further show that the main gain arises when property-bearing position information coordinates both training corruption and sampling commitment.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide Design, Masked Discrete Diffusion, Controllable Generative Modeling, Property-Guided Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pVYxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper propose PPDIFF, a generation framework for controllable peptide geenration. The framework adopts a lightweight random forest to predict in silico mutagenesis to estimate residue importance and uncertainty. The leveraging the uniqueness of discrete diffusion model, the framework remasks the generated sequences for improving certain characteristics with guidance. This framwork is flexible on backbones and guidance. Experimental result demonstrate the effectiveness of PPDIFF.

Critical strengths
  • The idea of priority-gating sampling using a shared prior is an interesting idea and mathmatically sound. The result also demonstrate effectiveness of such method.

  • The experiments are comprehensive and complete. The paper not only demonstratre the effectiveness of the model but also illustrate the random forest performance.

Critical weaknesses
  • The motivation of controllable peptide generation is unclear to me. From my perspective, it is still considered as a broader view of conditioned generation or guided generation. So how to distinguish this work and existing work such as [1] and [2]? Except the technical difference, what is the advantage of the proposed solution?

  • The presentation of this paper needs a significant polish, especially the method section. I noted several symbols are not defined well. For instance, delta in Eq. 1. And typos such as Eq. 5 x_{\i}. Furthermore, the flow and illustration of this section is not clear to me. I found it is very hard to understand the logic and mechanism of each components.

  • This is not technically a weakness. Since ISM is adopted in inference stage, it is interesting to see the extra computational overhead of the RF model.

  • From table 28, It looks like PPDIFF lowers the validity, uniqueness and diversity comparing with baselines. I think this might because the bias introduced by the RF model. The author should acknowledge or conduct further analysis.

[1] Tang, Sophia, Yinuo Zhang, and Pranam Chatterjee. "Peptune: De novo generation of therapeutic peptides with multi-objective-guided discrete diffusion." ArXiv (2025): arXiv-2412.

[2]Chen, Leo Tianlai, Zachary Quinn, Madeleine Dumas, Christina Peng, Lauren Hong, Moises Lopez-Gonzalez, Alexander Mestre et al. "Target sequence-conditioned design of peptide binders using masked language modeling." Nature Biotechnology 44, no. 6 (2026): 1002-1010.

Questions for authors
  • What is the motivation of controllable peptide geenration? And the difference between existing guided generation model?

  • How much computational overhead introduced by the RF model? Especially in the inference stage.

  • Does the RF model introduce bias that eventually lowering the validity, uniqueness and diversity?

I put a weak rejection as a placeholder as of right now. But I am willing to increase my rating if the author address my concerns.

Reviewer_VRg2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PPDiff, a position-priority-aware masked diffusion framework for controllable peptide generation. The method uses external property predictors and in silico mutagenesis to estimate residue-level importance, which is then incorporated into both training and inference. During training, the model preferentially masks high-priority positions to emphasize their reconstruction. During inference, the estimated position importance is combined with denoiser confidence, prediction margin, and temporal stability to guide token commitment and remasking. The authors evaluate PPDiff on three peptide properties, including permeability, hemolysis safety, and nonfouling, and further extend the framework to target-specific peptide generation using PepMLM-based guidance. Experiments with multiple computational evaluators show improvements over several baselines, while factorial experiments and ablation studies investigate the contributions of priority-aware training and sampling.

Critical strengths
  1. The paper explores an interesting and relatively underexplored direction in masked diffusion models: using property-aware position importance to guide both training corruption and inference sampling. The overall motivation is clear, and the proposed framework is conceptually reasonable and does not require architectural modifications to the diffusion backbone.

  2. The experimental evaluation is relatively comprehensive, including comparisons with existing methods, extensive ablation studies, multi-seed experiments, and evaluations using separate property predictors. In particular, the factorial experiments and shuffled-priority controls provide useful evidence that incorporating property-relevant position information into both training and sampling can improve generation performance beyond applying either component alone.

Critical weaknesses
  1. The reliability of the estimated residue priorities during inference remains a concern. PPDiff computes position importance from intermediate full-sequence proposals generated by the denoiser. However, these proposals may be highly unreliable during early denoising, and the estimated importance of a residue can depend strongly on its surrounding sequence context. It is therefore unclear whether these priorities provide meaningful guidance throughout the generation process. Furthermore, although the evaluation models are separate from the Random Forest guides, they may still share similar data biases, label characteristics, and sequence-level shortcuts. This raises concerns about correlated prediction errors and potential reward hacking, making it difficult to determine whether the reported improvements reflect robust property optimization.

  2. The deterministic Top-K masking strategy substantially restricts the diversity of corruption patterns during training. For a given sequence and noise level, the same high-priority positions are always masked. This may limit the model's ability to learn low-priority positions conditional distributions. More importantly, the effectiveness of this strategy depends heavily on the accuracy of the property predictors, since both training and inference rely on their estimated residue priorities. Inaccurate priorities could therefore affect both stages and potentially reinforce the same errors. Although the paper includes guide-degradation experiments, these do not separately quantify the sensitivity of training and inference to predictor errors, nor do they adequately investigate robustness under realistic distribution shifts. This leaves the reliability and generalizability of the proposed mechanism insufficiently established.

Questions for authors
  1. What motivates the choice of Random Forest models as property predictors instead of more expressive alternatives, such as pretrained protein language models? How would different predictor architectures and their accuracy affect the estimated residue priorities and downstream generation performance?

  2. The method introduces substantial computational overhead due to repeated in silico mutagenesis and predictor evaluations, especially for target-conditioned generation with PepMLM. How do the authors justify this additional cost, and are there more efficient alternatives?

  3. The experimental comparisons focus heavily on the DPLM backbone and ablation studies. Although PAPL, PepTune, and classifier guidance are included, could the authors provide broader comparisons with other closely related controllable-generation methods to better establish the advantages of PPDiff?

I am open to raising my rating if the authors could address the comments above.

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

This paper presents PPDiff and provides an interesting approach to control the training and generation trajectories of masked diffusion models. It uses a property predictor to determine which positions matter most and uses the selected positions to prioritize the mask ordering. Such procedure encourages the diffusion model to focus more on the property-relevant positions.

Critical strengths
  1. PPDiff provides a straightforward way to integrate property info in training and sampling, without modifying the diffusion backbone.
  2. The abalation study is thorough and demonstrate the usefulness of the proposed strategy.
Critical weaknesses
  1. The criticality formulation is not robust enough. As defined in Eqn. 2, the criticality is formulated as the negative mean property score. However, an important position with, say, different substitutions that have large but opposing effects can lead to a low priority score. The author did not explore alternative candidates like mean absolute score, or some variance adjusted scores.
  2. It is unclear whether the training and sampling distributions are truly aligned. Training encourage the model to treat positions with high priority as reconstruction targets, while the same position tends to be treated as the context in sampling. Although remasking may mitigate this issue to some extent, there is no gurantee that the training and sampling distribution are aligned as the authors claim. I is also unclear that how reliable is the priority determined based on masked sequence in sampling, since it seems the priority is always computed from the raw sequence in training.
  3. The computational cost of the planner limits scalability. PPDiff repeated query external planners, from relative easy RF to expensive PepMLM. The computational cost, espefically for target conditioned tasks, make the current model hard to scale to longer sequence or larger sets of protein targets.
Questions for authors
  1. Can the authors compare different criticality formulation options and justify the current choice?
  2. Can the authors justify the alignment of training and sampling distributions?
  3. How reliable is the priority determined in the sampling process?
  4. Can the authors provide a computational cost comparison with different peptide lengths, especially for target-conditioned tasks?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.