An Agentic Retrobiosynthesis Framework with Learned Frontier Selection

arXiv:2608.30702v1 Announce Type: cross Abstract: Large language models are increasingly used as agents for multistep retrosynthesis, raising the question of how much their search policy contributes independently of the underlying reaction model. We investigate this question in a biological setting through rule-based retrobiosynthesis: a deterministic biochemical engine generates the same validated transitions for every method, searching for routes that terminate in metabolites available to an Escherichia coli chassis, while the policy only selects which frontier molecule to expand next. Prompted and LoRA-tuned Qwen2.5-7B policies use a strict choice-only interface. The fine-tuned policy reaches 65pm1% solve rate at 10 expansions on LASER versus 59% for MCTS, and at 200 expansions reaches 78pm1% versus 75% on LASER, 88pm3% versus 80% on the RetroPath RL Golden benchmark, and 63pm2% versus 45% on the BioNavi-NP benchmark. Fine-tuning also consistently outperforms direct prompting. These results show that route-supervised frontier selection can improve budgeted search without altering biochemical generation, although performance remains dependent on frontier construction and reaction ranking.

Related

• AutoCedar: An Agentic Framework for Verifier-Guided Access Control Policy Synthesis
• Rethinking Agentic Search with Pi-Serini: Is Lexical Retrieval Sufficient?
• Quantifying Frontier LLM Capabilities for Container Sandbox Escape
• Supervising the search process produces reliable and generalizable information-seeking agents

Source: arXiv cs.AI | 2026-09-01