acceptodds
Withdrawn from ICLR 2027

Pepti-drift: Scalable Safe-Active Peptide Generation Without Inference-Time Guidance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Therapeutic peptides are a promising drug modality, but their generation must satisfy multiple therapeutic constraints. We introduce BindSafe-PepBench, a fixed-budget benchmark that jointly evaluates target binding and four major safety metrics on the same generated candidates. We reveal that peptide length is a major confounder of joint binding–safety evaluation: longer peptides tend toward stronger predicted binding but less favorable predicted safety. This creates an apparent trade-off and can bias comparisons among models with different output-length distributions. We report absolute Safe-Active yield and exact-length-matched gains to distinguish generative improvements from output-length effects. High Safe-Active yield remains challenging, while the strongest multi-property methods rely on costly inference-time guidance. We therefore introduce Pepti-drift, a one-step generation framework that incorporates attraction toward target-specific binders and repulsion from liability-associated regions, requiring a single latent refinement followed by parallel decoding without inference-time guidance. Across 88 held-out targets, Pepti-drift achieves an 18.37% predicted Safe-Active yield while retaining positive exact-length-matched gains. The resulting gains are competitive with multi-property-guided baselines while requiring 468× lower generation cost, enabling scalable and fair high-throughput peptide design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI4Science, Therapeutic Peptide Design, Drug Discovery, Generative Models for Biology

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WLV7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces BindSafe-PepBench, a benchmark for evaluating target-specific peptide generation under joint binding and safety constraints, and Pepti-drift, a one-step generative framework intended to eliminate expensive inference-time guidance.

The authors identify peptide length as a major confounding factor in predicted binding–safety trade-offs and introduce exact-length-matched enrichment alongside raw Safe–Active yield. Pepti-drift uses frozen ESM-2 representations, an attraction–repulsion objective, and a non-autoregressive Transformer decoder. On 88 held-out targets, the method achieves 18.37% predicted Safe–Active yield, comparable to PepTune's 18.45%, while reporting approximately 468× lower generation cost. The experiments include comparisons with existing generators, component ablations, alternative prediction oracles, and predicted structural analyses.

Critical strengths

1. A useful joint binding–safety benchmark.

BindSafe-PepBench evaluates binding and multiple safety constraints on the same candidates under a fixed generation budget. The length-confounding analysis is particularly valuable, and the robustness checks across alternative prediction oracles strengthen this observation.

2. Competitive predicted performance with substantially lower generation cost.

Pepti-drift achieves nearly the same predicted Safe–Active yield as PepTune while requiring much less generation time. Avoiding repeated inference-time guidance is practically appealing for large-scale candidate generation.

Critical weaknesses

W.1 Limited methodological contribution. Removing the drifting block barely changes Safe–Active yield (17.80% vs. 17.31%, Table 2), questioning whether the proposed mechanism provides meaningful improvements.

W.2 Questionable repulsion mechanism. The distance normalization and temperature in Equations 20–21 make hard-negative weights nearly uniform, weakening the claimed distance-adaptive repulsion.

W.3 Unreliable evaluation. The benchmark relies entirely on computational predictors with limited accuracy. Even length-matched random peptides achieve 10.27% Safe–Active yield, raising concerns about the endpoint's biological relevance.

W.3 Insufficient experimental evidence. The reported gains are modest (+2.65 percentage points over length-matched random controls), and the main comparisons lack multiple training seeds and statistically convincing evidence of superiority.

Questions for authors

Q.1 Can the authors explain why removing the drifting block has almost no effect on performance (17.80% vs. 17.31%)?

Q.2 Given the distance normalization in Equations 20–21, how does the repulsion mechanism effectively distinguish hard negatives?

Q.3 How reliable is the Safe–Active benchmark when random peptides already achieve 10.27% success? Can the authors provide independent validation?

Q.4 Can the authors provide multi-seed experiments and stronger statistical evidence to support the modest performance improvements?

Reviewer_6fAoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The main argument of the paper is very clear: therapeutic peptide generation is not only about generating candidates that bind to specific sites of disease-related proteins, but also about ensuring that the generated candidates satisfy certain safety requirements, such as toxicity, hemolysis, etc. The generation task should account for both (safe-active): a safe but non-binding peptide is ineffective, and a toxic binding peptide is not a viable drug candidate. This motivates BindSafe-PepBench, as a new benchmark, which evaluates binding and four major safety liabilities jointly on the generated candidates. Authors then argue that existing methods which are tailored to optimize generation to multiple therapeutic objectives (through guidance, steering, etc) are computationally expensive at inference time, and introduce Pepti-drift, a framework that integrates multi-objective control into training to achieve fast inference. The idea is simply through attraction toward target specific binders and repulsing from safety liabilities.

Critical strengths
  • Meaningful, important, and timely study.
  • BindSafe-PepBench benchmark and the accompanying evaluation potentially very useful for the broader community.
  • The observation of peptide length as a confounder is important, which may be obvious to domain experts, but it doesn’t seem to be highlighted in the AI literature or benchmarks so far.
  • Pepti-drift inference is very fast, more than two orders of magnitude faster than existing guidance-based methods.
  • While simple and straightforward, the attraction–repulsion training objective is a meaningful approach.
Critical weaknesses
  • The approach is somewhat incremental and the results are not better in any meaningful way than some strong baselines in safety and binding (e.g., PepTune). It is much faster at inference time, which is the core benefit, but the presented results are not very strong in terms of joint predictive performance. I find this valuable, but the ablations are not convincing in my opinion. For example, removing the drifting block does not seem to present a substantial degradation based on a single seed run (according to Table 2). Since the drifting idea is central to the paper, its role needs to be rigorously validated. Also, in my understanding, removing drifting block does not necessarily mean removing the attraction-repulsion objective (Appendix A4). So it could very well be that this objective is key in the design and drifting block is not adding much to the picture. I would have expected a controlled study, i.e., removing the drift module, alignment loss, and repulsion term, with several training seeds and paired comparisons. All in all, the experiments do not currently support the claim that drift helps with then joint-objective training, and the architectural choices are not fully validated/justified through the presented experiments.
  • Pepti-drift achieves an 18.37% SAFE–ACTIVE yield with a +2.65 percentage-point exact-length-matched enrichment. This means that for every 100 candidates, Pepti-drift produces about 18 predicted successes, while its matched random controls produce about 16. This makes me feel like the results are strongly overstated. In the appendix, authors state that “The exact-length-matched random control reaches 10.27% raw SAFE–ACTIVE yield”, but it is not clear to me to which random baseline this refers to, and how it relates to the baseline in the main result. Each method is compared with a different lenght-matched control, which makes sense, but it is difficult to understand the full picture from the reporting.
Questions for authors

Please see the weaknesses.

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

This paper introduces BindSafe-PepBench and Pepti-drift for target-conditioned peptide generation with joint binding and safety objectives. The benchmark evaluates whether the same candidate satisfies predicted binding criteria and four safety constraints under a fixed generation budget. It also introduces exact-length-matched random controls and finds that peptide length largely explains the observed negative association between predicted binding and safety. Pepti-drift combines frozen protein language model representations, a target-conditioned latent generator, a single learned drift update, and a parallel sequence decoder. Training uses attraction toward reference binders and repulsion from toxicity- and hemolysis-associated peptides, without property-model guidance during inference. Across 88 held-out targets, the model achieves an 18.37% predicted SAFE-ACTIVE yield, close to PepTune’s 18.45%, with positive length-matched enrichment and approximately 468-fold lower measured generation cost.

Critical strengths
  1. BindSafe-PepBench evaluates predicted binding and four safety liabilities on the same candidate under a fixed generation budget. Counting unsuccessful and unusable outputs against that budget makes the reported yield more informative than separate marginal property scores.
  2. The paper shows that the apparent negative association between predicted binding and safety is substantially reduced after exact-length stratification. The consistency of this observation across multiple evaluator configurations strengthens the motivation for length-controlled comparisons.
  3. A favorable empirical quality–generation-cost trade-off. Pepti-drift achieves a predicted SAFE–ACTIVE yield close to PepTune while requiring substantially less measured generation time and retaining positive exact-length-matched enrichment. This demonstrates the practical potential of learning property-aware generation without iterative inference-time guidance.
Critical weaknesses
  1. The independent contribution of drifting is not sufficiently established. In Appendix A.5, removing the drifting block reduces the SAFE–ACTIVE rate only from 17.80% to 17.31%. This single-seed comparison does not establish a consistent benefit beyond the underlying generation architecture. A parameter-matched non-autoregressive baseline and paired comparisons across multiple training seeds are needed to determine whether the improvement is attributable to drifting rather than additional model capacity or training variability.
  2. The robustness of the joint-performance ranking to safety-evaluator choices remains unclear. The cross-oracle analyses provide useful evidence that the length-confounding observation persists across evaluator configurations. However, they do not establish that the reported SAFE–ACTIVE yields or method rankings are similarly robust. Given the moderate performance of some safety classifiers and the reported disagreement between evaluators, the joint endpoint may be sensitive to the selected predictors and decision thresholds. Recomputing the main comparison under alternative evaluators and thresholds would strengthen the evidence for the claimed quality advantage.
Questions for authors
  1. Does drifting provide a consistent improvement over a parameter-matched non-autoregressive model across multiple training seeds? Please report paired differences in SAFE–ACTIVE yield and length-matched enrichment, with uncertainty estimates that account for training variability.
  2. Does the implementation divide squared distances between normalized vectors by 1024 and use a softmax temperature of 0.1, as specified in the paper? If so, how concentrated are the resulting negative-reference weights, and does distance weighting improve performance over uniform weighting of the selected hard negatives?
  3. How stable are SAFE–ACTIVE yields and method rankings when the safety evaluators or decision thresholds are changed? Please distinguish robustness of the length-confounding finding from robustness of the model comparison, and report whether Pepti-drift’s relative performance persists under these alternative evaluation settings.

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

Related papers

Discussion (0)

Sign in to comment.