Skip to content

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

🕒 Published (v1): 2026-03-27 14:21 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ATOM (Agents on a Tree for multi-Objective Molecular optimization) is a multi-agent framework that maps multi-objective molecular optimization onto a Monte Carlo Tree Search (MCTS) structure, where each tree node hosts a specialized LLM agent targeting one molecular property. Agents coordinate along distinct search paths rather than converging to a single global policy, enabling explicit comparison of alternative optimization trajectories. On four-objective benchmarks (QED, SA, GSK3\(\beta\), JNK3), ATOM achieves a hypervolume sum of 4.351, outperforming the strongest baseline (GRAPH GA, 3.969) and the best prior multi-agent method (EAG, 3.752).

Problem

Multi-objective molecular optimization requires navigating an exponentially large discrete chemical space under conflicting property objectives (e.g., druglikeness vs. synthetic accessibility vs. kinase inhibition). Existing approaches either collapse objectives via fixed scalarization, rely on expensive two-stage Pareto sampling, or apply monolithic single-policy LLM generation—none of which can explicitly model inter-objective trade-offs or maintain diverse optimization trajectories simultaneously.

Method

ATOM formulates molecular optimization as a population-based MCTS where each node stores a population of molecules (as SMILES strings) and hosts \(K\) specialized expert agents, each implemented as a GPT-4o mini instance with domain-specific prompt templates and tool-calling access (RDKit, learned property oracles).

Tree operations: - Selection: UCT-style scoring \(\text{UCT}(N) = V(N) + c\sqrt{\log N_\text{parent} / N_\text{visit}(N)}\), where \(V(N) = \lambda \cdot S_\text{attr}(N) + (1-\lambda) \cdot \widehat{HV}(N)\) balances directional property improvement against Pareto hypervolume diversity. - Expansion: Each selected node is expanded by the designated lead agent, augmented via lateral knowledge exchange: \(\tilde{S}_q(N) = S_q(N) \cup \bigcup_{j \neq q} \omega_{j \to q} S_j(N)\), blending ranked molecule subsets from peer agents. - Backup: Hierarchical averaging propagates node values upward: \(V_\text{parent} \leftarrow \frac{1}{2}(V_\text{parent} + \frac{1}{|C|}\sum_i V_\text{child}^{(i)})\). - Global memory: A dynamic store \(\mathcal{M}\) aggregates top-scoring molecules across all paths, retrieved to bias local agent scoring: \(S_\text{attr}^{(\mathcal{M})}(N) = (1-\gamma)S_\text{attr}(N) + \gamma S_\text{attr}(\text{Retrieve}(\mathcal{M}))\).

Theoretical analysis shows that agent synergy reduces effective branching factor by concentrating probability mass on the orthogonal agent resolving the current "property bottleneck," with mis-selection probability decaying as \(O(\exp(-\delta^2/\sigma_d^2))\). Knowledge-base conditioning replaces physical edit depth \(L\) with effective depth \(L_{KB} < L\), contracting sample complexity.

Key Contributions

  • Pathwise MCTS formulation of multi-objective molecular optimization that preserves alternative evolution trajectories rather than collapsing to a single policy.
  • Knowledge-mediated coordination mechanism combining lateral agent exchange (Eq. 15), hierarchical parent-to-child molecule propagation (Eq. 16), and global dynamic memory (Eqs. 17–18).
  • Theoretical bounds showing agent synergy prevents local-optima stagnation and knowledge-base grounding contracts MCTS sample complexity.
  • Empirical state-of-the-art across six multi-objective benchmarks (2–4 objectives over QED, SA, GSK3\(\beta\), JNK3), with HV sum 4.351 vs. 3.969 for GRAPH GA and 3.752 for EAG.

Results

  • ATOM HV sum (all 6 tasks): 4.351 vs. GRAPH GA 3.969, EAG 3.752, GPT-4o-mini 3.678, GP BO 3.822.
  • QED+SA (2-obj): ATOM 0.941±0.001 (best); SMILES LSTM 0.925, GRAPH GA 0.899.
  • JNK3+GSK3β (2-obj, biological): ATOM 0.790±0.017 vs. GP BO 0.783, GRAPH GA 0.725.
  • QED+SA+GSK3β+JNK3 (4-obj, hardest): ATOM 0.606±0.041 vs. EAG 0.482, GPT-4o-mini 0.433.
  • Performance gap vs. baselines widens as number of objectives increases, with traditional methods (SMILES LSTM, STONED) failing most severely on biological objectives.

Limitations

  • Relies on GPT-4o mini as the LLM backbone—performance is coupled to API availability, latency, and cost at scale.
  • Property oracles (random forest models for GSK3\(\beta\)/JNK3) are proxy models; real wet-lab validation is absent.
  • MCTS tree depth and branching factor hyperparameters require tuning per task; no ablation of \(\lambda\), \(\gamma\), \(c\) sensitivity is reported.
  • Evaluated only on SMILES-based ZINC20 starting points; generalization to 3D or graph-native representations is unexplored.
  • Theoretical analysis relies on assumptions (decaying noise, optimistic pruning) whose empirical validity under realistic oracle noise is unverified.

Relevance to Harnesses / Meta-Harnesses

ATOM is directly a meta-harness pattern: a fixed orchestration backbone (MCTS) that dynamically schedules, coordinates, and aggregates outputs from a pool of specialized sub-agents, each with its own tool-calling stack. The global memory and UCT policy function as the meta-level state and routing policy, analogous to a harness that maintains a shared context store and dispatches tasks to specialist workers based on current bottlenecks. The lateral knowledge exchange and hierarchical propagation mechanisms are concrete implementations of cross-agent information sharing—a central design challenge in any multi-agent harness. For practitioners building harnesses for complex multi-objective optimization, ATOM demonstrates how MCTS can serve as a principled orchestration primitive that balances exploration/exploitation across specialist agents without requiring a monolithic coordinator policy.