Organizations: State Key Laboratory of Analytical Chemistry for Life Science, Kuang Yaming Honors School, Chemistry and Biomedicine Innovation Center (ChemBIC), ChemBioMed Interdisciplinary Research Center at Nanjing University, Nanjing University, Nanjing 210023, China · School of Medicine, Guangxi University, Nanning, 530004, China
Computer-aided retrosynthesis has advanced through single-step prediction, tree search and route scoring, but complex targets still require judgement over the whole route, where stereochemistry, functional-group compatibility, cost and greenness interact. Here we present \mbox{ReTRACE} (REtrosynthesis with Tool-orchestrated Reasoning And Context-aware Evidence), an agent that places a large language model (LLM) in a chemistry harness that integrates molecular tools, complementary single-step prediction engines, a purchasable starting-material library, a route memory and a multi-criterion route evaluator. Routes are judged by a new metric that combines programmatic checks with adversarial LLM review, and its verdicts are compared with a synthetic chemist's. The agent learns continually from evaluator feedback during search, diagnosing weak routes and backtracking to revise them. With Claude Opus 4.8, at least one of the top 10 routes proposed by ReTRACE passed the audit for 87.4%, 95.0% and 86.7% of targets on the three benchmarks. Two prospective wet-lab syntheses show that ReTRACE route proposals can be converted into executable procedures after expert revision.
Figures & tables
Figure 1: ReTRACE harness, two decision instances from real search trajectories and the response to score feedback within a run. a , An LLM coordinator works inside a harness of heterogeneous chemistry tools, deterministic gates that check every tool call, tool result and saved route, and a route memory that carries partial routes, past decisions and abandoned branches into the next round. Route scores return to the coordinator as feedback within a run. Ranked routes are judged by Audit@k as passed or failed, and selected routes go to chemist review and wet-lab synthesis. b , Adjudicating conflicting single-step suggestions. The coordinator keeps both a late-stage nitro reduction and a glycidyl N-alkylation open, checks whether the electrophile can be purchased and expands the shared lactam core with both engine classes before choosing. c , Diagnosis and backtracking. Route B scores 0.75 against 0.80 for the best route. The coordinator attributes the gap to lower cost and feasibility from the added steps, saves route B as an alternative and backtracks to shorten the linear sequence. Monologues in b and c are verbatim excerpts from run trajectories. d , Score feedback within a run in 670 completed Retro*-190 runs. Mean score of the best route so far (dark green) and of the j-th scored route (light green; left axis), and percentage of runs in which at least one of the first j scored routes passed the audit (grey, dashed; right axis).
Figure 2: Validity, reference recovery and complementarity of ten single-step engines, and engine agreement in planned routes. a–c, 50,000 USPTO-50K reactions, with engines ordered by top-10 recovery and reference matching. a , Fraction of reactions with output, with a rank-1 prediction that passes RDKit screening, with a rank-1 prediction that passes the LLM audit (95% Wilson interval), and with an exact top-1 reference match. b , Reactions recovered within the top 10 by that engine alone. c , Percentage of the top-10 misses of the row engine that the column engine recovers within its top 10, and the right column gives the top-10 miss rate of each engine. d , Route steps in the Retro*-190 runs that passed or failed the audit, by the number of runtime engines that proposed the step.
Controller / baseline
Run success (%)
Audit@ k (%) k=1k=3k=5k=10
Rank-1 route length vs reference (%) Shorter Equal Longer None
Retro*-190 (N = 190)
Claude Opus 4.8
97.9
70.0
78.4
83.7
87.4
65.8
15.8
18.4
0.0
GPT-5.6 Sol fast
100.0
64.2
76.8
78.4
80.0
64.2
16.3
19.5
0.0
GPT-5.6 Sol standard
100.0
65.8
75.3
79.5
82.6
64.7
11.6
23.7
0.0
GPT-5.6 Sol thorough
100.0
63.7
72.6
79.5
82.1
60.0
15.8
24.2
0.0
Qwen3.6-35B-A3B
91.6
23.7
33.2
33.7
34.7
74.7
12.1
8.4
4.7
Table 1: Route-planning performance of coordinating models and unified audit baselines across three datasets, with Claude Opus 4.8 tool-group ablations on Retro*-190.
Figure 3: Conventional tree-search control and ReTRACE route-level control, with one real decision trajectory. a , Conventional MCTS or Retro* controllers repeatedly select, expand, evaluate and backpropagate within a fixed candidate space. ReTRACE instead runs a agent loop that diagnoses weak steps, introduces alternative precursors and restructures a route before saving it. b , Five routes saved for one representative target, with their total scores. Routes 2 and 3 only reorder the steps of Route 1, Route 4 uses a Boc-last assembly, and Route 5 changes the disconnection class and reaches the highest score (0.8316) after its unpurchasable BPin–aryl ether leaf is disconnected further.
Figure 4: Prospective wet-lab validation of ReTRACE-assisted route design. a , Compound C1. b , Compound C2. Each panel compares the initial ReTRACE route with the chemist-revised experimental route. C1 required a revised protection strategy and an earlier Suzuki coupling, whereas C2 retained the core sequence with localized substrate and condition changes.
Figure 5: Chemical failure modes identified by the route-level adversarial audit. a , Operational grouping of the 497 audit-failed routes of GPT-5.6 Sol (standard) on Retro*-190. Each route is assigned the first family A–D that occurs among its fatal steps (filled bars); hatched extensions add routes that also contain that family but were assigned an earlier one. b , a Knoevenagel condensation should keep all 19 reactant carbons; the drawn product has no counterpart of carbon 2 and contains 18. c , an intramolecular oxa-Michael addition should join alcohol oxygen 3 to β -carbon 1 and close a seven-membered ring; the drawn product joins it to α -carbon 2 and turns carbon 1 into a methyl group, giving a six-membered ring. d , a one-step construction of a δ -carboline (pyrido[3,2- b ]indole) from a phenylenediamine and a dibromopyridine ester forms new C–N and C–C bonds while the nitrogen count falls from five to four, with no nitrogen-containing leaving group.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1: NMR and HRMS spectra of compound C1 and its intermediates. Compound numbers follow Figure 4 ; spectra are reproduced as recorded.
Figure A2: Continued. NMR and HRMS spectra of compound C1 and its intermediates.
Figure A3: Continued. NMR and HRMS spectra of compound C1 and its intermediates.
Figure A4: Continued. NMR and HRMS spectra of compound C1 and its intermediates.
Figure A5: Continued. NMR and HRMS spectra of compound C1 and its intermediates.
Figure A6: NMR and HRMS spectra of compound C2 and its intermediates. Compound numbers follow Figure 4 ; spectra are reproduced as recorded.
Figure A7: Continued. NMR and HRMS spectra of compound C2 and its intermediates.
Figure A8: Continued. NMR and HRMS spectra of compound C2 and its intermediates.
Multi-step retrosynthesis planning seeks to decompose a target molecule into commercially available building blocks through a sequence of feasible reactions. The vast combinatorial search space makes this task challenging even for expert chemists. Traditional methods combine tree search with offline-trained value networks that score candidates in isolation, without reasoning about complete multi-step routes. Recent work leverages Large Language Models (LLMs) for this task, but relies on simple interfaces that limit exploration of the full search space. We introduce RetroAgent, an LLM agent that bridges symbolic search and neural reasoning through a harness with structured memory. Through memory and chemistry tools, the agent observes the full search state, including explored routes, available alternatives, and properties of intermediates, enabling informed decisions grounded in both global progress and domain knowledge. Experiments on in-distribution and out-of-distribution benchmarks demonstrate that RetroAgent delivers strong performance and generalization.
Multi-step retrosynthesis planning is often considered solved once a route to purchasable starting materials is found, but such a route should also be practical in real-world chemistry. Recent LLM-based planners formulate this process as sequential decision-making over an AND-OR graph, yet typically expand a single molecule per turn and are optimized for solvability alone, leaving alternative routes uncompared across the multiple objectives that determine practicality before the search budget is committed. We introduce RetroPlanner, an LLM-based planner that jointly compares multiple routes to steer the search toward higher synthesizability, as measured by reaction plausibility, round-trip consistency, and starting-material price. RetroPlanner is trained to reason over chemical evidence: structural progress from the target molecule, bond changes and reaction classes, where candidate routes diverge from one another, and how they rank across multiple objectives. Experiments on multi-step retrosynthesis benchmarks demonstrate that RetroPlanner recovers more alternative routes than prior LLM-based planners while improving synthetic route quality and maintaining a high success rate.
Ye Rin Kim, Yongjun Jeong, Seongjun Yun +4
Korea University, Department of Artificial Intelligence · Korea University, Department of Statistics
Retrosynthesis prediction aims to identify reactants that can synthesize a given product molecule. Although molecular large language models (LLMs) have recently shown promising results, most existing methods either generate reactants directly or provide only generic product-level analysis, without explicitly reasoning about bond-disconnection strategies that justify specific reactant choices. This paper proposes RetroReasoner, a retrosynthetic reasoning model that captures chemists' strategic disconnection-based thinking. RetroReasoner is trained with supervised fine-tuning and reinforcement learning. For supervised fine-tuning, SyntheticRetro generates structured disconnection rationales paired with reactant predictions. For reinforcement learning, a round-trip reward evaluates predicted reactants by passing them through a forward synthesis model and rewarding predictions that reconstruct the original product. RetroReasoner can also be applied to multi-step retrosynthetic planning by incorporating it into a parallelized Monte Carlo tree search framework, reducing search time while increasing the number and diversity of valid synthetic pathways. Experimental results show that RetroReasoner outperforms prior baselines, including not only molecular LLMs but also retrosynthesis-specific expert models, and generates a broader range of feasible reactant proposals, especially for challenging reaction instances. The code is available at https://github.com/KU-AGI/RetroReasoner.
Hanbum Ko, Chanhui Lee, Ye Rin Kim +4
Department of Artificial Intelligence, Korea University · Materials Intelligence Lab, LG AI Research · Department of Statistics, Korea University