MMORF: A Multi-agent Framework for Designing Multi-objective Retrosynthesis Planning Systems

A research team led by Frazier N. Baker from Ohio State University has introduced MMORF, a novel framework for constructing multi-agent systems (MAS) specifically designed for multi-objective retrosynthesis planning in chemistry. Retrosynthesis—the process of working backward from a target molecule to identify viable synthetic routes—traditionally focuses on yield optimization, but real-world applications require balancing competing objectives like safety, cost, and environmental impact. MMORF addresses this by providing modular components that can be configured into different agent-based systems, enabling researchers to systematically explore how specialized AI agents can collaborate to incorporate multiple constraints into synthesis planning.

Using MMORF, the team built two representative systems: MASIL, which excels at soft-constraint optimization (balancing trade-offs), and RFAS, designed for hard-constraint satisfaction (meeting strict requirements). They evaluated these systems on a newly curated benchmark of 218 multi-objective retrosynthesis tasks. RFAS achieved a 48.6% success rate on hard-constraint problems, significantly outperforming existing state-of-the-art baselines. Meanwhile, MASIL demonstrated strong performance on soft-constraint tasks, frequently producing routes that Pareto-dominated baseline approaches—meaning they were better across multiple objectives without sacrificing others. The framework's modular design allows for principled comparison of different system architectures, advancing research into how multi-agent AI can tackle complex scientific optimization problems.

The research represents a significant step toward practical AI-assisted chemistry where real-world constraints must be considered alongside synthetic feasibility. By making both code and data publicly available, the team enables broader exploration of multi-agent approaches in scientific domains. This work demonstrates how specialized AI agents, each focusing on different objectives (cost, safety, yield), can collaborate to find optimal solutions that single-objective systems would miss, potentially accelerating drug discovery and materials science while reducing experimental costs and safety risks.