Advancing Protein Design via Multi-Agent Reinforcement Learning with Pareto-Based Collaborative Optimization

Mingming Zhu, Jiahua Rao, Xiaoyu Chen, Qianmu Yuan, Yuedong Yang2026 · OpenReview

TL;DR

This paper introduces MAProt, a novel multi-agent framework that synergistically combines structure-based and protein language models for protein design. By employing a Pareto-based negotiation module, MAProt effectively balances conflicting objectives like structural stability and functional performance, achieving state-of-the-art results across various protein design tasks.

What Problem Does This Paper Address?

Designing proteins with specific functions and stable structures is a cornerstone of biotechnology, but it's incredibly challenging. Current methods often fall into two camps: structure-based models (like ProteinMPNN) are great at ensuring a protein folds correctly but might miss crucial functional aspects, while protein language models (like ESM) capture evolutionary and functional signals but can generate sequences that are structurally unstable. The core problem is that these two objectives—structural integrity and functional performance—often conflict. Optimizing one can negatively impact the other, and existing approaches struggle to find a harmonious balance, especially given the vast combinatorial space of possible protein sequences.

Methods

MAProt tackles this challenge by framing protein design as a multi-agent reinforcement learning problem. The framework comprises specialized "agents," each focusing on a distinct aspect of protein design:

  • Structure-based Agent (ProteinMPNN): Ensures generated sequences are compatible with a given protein backbone structure.
  • Protein Language Model Agents (ESM, SaProt): Capture evolutionary plausibility and functional potential by analyzing sequence patterns and, in SaProt's case, integrating structural information.

The key innovation is the Pareto-based negotiation and conflict resolution module. This module allows the agents to "negotiate" their potentially conflicting objectives. It quantifies inter-agent agreement using a consensus score derived from preference margins and structural/continuous consensus measures. This enables the framework to identify "consensus regions" that represent Pareto-optimal solutions, balancing multiple desirable properties. The system then uses this consensus information to construct preference pairs for training, guiding the agents towards solutions that satisfy multiple criteria simultaneously. For sequence generation, MAProt employs a logit fusion strategy, combining the outputs of all agents to produce a more informed and balanced prediction.

Datasets

MAProt was evaluated on three established benchmarks:

  • Megascale: Used for stability optimization, containing approximately 1.8 million variants.
  • GFP (Green Fluorescent Protein): Used for functional phenotype design, with 56,806 variants.
  • AffinityDesign: Used for high-affinity antibody-antigen engineering, involving 47 complexes.

The paper does not introduce new benchmarks but leverages existing ones to demonstrate MAProt's generalizability.

Key Insights and Ideas

  • Synergy through Negotiation: The most significant insight is that by explicitly modeling and resolving conflicts between different protein design objectives (structure vs. function vs. evolution), MAProt achieves superior performance compared to methods that optimize these objectives in isolation or with simpler integration strategies.
  • Pareto Optimality as a Guiding Principle: Applying Pareto optimality to multi-agent coordination provides a principled way to find trade-offs in complex, multi-objective optimization problems like protein design.
  • Complementary Strengths: The paper reinforces the idea that structure-based and language model-based approaches are complementary, and their integration, when done effectively, unlocks new design capabilities.
  • Generalizability: MAProt's architecture is flexible, allowing for the integration of various models and scales, suggesting its potential for broad applicability in protein engineering.

Key Results

MAProt demonstrated state-of-the-art performance across all tested benchmarks:

  • Stability Design (Megascale): Achieved a +14.1% improvement in success rate and a +43.1% improvement in median predicted ΔΔG over the DRAKES baseline.
  • Functional Protein Design (GFP): Outperformed all baselines, achieving a fitness score of 0.88 (medium difficulty) and 0.87 (hard difficulty), with up to 12% higher fitness than the next best method on the hard task.
  • High-Affinity Antibody Design (AffinityDesign): Achieved a 32.1% success rate, a high sequence recovery rate (0.763), and a strong binding affinity (mean ΔAffinity of 0.153), significantly outperforming competitors.
  • Ablation studies confirmed the critical role of the multi-agent collaboration and consensus module, showing that MAProt with full integration achieved the highest success rate (89.7%) on the Megascale benchmark.

Limitations and Future Work

  • Limitations: While MAProt shows remarkable performance, the paper notes that the preference sharpness parameter (β) requires careful tuning to balance structural quality and fitness. Low β can lead to structural deviations, while high β can reduce fitness. The current framework focuses on sequence design given a fixed backbone.
  • Future Work: The authors propose extending MAProt to enable joint structure and sequence co-design, which would be a significant advancement. They also aim to apply the framework to more complex protein assemblies and de novo design tasks, further expanding its impact on protein engineering and therapeutic discovery. Integrating experimental feedback and active learning is also suggested to enhance practical applications.

Summary updated

Abstract

h3>Abstract</h3> <p>Protein design is revolutionizing biotechnology, yet existing approaches struggle to balance structural foldability with functional performance. Structure-based models excel at generating stable protein backbones but often overlook critical functional properties, while protein language models capture evolutionary and functional signals but frequently predict sequences lacking structural stability. Integrating these complementary approaches remains challenging due to their inherently conflicting objectives. We present MAProt, a multiagent framework that synergistically combines structure-based and protein language model-based methods for protein design. Each agent specializes in a distinct aspect of the design objective: the structure-based agent (e.g., ProteinMPNN) ensures compatibility with the target backbone, while protein language model-based agents (e.g., ESM, SaProt) capture evolutionary plausibility and functional potential. To reconcile conflicts and achieve optimal trade-offs, we introduce a Pareto-based negotiation module that enables effective multi-objective coordination and consensus among agents. Extensive experiments on benchmark datasets demonstrate that MAProt achieves a remarkable improvement over state-of-the-art baselines, and generalizes robustly across a range of tasks, including thermodynamic folding stability design, functional protein design, and high-affinity antibody design. These results highlight the power of collaborative optimization for advancing rational protein engineering.</p><h3>Code</h3> <p>https://github.com/biomed-AI/MAProt</p>
BibTeX
@misc{zhu2026advancing,
  author = {Mingming Zhu and Jiahua Rao and Xiaoyu Chen and Qianmu Yuan and Yuedong Yang},
  title = {{Advancing Protein Design via Multi-Agent Reinforcement Learning with Pareto-Based Collaborative Optimization}},
  year = {2026},
  url = {https://openreview.net/forum?id=T26RQWo9ek},
}
Figure