Research map / Multi-agent systems and games / LLM agents

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

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MAC-AMP, a closed-loop multi-agent LLM system (simulated peer review feeding PPO reward redesign) that generates multi-objective antimicrobial peptides for target bacteria, outperforming AMP-Designer, BroadAMP-GPT, PepGAN and Diff-AMP on in-silico activity, toxicity and structural reliability.

Problem

AMP design models struggle to balance activity, toxicity, novelty and stability with rigid scoring, and LLM multi-agent systems are open-loop and produce non-executable outputs rather than reward signals for optimization.

Approach

Closed-loop multi-agent system in which simulated peer reviewers (LLM agents with weighted lexicon tags and an area chair) convert multi-criteria evaluations into machine-actionable reward functions that are iteratively redesigned by CS and biomedical reward-design agents and validated in sandbox runs, driving PPO training of a GPT-2 peptide generator.

Key results

• MAC-AMP reaches 0.943 ± 0.008 antibacterial activity and 0.154 ± 0.008 toxicity on E. coli, versus 0.807–0.831 activity and 0.235–0.251 toxicity for LLM and traditional baselines (Table 1).
• Structural reliability of 0.873 ± 0.009 on E. coli is the highest among compared models, while AMP likelihood (0.797) is below BroadAMP-GPT (0.821).
• 36 of 90 (40%) generated E. coli AMPs are predicted broad-spectrum active against the five additional ESKAPE pathogens by APEX 1.1.

Surprising findings

• Diff-AMP and PepGAN achieve higher S. aureus and P. aeruginosa activity than MAC-AMP in some cases (e.g., Diff-AMP 0.926 vs 0.931 is close; PepGAN 0.901 on S. aureus), and MAC-AMP's broad-spectrum S. aureus score (0.81) is lower than baselines.
• BroadAMP-GPT has highest AMP likelihood on S. aureus and P. aeruginosa while showing toxicity of ~0.45, which the authors note as a trade-off.

Limitations

• Evaluation relies entirely on in-silico predictors (MIC predictor, Macrel, ToxinPred 3.0, OmegaFold) without wet-lab validation of generated peptides.
• Reported results are for a single generator configuration and the reward/review pipeline depends on proprietary LLM APIs (GPT-5, Gemini 2.5, Perplexity), limiting reproducibility.
• Motif interpretations (e.g., KFLKGA, WLLGKW) are not experimentally validated.

Experimental setup

Datasets: DBAASP v3 dbAMP 3.0 UniProtKB reference proteomes

Benchmarks: E. coli AMP generation (antibacterial activity, AMP likelihood, toxicity, structural reliability) S. aureus AMP generation P. aeruginosa AMP generation E. coli broad-spectrum testing with APEX 1.1

Metrics: predicted MIC activity score Macrel AMP likelihood ToxinPred 3.0 toxicity OmegaFold pLDDT structural reliability Needleman-Wunsch similarity

Baselines: AMP-Designer BroadAMP-GPT PepGAN Diff-AMP Real-World top-K

Builds on: AMP-Designer Eureka AI Virtual Lab PPO

Closest papers by meaning

• Protein Design with Agent Rosetta: A Case Study for Specialized Scientific Agents (ICML 2026 · similarity 0.85)
• Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.84)
• A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification (NeurIPS 2025 · similarity 0.84)
• PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · similarity 0.83)
• ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design (NeurIPS 2026 · similarity 0.82)
• Searching for High-Value Molecules Using Reinforcement Learning and Transformers (ICLR 2024 · similarity 0.82)