Research map / Large language models / Agents, code and math
LLM-Augmented Chemical Synthesis and Design Decision Programs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LLM-Syn-Planner, an evolutionary search where GPT-4o or DeepSeek-V3 generates and mutates whole multi-step retrosynthesis routes in a linear format with partial SC-score rewards, plus an LLM-Syn-Designer for synthesizable molecular design.
Problem
Multi-step retrosynthesis planning is a tightly constrained sequential decision problem with an exponentially large search space; it is unclear whether LLMs, without fine-tuning, can navigate it and also support synthesizable molecular design.
Approach
Instead of having the LLM predict one reaction per step inside a search, the LLM generates entire multi-step routes in a linear step format, which are then improved by an evolutionary loop with three-level feedback (molecule, reaction, route), partial rewards based on SC score, retrieval of similar-molecule routes, and template-matching of proposed reactions.
Key results
- On USPTO-190 with N=500 model calls, LLM-Syn-Planner (DS) reaches 92.6% solve rate versus 88.9% for RootAligned(Retro*).
- On Pistachio Hard at N=500, LLM-Syn-Planner (GPT) reaches 87.0% and (DS) 86.0%, above the best single-step search baseline of 85.0%.
- Direct LLM querying solves only 4.0% (GPT-4o) and 0.0% (GPT-4o) of Pistachio Hard routes, versus 72.0% for LLM-Syn-Planner (GPT-4o) at N=100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Generate and optimize complete routes rather than use the LLM as a single-step predictor in search | LLM-Syn-Planner (GPT) vs LLM (MCTS/Retro*) · Pistachio Hard, N=100 | solve rate (%): 0.0 -> 72.0 | Helped |
| Use sequential route format instead of tree format (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 65.5 -> 91.0; 13.1 -> 64.7 | Helped |
| Use textual generation with extraction instead of tree format (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 65.5 -> 84.5; 13.1 -> 48.9 | Helped |
| Add retrieved example routes instead of providing no examples (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 51.0 -> 91.0; 20.0 -> 64.7 | Helped |
| Add random example routes instead of providing no examples (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 51.0 -> 84.0; 20.0 -> 52.1 | Helped |
| Use retrieved rather than random example routes (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 84.0 -> 91.0; 52.1 -> 64.7 | Helped |
| Add partial reward instead of using only final reward (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 63.5 -> 91.0; 15.3 -> 64.7 | Helped |
| Add an explicit Explanation section to the prompt (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 79.5 -> 91.0; 55.3 -> 64.7 | Helped |
| Add the Rational section to the prompt (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 88.5 -> 91.0; 62.6 -> 64.7 | Helped |
| Replace DeepSeek-V3 with DeepSeek-R1 (ablation) | LLM-Syn-Planner · USPTO Easy and USPTO-190, N=100 | solve rate (%): 93.0 -> 95.0; 62.1 -> 57.9 | Mixed |
| Use direct target queries rather than LLM-Syn-Planner | GPT-4o · USPTO Easy and Pistachio Hard, N=100 | solve rate (%): 91.0 -> 4.0; 72.0 -> 0.0 | Hurt |
Surprising findings
- Tree-format route storage performed far worse than a linear step format (USPTO-190 13.1% vs 64.7%).
- DeepSeek-R1 performed worse than DeepSeek-V3 on USPTO-190 (57.9% vs 62.1%) while costing roughly three times more.
Limitations
- Relies on a fixed reaction template set, so proposed routes inherit template limitations and show reactivity/selectivity issues in case studies.
- LLMs struggle with sparse rewards; performance depends on partial rewards and retrieved sub-trajectories.
- Higher inference cost than specialized single-step models such as RootAligned.
Experimental setup
Datasets: USPTO USPTO-190 Pistachio eMolecules building blocks
Benchmarks: USPTO Easy USPTO-190 Pistachio Reachable Pistachio Hard
Metrics: solve rate top-1 expected property
Baselines: Graph2Edits RootAligned LocalRetro Graph-GA REINVENT MolLEO MARS
Builds on: Monte Carlo tree search Retro* MolLEO SC score
Closest papers by meaning
- URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment (NeurIPS 2026 · similarity 0.90)
- Double-Ended Synthesis Planning with Goal-Constrained Bidirectional Search (NeurIPS 2024 · similarity 0.90)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.89)
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.89)
- Multimodal Large Language Models for Inverse Molecular Design with Retrosynthetic Planning (ICLR 2025 · similarity 0.89)
- Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.