An Agentic Retrobiosynthesis Framework with Learned Frontier Selection

TL;DR - This paper isolates the impact of LLM-guided search in rule-based retrobiosynthesis by keeping biochemical reaction generation fixed. A LoRA-tuned Qwen2.5-7B frontier-selection policy outperforms prompting and MCTS across multiple benchmarks, especially under constrained search budgets.

• The LLM only selects the next frontier molecule to expand through a strict choice-only interface; a deterministic engine supplies identical validated biochemical transitions.
• On LASER, the fine-tuned policy achieves a 65±1% solve rate with 10 expansions versus 59% for MCTS, and 78±1% versus 75% with 200 expansions.
• At 200 expansions, it reaches 88±3% versus 80% on RetroPath RL Golden and 63±2% versus 45% on BioNavi-NP.
• Gains depend on frontier construction and reaction ranking, indicating that search-policy improvements do not eliminate upstream design sensitivities.