Title: Agents on a Tree: Pathwise Coordination for Multi-Objective Molecular Optimization

Abstract:

Navigating the expansive chemical space under conflicting goals is the core challenge of multi-objective molecular optimization, where initial design choices heavily dictate subsequent possibilities. Current approaches often depend on a unified policy or static scalarization methods, which restricts their capacity to capture varied trade-offs and investigate several viable design pathways. To address this, we introduce ATOM, a multi-agent framework that treats molecular optimization as a search process structured around a tree. In this architecture, each node represents an atomic operation and is managed by an agent tailored to specific objectives or decision contexts. Rather than seeking a global consensus, these agents coordinate along distinct tree paths, allowing the system to preserve and evaluate multiple alternative trajectories for molecular evolution. This approach is bolstered by a global memory that tracks previous optimization behaviors, facilitating a balanced balance between exploration and exploitation across different objectives. By leveraging this tree-structured interaction, the method can effectively reason through the long-horizon dependencies characteristic of molecular design. Our evaluations on rigorous multi-objective benchmarks—covering activity, synthesizability, and ADMET-related properties—demonstrate that ATOM consistently outperforms strong baseline models in terms of Pareto coverage and hypervolume. These findings highlight the efficacy of pathwise multi-agent coordination in molecular optimization. The source code is accessible at https://anonymous.4open.science/r/ATOM-41CE.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC