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

Source: arXiv - 2602.14926v1

Overview

The paper presents MAC‑AMP, a closed‑loop system that lets multiple large‑language‑model (LLM) agents collaborate to design antimicrobial peptides (AMPs) that simultaneously satisfy several competing goals—high antibacterial activity, low toxicity, structural plausibility, and novelty. By framing peptide design as a peer‑review‑style reinforcement‑learning loop, the authors show how AI can move beyond single‑objective “black‑box” generators toward explainable, multi‑objective molecular engineering.

Key Contributions

• Closed‑loop multi‑agent framework: Introduces a peer‑review‑inspired reinforcement‑learning loop where LLM agents generate, critique, and refine peptide sequences autonomously.
• Multi‑objective optimization: Simultaneously optimizes activity, toxicity, novelty, and structural reliability without hand‑crafted scalar scoring functions.
• Explainability by design: Each agent’s feedback is textual and interpretable, allowing developers to trace why a peptide was accepted or rejected.
• Domain‑agnostic architecture: The system only needs a high‑level task description and a small example dataset, making it transferable to other molecular design problems.
• Empirical superiority: Outperforms state‑of‑the‑art AMP generators on benchmark metrics for antibacterial potency, AMP‑likeness, toxicity compliance, and structural validity.

Methodology

1. Task Specification – The user supplies a concise natural‑language description of the design goal (e.g., “design 12‑mer peptides with strong Gram‑negative activity, <5 % hemolysis, and novel sequences”).
2. Seed Dataset – A modest collection of known AMPs (≈ 200–500 sequences) is provided to bootstrap the agents.
3. Agent Roles
   • Generator Agent: Uses an LLM fine‑tuned on peptide data to propose candidate sequences.
   • Reviewer Agents: Three specialized agents evaluate the candidate on activity, toxicity, and structural feasibility, each producing a textual critique and a numeric score.
   • Editor Agent: Synthesizes the reviewers’ comments, decides whether to accept, reject, or request revisions, and feeds the outcome back to the generator.
4. Reinforcement Loop – The system treats the reviewer feedback as a reward signal. The generator updates its prompting strategy via a lightweight policy‑gradient method, gradually biasing generation toward higher‑scoring peptides.
5. Termination – After a fixed number of iterations (or when a performance plateau is reached), the best‑scoring peptides are output together with the full review transcript for human inspection.

The entire pipeline runs autonomously, requiring no hand‑crafted scoring functions or external simulators beyond the LLMs themselves.

Results & Findings

Key takeaways

• The multi‑agent loop consistently nudges the generator toward regions of sequence space that satisfy all objectives, rather than over‑optimizing a single metric.
• The textual reviews provide a transparent audit trail, which is missing in typical GAN‑ or VAE‑based peptide generators.
• Even with a tiny seed dataset, the system discovers peptides that are predicted to be both potent and safe, demonstrating strong data efficiency.

Practical Implications

• Accelerated R&D pipelines – Pharmaceutical teams can plug MAC‑AMP into early‑stage discovery workflows to generate candidate AMPs that already respect toxicity constraints, reducing the number of costly wet‑lab validation rounds.
• Rapid prototyping for niche pathogens – By simply swapping the activity description (e.g., “target Pseudomonas aeruginosa”), the same system can produce tailored peptide libraries without retraining a new model.
• Explainable AI compliance – Regulatory environments increasingly demand traceability of AI‑generated designs. MAC‑AMP’s reviewer comments serve as a built‑in documentation layer, easing audit and IP‑ownership discussions.
• Cross‑domain reuse – The closed‑loop architecture can be repurposed for other sequence‑based design tasks such as enzyme engineering, DNA aptamer discovery, or even non‑biological code generation, requiring only a new task prompt and example set.
• Developer‑friendly integration – The system is built on open‑source LLM APIs and lightweight RL code, making it straightforward to embed in CI/CD pipelines for continuous molecular design.

Limitations & Future Work

• Reliance on LLM prediction quality – The reviewers’ judgments are only as good as the underlying language model’s understanding of peptide biophysics; occasional mis‑scorings were observed for rare amino‑acid motifs.
• Simulation vs. wet‑lab gap – All evaluations are in silico (MIC predictors, AlphaFold confidence). Real‑world synthesis and activity assays may reveal unforeseen failures.
• Scalability of the review loop – As the number of objectives grows, the loop can become computationally expensive; future work could explore hierarchical reviewer structures or distilled reward models.
• Benchmark diversity – Experiments focused on a limited set of bacterial strains; extending to fungi, viruses, and biofilm contexts would broaden applicability.
• Human‑in‑the‑loop extensions – Incorporating expert chemist feedback alongside LLM reviewers could further improve robustness and trustworthiness.

Bottom line: MAC‑AMP showcases how multi‑agent LLM collaboration can turn peptide design from a black‑box optimization problem into an interpretable, multi‑objective engineering workflow—an approach that could reshape AI‑driven drug discovery in the years ahead.

Authors

• Gen Zhou
• Sugitha Janarthanan
• Lianghong Chen
• Pingzhao Hu

Paper Information

• arXiv ID: 2602.14926v1
• Categories: cs.AI
• Published: February 16, 2026
• PDF: Download PDF