acceptodds
Under review as a conference paper at ICLR 2027

ApexMO: Edit-Constrained Multi-Objective Optimization of Generative Peptide Proposals

Abstract

Generative sequence models produce diverse peptide proposals, but refining a proposal requires balancing competing properties while controlling how far its sequence changes. We present ApexMO, a computational framework that combines length-conditioned discrete flow matching with anchor-relative, edit-constrained Pareto optimization. Generated peptides provide fixed anchors for local refinement of predicted antimicrobial activity, toxicity, and hemolysis. Under matched raw-attempt budgets, the flow-matching generator yielded approximately 1.8 times as many sequences meeting the specified novelty, predicted-activity, and ensemble-dispersion criteria as each of the two tested baseline systems. Across five refinement runs, all-survivor hypervolume of predicted activity, toxicity, and hemolysis increased by 26.8–27.9%. Under equal evaluation budgets, Pareto-based search achieved higher hypervolume than random local search, both for the final population and for all feasible solutions evaluated during a run. Adding crossover did not improve the median relative to mutation-only search. An edit-budget ablation quantified the trade-off between predicted objective improvement and anchor proximity. A screening cascade prioritized 35 DBAASP-predicted AMP candidates from 25 generated anchors, with recorded sequence ancestry. These results are based on model predictions and do not establish experimental antimicrobial activity or safety.

Primary area:
optimization
Keywords:
Peptide design, multi-objective optimization, discrete flow matching, constrained sequence optimization

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

This paper proposes ApexMO, an framework to generate, filter, and refine peptide. ApexMo adopts a length-conditioned discrete flow matching model to generate 63,000 proposals. Then an anchor selection algorithm to filter and select 500 anchor candidates. A Refinementi algorithm iteratively refine the selectaed anchors. Finally, a screen algorithm ranks and clusters to have the 25 candidates.

In the experiemnt, the paper benchmarks 3 different generation algorithm, flow matching, AMP-Diffusion, and ARCADIAMP in generation quality and generation throughput. Then it evaluate the MIC predictor performance. The paper finaly benchmarks the pareto refinement improvement and the edit budget constraints.

Critical strengths
  • The general idea of generate -> anchor select -> Refine -> ranking is an interesting pipeline. This looks to me can be extended into a framework for drug discovery.

  • Comparing different generation algorithm provides an informative perspective to the community. It shows that flow matching balances the throughput and quality.

Critical weaknesses

The motivation of certain components in the pipelien of this paper remains unclear to me.

  1. Why budget constraint refinement is necessary. Since the process is completely in silico, it does not makes sense to me to save budget in this process.

  2. The second and third research questions are not specific and deep. For instance, "does constrained multi-objective selection improve the final population relative to matched local-search controls?" This looks to me can be universally applied to any objective-condition generation task. So it is important to explain why this research is important. What we can learn from studying this research question. Similar argument can be applied to the third research question.

  3. The method section lacks a high-level explanation to the proposed workflow (figure 1). I found it hard to understand the bigger picture and background of this section. Tons of notations lacks definition and explanation, such as f(x) in Eq. 1. Is it a energy function or loss to minimze? Similarly what is B(L) in Eq 2?

  4. The paper mentions the flow matching model is condifitoned on peptide length but did not mention how to achieve this or cite any paper.

  5. The refinement process is a mixing of insertion, substituting, and replacing in a random manner. This looks very brutal force and lacks insignt? For instance, there is no explanation for replacing glutamic acid into Leucine (Figure 4). However, scientis perform such replace for a certain reason, i.e., increasing half life.

Questions for authors

Please refer to my weakness section. In the rebuttal, I would like to see more discussion in motivation and the justification of the key research questions.

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

The paper studies methods for multi-objective optimization of peptide generative modeling under the edit distance constraint. The proposed method include a peptide generative model based on discrete flow matching to generate initial batch of candidate modelcules. The generation processes are constrained by the maximum edit distance. NSGA-II is used for multi-objective optimization while generating new candidates via evolutionary action such as crossover, where the goal of is to optimise multiple objectives including activity, low toxicity, hemolysis.etc.

I the experiments, the authors show that NSGA-II improves the parreto optimal solution via increase of the hyper volume compared to basic baselines such as local search.

Critical strengths

The given application is important and practical but the proposed methods are not novel.

Critical weaknesses

The proposed method is a combination of known techniques including discrete flow matching generative models, NSGA-II evolutionary optimization, therefore in terms of novelty the paper has limited contribution. For a ML conference I would expect more on proposal of new concepts and methodologies that are non-trivial combination of existing approaches.

The work demonstrated some improvement over local search by using NSGA-II, this result is not surprising and thus stronger baselines should be considered such as search based on Bayessian optimizations or Reinforcement learning coupled with a generative model.

Questions for authors

Please clarify novelty, and add comparison to strong baselines.

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

The paper proposes an antimicrobial peptide design process called ApexMO, which involves”generation,anchoring,multi-objective-optimization,screening”. First, a length-conditional discrete flow matching model is used to generate peptide sequences, and 500 fixed anchor points are selected based on criteria such as novelty, applicability, and prediction uncertainty. Then, within the Levenshtein editing budget for each anchor point, NSGA-II is used to simultaneously optimize the prediction of MIC, toxicity, and hemolytic activity. Finally, clustering, cross-predictor screening, DBAASP discrimination, and lineage recording are performed, resulting in 35 candidates. Under equal evaluation budgets, Paretobased search achieved higher hypervolume than random local search, both for the final population and for all feasible solutions evaluated during a run。These results are based on model predictions and do not establish experimental antimicrobial activity or safety.

Critical strengths

The article is clear and easy to read, and the experimental design is simple and easy to reproduce.

Critical weaknesses

1、The paper does not adequately demonstrate substantial method gains relative to existing strong methods. Flow matching, mutation with edit constraints, NSGA-II, Pareto sorting, and clustering are all mature components, and related works also cite ApexGO, LaMBO, MosPro, PepTune, and MP2D, but the experiments only compared random local search with mutation-only NSGA-II.The paper's core claim—what exactly is added by anchor points relative to edit-constrained Pareto optimization—is not proven by strong baselines and ablation analysis.

2、The optimization objectives and most evaluation metrics rely on the APEX series of predictors; furthermore, the tools subsequently employed—APEX-pathogen, DBAASP, ToxinPred3, and HemoPI2—share correlations with these primary predictors in terms of training data or underlying knowledge sources. Consequently, an increase in hypervolume merely indicates that the optimizer has successfully navigated the search space defined by a fixed surrogate model; it does not constitute proof that the refined peptides exhibit genuine improvements in antimicrobial activity or safety. Independent predictors are required to verify whether these computational gains translate into actual biological performance.

Questions for authors

1、The core methodological claim of the paper is that the performance gains stem from anchor-relative, edit-constrained Pareto optimization. However, the current search baselines are limited to mutation-only NSGA-II and random local search. Could the authors provide a comparison—under the same fixed-anchor constraints and evaluation budget—against at least one existing, robust multi-objective sequence optimization method, such as an adapted version of ApexGO, LaMBO, MosPro, PepTune, or MP2D? Furthermore, could the authors conduct a factorial ablation study—isolating the parent selection and survival mechanisms—to determine whether the observed gains are specifically attributable to Pareto survival, tournament selection, edit constraints, or the anchor-preservation mechanism?

2、The performance gains reported in the paper are primarily measured using the APEX family of predictors employed during the construction of the optimization objective. Could the authors conduct a post-optimization evaluation using a truly independent predictor—one that shares no overlap with the optimization surrogate in terms of training data, model family, or features? If conducting prospective experiments is not currently feasible, could the authors at least report whether the optimized population still outperforms the corresponding anchor points when evaluated using this external model, and whether the ranking consistency—relative to the internal predictor—is maintained?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.