acceptodds
Under review as a conference paper at ICLR 2027

Learning from Experience for Multi-Objective Antimicrobial Peptide Design

Abstract

Antimicrobial peptides (AMPs) are promising candidates for drug-resistant infections, but practical design must balance antimicrobial activity, pathogen-specific potency, hemolytic safety, and toxicity. Conflicts among these objectives make it difficult to identify peptides with balanced property profiles. Navigating such trade-offs motivates the reuse of prior design experience under different objective bottlenecks. However, existing methods do not explicitly leverage successful and failed design cases as reusable experience during iterative optimization. To address this gap, we introduce PepMind, an experience-guided framework for iterative multi-objective AMP optimization. PepMind converts evaluated designs into structured records containing sequences, objective outcomes, design rationales, and failure diagnoses. A learned router is used to select the type and amount of experience according to the current optimization state for targeted sequence refinement, with new outcomes subsequently added to memory. Experimental results consistently demonstrate PepMind's effectiveness: it achieves the best performance in 19 of 20 evaluation settings and improves four-objective hypervolume (HV) by an average of 67.1% over the strongest external baseline. Further analyses confirm cross-pathogen experience transfer and favorable predicted properties of the generated peptides. These results support persistent, state-conditioned experience reuse as an effective strategy for adaptive multi-objective peptide design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antimicrobial Peptide Design, Multi-Objective Optimization, AI Agents, Long-Term Memory, Experience-Guided Optimization, Biological Sequence Design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nFj5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper proposes a lightweight and cheap generative framework for designing AMPs with multiple desired properties. The paper is well written overall, provides many experiments, and presents some strong results. However, several things need to be addressed. The novelty lies in the addition of the router, which is a small MLP that provides which category of past experience to retrieve and how many records to retrieve from it. The strong results can not be fully supported because several baselines, ablation studies, and justifications are missing.

Critical strengths
  • The method is cheap since the generator is frozen. They train only one small MLP.
  • Performance is shown for many pathogens across several property subsets.
  • The cross-pathogen transfer experiment is very interesting and should be highlighted more in the paper.
Critical weaknesses
  • The only extra component compared to the SOTA is the router, and its type head seems to favor potency failure in many states, as given in the appendix. I believe the authors should explore whether a simple rule that always retrieves potency failure experience could give the same result.

  • The paper claims that the structured experience is what gives the improvement. However, in the ablation they provide comparisons only between full memory and no memory at all. This removes the whole history, not only the structure. I would suggest that the authors give the model the previous sequences and their scores as plain text, without the six fields, without the rationale and without the router. Without this, the results can also be explained by the LLM simply seeing what was already tried.

  • The authors should include in the baselines (i) a protein language model with rejection sampling under the same predictors and budget, and (ii) the same model fine-tuned on AMPs.

  • The thresholds are fixed and they also seem quite low, given that the score is in [0,1]. A sensitivity analysis is needed.

  • All the properties are scored from external predictors and the system keeps editing the sequence until the predictors are satisfied. So the whole method is optimized against them. These predictors also have their own error, and since PepMind makes repeated targeted edits against their scores, part of the improvement may come from exploiting their errors rather than from better peptides. The LLM evaluation does not add much either because the evaluators are asked to look at the features that were deliberately engineered, which is probably why the scores are high.

  • It would also help to compare the generated peptides against a set of known AMPs from DBAASP, on pLDDT, length, net charge, hydrophobicity and amino acid composition, with a statistical test on the distributions.

  • Finally, I would suggest providing an analysis of novelty and diversity to known AMPs and sequence similarity within the generated data. Without those, it is not clear whether the method produces novel peptides or rewrites known ones, and whether the 300 candidates per task are diverse or variations of a few templates. Diversity is mentioned, but it is not discussed how it is computed.

Questions for authors
  1. What is the overlap in the training data between the predictors used during the optimization and the ones used for the external validation of the same property?

  2. How often is each memory category selected across all decision points?

8 comments
AuthorsOpenReview ↗

W1, W2, W3, W4

We appreciate your constructive feedback. We will incorporate the additional discussion and results into the revised manuscript.

W1&W2: Whether the router outperforms a simple rule that always retrieves potency-failure experience. Whether the improvement comes from structured experience rather than access to history, requiring a plain-text baseline with previous sequences and scores but no six-field structure, rationale, or router.

Response: We compared PepMind with two alternatives: unstructured, text-only experience and fixed retrieval of potency-failure experience (Table R1). PepMind achieves higher HV in all seven species against both alternatives, supporting the benefits of structured experience representation and adaptive category selection for multi-objective design. Fixed potency-failure retrieval also produces higher predicted hemolysis and toxicity than PepMind, indicating that potency-focused experience alone does not ensure safety.

ConfigurationMean HV ↑AMP ↑Hemolysis ↓Toxicity ↓MIC(μM) ↓
PepMind0.86030.89930.30830.053747.04
w/o Structured Experience Format (text)0.67340.76160.30580.4385155.98
w/o Category Adaptation (fixed potency)0.67920.75330.32600.4095141.62

Table R1. Ablation of structured experience representation and adaptive category selection.

W3: The authors should include in the baselines (i) a protein language model with rejection sampling under the same predictors and budget, and (ii) the same model fine-tuned on AMPs.

Response: PepMind can generate peptides with stronger predicted antimicrobial activity and potency while keeping predicted hemolysis and toxicity lower than both ESM-2 baselines. As shown in Table R2, it achieves the best seven-species means across all four properties, with a four-objective HV of 0.6895, compared with 0.1854 for base ESM-2 and 0.4254 for AMP-tuned ESM-2.

ModelEvaluated setNAMP ↑Hemolysis ↓Toxicity ↓MIC (μM) ↓HV ↑
PepMindAll evaluated1000.89930.30830.053747.040.6895
Base ESM-2All generated1000.31080.46890.1228469.640.1854
AMP-tuned ESM-2All generated1000.58740.46950.3769397.000.4254
Base ESM-2After rejection0NANANANANA
AMP-tuned ESM-2After rejection280.79370.55960.3818309.920.4056

Table R2. Comparison of PepMind with base and AMP-tuned ESM-2 before and after rejection sampling.

W4: The thresholds are fixed and they also seem quite low, given that the score is in [0,1]. A sensitivity analysis is needed. Response: We evaluated sensitivity on fixed candidate sets by varying property cutoffs from 0.3 to 0.9 and MIC-scale multipliers from 0.5 to 2. Tables R3 and R4 summarize the results; further details are provided in Appendix Q and Appendix S, Figure 28.

MIC-scale multiplierPepMind HV ↑
0.50.6194
0.750.6642
1.00.6895
1.50.7170
2.00.7317

Table R3. Sensitivity of PepMind’s seven-species mean four-objective HV to MIC-scale multipliers.

As shown in Table R3, PepMind maintains the highest four-objective HV across all tested MIC scales, supporting the robustness of its ranking to MIC normalization.

PropertyCutoff changePassing fraction change
Activity≥0.7 → ≥0.999.24% → 60.24%
Hemolysis≤0.4 → ≤0.383.79% → 58.05%
Toxicity≤0.4 → ≤0.399.00% → 98.69%

Table R4. PepMind’s single-property passing fractions under stricter cutoffs.

As shown in Table R4, stricter cutoffs substantially reduce activity and hemolysis passing fractions, while the toxicity passing fraction remains comparatively stable.

AuthorsOpenReview ↗

W5, Q1, W6

W5 (a): Whether repeated optimization exploits predictor errors rather than improves peptide quality.

Response: As described in the manuscript, we used an additional external predictor for each property to provide secondary validation. We analyzed correlations between predictors of the same property and quantified overlap between their training datasets.

As shown in Table R5, agreement between optimization and external predictors is limited, with Spearman correlations ranging from 0.0792 to 0.1656. Despite this limited agreement, PepMind-generated sequences consistently achieve the best performance under the external predictors. This consistency supports the robustness of PepMind’s improvements across scoring models and provides evidence against the gains arising solely from exploiting errors specific to the optimization predictors.

Predictor pairSpearman ρPearson r
amPEPpy / AMP-CLIPNANA
APEX / AMP-READ0.07920.1405
HemoPI2 / ConsAMPHemo0.10520.0856
ToxinPred3 / ToxTeller LR0.16560.1278

Table R5. Correlations between predictors evaluating the same peptide property.

W5 (b): Whether the LLM evaluation provides independent validation.

Response: We redesigned the prompt. It removes explicit feature-based guidance that could favor PepMind’s designs.

As shown in Table R6, under the revised prompt, PepMind receives the highest mean scores from GLM-5.2 Kimi-K2.6 and GPT-5.6-Luna, while AMP-Designer leads under Qwen3.5-Plus. These results show that PepMind remains competitive. We therefore use LLM scores alongside external predictor analyses as supplementary checks to reduce reliance on the optimization predictors and mitigate the risk of exploiting their scoring biases.

Prompt:
{
Please rate this antimicrobial peptides sequence on a scale from 0 to 10.
Sequence: {sequence}
Return only a single number.
}
MethodsGLM-5.2Kimi-K2.6Qwen3.5-PlusGPT-5.6-Luna
PepMind7.75166.52198.05288.4884
AMP-Designer7.72756.49808.26578.2580
BroadAMP-GPT7.24566.16597.94688.4783
PepGAN7.47226.41767.93588.4210
diff-AMP6.86256.15677.62816.7104

Table R6. Mean LLM evaluation scores under the revised prompt without explicit feature criteria.

Q1: The extent of training-data overlap between optimization and external-validation predictors for the same property.

Response: The overlap in the training data between the predictors:

Predictor pair (A / B)Sequences in ASequences in BSharedShared / AShared / B
amPEPpy / AMP-CLIP6,33545,7312,91045.94%6.36%
APEX / AMP-READUnavailable7,434NANANA
HemoPI2 / ConsAMPHemo1,5401,7861439.29%8.01%
ToxinPred3 / ToxTeller LR8,8284,1291,78620.23%43.26%

Table R7. Overlap in the training data between the predictors.

As shown in Table R7, most sequences in each available dataset are not shared with its paired predictor. Overlap is relatively low for HemoPI2/ConsAMPHemo (9.29%/8.01%) and asymmetric for amPEPpy/AMP-CLIP (45.94%/6.36%) and ToxinPred3/ToxTeller LR (20.23%/43.26%). These results show substantial distinct sequence coverage, although the datasets are not fully independent.

W6: Do generated peptides differ significantly from DBAASP AMPs in structural and physicochemical properties?

Response: As shown in Table R8, PepMind candidates share the short, positively charged, and hydrophobic character of the 2,859 curated DBAASP reference peptides, supporting their physicochemical plausibility. PepMind favors somewhat shorter, more positively charged, and more hydrophobic sequences, with higher mean predicted structural confidence (pLDDT: 83.17 versus 78.34). Figure 15 in Appendix S further presents the property distributions and amino acid composition.

PropertyPepMind meanDBAASP mean
Length (residues)15.5618.27
Net charge (pH 7.4)6.124.14
Hydrophobic residue fraction0.51250.4463
pLDDT83.1778.34

Table R8. Peptides generated by Pepmind compared to peptides from dbaasp.

We obtained 2,859 unique DBAASP reference sequences by screening 4,923 initially selected records for monomeric peptides of 8–35 standard amino acids, with no recorded terminal modifications or crosslinks, at least one MIC measurement against a species in our seven-species panel, and a publication reference.

W7: How are novelty and diversity quantified, and do the 300 candidates per task extend beyond a few sequence templates?

Response: Table R9 shows substantial within-task sequence variation: candidates are 100% unique, with a mean pairwise distance of 0.4169 and 4.31 sequence families per task. Global uniqueness is 80.36%. These results support sequence diversity while revealing family-level clustering. Detailed analyses have been added to Appendix R, Table 15, and Figure 27.

Novelty/diversity measurePepMind
Exact matches to DBAASP0.00%
Mean nearest-DBAASP sequence similarity0.5377
Sequences with nearest-DBAASP similarity <0.899.95%
Within-task unique sequences100.00%
Mean within-task pairwise sequence distance0.4169
Mean sequence families per task (similarity ≥0.8)4.31
Global unique-sequence fraction80.36%

Table R9. Novelty relative to DBAASP and sequence diversity of PepMind-generated peptides.

Q2: How often is each memory category selected across all decision points?

Response: PepMind draws on complementary failure experiences to guide refinement across antimicrobial activity, potency, and safety:

Experience categoryCalls receiving this category
Activity-failure experience16.41%
Hemolysis-failure experience36.69%
Toxicity-failure experience63.32%
Potency-failure experience55.28%

Table R10. Failure-experience retrieval frequencies across conditional-retrieval decisions.

As shown in Table R10, Potency-related guidance appears in 55.28% of calls, while toxicity and hemolysis evidence appears in 63.32% and 36.69%. This demonstrates PepMind’s ability to bring both efficacy-related and safety-related experience into refinement decisions, providing evidence for anticipating trade-offs when improving a peptide.

We thank the reviewer for suggesting these meaningful experiments. The additional analyses strengthen our assessment of PepMind’s robustness, novelty, and diversity. We will incorporate these experiments and findings into the revised main text.

Reviewer_w7ZHOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies methods for multi-objective antimicrobial peptide design. The main contribution of the proposed approach is a new way to leverage experience observed during the optimization process to enhance the model. The method include a learning to retrieve relevant experience records and use them for test time enhancement of the next generation steps. The authors demonstrated the advance of their methods against other baseline agentic methods that do not make use of experience information during their discovery processes.

Critical strengths

The proposed ideas are interesting and might be useful for practical antimicrobial peptide discovery processes. The experiments were conducted against interesting and strong agentic systems even they might not be the right baseline to compare to. The paper was well written.

Critical weaknesses

I have some concerns about the novelty of the proposed approach. In practice, during the discovery process, one experiments were carried out either by very expensive computational methods or by expensive wet lab experiments, the data is usually not thrown away but used as additional information to train and update the models. This process is used in active drug discovery and thus making use of experience is not a novel contribution that the paper can focus on.

Now the main novel technical contribution may concern the method to use experience as test-time information plus a method to retrieve relevant experience for test time inference. In that case the authors need to provide the evidence of why not turn the experience as additional labelled data or at least in the form of preference pairs so that LLM can be fine-tuned with methods like DPO to enhance their ability to improve itself using the feedback during the optimization process. Comparison directly to those baselines would streghthen the motivation behind the proposed retrieval based test-time inference methods.

The experiments limit the number of candidates to 300 and thus whether that small number of steps require a retrieval method, why not using all the information for enhancement? The chosen baselines agentic system were not finetuned with additional feedback data so it is understandable that they do not work better than the methods that use those explicit knowledge so the baselines should be methods that leverage feedbacks to improve their own search, RL-based approaches with carefully design of reward function should be considered as baselines for comparison.

Questions for authors

Please look at the weaknesses and the questions I raised in the weakness section.

4 comments

We appreciate your careful review and suggestions. Below, we clarify PepMind’s contribution and provide additional fine-tuning and reinforcement-learning comparisons to support the rationale for test-time retrieval.

W1: What is PepMind’s novel contribution, and why does its use of experience constitute an advance over existing data-reuse practices in drug discovery?

Response: PepMind’s novelty lies in classifying different failure types and using them selectively to address the current optimization bottleneck. Effectively using the data generated during drug discovery remains an open challenge [1]. PepMind structures failure data by type and retrieves relevant experience to guide targeted edits. Our contribution is therefore a concrete mechanism for using heterogeneous failures in multi-objective design.

W2: What advantage does test-time retrieval offer over supervised fine-tuning or DPO using the same historical experience?

Response: PepMind enables immediate, bottleneck-specific use of new experience: each round’s outcomes can guide the next decision, whereas incorporating them through SFT or DPO requires additional training. We have also added fine-tuned baselines in Table R1 to address your concern for additional comparisons. PepMind achieves the highest mean HV (0.6895), exceeding the strongest baseline, AMP-RL (0.5342), by 29.1%.

ConfigurationMean HV ↑AMP ↑Hemolysis ↓Toxicity ↓MIC, μM ↓
PepMind0.68950.89930.30830.053747.04
AMP-Designer [2]0.50170.69680.49110.3301239.40
ESM-2 base [3]0.18540.31080.46890.1228469.64
ESM-2 AMP-tuned [3]0.42540.58740.46950.3769397.00
AMP-RL [4]0.53420.71560.42980.4024240.59
ProDCARL [5]0.28910.55170.58280.1283439.06

Table R1. Comparison of PepMind with pretrained, fine-tuned, and RL-based baselines.

W3 (a): How does PepMind compare with feedback-driven baselines, particularly RL-based methods?

Response: We expanded the comparison to include AMP-Designer [2], ESM-2 AMP-tuned [3], AMP-RL [4], and ProDCARL [5]. These comparisons cover fine-tuned models and methods that leverage feedback to improve their own search, addressing the reviewer’s concern for RL-based approaches with explicitly designed reward functions.

ConfigurationMean HV ↑AMP ↑Hemolysis ↓Toxicity ↓MIC, μM ↓
PepMind0.68950.89930.30830.053747.04
AMP-Designer [2]0.50170.69680.49110.3301239.40
ESM-2 base [3]0.18540.31080.46890.1228469.64
ESM-2 AMP-tuned [3]0.42540.58740.46950.3769397.00
AMP-RL [4]0.53420.71560.42980.4024240.59
ProDCARL [5]0.28910.55170.58280.1283439.06

Table R1. Comparison of PepMind with pretrained, fine-tuned, and RL-based baselines.

As shown in Table R1, PepMind achieves the highest mean four-objective HV (0.6895), exceeding the strongest baseline, AMP-RL (0.5342), by 29.1%. It also achieves the highest activity and lowest toxicity and MIC, supporting its effectiveness in multi-objective AMP design compared with the evaluated fine-tuned and RL-based baselines.

W3 (b): Under a 300-candidate budget, why retrieve selected experience instead of using all history, and For why not use all history?

Response: Retrieval is needed to match experience to the current bottleneck, regardless of the candidate budget. Different peptide sequences require different editing guidance, and including all history can introduce irrelevant or conflicting evidence. Supporting relevance-based selection, selective retrieval achieves a mean HV of 0.7011, compared with 0.5903 for random retrieval (Appendix I, Table 7). PepMind retains the full library but selectively retrieves categorized experience for targeted refinement. The 300-candidate limit concerns new evaluations, not the size of this library.

We thank the reviewer for recommending these strong baselines. The added fine-tuning and reinforcement-learning comparisons strengthen the evaluation of PepMind’s effectiveness. We will include these baselines and results in the revised main text.

[1] Bender et al. Artificial intelligence in drug discovery — what it is, where we stand and the path forward. Nature Reviews Drug Discovery. 2026; DOI: 10.1038/s41573-026-01496-2.

[2] Wang et al. Discovery of antimicrobial peptides with notable antibacterial potency by an LLM-based foundation model. Science Advances. 2025; DOI: 10.1126/sciadv.ads8932.

[3] Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023; DOI: 10.1126/science.ade2574.

[4] Park et al. Reinforcement learning with low-rank adaptation for targeted antimicrobial peptide design. Briefings in Bioinformatics. 2025; DOI: 10.1093/bib/bbaf641.

[5] Sheng et al. ProDCARL: Reinforcement Learning-Aligned Diffusion Models for De Novo Antimicrobial Peptide Design. arXiv. 2026; URL: https://arxiv.org/abs/2602.00157.

Reviewer_w7ZHOpenReview ↗

Thank you for the response.

Thank you the authors for additional explanation of the novelty contribution and additional experiments. The classification of the failure type is an interesting way to use the experience for learning but in general I do not see that contribution is novel enough to be considered as a paper published at the top conference like ICLR.

AuthorsOpenReview ↗

Thank You for Your Follow-up: Clarifying Our Contributions

We would like to clarify PepMind’s contribution at two complementary levels:

1. Contribution to antimicrobial peptide design. PepMind addresses the practical challenge of optimizing antimicrobial activity and potency alongside hemolysis and toxicity under a limited candidate-evaluation budget. It makes historical design experience actionable by connecting specific sequence edits with their observed benefits and trade-offs, then selecting relevant evidence to guide subsequent modifications. This supports an iterative design process that considers both improvement of deficient properties and preservation of satisfactory ones. Computational comparisons and recorded editing trajectories provide evidence of its utility and show how the framework responds to changing safety–potency trade-offs.

2. Contribution to AI4Science and broader agent research. PepMind develops a concrete mechanism for using accumulated scientific experience in sequential decision-making: represent the context, action, and multi-objective consequences of previous attempts; retrieve evidence relevant to the current bottleneck; and revise decisions as new feedback arrives. This combination of methodological development and evaluation on a substantive scientific problem aligns with the scope of AI4Science. It also offers insights for broader agent research into how memory should be organized and used when the relevance of past experience changes with the current state. In particular, successes and failures can supply complementary evidence for choosing actions and anticipating adverse effects under competing constraints. These principles motivate further investigation beyond peptide design.

We hope the paper’s novelty and significance will be considered in light of both its contribution to peptide design and its methodological implications for experience-driven scientific agents.

Reviewer_1mDSOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the trade-offs among antimicrobial activity, target-specific potency, hemolytic safety, and toxicity safety in antimicrobial peptide design. The study proposes an experience-driven optimization framework, PepMind. The method stores sequences, predicted outcomes, design rationales, and failure diagnoses from previous designs as structured memory. A router conditioned on the current optimization state is trained using counterfactual retrieval gains to select the category and amount of experience for subsequent candidate refinement. The authors compare methods across multiple pathogen-specific tasks and examine performance and the role of experience reuse through memory ablations, router comparisons, reevaluation with independent predictors, and cross-pathogen transfer.

Critical strengths
  1. The research problem is clearly defined, and the method is closely aligned with the need to manage multi-objective trade-offs. The authors treat currently unmet objectives as optimization bottlenecks, and convert successful and failed design records into reusable experience. Learning the experience category and retrieval amount from counterfactual retrieval gains provides an explicit optimization basis for experience selection, and also represents a methodological design with good novelty.

  2. The experiments are reasonably comprehensive, and the writing and presentation of results are clear. Under a common budget of 300 candidates per task, the main experiments achieve the best results in 19 of 20 representative settings across four pathogens, with complete comparisons across seven pathogens provided in the appendix. The main paper and appendix also provide detailed explanations for their results.

  3. The evaluation extends beyond the predictors used during optimization. Independent predictors strengthen the computational evidence and reduce the possibility that the gains arise solely from exploiting a particular set of optimization predictors. The cross-pathogen experiments further demonstrate the potential of reusing design experience.

Critical weaknesses
  1. The incremental value of the individual contributions could be clarified further. The respective contributions of structured diagnosis, routing learning, and adaptation to multi-objective design to the overall improvements need to be linked more clearly to the existing experimental results.

  2. The specific roles of state-conditioned retrieval have not been fully disentangled. The existing results do not clearly distinguish the gains from selecting experience categories from those arising from adjusting the retrieval amount.

  3. The broader methodological contribution of PepMind could be articulated more clearly. Building on its state-conditioned routing design, the paper should explain how the framework determines which past experiences remain useful as multi-objective trade-offs change, and distinguish its generalizable decision principles from AMP-specific assumptions.

Questions for authors
  1. Given the clearly stated methodological distinction of state-conditioned routing, what evidence do the existing results provide for the incremental value of structured diagnosis, routing learning, and adaptation to multi-objective design, respectively? The authors may also explain and compare the relationship to methods specifically developed for multi-objective antimicrobial peptide design as discussed in Related Work.

  2. The authors point out that the type head often favors potency-failure experience, whereas retrieval size shows clearer state dependence. Using the existing optimization trajectories, could the authors explain whether the selected experience category changes when the main optimization bottleneck shifts? What additional value does adaptive category selection provide over consistently prioritizing potency-failure experience while dynamically adjusting only the retrieval size?

  3. Equation (9) shows that the subsequent LLM already receives the current candidate sequence, its objective scores, and its unmet objectives, while the memory ablations support the overall value of historical experience. Could the authors explain what useful information retrieved experience adds beyond the current scoring feedback, and how this information helps the LLM determine specific sequence modifications, with a representative example?

  4. How can we explain the different changes observed for E. faecalis and E. faecium after adding external experience and removing source MIC information? Do these results reflect different transfer conditions for general activity/safety experience and target-specific potency experience, and how does the current workflow handle source potency information that may not be applicable?

  5. In the forumulas, there are lots of constant values (too many and it is hard for me to list them here) without justifications. Figure 2 looks generated by AI.

5 comments
AuthorsOpenReview ↗

W1, Q1, W2

We thank you for your thoughtful and constructive feedback. We address the raised concerns point by point and provide additional analyses below.

W1: Concerns about the incremental contributions of PepMind’s components—structured diagnosis, router learning, and adaptation to multi-objective design.

Response: We have added ablation experiments that cover all the concerns you mentioned and verify the incremental value of each module. As shown in Table R1, PepMind achieves a mean HV of 0.8603, exceeding the variants without routing learning (+0.1907), structured diagnosis (+0.1869), diagnosis-based matching (+0.1881), retrieved experience (+0.1311), preservation margins (+0.2244), and category adaptation (+0.1811). These comparisons support the complementary contributions of structured diagnosis, routing learning and multi-objective design, with the largest observed HV gap associated with removing preservation margins.

ConfigurationMean HV ↑AMP ↑Hemolysis ↓Toxicity ↓MIC(μM) ↓
PepMind0.86030.89930.30830.053747.04
w/o routing learning0.66960.74700.29290.4565161.23
w/o Diagnosis-based Matching0.67220.73530.31020.4103159.21
w/o Retrieved Experience0.72920.73960.28950.3953143.86
w/o Multi-objective Preservation Margin0.63590.73550.30880.4420145.13
w/o Structured diagnosis (text)0.67340.76160.30580.4385155.98
w/o Category Adaptation (fixed potency)0.67920.75330.32600.4095141.62

Table R1. Ablation study of PepMind.

Q1: Insufficient clarification of PepMind's relationship to existing methods. Response: PepMind’s advantage is targeted, adaptive, and traceable refinement: it retrieves concrete editing experience for the current bottleneck, helping improve unmet objectives while limiting disruption to satisfactory ones. MODAN, HMAMP, and MOFormer guide optimization through Bayesian search, hypervolume-driven learning, and Pareto-based selection. PepMind additionally reuses historical contexts, rationales, and observed trade-offs as explicit guidance at test time. These distinctions and advantages are already discussed in our Related Work section.

W2 (a): Clarification of the role of state-conditioned retrieval.

Response: State-conditioned retrieval is important for balancing design quality and context cost. We examined its two components through separate ablations. For retrieval quantity, we fixed K, the maximum number of experience records retrieved per decision, to 1, 2, 4, or 8. For experience category, we restricted retrieval to one of five types: success, activity failure, hemolysis failure, toxicity failure, or potency failure. Results are showed in Table R2.

ConfigurationMean HV ↑AMP ↑Hemolysis ↓Toxicity ↓MIC (μM) ↓
PepMind Full0.83250.90240.29270.048045.29
K=10.82760.89920.29760.052146.14
K=20.82780.90220.29820.044845.45
K=40.82560.89630.29990.059046.70
K=80.83100.89880.29470.051845.60
success0.82960.90140.29510.050746.44
activity failure0.82360.89880.30090.053046.18
potency failure0.82870.89800.30000.055746.49
hemolysis failure0.82150.89880.30250.044446.09
toxicity failure0.82880.90000.29950.052646.34

Table R2. Ablation study of state-conditioned retrieval.

As shown in Table R2, PepMind achieves the best mean activity, hemolysis, and MIC, and HV, highlighting the importance of state-conditioned retrieval for providing relevant guidance and balancing competing objectives.

W2 (b): How retrieved experience categories change as optimization bottlenecks shift.

Response: Retrieval adapts the evidence used for sequence editing to the current bottleneck. As shown in Table R3, activity- and hemolysis-failure retrieval peak under their corresponding bottlenecks (20.71% and 67.34%). Other failure categories are also frequently retrieved, providing guidance about competing objectives during targeted refinement.

Current bottleneckActivity failureHemolysis failureToxicity failurePotency (MIC) failure
Activity20.71%3̲2̲.̲3̲2̲%̲59.93%5̲3̲.̲8̲0̲%̲
Hemolysis12.85%67.34%65.12%55.42%
Toxicity12.02%31.05%6̲0̲.̲8̲0̲%̲50.79%
Potency (MIC)1̲2̲.̲9̲4̲%̲25.92%53.59%51.68%

Table R3. Failure-experience retrieval frequencies by current optimization bottleneck.

AuthorsOpenReview ↗

W3, Q2, Q3, Q4

Q2: Its comparison with potency-only retrieval.

Response: Adaptive category selection adds value by balancing potency and safety. As shown in Table R4, PepMind achieves higher HV (0.8603 vs. 0.6792) and better mean values across all four properties than fixed potency-failure retrieval.

ConfigurationMean HV ↑AMP ↑Hemolysis ↓Toxicity ↓MIC(μM) ↓
PepMind0.86030.89930.30830.053747.04
Fixed potency0.67920.75330.32600.4095141.62

Table R4. Comparison of PepMind and fixed potency-failure retrieval.

W3 & Q3: How experience relevance is determined as multi-objective trade-offs change, with a representative example showing what retrieved experience adds beyond current scores and how it guides specific sequence edits.

Response: Current feedback identifies the optimization bottleneck, while retrieved experience provides evidence about specific edits and their trade-offs. Table R5 traces how PepMind translates this evidence into three successive sequence modifications.

Process / resultStep 1: delete L6Step 2: append F12Step 3: append K13
1. Current parent sequenceKRLWKLWRLLKRKRLWKWRLLKRKRLWKWRLLKRF
2. Experiences deliveredhemolysis-failure experiencessuccess, hemolysis-failuresuccess, hemolysis-failure
3. Experience cited and information addedInternal-L deletion: historical median Δhemolysis −0.150C-terminal F insertion: historical median Δlog₂(MIC/μM) −0.449C-terminal K insertion: historical median Δlog₂(MIC/μM) −0.256 and Δhemolysis −0.015
4. Recorded LLM decisionDelete L6 to reduce hemolysis, expecting the parent's MIC margin to absorb a modest cost.Append F12 to repair the new MIC bottleneck, using the remaining AMP and hemolysis margins.Append K13 to further improve MIC while maintaining low hemolysis.
5. Edited sequenceKRLWKWRLLKRKRLWKWRLLKRFKRLWKWRLLKRFK
6. Rescored AMP ↑0.922 → 0.8900.890 → 0.9240.924 → 0.931
Rescored hemolysis ↓0.637 → 0.1860.186 → 0.2380.238 → 0.198
Rescored toxicity ↓0.215 → 0.2600.260 → 0.2150.215 → 0.245
Rescored MIC (μM) ↓5.63 → 111.06111.06 → 8.108.10 → 4.38

Table R5. Experience-guided refinement of an E. faecium peptide across three optimization steps.

As shown in Table R5, deleting L6 reduces hemolysis but creates a potency deficit. PepMind then appends F12 to recover potency and K13 to improve it further. The final peptide has lower MIC (5.63→4.38 μM) and hemolysis (0.637→0.198) than the initial parent. This illustrates PepMind’s ability to turn historical experience into concrete edits and adapt refinement to shifting bottlenecks through feedback. This case study now is added in Appendix L (Figure 12) clearly.

Q4: Why E. faecalis and E. faecium respond differently to external experience and source-MIC removal, and how the workflow separates transferable activity/safety guidance from target-specific potency information.

Response: We added a controlled experiment in Table R6 to examine target-dependent transfer and isolate the effect of source MIC: A, no external experience; B, experience with source MIC; and C, the same experience without source MIC.

SpeciesSettingAMP ↑Hemolysis ↓Toxicity ↓MIC (μM) ↓HV ↑
E. faecalisA0.78090.35730.4975299.590.3555
E. faecalisB0.77710.45450.4272267.760.4268
E. faecalisC0.82050.44960.3295205.770.5383
E. faeciumA0.79820.22120.079045.450.9005
E. faeciumB0.76970.23510.075936.220.7575
E. faeciumC0.73220.24940.074246.180.6120

Table R6. Target-dependent effects of external experience and source MIC.

Removing source MIC improves all four mean properties for E. faecalis, increasing HV from 0.4268 to 0.5383. For E. faecium, it worsens activity, hemolysis, and MIC, reducing HV from 0.7575 to 0.6120. These results suggest that source-potency information can help or hinder depending on the target; general editing guidance is not universally transferable either. Our workflow preserves source-species labels, treats source MIC as context rather than target potency, and rescores every candidate against the current target.

Q5: Constant values without justifications and Figure 2 looks generated by AI.

Response: Most constants follow reference values provided in the official GitHub repositories of the corresponding tools or reflect practical choices such as generation budgets under available computational resources. Within each comparison, candidate budgets and evaluation criteria are matched to ensure fairness. We have added Table 8 in Appendix L, which lists all constants mentioned in the manuscript and explains their meanings and roles. Some visual elements in Figure 2 were generated using AI tools, while the overall figure was carefully designed and assembled manually. We have uploaded the original editable PowerPoint (.pptx) file to the anonymous GitHub repository for your review.

We thank the reviewer for the constructive suggestions. The added ablations and case analyses clarify how PepMind uses experience to guide multi-objective refinement. We will incorporate these results into the revised main text.

Comments on the paper

1 comment

Author Final Remarks

Dear AC and Program Committee,

We sincerely thank the AC and the Program Committee for the time and care they have devoted to the review process. We are grateful to the reviewers for their thoughtful feedback, their shared interest in PepMind, and their recognition that it addresses a practical challenge in antimicrobial peptide design. Below, we summarize our main contributions and the revisions we made in response to their feedback.

Key Strengths Highlighted by Reviewers

  • Making failed designs informative. A candidate that falls short on one objective can still reveal how an edit affects several others. PepMind preserves the context, rationale, and observed outcome of past edits, so that an unsuccessful design can inform a later decision. The reviews shared an interest in how much this experience contributes beyond the current candidate’s scores.
  • Adapting guidance as design priorities change. The property that limits a candidate can change after a single edit. PepMind selects experience according to the current bottleneck and records how that evidence informs the next decision. This makes it possible to trace how the method responds when improving one property comes at a cost to another.

Concerns and Our Clarifications

  • Methodological novelty. We respectfully disagree with Reviewer w7ZH’s assessment that PepMind offers limited novelty. PepMind records which objective a past edit failed to meet and uses the current candidate’s bottleneck to determine which prior failures can inform the next design decision. We found it discouraging that, even after we clarified this mechanism and tested it against simpler alternatives, the assessment continued to characterize the contribution as routine data reuse. We believe that developing a way to learn from unsuccessful designs as scientific objectives compete is itself a methodological contribution to AI for Science. We would be grateful if the Area Chair could consider this contribution alongside the experiments examining when and how it helps.
  • Ablations and simpler explanations. We conducted extensive additional ablations and comparisons to address every concern raised by Reviewers 1mDS and nFj5. PepMind consistently outperformed both simpler alternatives across all evaluated species. We appreciated Reviewer 1’s positive assessment. Reviewer nFj5’s remaining questions focused on details of the ablations and their interpretation, and we responded to each in turn.
  • Stronger baselines and matched comparisons. We evaluated PepMind against five additional baselines, including those proposed by the reviewers. These comparisons covered pretrained and AMP-tuned protein language models as well as reinforcement-learning methods. PepMind continued to achieve the highest mean hypervolume across the evaluated species.
  • Robustness, validation, and diversity. We also addressed Reviewer nFj5’s questions about threshold choices, predictor dependence, and sequence diversity. PepMind retained the highest HV. We reported agreement and training-data overlap between predictors, and we treated LLM ratings as supplementary evidence.

We appreciate the reviewers’ questions, which helped us sharpen the contribution and test simpler explanations for the observed gains. We sincerely appreciate the reviewers’ feedback and respectfully hope that the strengthened manuscript merits positive consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.