MAC-AMP: A Closed-Loop Multi-Agent Collaboration System for Multi-Objective Antimicrobial Peptide Design

Gen Zhou; Sugitha Janarthanan; Lianghong Chen; Pingzhao Hu

cs.AIFeb 16, 2026

MAC-AMP: A Closed-Loop Multi-Agent Collaboration System for Multi-Objective Antimicrobial Peptide Design

Authors: Gen Zhou, Sugitha Janarthanan, Lianghong Chen, Pingzhao Hu

Organizations: Department of Computer Science, Western University, London, ON, Canada · Department of Biochemistry, Western University, London, ON, Canada

Abstract

To address the global health threat of antimicrobial resistance, antimicrobial peptides (AMP) are being explored for their potent and promising ability to fight resistant pathogens. While artificial intelligence (AI) is being employed to advance AMP discovery and design, most AMP design models struggle to balance key goals like activity, toxicity, and novelty, using rigid or unclear scoring methods that make results hard to interpret and optimize. As the capabilities of Large Language Models (LLM) advance and evolve swiftly, we turn to AI multi-agent collaboration based on such models (multi-agent LLMs), which show rapidly rising potential in complex scientific design scenarios. Based on this, we introduce MAC-AMP, a closed-loop multi-agent collaboration (MAC) system for multi-objective AMP design. The system implements a fully autonomous simulated peer review-adaptive reinforcement learning framework that requires only a task description and example dataset to design novel AMPs. The novelty of our work lies in introducing a closed-loop multi-agent system for AMP design, with cross-domain transferability, that supports multi-objective optimization while remaining explainable rather than a 'black box'. Experiments show that MAC-AMP outperforms other AMP generative models by effectively optimizing AMP generation for multiple key molecular properties, demonstrating exceptional results in antibacterial activity, AMP likeliness, toxicity compliance, and structural reliability.

Figures & tables

Figure 1: MAC-AMP Framework. (a) Overview of the closed-loop workflow that iteratively guides AMP design from input to output. (b) Schematic of the MAC-AMP pipeline, showing its modules and their interactions.

Figure 2: Overview of the Artificial Intelligence-simulated Peer Review module. Green text indicates the injectable section, while the green boxes denote content obtained from the separate preparatory meeting.

Figure 3: Overview of the Reinforcement Learning Refinement module.

Model	Antibacterial Activity ( $\uparrow$ )	AMP Likelihood ( $\uparrow$ )	Toxicity ( $\downarrow$ )	Structural Reliability ( $\uparrow$ )
Escherichia coli (E. coli)
MAC-AMP	0.943 ± 0.008	0.797 ± 0.012	0.154 ± 0.008	0.873 ± 0.009
AMP-Designer	0.807 ± 0.021	0.811 ± 0.011	0.251 ± 0.024	0.817 ± 0.017
BroadAMP-GPT	0.831 ± 0.025	0.821 ± 0.018	0.246 ± 0.033	0.763 ± 0.023
PepGAN	0.823 ± 0.023	0.572 ± 0.035	0.247 ± 0.064	0.637 ± 0.026
Diff-AMP	0.822 ± 0.006	0.554 ± 0.036	0.235 ± 0.072	0.752 ± 0.020

Table 1: Property scores of antimicrobial peptides (AMP) generated by MAC-AMP compared to baseline models and real-world AMP datasets across bacterial species

Model	E. coli	S. aureus	P. aeruginosa	K. pneumoniae	E. faecium
MAC-AMP	0.94 ± 0.01	0.81 ± 0.03	0.94 ± 0.00	0.98 ± 0.00	0.95 ± 0.01
AMP-Designer	0.81 ± 0.02	0.81 ± 0.02	0.85 ± 0.01	0.96 ± 0.01	0.96 ± 0.01
BroadAMP-GPT	0.83 ± 0.02	0.82 ± 0.03	0.87 ± 0.02	0.96 ± 0.01	0.97 ± 0.01
PepGAN	0.82 ± 0.02	0.89 ± 0.02	0.91 ± 0.01	0.98 ± 0.00	0.96 ± 0.01
Diff-AMP	0.82 ± 0.01	0.91 ± 0.01	0.94 ± 0.01	0.98 ± 0.00	0.93 ± 0.01

Table 2: Antibacterial activity scores of Escherichia coli -targeted antimicrobial peptides against other bacterial strains

Figure 4: (a) UMAP of the peptide chemical space for Escherichia coli (E. coli) inhibition. Gray dots show UniProt peptides, orange dots show real-world E. coli AMPs, and red dots are MAC-AMP candidate AMPs. Nested blue contours indicate increasing kernel density from these AMPs. (b,c) Sequence diagrams of two example MAC-AMP designed E. coli AMPs.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Three-stage reward–episode learning curves for MAC-AMP. The curves illustrate how the total reward and its components ( $S_{a}$ , antibacterial activity score; $S_{b}$ , AMP likelihood score; and $S_{c}$ , average meta score) evolve during (a) the early stage, (b) the middle stage, and (c) the late stage. The light blue area indicates the moving variance of the total reward.

Figure A2: Predicted Minimum Inhibitory Concentrations (MICs) of the top 90 MAC-AMP-generated anti- Escherichia coli (E. coli) peptides against E. coli strains and ESKAPE pathogens using APEX 1.1.

Figure A3: Frequency of the top ten motifs identified in broad-spectrum MAC-AMP–generated anti- Escherichia coli antimicrobial peptides.

Figure A4: Distribution of (a) average and (b) highest normalized Needleman–Wunsch global-alignment similarity scores for the top 90 MAC-AMP–generated anti- Escherichia coli (E. coli) antimicrobial peptides (AMPs), evaluated against an external dataset of real-world E. coli AMPs.

Figure A5: Normalized Needleman-Wunsch global-alignment–based similarity scores between top 90 anti- Escherichia coli generated antimicrobial peptides by MAC-AMP.

AMP	RMSD (Å)
1	2.22 ± 0.79
2	2.93 ± 0.97
3	2.16 ± 1.12
4	3.40 ± 1.85
5	2.17 ± 0.68
6	2.66 ± 0.53

Appendix

Table A1: Backbone root mean square deviation (RMSD) of molecular dynamics simulations of MAC-AMP-generated anti- Escherichia coli antimicrobial peptides (AMPs) (mean

\pm

standard deviation)

Sequence	Length	K/R Fraction	K/R Fraction Positions	Identity Fraction
FRVFGFIAKKVKKLVKKI	18	0.389	1,8,9,11,12,15,16	0.556
VRGGAIKKIAKILAKLLAR	19	0.316	1,6,7,10,14,18	0.579
VGLVKKWFKSVIKKVAKS	18	0.333	4,5,8,12,13,16	0.500
RIFKFLKRAFGIIGLFKRRIKS	22	0.364	0,3,6,7,16,17,18,20	0.500
KIWKLLKKVLAKVAK	15	0.400	0,3,6,7,11,14	0.600
IIGKLVLKKVGKIIKKILKKKA	22	0.409	3,7,8,11,14,15,18,19,20	0.500

Appendix

Table A2: Summary of sequence-level features for six generated antimicrobial peptides by MAC-AMP for Escherichia coli

Sequence	GRAVY	Hydrophobic Moment ${}_{\alpha\text{-helix}}$	Net Charge ${}_{\text{pH7}}$	pI
FRVFGFIAKKVKKLVKKI	0.406	0.734	7	11.95
VRGGAIKKIAKILAKLLAR	0.574	0.375	6	12.52
VGLVKKWFKSVIKKVAKS	0.189	0.546	6	11.25
RIFKFLKRAFGIIGLFKRRIKS	0.155	0.578	8	12.97
KIWKLLKKVLAKVAK	0.240	0.721	6	11.25
IIGKLVLKKVGKIIKKILKKKA	0.373	0.427	9	11.45

Appendix

Table A3: Summary of biophysical properties for six generated antimicrobial peptides by MAC-AMP for Escherichia coli

Figure A6: Comparative distributions of sequence and biophysical properties across 2000 random antimicrobial peptides (AMPs) generated by MAC-AMP, five baseline models, and a real-world AMP dataset targeting Escherichia coli . (a) Amino-acid composition of AMPs, normalized to per-model fractions. (b) Global charge at pH 7.4 for all AMPs. (c) Grand Average of Hydropathy (GRAVY) of AMPs, representing overall hydrophobicity. (d) Predicted helix fraction of AMPs, indicating their propensity to adopt $\alpha$ -helical structure. (e) AMPs length in amino acids. (f) Eisenberg hydrophobic moment of AMPs, reflecting amphipathicity. (g) Instability index of AMPs, estimating their predicted stability. (h) Global hydrophobicity of AMPs.

Model	Antibacterial Activity ( $\uparrow$ )	AMP Likelihood ( $\uparrow$ )	Toxicity ( $\downarrow$ )	Structural Reliability ( $\uparrow$ )
MAC-AMP	$0.943\pm 0.008$	$0.797\pm 0.012$	$0.154\pm 0.008$	$0.873\pm 0.009$
Drop_ $V_{b}$	$0.904\pm 0.028$	$0.799\pm 0.009$	$0.157\pm 0.021$	$0.818\pm 0.012$
Drop_ $V_{a}$	$0.946\pm 0.045$	$0.782\pm 0.016$	$0.231\pm 0.028$	$0.799\pm 0.017$
Drop_ $S_{b}$	$0.923\pm 0.037$	$0.742\pm 0.006$	$0.164\pm 0.010$	$0.830\pm 0.014$
Drop_ $V_{a}$ _ $V_{b}$	$0.904\pm 0.043$	$0.764\pm 0.023$	$0.183\pm 0.016$	$0.758\pm 0.015$
Drop_ $V_{b}$ _ $S_{b}$	$0.889\pm 0.040$	$0.745\pm 0.035$	$0.134\pm 0.087$	$0.763\pm 0.017$

Appendix

Table A4: Ablation studies of the Property Prediction module: property scores of antimicrobial peptides generated by MAC-AMP and its variants. Here,

V_{a}

denotes the toxicity predictor (ToxinPred 3.0),

V_{b}

the structural-reliability property predictor (OmegaFold v1), and

S_{b}

the AMP-likelihood predictor (Macrel 1.5).

Figure A7: Ablation studies of the Property Prediction module. (a) API costs for each ablation model, broken down by the percentage contribution of each Reviewer Agent: GPT-5, Perplexity, and Gemini 2.5. (b) GPU hours, total number of API calls (further separated into percentages from the Reinforcement Learning Refinement module versus the Artificial Intelligence-simulated Peer Reviewer module), and peak memory usage (MB) for each ablation model. Here, $V_{a}$ denotes the toxicity predictor (ToxinPred 3.0), $V_{b}$ the structural-reliability property predictor (OmegaFold v1), and $S_{b}$ the AMP-likelihood predictor (Macrel 1.5).

Model	Antibacterial Activity ( $\uparrow$ )	AMP Likelihood ( $\uparrow$ )	Toxicity ( $\downarrow$ )	Structural Reliability ( $\uparrow$ )
MAC-AMP	0.943 $\pm$ 0.008	0.797 $\pm$ 0.012	0.154 $\pm$ 0.008	0.873 $\pm$ 0.009
w/o RA3	0.857 $\pm$ 0.016	0.840 $\pm$ 0.008	0.133 $\pm$ 0.023	0.731 $\pm$ 0.017
w/o RA2	0.884 $\pm$ 0.015	0.780 $\pm$ 0.008	0.164 $\pm$ 0.023	0.734 $\pm$ 0.021
w/o RA1	0.953 $\pm$ 0.010	0.742 $\pm$ 0.016	0.181 $\pm$ 0.019	0.783 $\pm$ 0.017
w/o RA2 + RA3	0.870 $\pm$ 0.013	0.806 $\pm$ 0.006	0.130 $\pm$ 0.015	0.740 $\pm$ 0.011
w/o RA1 + RA3	0.979 $\pm$ 0.010	0.700 $\pm$ 0.019	0.185 $\pm$ 0.019	0.758 $\pm$ 0.015

Appendix

Table A5: Ablation studies of the AI-simulated Peer Review module: property scores of antimicrobial peptides generated by MAC-AMP and its variants. Here, RA1 denotes Reviewer agent 1 (GPT-5), RA2 the Reviewer agent 2 (Perplexity), and RA3 the Reviewer agent 3 (Gemini 2.5)

Model	Antibacterial Activity ( $\uparrow$ )	AMP Likelihood ( $\uparrow$ )	Toxicity ( $\downarrow$ )	Structural Reliability ( $\uparrow$ )
MAC-AMP	0.943 $\pm$ 0.008	0.797 $\pm$ 0.012	0.154 $\pm$ 0.008	0.873 $\pm$ 0.009
Without RL Decision Agent	0.935 $\pm$ 0.011	0.791 $\pm$ 0.013	0.260 $\pm$ 0.022	0.879 $\pm$ 0.010
Without Adaptive Optimization	0.955 $\pm$ 0.003	0.752 $\pm$ 0.005	0.279 $\pm$ 0.006	0.869 $\pm$ 0.013
Without RL Decision Agent & Adaptive Optimization	0.818 $\pm$ 0.021	0.737 $\pm$ 0.009	0.200 $\pm$ 0.022	0.768 $\pm$ 0.020
Replaced by Human-Designed RL	0.900 ± 0.014	0.776 ± 0.009	0.175 ± 0.018	0.759 ± 0.015

Appendix

Table A6: Ablation studies of the Reinforcement Learning Refinement module: property scores of antimicrobial peptides generated by MAC-AMP and its variants

Figure A8: Substitution analyses on the number of Reinforcement Learning module training epochs. (a) API costs for each substitution variant, broken down by the percentage contribution of each Reviewer Agent: GPT-5, Perplexity, and Gemini 2.5. (b) GPU hours, total number of API calls (further separated into percentages from the Reinforcement Learning module versus the AI-simulated Peer Review module), and peak memory usage (MB) for each substitution variant.

Figure A9: Substitution analyses on the maximum length of generated peptides. (a) API costs for each substitution variant, broken down by the percentage contribution of each Reviewer Agent: GPT-5, Perplexity, and Gemini 2.5. (b) GPU hours, total number of API calls (further separated into percentages from the Reinforcement Learning module versus the Artificial Intelligence-simulated Peer Review module), and peak memory usage (MB) for each substitution variant.

Model	Antibacterial Activity ( $\uparrow$ )	AMP Likelihood ( $\uparrow$ )	Toxicity ( $\downarrow$ )	Structural Reliability ( $\uparrow$ )
MAC-AMP	0.943 $\pm$ 0.008	0.797 $\pm$ 0.012	0.154 $\pm$ 0.008	0.873 $\pm$ 0.009
RA_AC_Llama	$0.944\pm 0.007$	$0.713\pm 0.012$	$0.201\pm 0.020$	$0.896\pm 0.012$
RA_AC_Qwen	$0.949\pm 0.009$	$0.683\pm 0.010$	$0.308\pm 0.013$	$0.913\pm 0.016$
RA_API_AC_Qwen	$0.865\pm 0.014$	$0.783\pm 0.011$	$0.206\pm 0.013$	$0.863\pm 0.007$
RA_Qwen_AC_API	$0.835\pm 0.023$	$0.734\pm 0.017$	$0.211\pm 0.030$	$0.854\pm 0.015$

Appendix

Table A7: Substitution analyses on the Artificial Intelligence-simulated Peer Review Module Agents: property scores of antimicrobial peptides generated by MAC-AMP and its variants. RA_AC_Llama: all Reviewer agents (RA) and the Area Chair (AC) agent are replaced with Llama 3.1 8B, RA_AC_Qwen: all RA and the AC agent are replaced with Qwen 2.5 7B-instruct, RA_API_AC_Qwen: only the AC agent was replaced with Qwen, RA_Qwen_AC_API: only the RA are replaced with Qwen

Figure A10: Substitution analyses on the AI-simulated Peer Review Module Agents. (a) API costs for each comparative model, broken down by the percentage contribution of each Reviewer Agent: GPT-5, Perplexity, and Gemini 2.5. (b) GPU hours, total number of API calls (further separated into percentages from the Reinforcement Learning module versus the Artificial Intelligence-simulated Peer Review module), and peak memory usage (MB) for each comparative model.

Predictor	Accuracy	Weighted Precision	Weighted Recall	Weighted F1-score
ToxinPred 3.0	0.828	0.849	0.828	0.828
CAPTP	0.770	0.787	0.77	0.770

Appendix

Table A8: Performance comparison between ToxinPred 3.0 and CAPTP

Predictor	$R^{2}$	Pearson $r$	Spearman $\rho$
MAC-AMP MIC Predictor	0.572	0.758	0.724
APEX 1.1	0.546	0.728	0.607

Appendix

Table A9: Performance comparison between MAC-AMP MIC Predictor and APEX 1.1

Model	Antibacterial Activity ( $\uparrow$ )	AMP Likelihood ( $\uparrow$ )	Toxicity ( $\downarrow$ )	Structural Reliability ( $\uparrow$ )
MAC-AMP	0.943 $\pm$ 0.008	0.797 $\pm$ 0.012	0.154 $\pm$ 0.008	0.873 $\pm$ 0.009
Lexicon Weight + 0.1	$0.878\pm 0.023$	$0.794\pm 0.009$	$0.156\pm 0.012$	$0.836\pm 0.013$
Lexicon Weight - 0.1	$0.938\pm 0.019$	$0.804\pm 0.006$	$0.171\pm 0.012$	$0.871\pm 0.013$

Appendix

Table A10: Substitution analyses on the lexicon weights: property scores of antimicrobial peptides generated by MAC-AMP and its variants

Data Subset	Model	BLEU	PARENT	BLEURT
Overall	T5 Baseline	44.6	56.0	0.179
Overall	T5-based MAC-AMP	46.2	58.0	0.208
Overlap	T5 Baseline	52.6	60.7	0.311
Overlap	T5-based MAC-AMP	54.1	62.4	0.338
Non-Overlap	T5 Baseline	36.8	51.4	0.051
Non-Overlap	T5-based MAC-AMP	38.5	53.7	0.083

Appendix

Table A11: Comparison of the T5 baseline and T5-based MAC-AMP agent framework for the English table-to-text generation task

Molecular Generation

Molecular generation develops computational methods to create chemical structures and conformations for discovery, optimization, and structure elucidation.

See all 234 papers

Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3
Jun 15, 2026Jay Jung, Xiaohan Zhang, Shenghan Song +8AI Agents for Scientific Discovery Molecular Generation
Department of Computer Science, University of Vermont, Burlington, VT, USA · Borch Department of Medicinal Chemistry and Molecular Pharmacology, Purdue University, West Lafayette, IN, USA · Department of Comparative Pathobiology, Purdue University, West Lafayette, IN, USA +3
OmegAMP: Targeted AMP Discovery via Biologically Informed Generation
Apr 24, 2025Diogo Soares, Leon Hetzel, Paulina Szymczak +6Molecular Generation Conditional Generative Modeling
Institute of AI for Health, Helmholtz Munich · School of Computation, Information and Technology, Technical University of Munich · Institute of Computational Biology, Helmholtz Munich +8
Pepti-Agent: An AI Agent for Peptide Design and Optimization
Jun 13, 2026Houxu Chen, Achuth Chandrasekhar, Amir Barati FarimaniAI Agents for Scientific Discovery Large Language Model-Guided Optimization
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

Jun 15, 2026J/K move · Enter open · S save

Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3

Jay Jung, Xiaohan Zhang, Shenghan Song, Mahmoud Sayedahmed+7

Department of Computer Science, University of Vermont, Burlington, VT, USA · Borch Department of Medicinal Chemistry and Molecular Pharmacology, Purdue University, West Lafayette, IN, USA · Department of Comparative Pathobiology, Purdue University, West Lafayette, IN, USA+3

AI Agents for Scientific Discovery Molecular Generation Conditional Generative Modeling Generative Adversarial Networks Drug Discovery

Open paper AskPDF arXiv

Antimicrobial resistance causes to over a million deaths annually. Antimicrobial peptides (AMPs) are a promising solution, but generative AMP models are not yet ready to design peptides with non-natural amino acids and/or chemical modifications, which are essential for real-world peptide drugs. We present AMPGAN v3, a multi-objective conditional GAN that expands the generative vocabulary to D-amino acids and N/C-terminus modifications such as amidation. By separating adversarial and activity-aware supervision across two specialized discriminators, AMPGAN v3 substantially improves training stability and outperforms prior generative AMP models on external classifiers. We validated five candidates spanning three structural classes in vitro; two showed activity against Gram-positive strains, with the best candidate reaching MIC 8 μg/mL against B. subtilis. To support downstream curation, we further present PepCraft, a multi-agent framework for end-to-end AMP discovery in which a Planning Agent orchestrates specialized executors for generation, filtering, and verification. Its prioritization recommendations align with our in vitro outcomes. Together, these contributions let us examine, on a small but real scale, how generative and agentic AI compose in therapeutic peptide discovery. Code: https://github.com/marszzibros/AMPGANv3

Abstract

Molecular Generation

Explore similar work

Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3

OmegAMP: Targeted AMP Discovery via Biologically Informed Generation

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