acceptodds
Under review as a conference paper at ICLR 2027

Can Agent Understand the Differences Between Functionally Distinct Peptides? A Self-Evolving Bifunctional Peptide Discovery Engine

Abstract

Designing peptides that both bind a protein target and kill bacteria means navigating trade-offs no single objective captures: the cationicity that improves antimicrobial potency also drives hemolysis, and edits that raise membrane activity tend to erode binding. We present a self-evolving agent that pairs LLM-driven sequence redesign with physical evaluators it cannot argue past — interface energetics, MIC prediction, and hemolysis prediction. Starting from de novo binder backbones, the agent diagnoses deficits along three axes, proposes edits with explicit predicted outcomes, and accumulates transferable rules in persistent memory. Because it commits to a prediction before each evaluation, its stated reasoning can be audited rather than assumed.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Agent for Peptide Design

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_8y5KOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes JANUS-PEP, a multi-agent framework for designing bifunctional antimicrobial peptides. Starting from a natural-language description of an infection scenario, the system sequentially determines a second therapeutic function, a molecular target, and a structural epitope, and then generates peptide sequences that simultaneously satisfy target binding, antimicrobial activity, and hemolytic safety constraints. The framework contains four modules: planning and generation, diagnosis, redesign, and external physics-based scoring. Through persistent memory, it stores candidate sequences, diagnoses, pre-registered directional predictions, and subsequent scoring results. For targets with an available structural pocket, the system uses the structure-based path and interface energetics; for targets lacking a structural pocket, it uses the sequence-based path and interface confidence.

The authors further perform supervised fine-tuning and GRPO optimization of the Redesigner using a decision corpus constructed from experimental measurements. The corpus is drawn mainly from SKEMPI 2.0, DBAASP, and GRAMPA, with general biomedical tasks also included; the authors report entity-level splits by complex or parent peptide, class balancing, and shortcut audits. The optimized Qwen3-14B achieves a mean accuracy of 73.83% across six tasks, an improvement of 25.64 percentage points over the zero-shot result of the same backbone. For the clinical scenarios listed in Table 2, the system produces at least one passing candidate under the predicted binding, MIC, and HC50 metrics; after re-scoring with external predictors, 73% of candidates still satisfy both activity and safety criteria.

The main value of the paper lies in organizing scenario parsing, candidate redesign, external scoring, and evidence memory into an auditable computational pipeline, and in attempting to train the Redesigner with decision tasks built from experimental measurements. At the same time, the current evidence still comes mainly from computational predictors and an internal benchmark; the paper lacks comparisons with comparable work and does not provide wet-lab validation.

Critical strengths

S1. The problem setting has identifiable novelty: incorporating "inferring a second therapeutic function from a natural-language scenario and resolving it to a target" into the peptide-design pipeline differs from existing AMP multi-objective optimization work (which presets activity/toxicity/novelty objectives) in its inputs and intermediate representations (Page 1-2, Fig. 1).

S2. The experimental design is rich and the workload is substantial; the experiments provide a fairly comprehensive analysis of individual pathways and modules within the agent framework as well as the overall capability of the framework. Breadth of scenario coverage: the 26 scenarios span three major categories—host response, pathogen virulence, and tissue repair—and the targets include both host proteins and microbial proteins (Appendix C, Fig. 9).

Critical weaknesses

W1. There is a lack of external baseline comparisons with comparable work. The comparison in Table 1 is between the same backbone model in "zero-shot" mode and embedded within the framework; there is no head-to-head comparison with any other agent framework (such as MAC-AMP, PABLO, or EASER) or AMP generation model (such as PepTune or AMPGAN v3) under a unified predictor setting. Readers cannot judge whether the JANUS-PEP architecture is superior to existing methods (Page 7, Table 1).

W2. Some claims lack direct experimental support and are not directly tested. The paper claims that "bifunctionality is not a simple sum of two objectives," but does not include a "single-function AMP plus post hoc targeting filtering" control; it claims that target resolution can infer targets from scenarios, but does not include an ablation of "manually specified targets vs. scenario-inferred targets." The absence of these two controls leaves the "necessity of joint optimization" and the "non-substitutability of target resolution" at the level of claims (Page 2, Section 3.2).

Questions for authors

Q1. Could you add a comparison with at least one comparable agent framework (such as the public implementation/protocol of MAC-AMP or PABLO) under a unified predictor setting? If direct execution is not possible, could you provide a qualitative or quantitative protocol-level comparison? Q2. Could you conduct a "single-function AMP baseline plus post hoc targeting filtering" control to demonstrate the necessity of joint bifunctional optimization? Q3. In the caption of Figure 2, could the regions corresponding to (A), (B), (C), and (D) be labelled in the figure? Should the reference to "Fig. 1a" at the beginning of Section 2 be "Fig. 2"? Q4. Was the selection of second-function targets across the 26 scenarios validated by independent experts? Q5. What is the exact composition of the eight scenarios in the factorial experiment in Table 3? Do they cover both the structure-based and sequence-based paths, as well as host and pathogen targets?

3 comments

We thank the reviewer for the constructive feedback and for recognizing the novelty of scenario-grounded bifunctional peptide design. We appreciate the suggestions for strengthening the baseline comparisons and experimental validation. We address each question below.

Q1. Comparison with Existing Agent Frameworks

We thank the reviewer for this suggestion. JANUS-PEP has two stages: resolving the second function and its target from the infection scenario, and then optimizing one peptide against binding, potency and hemolysis jointly. MAC-AMP and PABLO address the second stage only, since both start from objectives the user fixes in advance, so the comparisons below hold the first stage fixed and compare the redesign loop.

PABLO under unified predictors. PABLO (Maus et al., 2026) is a general-purpose agentic black-box optimizer that supports custom objectives and constraints. We adapt it to our task without changing its core optimization framework. We use the same evaluators as JANUS-PEP: PyRosetta ΔΔG\Delta\Delta G for structure-based binding, Boltz ipTM for sequence-based binding, APEX for MIC, and HemoPI2 for HC50.

Both methods use the same starting peptides, acceptance thresholds, and total candidate-evaluation budget (150 evaluations per scenario). We evaluate six scenarios, including three structure-based and three sequence-based cases. For PABLO, we minimize predicted MIC subject to the same binding and hemolytic safety thresholds used by JANUS-PEP. Both methods are evaluated using the same joint acceptance criteria. On the structure-based path, both methods operate on the same complex with a fixed backbone. Only amino acid substitutions are allowed.

MAC-AMP comparison. MAC-AMP (Zhou et al., ICLR 2026) optimizes antimicrobial activity, AMP likelihood, toxicity, and structural reliability. Its original framework targets bacterial species rather than specific protein-binding interfaces and does not directly include a receptor-binding objective.

We therefore evaluate an adapted version on antimicrobial activity and hemolytic safety. We replace its MIC predictor and ToxinPred 3.0 with APEX and HemoPI2 during optimization while retaining the remaining optimization framework. This ensures consistent activity and hemolysis evaluation.

Since MAC-AMP generates peptides de novo rather than editing a fixed starting peptide, we report it as a complementary reference rather than a matched-budget comparison.

MethodOptimization ObjectivesJoint SuccessMIC (µM) ↓HC50 (µM) ↑Selectivity ↑
PABLOBinding + MIC + HC504/615.8109.16.9
MAC-AMPAMP activity + safetyn/a14.688.36.0
JANUS-PEPBinding + MIC + HC506/617.3142.28.2

MAC-AMP optimizes no binding objective, so the joint criterion is undefined for it.

JANUS-PEP delivers a passing candidate in all six scenarios and PABLO in four. With n=6 we do not read much into the gap itself. The point is that the two methods are in the same range on each individual axis, and separate on the joint criterion.

The MAC-AMP comparison provides a reference for sequence-only AMP optimization. It achieves lower predicted MIC but also lower HC50, highlighting the trade-off between antimicrobial potency and hemolytic safety. Unlike JANUS-PEP, this comparison does not assess target-specific binding.

Q2. Joint Optimization versus Post-hoc Filtering

We thank the reviewer for this suggestion. We compare JANUS-PEP with an AMP-only baseline that optimizes MIC and HC50 without binding feedback during redesign. All generated candidates are subsequently evaluated using the same binding criterion. Both settings use the same starting peptides, prediction tools, and maximum candidate-evaluation budget (150 evaluations per scenario). The key difference is whether binding feedback is incorporated during iterative redesign.

MethodJoint SuccessMIC (µM) ↓HC50 (µM) ↑Binding Pass Rate
AMP-only + Post-hoc Binding Filter1/611.2118.40.12
JANUS-PEP (Joint)6/617.3142.21.00

JANUS-PEP's binding pass rate is 1.00 by construction, since a candidate is delivered only when all three criteria hold. The column is included to make the difference between the two procedures visible rather than as a performance figure.

The results show that AMP-only optimization achieves a lower predicted MIC, but only 1/6 scenarios produce candidates satisfying all three criteria after binding filtering. In contrast, JANUS-PEP succeeds in 6/6 scenarios, with a higher predicted HC50.

These results highlight the importance of joint optimization. Improving antimicrobial potency and hemolytic safety alone does not guarantee target binding. Incorporating binding feedback during redesign helps the agent identify peptides that satisfy all three objectives. Together with Q1: existing frameworks optimize objectives that are given, and this control shows that even when all three objectives are given, satisfying them in sequence does not work.

Q3. Figure Labels and References

Thank you for identifying these issues. We will add the (A), (B), (C), and (D) labels to Figure 2 and correct the reference from "Fig. 1a" to "Fig. 2" in Section 2.

Q4. Validation of Second-Function Target Selection

No independent expert review was conducted for the 26 selected targets. Instead, target selection was performed by the Planner and supported by literature evidence and three validation gates: extracellular accessibility, mechanistic relevance (effector rather than correlate), and intervention precedent. Appendix C documents the biological rationale and supporting literature for each target, making the selection process traceable to published evidence rather than relying solely on model judgment.

We further evaluate the specificity of target selection in Table 5. Fifteen deliberately inappropriate targets from three failure categories were embedded in otherwise plausible scenarios. The system rejected 14 of the 15 targets. The single incorrectly admitted target was a disease correlate rather than a causal effector, suggesting that distinguishing effectors from correlates remains the more challenging criterion.

We acknowledge that literature-grounded selection and these validation gates do not replace independent expert assessment. We will explicitly state this limitation in the revised manuscript and identify independent expert review as an important direction for future validation, alongside experimental testing of the designed peptides.

Q5. Composition of the Eight Scenarios in Table 3

The factorial experiment uses eight scenarios from the 26 clinical cases, covering both design paths and host and pathogen targets.

ScenarioTargetSecond FunctionPathTarget Type
S1TLR4 (O00206)Endotoxin neutralizationStructureHost
S2IL-1β (P01584)Anti-inflammatoryStructureHost
S4α-toxin (P0C216)Anti-virulenceStructurePathogen
S5CagA (P55980)Anti-virulenceStructurePathogen
Q2IL-17A (Q16552)ImmunomodulationSequenceHost
Q3Enterotoxin B (A7Y1W2)Anti-toxinSequencePathogen
Q11MrkA (D5FKS5)Anti-biofilmSequencePathogen
Q13Angiopoietin-2 (O15123)Pro-angiogenicSequenceHost

The set contains four structure-based and four sequence-based scenarios, with four host and four pathogen targets. It also covers seven distinct second-function labels.

All four experimental conditions use identical starting peptides, batch sizes, and round budgets. In the memory-off conditions, neither the page store nor the memory index is available to the agent during redesign. This controlled comparison isolates the effects of memory and Redesigner optimization under matched experimental settings.

I appreciate the authors’ response and the additional experiments. Using consistent predictors and a fixed 150-evaluation budget across all six test scenarios, JANUS-PEP successfully solves 6 out of 6 cases, versus 4 out of 6 for PABLO. It yields higher predicted HC50 (142.2 µM vs. 109.1 µM) and selectivity (8.2 vs. 6.9), though with a marginally higher MIC (17.3 µM vs. 15.8 µM). The post-hoc filtering control in Q2 helps isolate the contribution of the joint redesign framework. When optimizing potency and safety without binding feedback, the MIC drops to 11.2 µM, but the binding pass rate falls to 0.12, and overall joint success drops to just 1 out of 6 cases. Table 3’s breakdown of the eight scenarios confirms a balanced split by design paradigm (four structure-based, four sequence-based) and target category (four host, four pathogen). Adding (A-D) panel labels to Figure 2 and correcting the reference in Section 2 should address the remaining presentation issues.

My main remaining concern lies in the limited scope of external comparisons and independent validation. The budget-matched comparison against PABLO only covers six scenarios, and MAC-AMP does not incorporate a binding objective, making it a sequence-only baseline rather than a direct head-to-head comparator. As I noted in Q4, the 26 second-function targets were selected via literature-based criteria without independent expert review, and all candidate peptides generated are still at the computational prediction stage, with no wet-lab experimental validation. Although the internal benchmark results are consistent and well-controlled, the breadth of external validation against comparable design frameworks is still fairly limited.

Overall, the architectural design and empirical controls are sufficiently robust to support the paper’s core claims, provided the additional baselines, ablation studies, and discussion of limitations are included in the revised manuscript. I therefore maintain my original rating: Rating 3, Confidence 3

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

This paper introduces JANUS-PEP, a multi-agent framework for designing bifunctional antimicrobial peptides that simultaneously exhibit antimicrobial activity and an additional therapeutic function determined by the infection context. The system translates natural-language infection scenarios into molecular targets and design objectives through a planning and evidence-retrieval stage, then selects either structure-based or sequence-based peptide generation depending on the available target information. An iterative redesign framework separates diagnostic reasoning, candidate generation, and computational evaluation, using persistent memory to record previous candidates, predicted effects, and measured-by-model outcomes.

Critical strengths

The paper addresses an interesting and practically relevant problem: integrating therapeutic objective selection with multi-objective peptide optimization, rather than assuming that all therapeutic requirements are specified at the molecular level beforehand. The proposed architecture is thoughtfully modular, separating planning, diagnosis, sequence redesign, and external evaluation. This separation improves interpretability and provides a mechanism for inspecting whether the reasoning behind a proposed modification agrees with subsequent computational evaluations. The persistent memory design is another valuable component, particularly because it records unsuccessful candidates and supports the reuse of previous optimization experience. The authors provide a useful factorial comparison of memory and Redesigner optimization, showing that the two components contribute complementary benefits.

Critical weaknesses

The central weakness is the lack of experimental validation of the generated bifunctional peptides, which substantially limits the strength of the biological conclusions. Although the benchmark used for Redesigner training contains experimentally determined labels, the actual peptide-design results rely on computational predictors for binding, antimicrobial activity, and hemolysis. Consequently, satisfying these predicted constraints does not establish that a peptide possesses both therapeutic functions experimentally. Independent computational rescoring helps assess consistency but cannot rule out shared model biases, distribution shifts, or optimization exploiting predictor weaknesses. Moreover, target binding alone does not demonstrate the intended functional consequence, such as toxin neutralization, inhibition of virulence, or promotion of wound healing.

Questions for authors
  1. Experimental validation: Can the authors experimentally evaluate a representative subset of the generated peptides for antimicrobial potency, hemolysis, target binding, and the intended complementary biological function? Without such validation, how should readers interpret the reported success across all 26 scenarios?

  2. Reliability of computational evaluators: Since the optimization process repeatedly queries the same activity, toxicity, and binding predictors, how do the authors rule out predictor exploitation or out-of-distribution optimization? Could they evaluate candidates using substantially different scoring models or independent experimental datasets?

  3. Biological relevance of target engagement: How does the framework establish that binding a selected target produces the desired therapeutic effect rather than merely forming a stable interaction? In particular, how are intracellular effectors such as CagA handled under the extracellular-accessibility requirement?

  4. Autonomous second-function selection: Appendix C indicates that each scenario already specifies an additional therapeutic requirement. Could the authors evaluate scenarios in which only the infection context is supplied and the system must independently identify, justify, and prioritize the complementary therapeutic objective?

  5. Contribution of individual agents: Could the authors provide controlled ablations removing the Diagnoser, separating versus combining diagnosis and proposal, disabling target-aware planning, and replacing the Redesigner with a simpler optimization algorithm? How much improvement is attributable to each component under an identical candidate-evaluation budget?

  6. Fairness of language model comparisons: How would the conclusions change if the comparison focused only on peptide-specific benchmarks and used matched training or adaptation conditions? Can the authors distinguish the gains from supervised fine-tuning, GRPO, and iterative agentic search more explicitly?

  7. Generalization of memory: The paper argues that memory transfers useful experience between targets. Could the authors evaluate this using held-out target families, randomized target order, and controlled memory resets to determine whether the accumulated rules genuinely transfer beyond closely related design tasks?

  8. Computational efficiency and practical usability: What are the total runtime, number of evaluator calls, token consumption, and computational costs per successful candidate? How does this compare with established peptide optimization pipelines under matched computational budgets?

7 comments

Response to Reviewer NHFg

We thank the reviewer for the constructive feedback and for recognizing the novelty of scenario-grounded bifunctional peptide discovery, the modular agent architecture, and the complementary contributions of memory and Redesigner optimization. We appreciate the suggestions for strengthening the experimental evaluation and biological validation. We address each question below.

Q1. Experimental Validation of Bifunctional Peptide Design

We thank the reviewer for this important question. JANUS-PEP is a computational discovery framework that integrates scenario-grounded target selection with multi-objective peptide optimization. Across 26 infection scenarios, under the function-specified setting, the framework successfully identifies candidates satisfying the predicted binding, antimicrobial potency, and hemolytic safety criteria.

We further assess the reliability of these candidates using independent computational evaluations. Independent predictors achieve Pearson correlations of 0.84 for hemolysis and 0.86 for antimicrobial potency, with 73% of candidates retaining both activity and safety criteria. Binding specificity analysis also shows preferential interactions with intended targets compared with eight decoy receptors.

Importantly, we identify a strong connection between our computational results and an independent experimental study. Liu et al. (Theranostics, 2025) investigated bacterial pneumonia and sepsis, designing a bifunctional peptide R18 targeting the TLR4/MD-2 and TLR2 inflammatory pathways. Although our S1 scenario concerns meningococcal meningitis with endotoxemia, JANUS-PEP resolved the specified anti-endotoxin requirement to TLR4/MD-2, the same pathway targeted in that study.

Starting from a 17-residue peptide (FFFAELLRAIIKNGGFM), JANUS-PEP redesigned it into RWRRELLRAIIKNGGRW. The redesign increased arginine residues from 1 to 5 and introduced two tryptophan residues while retaining hydrophobic leucine and isoleucine residues. The resulting peptide achieved a predicted MIC of 4.5 µM, HC50 of 118.6 µM, and binding ΔΔG of −5.1 REU, satisfying all three criteria. Notably, this design shares the cationic and hydrophobic characteristics of R18, whose antibacterial activity, TLR4/MD-2 interaction, and inhibition of downstream inflammatory signaling were experimentally demonstrated.

This independent experimental precedent supports the biological relevance of JANUS-PEP's target selection and peptide design rationale. The 26/26 success rate represents computational design success, rather than experimental confirmation of therapeutic efficacy. Our central contribution is a unified framework that connects infection context, therapeutic target selection, and constrained peptide redesign, supported by systematic computational evaluation across diverse scenarios. Experimental validation of the designed peptides will be an important next step to further establish their biological activity and therapeutic potential.

Q2. Reliability of Computational Evaluators and Predictor Exploitation

We thank the reviewer for raising this concern. To assess whether JANUS-PEP exploits the predictors used during optimization, we perform several complementary evaluations involving independent predictors, alternative targets, and residue-level analysis.

(1) Independent predictor validation.

JANUS-PEP uses APEX and HemoPI2 to evaluate antimicrobial activity and hemolysis during optimization. We independently re-evaluate the final peptides using predictors trained on different datasets that were not involved in candidate selection.

The results show strong agreement, with Pearson correlations of 0.86 for antimicrobial potency and 0.84 for hemolysis. Moreover, 73% of the designed peptides continue to satisfy both activity and safety thresholds under independent rescoring.

These results indicate that the improvements are not limited to the original optimization predictors.

(2) Binding specificity and independent residue-level analysis.

We further evaluate binding specificity against eight decoy receptors, including three same-family receptors and five abundant host proteins. The designed peptides show substantially more favorable interface energies against their intended targets, averaging −15.2 REU, while the mean energies against decoys are above zero.

We also examine the TLR4 redesign trajectory using AmpLyze, an independent hemolysis predictor with residue-level attribution. In the first two redesign rounds, all edited positions overlap with independently identified toxicity hotspots, and their predicted toxicity contributions decrease after redesign. From round 3 onward, a small number of edits fall outside the hotspot set; these serve the antimicrobial and binding axes, which AmpLyze does not score, while the highest-attribution positions that remain unedited lie at the receptor interface.

Q3. Biological Relevance of Target Engagement

We thank the reviewer for this important question. JANUS-PEP does not select targets solely based on predicted binding affinity. Instead, target selection is guided by three evidence gates: target accessibility, mechanistic relevance, and intervention precedent. The Planner further identifies functionally relevant binding regions based on established biological mechanisms. For example, the TLR4 scenario targets a receptor interface involved in inflammatory signaling, while the TcdB scenario targets a region associated with toxin neutralization.

These mechanisms provide a biological basis for connecting target engagement with the intended therapeutic function. Importantly, JANUS-PEP aims to identify peptides that engage functionally relevant interfaces rather than merely form stable protein interactions. While predicted binding does not directly establish functional inhibition, the selected interfaces are supported by published evidence of their roles in the corresponding pathological processes.

Regarding CagA accessibility, although CagA primarily exerts its pathogenic effects inside host epithelial cells, experimental studies have demonstrated that it also undergoes an extracellularly accessible stage during translocation. Murata-Kamiya et al. (Cell Host & Microbe, 2010) showed that CagA is exposed on the bacterial surface through the type IV secretion system before entering host cells. Kumar et al. (PLOS ONE, 2013) further investigated the bacterial surface localization of CagA.

More importantly, Sain et al. (3 Biotech, 2025) experimentally demonstrated that extracellular treatment with CagA-specific antibodies reduces CagA translocation into gastric epithelial cells. This provides direct experimental evidence that CagA can be targeted before entering host cells, supporting the biological rationale for considering CagA in our target-selection framework.

Our structure-based design uses the CagA–ASPP2 interface (PDB 4IRV), which is involved in CagA's intracellular pathogenic activity. Although the extracellular exposure of CagA is experimentally supported, the accessibility of this specific interface during translocation requires further validation.

Overall, JANUS-PEP connects infection-associated biological mechanisms with functionally motivated target and epitope selection, rather than relying on predicted binding affinity alone. The existing experimental literature supports the biological rationale behind these design choices, while direct functional validation of the generated peptides will be an important next step.

Q4. Autonomous Second-Function Selection

We thank the reviewer for this suggestion.

In the original 26 scenarios the complementary function is given at a high level, such as anti-inflammatory, anti-toxin, or anti-virulence activity, while the molecular target and functional epitope are not. The reviewer is right that this makes the 26 scenarios a test of target resolution rather than of function selection, and our framing in the abstract and in Figure 1 does not draw that line clearly enough. We have run the experiment the reviewer describes.

Experimental setup. We construct infection-context-only variants of ten scenarios, retaining only the pathogen and clinical presentation while removing both the complementary function and the molecular target. The Planner must independently propose candidate secondary functions, rank them by clinical relevance, resolve the top-ranked function into a molecular target, and evaluate that target using the three admission gates. The downstream peptide redesign process remains unchanged. Because a single infection admits several reasonable therapeutic objectives, we do not require the Planner to reproduce the function we originally specified; we report that agreement as one measure among several, and otherwise assess whether the proposed functions are supported by published evidence.

MetricInfection-context-only
Scenarios evaluated10
Candidate functions proposed per scenario (mean)3.4
Top-ranked function with mechanistic support9 of 10
Top-ranked function matching the original requirement6 of 10
Resolved target passing all three admission gates8 of 10
Joint computational design success7 of 10

Results. The Planner proposes an average of 3.4 candidate functions per scenario, and its top-ranked choice is supported by published biological evidence in 9 of 10 cases. In 6 of 10 it independently recovers the therapeutic function we had specified by hand.

A single infection may admit several reasonable therapeutic objectives, and the divergences reflect this. In H. pylori chronic gastritis the Planner prioritizes anti-inflammatory intervention over the anti-virulence objective we originally specified, and for the diabetic wound scenario it proposes anti-biofilm activity where we had specified pro-angiogenic support. Both alternatives are defensible given the presentations supplied. In two further cases the proposed target fails the mechanistic relevance gate, consistent with the pattern in Table 5, where distinguishing effectors from correlates is the weaker of the two criteria.

End to end, 8 of 10 scenarios yield a target satisfying all three evidence gates and 7 of 10 produce a candidate meeting all three design criteria, against 26 of 26 when the function is supplied. The drop locates where autonomy currently costs the system: not in redesign, which is unchanged, but in resolving an underdetermined requirement into a single admissible target.

Reviewer_Rc1uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3(+1)
Read review
Paper summary

The paper introduces an agentic framework (JANUS-PEP) for peptide discovery. JANUS-PEP contains structure and sequence-based design paths with domain-specific toolsets, specialzized agents (Planner, Builder, Diagnoser, and Redesigner), and a recursive memory underlying all agents that stores design history across turns and episodes. On general-biomedical and peptide-specific benchmark evaluation, JANUS-PEP efectively improves frontier models, while optimizing the Redesigner with SFT + GRPO further boost the discovery performance of Qwen3-14B. On 26 clinical scenarios, JANUS-PEP produces candidates that satisfy potency, hemolysis, and binding.

Critical strengths
  1. Timely problem with practical potential. The paper addresses a timely problem with substantial application significance, and the computational results suggest that JANUS-PEP could potentially support clinical bifunctional peptide discovery.
Critical weaknesses
  1. Missing experimental and training details. The experiments do not isolate the contributions of SFT, GRPO, and self-evolving memory. The main text also omits key details of training-data construction, including whether SFT targets contain reasoning traces and how General QA training samples are sourced and selected. Moreover, the optimized model outperforms all evaluated frontier backbones within the framework on the three author-constructed design benchmarks, but not on ProtGym.
  2. Presentation should be further improved. Figure 4 covers several aspects of design performance across the 26 infection scenarios, but the results and their interpretation are too compressed in the main text. A clearer explanation of the experimental settings, comparisons, and main findings would help readers follow the analysis. The main text would also benefit from a more substantive discussion of related agentic discovery systems, clarifying how JANUS-PEP contextualizes among related work.
Questions for authors

Q1. Could the authors provide SFT-only, GRPO-only, and self-evolving ablations under comparable budgets, distinguishing within-task memory from cross-target transfer?

Q2. Could the authors clarify how the SFT/GRPO corpus is constructed, including SFT target format, General QA sources and selection criteria, split sizes, and overlap checks against the evaluation benchmarks?

3 comments

We thank the reviewer for the constructive feedback and for recognizing the practical value of JANUS-PEP. We appreciate the suggestions for improving the experimental analysis and training details. We address each concern below.

Q1. Ablation of SFT, GRPO, and Self-Evolving Memory

(1) Contributions of memory and policy optimization.

Table 3 in the Appendix provides a controlled 2×2 ablation of memory and policy optimization. We evaluate eight scenarios using identical starting peptides, batch sizes, and round budgets.

MemoryRedesignerSolvedRounds to PassRepeat Rate
NoBase3/811.40.31
NoSFT+GRPO5/88.70.28
YesBase6/86.20.09
YesSFT+GRPO8/83.40.04

Without policy optimization, adding memory increases the number of solved scenarios from 3/8 to 6/8. It also reduces the repeated-failure rate from 0.31 to 0.09. Policy optimization improves success from 3/8 to 5/8 without memory. Combining both components achieves 8/8 and further reduces repeated proposals.

These results indicate that memory helps the agent avoid previously unsuccessful decisions, while policy optimization improves its ability to select effective edits.

(2) Separating SFT and GRPO.

We agree that the contributions of SFT and GRPO deserve a more detailed analysis. Figure 5 in the Appendix compares Zero-shot, SFT, and SFT+GRPO on directional prediction and joint prediction. To examine the contribution of each training stage more directly, we report the following ablation across the design benchmarks.

RedesignerBind-SelBind-DirSel-TIProtGym
Base (Frozen)45.3758.7841.0260.34
SFT-only63.2071.8060.2064.50
GRPO-only48.1073.4043.5065.80
SFT+GRPO68.0076.3063.7070.00

All settings use the same Qwen3-14B backbone, evaluation datasets, agent tools, and inference budgets.

SFT+GRPO achieves the best performance across all four benchmarks, improving the average accuracy by 4.58 percentage points over SFT-only and 11.80 points over GRPO-only. SFT-only provides larger gains on Bind-Sel and Sel-TI, while GRPO-only performs better on Bind-Dir. Their combination consistently outperforms either training stage alone.

Appendix D.2 discusses the role of SFT initialization. SFT provides a stronger starting policy for GRPO, which learns from reward differences among sampled responses. The ablation results support the effectiveness of this two-stage training strategy.

(3) Within-task memory versus cross-target transfer.

We thank the reviewer for suggesting a more fine-grained evaluation of memory. To distinguish within-task improvement from cross-target transfer, we consider three settings:

  • No Memory: No previous design records are available.
  • Within-task Memory: Records accumulate across redesign rounds but are cleared between targets.
  • Full Memory: Records persist both within tasks and across targets.

We hold the model, starting peptides, candidate budgets, and evaluation tools fixed.

Memory SettingSolvedRounds to PassRepeat RateJoint Hit
No Memory5/88.70.280.19
Within-task Only7/84.50.060.22
Full Memory8/83.40.040.27

Within-task memory improves search efficiency. Cross-target memory provides further improvements. These results demonstrate the complementary roles of the two memory mechanisms. Within-task memory helps avoid repeated failures and guides iterative refinement. Cross-target memory provides additional benefits by reusing knowledge from previous design tasks.

Q2. Training Corpus Construction and Data Leakage

We appreciate the request for more details on training data construction. We will expand Appendix E to provide the complete data sources, training format, and split statistics.

(1) Data sources, splits, and leakage prevention.

We construct a corpus of 2,000 questions, comprising 1,200 domain-specific and 800 general questions.

Domain-specific questions are derived from experimental measurements: Bind-Sel and Bind-Dir from SKEMPI 2.0 (5,410 mutation records after filtering), and Sel-TI from DBAASP and GRAMPA (984 peptides carrying both minimum inhibitory concentration, MIC\mathrm{MIC}, and hemolytic concentration, HC50\mathrm{HC}_{50}, measurements).

The 800 general questions are sampled from the official training partitions of MedQA (300), PubMedQA (300), and MATH (200). They serve as capability anchors for general biomedical and mathematical reasoning and carry no peptide-design labels. The MATH questions are not included in the evaluation benchmarks reported in Table 1.

To prevent data leakage, we partition the underlying entities before constructing any items: protein complexes for binding tasks and parent peptides for selectivity. Items are then constructed independently within each partition, ensuring that training and evaluation items share no source protein complex or peptide.

Entity-level splitting is particularly important for Sel-TI, where options are drawn from a shared peptide pool. An item-level split could allow the same peptide to appear in both training and evaluation sets in different combinations.

The held-out evaluation sets contain 200 Bind-Sel, 200 Bind-Dir, and 150 Sel-TI questions, as reported in Table 1. An additional 120 questions are reserved for validation. ProtGym is an external variant-effect benchmark and contributes no training items. For MedQA and PubMedQA, we use the official training partitions for data construction and the official held-out test sets for evaluation.

We further verify data disjointness through direct overlap checks. No protein complex or parent peptide is shared between the training and evaluation sets (0 shared entities). Exact-match comparison of normalized question text between the 800 general questions and the MedQA and PubMedQA test sets also identifies 0 overlapping questions.

(2) SFT target format and GRPO reward.

For SFT, each training sample consists of a multiple-choice question and its corresponding correct answer. The target contains only the correct answer. The training loss is computed only over the answer tokens, as described in Appendix D.2.

For GRPO, we use the same training corpus and sample eight responses per question. Each response receives a binary reward of 1 if the predicted answer matches the ground-truth answer and 0 otherwise. For domain-specific questions, the correct answers are derived from experimental measurements in SKEMPI 2.0, DBAASP, and GRAMPA. For General QA, we use the original benchmark answer keys.

Additional Comments

(1) Performance on ProtGym.

ProtGym evaluates protein variant effects, which differs from the binding and peptide selectivity tasks used to optimize the Redesigner. Despite this difference, SFT+GRPO improves Qwen3-14B's accuracy on ProtGym from 60.34% to 70.00%, a gain of 9.66 percentage points. This result is also close to the strongest frontier backbone.

More broadly, our optimized model achieves the best performance on three of the four peptide-specific benchmarks and the highest overall average accuracy (73.83%). These results demonstrate that the benefits of our optimization extend beyond the training task families.

(2) Related agentic discovery systems.

We thank the reviewer for this suggestion. We will expand the related-work discussion to include a more detailed comparison with Pepti-Agent, agentic D-enantiomeric AMP discovery, and AMPGAN v3.

JANUS-PEP differs from these systems in several aspects. It connects scenario-grounded target resolution with bifunctional peptide design. It separates diagnosis from candidate proposal and records predicted versus computed outcomes for subsequent redesign. We will highlight these distinctions more clearly.

We thank the reviewer again for the valuable suggestions. The additional analyses and implementation details will strengthen the experimental evidence and improve the clarity and reproducibility of our work.

I appreciate the authors’ detailed response. The additional ablation results and the clarification of how the benchmark was constructed address several of my concerns and give me greater confidence in the paper’s technical soundness. I am therefore raising my soundness score to 3.

Presentation clarity was a central concern in my original review, and it continues to affect my overall assessment. The introduction lacks a clear discussion on the technical/empirical contribution of the paper. I would also suggest adding the ablation results into the main text, since it shows what each component contributes to the system’s performance. Figure 4 is informative, but readers outside peptide design would benefit from a short explanation of which panels support each claim and how to interpret the metrics. Also there lacks a discussion of related agentic systems in the main text. Considering the broad and diverse audience of ICLR, the paper would benefit from further revisions to improve its readability before acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Experimental Evidence for Resolved Targets and the Bifunctional Design Objective

We thank the reviewers for this valuable suggestion. Experimental validation is ultimately what a framework like ours must answer to, and while wet-lab testing of our own candidates is still ahead of us, we were able to find independent experimental work on four of the targets JANUS-PEP resolved.

ScenarioJANUS-PEP Resolved TargetJANUS-PEP Predicted ResultsIndependent Wet-lab StudyCorrespondence with Our Design
S1 — Endotoxin neutralizationTLR4 / MD-2 dimerization surface (PDB 4G8A)ΔΔG −5.13 REU; MIC 4.51 μM; HC₅₀ 118.60 μMR18 (Liu et al., 2025): MIC 20 μM against E. coli; TLR4/MD-2 interaction by SPR; inhibition of TLR signaling. Papiliocin (Krishnan et al., 2022): TLR4/MD-2 Kd 3.7 μM; functional IC₅₀ 1.1 μM; confirmed antibacterial activitySame receptor, same dimerization interface; both peptides are also antimicrobial
S7 — Anti-toxinTcdB CROPs domain (PDB 4NP4)ΔΔG −18.74 REU; MIC 15.22 μM; HC₅₀ 163.82 μMAngie 5 (Lietz et al., 2025): growth inhibition of C. difficile; TcdB functional IC₅₀ 40.94 ± 1.08 μM; reduced cytotoxic damageSame toxin, same anti-toxin function; Angie 5 is also antimicrobial. Binding to the CROPs domain was not assayed
Q12 — Anti-biofilmOmpA, A. baumanniiipTM 0.752; MIC 30.42 μM; HC₅₀ 153.51 μMP92 (AAC, 2024): OmpA Kd 7.84 nM; inhibition of adhesion, invasion and biofilm formation; protection in a murine infection modelSame outer-membrane protein, same anti-biofilm function; P92 is a single-function binder
Q2 — ImmunomodulationIL-17AipTM 0.884; MIC 8.30 μM; HC₅₀ 121.94 μMHAP (Liu et al., 2016): IL-17A Kd 2.6 ± 0.5 nM by SPR; IL-17A/IL-17RA blockade IC₅₀ 0.9 ± 0.3 nM; suppression of downstream cytokinesSame cytokine, same receptor-blockade mechanism; HAP is a single-function binder

These four cases provide independent experimental support for the molecular targets identified by JANUS-PEP. Importantly, the protein targets were not specified in our scenario inputs; the Planner autonomously resolved them from high-level therapeutic requirements.

The four differ in what they add beyond the target choice. R18, papiliocin and Angie 5 are themselves antimicrobial, so S1 and S7 also support the feasibility of the bifunctional objective. P92 and HAP contribute direct nanomolar binding evidence for the resolved target, but are single-function binders.

Wet-lab validation of our candidates remains our planned next step.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.