Bridging the Evidence-to-Execution Gap:A Reflective Agent for Multi-Objective Peptide Design

Haosen Zhang; Yang Yang

cs.AIOct 5, 2026

Bridging the Evidence-to-Execution Gap:A Reflective Agent for Multi-Objective Peptide Design

Authors: Haosen Zhang, Yang Yang

Organizations: School of Agriculture and Biology, Shanghai Jiao Tong University · School of Computer Science, Shanghai Jiao Tong University

Abstract

Large language models (LLMs) can reason over scientific literature to devise design strategies, yet fail to reliably implement them for biological sequences. While protein generative models learn sequence patterns, they lack the capacity to incorporate literature evidence for multi-step reflective reasoning, forming an evidence-to-execution gap between scientific reasoning and sequence manipulation. We present EASER (Evidence-Aware Sequence Engineering with Reflection), a reflective agent bridging reasoning and sequence generation via a learned property interface of offline-trained, fixed low-rank matrices. The agent steers a diffusion generator by combining these matrices, proposing intervention hypotheses (anchors, editable positions, control coefficients) grounded in retrieved evidence, sequence context and past results. A Probe-and-Steer mechanism validates interventions and allocates samples according to predicted property responses, with outcome reflection informing subsequent decisions. Evaluated on multi-objective antimicrobial peptide design (optimizing activity, non-hemolysis and non-toxicity), explicit hypothesis formulation delivers better multi-objective performance than direct action generation under identical decision conditions. Ablation studies verify the importance of evidence retrieval, episodic history, reflection and Probe-and-Steer. Over six repeated trials, EASER obtains the highest mean hypervolume and lowest mean IGD+ on screened candidates compared with competing baselines. Our work demonstrates how an executable property interface and iterative feedback link scientific reasoning to targeted peptide sequence generation.

Figures & tables

Figure 1: EASER overview. Offline-trained matrices steer a frozen generator during sampling and refinement. Evidence and internal predictor feedback guide agent decisions; external predictors score candidates after search. The diagram uses schematic matrices, sequences, and plots; the molecular example is PDB 2MAG .

Figure 2: Evidence-informed agent loop. Specialists organize evidence; the LLM proposes interventions, Probe tests responses, and Steer allocates samples. Outcomes update the archive, reflection, and memory. Sequences, controls, and plots are schematic.

Method	Activity score $\uparrow$	Non-hemolysis score $\uparrow$	Non-toxicity score $\uparrow$
AMPGAN	$0.843(\pm 0.019)$	$0.704(\pm 0.008)$	$0.963(\pm 0.001)$
AMPGPT2	$0.908(\pm 0.012)$	$0.688(\pm 0.014)$	$0.958(\pm 0.001)$
AMPGen	$0.809(\pm 0.049)$	$0.668(\pm 0.016)$	$0.962(\pm 0.000)$
AMPDesigner	$0.884(\pm 0.017)$	$0.705(\pm 0.014)$	$0.963(\pm 0.001)$
AMP-adapted DPLM	$0.856(\pm 0.029)$	$0.689(\pm 0.016)$	$0.962(\pm 0.001)$
EASER	$0.901(\pm 0.067)$	$0.731(\pm 0.024)$	$0.963(\pm 0.003)$

Table 1: External-generator property scores. Mean

\pm

sample SD of per-seed top-ten means over six seeds, after screening and before penalty. Each property selects its own sequences.

Method	Seeds	Passed	HV $\uparrow$	IGD+ $\downarrow$
AMPGAN	$6$	$1,100$	$0.0362(\pm 0.0044)$	$0.0543(\pm 0.0182)$
AMPGPT2	$6$	$1,717$	$0.0330(\pm 0.0026)$	$0.0603(\pm 0.0090)$
AMPGen	$6$	$395$	$0.0301(\pm 0.0065)$	$0.0739(\pm 0.0307)$
AMPDesigner	$6$	$1,409$	$0.0356(\pm 0.0041)$	$0.0519(\pm 0.0106)$
AMP-adapted DPLM	$6$	$560$	$0.0350(\pm 0.0042)$	$0.0518(\pm 0.0131)$
EASER	$6$	$1,947$	$0.0450(\pm 0.0086)$	$0.0277(\pm 0.0177)$

Table 2: External-generator frontier quality. Each seed retains 2,560 candidates; Passed is the total. HV/IGD+ report mean

\pm

sample SD on screened, unpenalized scores with the fixed reference front.

Method	Exact HV $\uparrow$	IGD+ $\downarrow$
Fixed coefficients	$0.0282(\pm 0.0076)$	$0.0781(\pm 0.0329)$
Random actions	$0.0389(\pm 0.0068)$	$0.0414(\pm 0.0162)$
LinUCB	$0.0373(\pm 0.0087)$	$0.0540(\pm 0.0260)$
NSGA-II	$0.0344(\pm 0.0069)$	$0.0588(\pm 0.0240)$
EASER	$0.0450(\pm 0.0086)$	$0.0277(\pm 0.0177)$

Table 3: Search policies on a shared generator and learned controls. Mean

\pm

sample SD across six runs of 2,560 candidates. Metrics use screened, unpenalized scores and the fixed reference (Appendix A.2 ).

Figure 3: Pareto-front projections. Pairwise projections of screened three-objective fronts, using normalized scores before penalty. Each method contributes all six run-level fronts.

Policy	Mean $\Delta\mathrm{HV}$ $\uparrow$	Screened candidates $\uparrow$
Direct action generation	$0.001510$	$162$
Explicit hypothesis formulation	0.002447	261

Table 4: Hypothesis formulation with matched information. Mean HV gain over pre-existing candidates at 30 states; screened counts sum across 1,920 draws per protocol.

Condition	HV $\uparrow$	IGD+ $\downarrow$	Passed
EASER	$0.0462(\pm 0.0066)$	$0.0243(\pm 0.0069)$	$1,076$
Without reflection feedback	$0.0397(\pm 0.0042)$	$0.0408(\pm 0.0152)$	$1,025$
Without round-history details	$0.0392(\pm 0.0034)$	$0.0389(\pm 0.0186)$	$574$

Table 5: Targeted LLM-input ablations. Mean

\pm

sample SD across three seeds, paired with EASER. Passed is pooled across 7,680 candidates per condition. Appendix C.4 defines the removed inputs.

Method	HV $\uparrow$	IGD+ $\downarrow$	Pass
EASER	$0.0439(\pm 0.0117)$	$0.0310(\pm 0.0265)$	$871$
Without memory and anchor reuse	$0.0350(\pm 0.0084)$	$0.0558(\pm 0.0267)$	$524$
Without retrieval	$0.0359(\pm 0.0080)$	$0.0544(\pm 0.0321)$	$867$
Without Probe-and-Steer	$0.0293(\pm 0.0065)$	$0.0692(\pm 0.0299)$	$1,089$
All matrices randomized	$0.0345(\pm 0.0025)$	$0.0499(\pm 0.0128)$	$1,076$

Table 6: Component ablations and matrix replacement. Mean

\pm

sample SD across three seeds, paired with EASER. This panel and Table 5 use disjoint three-run subsets of the main comparison. Metrics use screened, unpenalized scores and the common reference; Pass is pooled across 7,680 candidates.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Target	Predictor	Training data and method
Internal feedback
Activity	LLAMP ( Bae et al., 2025 )	Released MIC-trained model combining peptide ESM-2 representations with bacterial genomic features.
Hemolysis	Calibrated ConsAMPHemo ( Xie et al., 2025 )	Released HC50 regression model; we fit a gradient-boosted hemolysis classifier using its score and physicochemical features.
Toxicity	ToxinPred3 ( Rathore et al., 2024 )	Released toxic/non-toxic sequence classifier; the native hybrid score supplies toxicity feedback.
External evaluation
Activity	MIC regression model	Augmented experimental MIC records from four bacterial species; ESM-2 and sequence features feed a neural regressor. Evaluated for E. coli .

Appendix

Table 7: Predictor overview. Released models remain frozen during search; locally trained evaluators are fitted offline.

Search-policy baselines
Method	Exact HV $\uparrow$	IGD+ $\downarrow$
Fixed coefficients	$0.0277(\pm 0.0079)$	$0.0801(\pm 0.0339)$
Random actions	$0.0373(\pm 0.0086)$	$0.0485(\pm 0.0222)$
LinUCB	$0.0373(\pm 0.0087)$	$0.0544(\pm 0.0261)$
NSGA-II	$0.0344(\pm 0.0069)$	$0.0591(\pm 0.0246)$
External generators

Appendix

Table 8: Sensitivity to score penalties. The search-policy samples match Table 3 ; AMPGPT2 samples match Table 2 . Only the evaluation coordinates change. Mean

\pm

sample SD across six seeds.

Figure 4: Screening and reference sensitivity. Left: vary the common screening threshold with the HV reference fixed at 0.45. Right: vary the common HV reference coordinate with screening fixed at 0.50. All HV calculations use normalized scores before penalty and exact volume. Points and error bars are run means and sample SD across six seeds.

Use	Source and processing
Generator adaptation	Reviewed UniProtKB/Swiss-Prot ( The UniProt Consortium, 2025 ) short proteins and mature peptides, followed by a curated AMP collection. Sequences are deduplicated and restricted to 7–50 residues; the Swiss-Prot pool is split by sequence clusters.
Activity preferences	Experimental MIC records from AMPBench-MT ( Zhou et al., 2026b ) , including DBAASP, GRAMPA, CAMPR4, and DRAMP sources ( Pirtskhalava et al., 2021 ; Witten and Witten, 2019 ; Gawde et al., 2023 ; Shi et al., 2022 ) . The cleaned E. coli subset supplies physicochemically matched preference pairs.
Safety preferences	A combined experimental HC50 collection and a curated binary toxicity collection. Sequences are cleaned, filtered, and paired by the corresponding safety labels and physicochemical features.
Activity evaluator	Integrated experimental MIC data, supplemented with GRAMPA ( Witten and Witten, 2019 ) , AMP-Diffusion data ( Torres et al., 2025 ) , and published in-vitro measurements. Sequence, species, and MIC records are consolidated for regression.
AMP reference	AMP-Diffusion ( Torres et al., 2025 ) , AMPBench-MT ( Zhou et al., 2026b ) , and curated peptide activity collections are merged and deduplicated. The 27,733 sequences of length 15–35 provide the physicochemical KDE reference.
Agent knowledge	Curated property evidence, amino-acid guidance, motif-risk cards, and intervention strategies are organized into retrievable entries.

Appendix

Table 9: Data sources and processing.

Setting	Stage 1 Short peptides	Stage 2 AMP-oriented
Train / validation records	7,453 / 848	8,725 / 1,255
LoRA rank / scaling alpha	8 / 16	4 / 8
LoRA dropout	0.05	0.10
Target projections	Query, value	Query, value
Learning rate	$10^{-5}$	$5\times 10^{-6}$
Warmup steps	300	150

Appendix

Table 10: Offline backbone adaptation. Training settings for the two adaptation stages. Counts denote sequence records.

Property	Preference pairs			Selected step	Parameters
	Train	Valid.	Test
Activity	$324$	$14$	$20$	$140$	$21,008$
Low hemolysis	$406$	$49$	$44$	$170$	$21,008$
Low toxicity	$200$	$58$	$30$	$210$	$21,008$

Appendix

Table 11: Property-matrix training data and configuration. Preference-pair counts, selected checkpoints, and trainable factor parameters for each property.

Figure 5: Residue preferences and contextual variation of output interventions. Top: mean centered logit shifts for learned matrices at 50% masking across 6,258 masked positions from 500 peptides. Bottom: fraction of squared shift energy captured by a constant mean vector, evaluated separately at each mask fraction. Each random control preserves the corresponding learned matrix’s spectrum.

$c$	Activity score	Non-hemolysis score	Non-toxicity score
$-1$	$0.24837$	$0.44792$	$0.66840$
$-0.5$	$0.26753$	$0.44699$	$0.71803$
$0$	$0.25784$	$0.45066$	$0.75425$
$0.25$	$0.25859$	$0.45472$	$0.77322$
$0.5$	$0.28017$	$0.46308$	$0.79254$
$0.75$	$0.28755$	$0.47216$	$0.79798$

Appendix

Table 12: Learned target means across the full signed dose grid. Normalized scores before penalty, 500 draws per condition; only the named property matrix is active.

Active matrix	Activity score	Non-hemolysis score	Non-toxicity score
Activity	$+0.03188$	$-0.00133$	$+0.02242$
Low hemolysis	$+0.02132$	$+0.03170$	$+0.04748$
Low toxicity	$-0.01280$	$+0.02733$	$+0.06996$

Appendix

Table 13: Cross-property effects at

c=1

. Learned

c=1

minus zero control, averaged over all 500 de novo draws before screening.

Matrix set	Activity score	Non-hemolysis score	Non-toxicity score
Learned	$+0.03188$	$+0.03170$	$+0.06996$
Random 1	$+0.00261$	$+0.00448$	$+0.01503$
Random 2	$+0.00461$	$+0.00701$	$+0.00057$
Random 3	$+0.00577$	$-0.00398$	$+0.01226$

Appendix

Table 14: Target responses under full de novo sampling. Mean target-score change before penalty from

c=0

c=1

, with the other coefficients zero. Each condition contains 500 draws; the three random sets preserve each learned matrix's singular-value spectrum and norm.

Matrix set	Activity score	Non-hemolysis score	Non-toxicity score
Learned	$+0.06525$	$+0.03074$	$+0.08807$
Random 1	$+0.01832$	$+0.01226$	$+0.01455$
Random 2	$+0.01296$	$+0.01353$	$+0.02000$
Random 3	$+0.01834$	$+0.01517$	$+0.01579$

Appendix

Table 15: Best objective-specific mean gains in the joint coefficient sweep. Changes in KDE-adjusted utilities relative to zero control. Each cell is the maximum over 125 combinations evaluated on the same 500-draw schedule. The three columns can select different combinations.

Matrix set	Activity score (%)	Non-hemolysis score (%)	Non-toxicity score (%)	Pooled HV
Learned	$41.9$	$26.6$	$44.4$	$0.04957$
Random 1	$10.5$	$2.4$	$8.9$	$0.03686$
Random 2	$4.0$	$3.2$	$12.9$	$0.03516$
Random 3	$14.5$	$3.2$	$12.9$	$0.03844$

Appendix

Table 16: Response coverage and pooled candidate quality in the full grid. The first three columns give percentages of 124 nonzero settings whose adjusted target mean increased by more than 0.01. HV pools 62,500 draws per set, screens on adjusted scores, and uses unpenalized normalized coordinates with reference

(0.5,0.5,0.5)

Figure 6: Distribution of control responses. Empirical cumulative distributions of adjusted mean changes from zero control across 124 nonzero settings, each with 500 draws. Learned controls produce larger beneficial and adverse responses.

Search-policy baselines
Method	Activity score $\uparrow$	Non-hemolysis score $\uparrow$	Non-toxicity score $\uparrow$
Fixed coefficients	$0.842(\pm 0.089)$	$0.667(\pm 0.036)$	$0.959(\pm 0.003)$
Random actions	$0.837(\pm 0.026)$	$0.717(\pm 0.039)$	$0.963(\pm 0.003)$
LinUCB	$0.861(\pm 0.027)$	$0.725(\pm 0.026)$	$0.961(\pm 0.002)$
NSGA-II	$0.830(\pm 0.060)$	$0.694(\pm 0.017)$	$0.961(\pm 0.001)$
Proposed framework

Appendix

Table 17: Search-policy top ten after screening. Each column ranks by its own normalized score before penalty within each seed. Mean

\pm

sample SD across six seeds; every included run had at least ten feasible candidates.

Method	Activity score	Non-hemolysis score	Non-toxicity score
AMPGPT2	$0.831(\pm 0.010)$	$0.640(\pm 0.007)$	$0.945(\pm 0.002)$
AMPGAN	$0.757(\pm 0.011)$	$0.651(\pm 0.005)$	$0.956(\pm 0.001)$
AMPGen	$0.667(\pm 0.038)$	$0.595(\pm 0.016)$	$0.945(\pm 0.004)$
AMPDesigner	$0.806(\pm 0.011)$	$0.653(\pm 0.008)$	$0.959(\pm 0.001)$
AMP-adapted DPLM	$0.723(\pm 0.025)$	$0.620(\pm 0.013)$	$0.946(\pm 0.003)$
EASER	$0.824(\pm 0.095)$	$0.686(\pm 0.038)$	$0.959(\pm 0.006)$

Appendix

Table 18: External generator Top-50 after screening. Mean

\pm

sample SD of six within-seed target means; each column selects its own sequences. All seeds contain at least 50 screened candidates.

Search-policy baselines
Method	Pareto sequences	Occupied cells	Coverage (%)
Fixed coefficients	$23.3(\pm 15.3)$	$15.2(\pm 7.9)$	$11.9(\pm 10.8)$
Random actions	$15.5(\pm 6.2)$	$13.8(\pm 3.9)$	$16.9(\pm 11.0)$
LinUCB	$26.0(\pm 12.6)$	$17.8(\pm 6.0)$	$18.3(\pm 8.3)$
NSGA-II	$12.7(\pm 4.2)$	$11.7(\pm 3.8)$	$11.2(\pm 4.5)$
External generators

Appendix

Table 19: Within-run Pareto counts and coverage. Mean

\pm

sample SD across six runs. Coverage uses

\epsilon=0.025

and is reported in percent; occupied cells have side length 0.05. Sequence counts are deduplicated within each run.

Panel of Figure 7 — Figure 7: Physicochemical profiles after screening. Three seeds and 7,680 candidate records per generator before screening; retained counts: EASER 871, AMPGPT2 846, AMPGAN 530, AMPGen 189, AMPDesigner 691, and AMP-adapted DPLM 293. The real-AMP reference has 2,000 sequences, separate from the 27,733-point KDE reference. Boxes show median and interquartile range; whiskers extend to 1.5 IQR and points mark remaining observations. Helix-prone fraction uses EMALK; hydrophobic moment uses the Eisenberg scale and $100^{\circ}$ angle.

Proposal	Activity score	Non-hemolysis score	Non-toxicity score	Steer samples
1	$0.338$	$-0.020$	$0.148$	$0$
2	$-0.036$	$0.642$	$0.283$	$0$
3	$0.226$	$0.544$	$0.528$	$40$
4	$0.279$	$0.628$	$0.917$	$120$

Appendix

Table 20: Recorded Probe feedback and subsequent budgets. Scores are internal feedback means under the online scoring convention; penalties can make values negative. Each proposal used 16 probes.

Score convention	Sequence	Activity score	Non-hemolysis score	Non-toxicity score
Before penalty	Parent	$0.6487$	$0.4792$	$0.9575$
Before penalty	Offspring	$0.6456$	$0.6811$	$0.9559$
After penalty	Parent	$0.6487$	$0.4792$	$0.9575$
After penalty	Offspring	$0.5368$	$0.5723$	$0.8471$

Appendix

Table 21: Parent and offspring under the common external evaluation. Before penalty denotes normalized scores before KDE subtraction; adjusted scores determine eligibility. The three utility dimensions are all maximized.

Control	Predicted CPP score	Score $\geq 0.5$ (%)
Zero	$0.4257(\pm 0.0344)$	$40.4(\pm 5.2)$
Matched random ( $+1$ )	$0.4427(\pm 0.0356)$	$42.2(\pm 3.9)$
Learned ( $+1$ )	0.7067 ± 0.0220	76.8 ± 0.9
Learned ( $-1$ )	$0.1970(\pm 0.0229)$	$13.5(\pm 3.7)$

Appendix

Table 22: CPP control-interface transfer. Mean

\pm

SD across three seeds, with 128 de novo samples per condition and seed.

Explore similar work

Jun 13, 2026cs.CL

Pepti-Agent: An AI Agent for Peptide Design and Optimization

Therapeutic peptides occupy a valuable design space between small molecules and biologics, but their development requires satisfying several competing constraints at once: solubility, hemolytic activity, and nonspecific surface fouling are governed by overlapping sequence features, so improving one property often degrades another. Computational design addresses this by pairing generative models with sequence-based property predictors, iteratively proposing and refining candidates. However, these components are typically wired together as monolithic scripts that are difficult to inspect, extend, or reuse, and they often refine sequences by natural-language reasoning rather than by tracking the evolving multi-property state of each candidate. We present Pepti-Agent, a closed-loop, peptide-specific framework that exposes generation, property prediction, and single-residue mutation as independently inspectable Model Context Protocol (MCP) tools. A large language model controller invokes these tools and consults live predictor output between calls, so refinement is guided by each sequence's current property profile rather than by language reasoning alone. Task-specific PeptideGPT models generate candidates, ProtBERT-based classifiers score solubility, hemolysis, and non-fouling, and two interchangeable mutation operators propose sequence edits. By recording a per-step trace of controller decisions, predictor outputs, and accepted mutations, Pepti-Agent offers a reproducible substrate for benchmarking multi-objective design strategies and for prioritizing candidates for experimental validation.

Houxu Chen, Achuth Chandrasekhar, Amir Barati Farimani

Biomedical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA · Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA · Chemical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA +2

Lattice Mutation Molecular Property Prediction Artificial Intelligence Agents

Ask PDF

Jun 1, 2026cs.AI

AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design

Protein language models (PLMs) are passive oracles: they generate sequences in a single forward pass with no mechanism to consult external biophysical feedback or redirect generation when a candidate violates thermodynamic or structural constraints. We introduce AgentPLM, which addresses this by equipping a pre-trained PLM with i) Reasoning-Augmented Decoding (RAD), which interleaves autoregressive generation with tool calls (ESMFold, FoldX, AutoDock Vina), and ii) Contrastive Agent Policy Optimisation (CAPO), a trajectory-level extension of direct preference optimisation that trains the policy end-to-end to learn when oracle feedback is informative rather than merely imitating high-fitness sequences. We evaluate AgentPLM on benchmark tasks spanning de novo enzyme design, antibody optimisation, thermostability, PPI interface design, and zero-shot fitness prediction with standardised oracle APIs and controlled sequence-identity splits. AgentPLM achieves state-of-the-art results with a gain in antibody top-10% hit rate over the strongest passive baseline, providing mechanistic evidence of online error correction without explicit backtracking.

Sahil Rahman, Maxx Richard Rahman

Bedford College, London, United Kingdom · Saarland University, Saarbrücken, Germany

Protein Design Antibody Language Modeling Search-Augmented Reasoning Physical Activities

Ask PDF

Apr 18, 2026q-bio.QM

ProtoCycle: Reflective Tool-Augmented Planning for Text-Guided Protein Design

Designing proteins that satisfy natural language functional requirements is a central goal in protein engineering. A straightforward baseline is to fine-tune generic instruction-tuned LLMs as direct text-to-sequence generators, but this is data- and compute-hungry. With limited supervision, LLMs can produce coherent plans in text yet fail to reliably realize them as sequences. This plan-execute gap motivates ProtoCycle, an agentic framework for protein design that uses LLMs primarily to drive a multi-round, feedback-driven decision cycle. ProtoCycle couples an LLM planner with a lightweight tool environment designed to emulate the iterative workflow of human protein engineering and uses LLM-driven reflection on tool feedback to revise plans. Trained with supervised trajectories and online reinforcement learning, ProtoCycle achieves strong language alignment while maintaining competitive foldability, and ablations show that reflection substantially improves sequence quality.

Yutang Ge, Guojiang Zhao, Sihang Li +7

Shanghai Jiao Tong University, School of Mathematical Sciences · DP Technology · University of Science and Technology of China +1

Protein Design Sequence Modeling Planning

Ask PDF