MAC-AMP: A Closed-Loop Multi-Agent Collaboration System for Multi-Objective Antimicrobial Peptide Design
Gen Zhou, Sugitha Janarthanan, Lianghong Chen, Pingzhao Hu
TL;DR
MAC-AMP is a novel multi-agent AI system that autonomously designs antimicrobial peptides by balancing multiple objectives, outperforming existing models in key properties like activity, toxicity, and structural reliability.
Contribution
Introduces a closed-loop multi-agent system using LLMs for multi-objective AMP design with explainability and cross-domain transferability.
Findings
Outperforms existing AMP generative models in key molecular properties
Effectively balances activity, toxicity, and structural reliability
Demonstrates superior optimization in multi-objective AMP design
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…
Peer Reviews
Decision·ICLR 2026 Poster
- The paper introduces a credible pipeline that operationalizes “multi-agent collaboration” beyond conversational coordination, producing quantitative, auditable training signals. - Transparent logging and role-based agent structure allow reproducibility and traceability uncommon in LLM-based systems. - The system’s iterative RL refinement aligns multi-agent consensus with executable objectives, avoiding reward hacking typical in static scoring systems. - Results across multiple bacterial target
- The quality of generated AMPs is constrained by the accuracy of property predictors (e.g., ToxinPred 3.0, OmegaFold). The paper acknowledges this but does not quantify its impact. - While broad-spectrum testing is reported, the reasoning behind cross-species generalization (beyond physicochemical similarity) could be more rigorously analyzed. - Although individual module ablations are provided, it remains unclear whether a simpler reward design or fewer agents would achieve similar performance
The paper introduces a closed-loop multi-agent collaboration framework that transforms AMP design into an autonomous, explainable reinforcement learning process. Its integration of AI-simulated peer review to generate executable reward signals is both novel and technically elegant. It is clearly written, with transparent modular design and reproducible details. Overall, the paper is significant, offering a transferable paradigm for multi-objective molecular design and broader autonomous scientif
While the framework is conceptually strong, several areas could be improved. 1. The novelty is quite limited. Using existing PPO and peptide generation model (Wang et al. 2025) 2. Experimental validation is limited to in silico analyses—no wet-lab or biophysical confirmation is provided to verify that the generated AMPs exhibit real-world antimicrobial activity, which would strengthen the biological significance. 3. Although the AI-simulated peer review concept is creative, the paper could inc
The paper introduces a novel closed-loop multi-agent framework that compiles structured peer-review consensus into PPO-executable rewards, enabling interpretable and auditable multi-objective optimization for antimicrobial peptide design. Empirically, it reports consistent computational improvements on activity, toxicity, and structural reliability metrics over several baselines, supported by a clearly described system architecture and informative ablation studies.
1) The work is difficult to reproduce at review time. Code and data are not publicly available; while the authors commit to releasing them upon acceptance, independent replication and verification cannot be performed during evaluation. 2) Core evaluation and constraints depend on external tools (e.g., ToxinPred 3.0, OmegaFold, Macrel, Foldseek). The reward and selection are tightly coupled to their outputs, yet there is no systematic robustness or bias analysis to assess how these components inf
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAntimicrobial Peptides and Activities · vaccines and immunoinformatics approaches · Machine Learning in Bioinformatics