PaperlayerView exact revision on arXiv

Source-linked AI summary

Rachel: A general-purpose language model directs and revises retrosynthetic routes

Qisheng Li, Shunchao Jiang, Chen Qi, Xin Su, Da Han, Guangyong Chen

arXiv:2609.25118v1physics.chem-phcs.AI

TL;DR

Retrosynthetic planners commonly combine model proposals with search or template procedures, leaving open whether a general-purpose LLM can direct and revise route strategy itself. Rachel provides a stateful environment for LLM-directed planning, and GPT-5.5 achieved strict closure on most targets in both evaluated cohorts.

  • Problem

    Existing planning approaches couple LLM proposals or blueprints to recursive planning, search, or template refinement, leaving open whether a general-purpose LLM can direct successive route decisions and revise strategy as the remaining molecular problems change.

    Source: Introduction · p. 3

  • Method

    Rachel is a stateful environment in which a general-purpose LLM proposes and compares transformations, selects terminals, and maintains or revises strategy as the route changes.

    Source: Introduction · p. 4

  • Results

    Rachel reached strict closure for 111 of 120 PaRoutes120 targets and 24 of 25 RF25 targets, compared with 95 and 15 for direct LLM, 95 and 14 for AOT*, and 83 and 6 for SyntheLite, respectively.

    Source: A general-purpose LLM constructs… · p. 7

  • Takeaways & Limitations

    Recorded examples show that revised strategies can specify new precursor-forming chemistry that is then incorporated into the route and shapes the next precursor decisions.

    Source: Introduction · p. 4, Chemical choices and route strategy… · p. 10

    Rachel’s structure checks do not establish reactivity under specific conditions, so independent chemical evidence is needed throughout planning.

    Source: Discussion and Conclusion · p. 14

Abstract

from arXiv · show

Retrosynthetic planning advances through decisions that reshape the remaining chemical problem: a locally plausible disconnection can leave precursors whose chemoselectivity constraints complicate the rest of the route. Existing planners often channel model proposals through search or template procedures, leaving open whether a general-purpose large language model (LLM) can itself sustain and revise route strategy. We developed Rachel, a stateful environment that executes and checks LLM-directed chemistry but prescribes neither a search policy nor a stopping rule. Without supplied reference routes or route-level solutions, GPT-5.5 achieved strict closure for 111 of 120 PaRoutes120 targets and 24 of 25 targets in the separate RF25 difficult-target cohort. RF25 was drawn largely from studies published after GPT-5.5's reported knowledge cutoff. Closure required complete routes and independent source resolution of every terminal precursor after planning. On a shared PaRoutes subset, forward-model support exceeded that of most comparator methods, and Rachel received the highest mean overall route score from both method-blinded LLM evaluators. Recorded trajectories showed continued model-proposed chemistry, with revised strategies carried into subsequent steps. Replacing LLM route decisions with fixed policies reduced strict closure to 6-15/120 despite continued local chemical execution; restricting planning support also reduced closure in RF25. Within Rachel, a general-purpose LLM coordinated successive chemical choices and revised its strategy as earlier decisions reshaped the remaining problems.

Supporting Information.

The Supporting Information defines the benchmark cohorts and identifies the records used to match targets, routes, and endpoint evidence. It also locates the machine-readable target inventory and supplementary methods.

1 Hangzhou Institute of Medicine, Chinese Academy of Sciences, Hangzhou, China

  • The listed institutional affiliations identify the participating research organizations in Hangzhou, Beijing, Shanghai, and other locations in China.The affiliation list includes the Hangzhou Institute of Medicine, Chinese Academy of Sciences, and Zhenzhida Biotechnology (Shanghai) Co., Ltd.Source: §1 · p. 28

S1.1 Cohort definition

  • PaRoutes120 contains 120 targets, while RF25 is a separately analysed cohort of 25 targets documented in molecular-probe and chemical-biology studies.The RF25 source records include 24 targets reported in 2026 and one in 2025; its molecular weights range from 231.3 to 2082.4 Da and its heavy-atom counts from 16 to 149.Source: S1.1 Cohort definition · p. 30

S1.2 Minimum target-record fields

  • Stable target identifiers link each input structure to cohort membership, generated routes, endpoint records, and provenance pointers.The RF25 registry also retains source publications and molecular descriptors; these fields support matching and audit, not a difficulty score or reconstruction of historical sampling.Source: S1.2 Minimum target-record fields · p. 30

S1.3 Supplementary-data location

  • The target inventory, RF25 registry, and descriptor records are provided as machine-readable Supplementary Data rather than reproduced in this section.GPT-5.5 route outputs are linked through target and route identifiers.Source: S1.3 Supplementary-data location · p. 30

S2. Retrosynthetic planning protocol

Rachel runs GPT-5.5 in a stateful planning session where the model proposes and accepts transformations, revises route strategy, and receives validation feedback. The protocol distinguishes starting instructions from the broader session context and describes lightweight structure checks.

S2. Retrosynthetic planning protocol

  • GPT-5.5 planned in Rachel-beta with reasoning effort set to High, receiving progressively updated context about the active molecule, candidate operations, validation results, and route strategy.The authors did not specify an additional temperature setting for these runs.Source: S2. Retrosynthetic planning protocol · p. 30
  • The LLM selects transformations, decides whether to accept them, and chooses whether to continue decomposition or accept a terminal, while Rachel retains the evolving route state.Inspection and candidate attempts do not add reactions; accepted transformations enter the route through commitment, with validation feedback informing acceptance and hard blocks preventing commitment.Source: S2. Retrosynthetic planning protocol · p. 30

Initialization example: English translation

S2.1 Operational state fields

S3. Route closure assessment

Strict Closure is a route-level endpoint requiring complete construction and source resolution for every terminal. The assessment uses archived source evidence and preserves unresolved terminals in the audit record.

S3. Route closure assessment

  • A terminal is source-resolved when both a PubChem CID and Chemical Vendors evidence are available, with registered in situ metal reagents also eligible through a preparation-source mapping.The mapped source form or every required source reagent must meet the same criterion; these records establish the declared source-evidence endpoint at audit time.Source: S3. Route closure assessment · p. 31
  • Strict Closure requires a completed route with every terminal source-resolved; missing or failed outputs and all other outcomes are Unresolved.The post hoc audit follows route freezing, is not fed back into planning, and is separate from internal stopping.Source: S3. Route closure assessment · p. 31

S3.1 Minimum terminal-evidence fields

  • The complete terminal-by-route evidence table, including CID, vendor records, source mappings, and unresolved entries, is supplied as Supplementary Data rather than reproduced in the Supplementary Methods.The table is titled “Minimum terminal-evidence fields.”Source: S3.1 Minimum terminal-evidence fields · p. 32, Table S2 · p. 32

S4. Reaction and route evaluation

The evaluation combines forward-model checks on recorded reactions with blinded whole-route appraisal, using explicitly defined samples and scoring procedures. RF25 forward-model support is reported on method-specific closed-route sets, whose populations differ.

S4. Reaction and route evaluation

  • The common58 comparison uses 58 PaRoutes120 targets strictly closed by Rachel, direct LLM, AOT*, and SyntheLite, with one frozen route per target and PaRoutes reference routes for comparison.The subset contains 24 n1 and 34 n5 targets; RF25 forward evaluation instead uses each method’s own strictly closed routes.Source: S4. Reaction and route evaluation · p. 32
  • Forward models predict products from recorded reactants without supplied reaction conditions, and Top-k support measures whether the recorded product appears among the first k predictions.Both@k requires both models to recover the recorded product for the same step; it measures predictive agreement under this protocol, not experimental correctness, and a nonmatch does not distinguish chemistry errors from predictor limitations.Source: S4. Reaction and route evaluation · p. 32

S4.2 Whole-route appraisal rubric

  • Whole-route appraisal presents four anonymous routes per target to GPT-5.6 Sol and Opus 5, which score routes separately and recommend whether they should progress to human synthetic review.Reviewers assess structures without external searches or attempts to infer route provenance; missing conditions and yields are reflected as uncertainty, and no human-expert assessment of RF25 routes was performed.Source: S4. Reaction and route evaluation · p. 32, S4.2 Whole-route appraisal rubric · p. 33

S4.1 Forward-model configuration

  • The recorded RXNGraphormer temperature of 1.0 applies to forward-model inference and is separate from Rachel’s LLM planning settings.Coverage counts and unavailable-step records for common58 are supplied as machine-readable Source Data.Source: S4.1 Forward-model configuration · p. 33

S4.2 Whole-route appraisal rubric

  • 67/308 (21.8%) Both@10 steps were supported for Rachel on RF25, the largest absolute supported-step count among methods, but not the highest fraction.The corresponding results were 19/104 (18.3%) for direct LLM, 10/65 (15.4%) for AOT*, and 9/33 (27.3%) for SyntheLite; each method was evaluated on its own strictly closed routes, so the sets do not establish a matched-target quality ranking.Source: S4.2 Whole-route appraisal rubric · p. 33

S5. Control and ablation experiments

The section specifies mechanical policies for choosing among successfully executed candidates and describes RF25 component-ablation conditions. It also identifies tables covering policy definitions, component interventions, and additional API settings.

S5. Control and ablation experiments

  • Mechanical policies considered successfully executed candidates, tried bond alternatives before functional-group interconversions, and skipped the active molecule when none was eligible.Greedy complexity, highest score, and random valid selected candidates using distinct recorded criteria; outcomes used the shared closure criterion.Source: S5. Control and ablation experiments · p. 34

S5.2 Component interventions

  • RF25 component ablations removed custom chemistry, persistent route planning, local sketch and continuation, or detailed validation feedback, using GPT-5.5 at high reasoning effort.Each condition used one route per target, with a 200-turn budget, maximum depth of 15, and 50-step limit; strict closure was assessed on all 25 targets against native Rachel.Source: S5. Control and ablation experiments · p. 34, Table S6 · p. 34

S5.1 Mechanical policy definitions

S5.3 Additional API settings

  • Table S7 is identified as the table of additional API settings.Source: Table S7 · p. 35

S6. Data analysis

The analysis links frozen route records to session histories and defines how plan revisions, route-position profiles, validation episodes, and inherited molecular states are counted. It also describes how route records and evaluation evidence are assembled into an offline Route Atlas.

S6. Data analysis

  • Route records were linked to session histories by target, event, reaction, and plan-version identifiers, while missing version history remained unavailable.Revision trajectories compare successive plan versions, and the chemical revision ledger counts retained, added, and removed entries.Source: S6. Data analysis · p. 36
  • Route-position profiles divide reaction order by route length into five bins and average target-level custom-transformation fractions with equal target weight.Pointwise 95% intervals use 10,000 bootstrap means; revised-plan profiles use the same binning and weighting, with 104 PaRoutes120 and 25 RF25 eligible routes.Source: S6. Data analysis · p. 36
  • Validation-state inventories can overlap, and hard-block episodes separately track another candidate admitted to the final route and no selection.The target audit reports episode counts and binary presence of both outcomes, which may occur within one target across different episodes.Source: S6. Data analysis · p. 36
  • Inherited-state analyses count each molecular-node identity once per route and report identity-level compositions separately from route-level fractions.Identities are classified by whether they are active at the next commit, active after an intervening commit, or retained as final endpoints.Source: S6. Data analysis · p. 36

S6.1 Statistical definitions

  • The RF25 illustrative case was selected by a deterministic score of stopping types, active conditions, chemistry-source combinations, depth, unresolved frontiers, and native closure.The score selected an illustrative case only; it was not a chemical-quality measure or additional closure endpoint.Source: S6.1 Statistical definitions · p. 37

S6.2 Route Atlas assembly

  • The Route Atlas presents target-method route records and endpoints, retaining partial routes and distinguishing evaluated variants from process trees.It includes separate PaRoutes references, with 1,180 display positions comprising 1,060 method positions and 120 references; forward-model evidence requires stereochemistry-retained canonical product and reactant-multiset matches.Source: S6.2 Route Atlas assembly · p. 37
  • The Atlas preserves hashes and reconciliation notes, distinguishing chemical agreement from byte identity when adopted evidence differs from historical files.It adds an inspection resource without rerunning planning or evaluation.Source: S6.2 Route Atlas assembly · p. 38

Data and code availability

Rachel's code and the Route Atlas are available through GitHub. The release includes offline route comparisons and supporting source data, with stated licensing terms.

Data and code availability

  • The Rachel Route Atlas is available on GitHub at repository version 57c1716, with offline comparisons for all 120 PaRoutes and 25 RF25 targets.The Atlas includes separate PaRoutes reference routes.Source: Data and code availability · p. 38
  • Figure-level data, machine-readable records, and supporting documentation are available in the Source Data release, version 2026-09-07.Source: Data and code availability · p. 38
  • Rachel-authored materials and the repository are distributed under CC BY-NC 4.0, while third-party terms remain in force.Source: Data and code availability · p. 38

Figure S1

Figure S1 shows the share of committed steps registered under a revised plan across relative route-position bins for PaRoutes120 and RF25. The profile is descriptive and does not establish a causal effect of revision.

Figure S1

  • Figure S1 bins reactions by order in the committed sequence divided by total route reactions, showing PaRoutes120 and RF25 separately.Amber points are equally weighted means of within-target revised-plan step fractions, with pointwise 95% intervals from 10,000 target-bootstrap resamples.Source: Figure S1 · p. 39
  • 104 PaRoutes120 routes and 25 RF25 routes had available plan histories, but targets without a step in a bin do not contribute to its mean.This eligible population differs from the revised-route subset in Fig. 4c.Source: Figure S1 · p. 39
  • Registered plan context in the profile does not establish a causal effect of revision.Source: Figure S1 · p. 39

Figure S2

The supplementary audits characterize hard-block decisions, RF25 restriction checks, and how often terminal complexity exceeded the planning threshold. These are recorded outcomes, not causal tests of gate efficacy.

Figure S2

  • 56/121 PaRoutes120 episodes admitted another candidate after a hard-block record, while 65/121 had no selection; RF25 counts were 10/10 and 0/10, respectively.Episodes are nested within target routes, and admission refers to the alternative candidate, not the hard-blocked record.Source: Figure S2 · p. 40

Figure S3

  • The PaRoutes120 replacement audit separates candidate records, decision opportunities, route endpoints, and combinations of audit failures without establishing a sequential funnel or mediation.Internal completion and audit failures are diagnostic categories; the main endpoint remains Strict Closure or Unresolved.Source: Figure S3 · p. 41

Figure S4

  • For the four RF25 restrictions, recorded use of the removed signal was 0/25 in each condition, while retained-control and execution-check counts varied by restriction.The caption lists the restriction-specific counts and notes that observed use differs from command availability.Source: Figure S4 · p. 42

Figure S5

  • 173/616 PaRoutes120 terminal nodes and 36/163 RF25 nodes exceeded their run-specific planning thresholds for molecular complexity.This planning threshold is distinct from the independent terminal-source criterion used for strict closure.Source: Figure S5 · p. 43

Figure S6

These supplementary figures describe route morphology, reviewer recommendations, route-simplification profiles, and recorded chemistry and plan revisions. Several views are conditional or descriptive rather than paired comparisons or feasibility assessments.

Figure S6

  • The morphology figures compare maximum reaction depth and mean terminal heavy-atom and ring counts across strictly closed routes, using independent common cohorts in PaRoutes120 and unmatched method-specific RF25 sets.These are descriptive conditional views, not measures of synthetic feasibility or paired method effects.Source: Figure S6 · p. 44

Figure S7

  • Each evaluator made 58 method-blinded whole-route advancement appraisals; recommendations concern progression to human synthetic review, not experimental feasibility.GPT-5.6 Sol and Opus 5 decisions are shown and analyzed separately, not pooled.Source: Figure S7 · p. 45

Figure S8

  • The reduction profiles compare target-to-terminal complexity and BERTz changes across strictly closed cohorts, with PaRoutes n1/n5 unpaired and RF25 route sets unmatched.Both panels describe route simplification rather than experimental feasibility.Source: Figure S8 · p. 46

Figure S9

  • The records cover 1,131 committed transitions across 145 routes and show operation type and plan context; revision trajectories cover 58 routes, 145 events, and 580 field comparisons.Horizontal position is committed-state or event order, not elapsed time; field turnover does not measure semantic distance or chemical importance.Source: Figure S10 · p. 48, Figure S11 · p. 50

Figure S12

Figure S12 details one RF25 route revision in which a local coupling strategy changed while the global route frame remained in place.

Figure S12

  • In RF25 route rf25_008, the hydroxypentyl-azole bond strategy changed from Negishi/ZnCl to Suzuki/BF3K between revisions 3 and 4.The global route frame was retained; exact-entry turnover does not measure semantic or chemical improvement.Source: Figure S12 · p. 52

Figure S13

The figures audit hard-block outcomes and inherited molecular identities, illustrate identity handoffs, and describe route morphology under RF25 planning restrictions.

Figure S13

  • Across 145 targets, the hard-block audit records episode counts and whether any such episode admitted another candidate or had no selection; these indicators can both occur for one target.The records do not report target-level strict closure.Source: Figure S13 · p. 53

Figure S15

  • The examples show consecutive identity handoffs in PaRoutes120 and an RF25 identity becoming active at Step 16 after four intervening registered steps.These cases illustrate identity linkage, not experimental validation or population prevalence.Source: Figure S15 · p. 56

Figure S16

  • Under RF25 restrictions, strictly closed route counts varied from 2 to 18 across F1-F4, versus 24 for native Rachel; these outcome-conditioned populations are not paired comparisons.Depth and terminal complexity describe route morphology, not efficiency or a restriction ranking.Source: Figure S16 · p. 57

Figure S17

The supplementary figures show terminal-identity matches for one illustrative target, selected route examples, and forward-model agreement on method-specific RF25 routes. The route examples are not representative, and terminal matches alone do not establish closure.

Figure S17

  • Rachel’s six strictly closed terminal identities for target 014 matched 0, 1, 1 and 3 terminals across the four restricted terminal sets; all restricted routes remained unresolved.A terminal match alone does not establish closure of the complete route.Source: Figure S17 · p. 58
  • The n1_5948 and rf25_004 route pairs were selected for legibility, not as a representative sample; the complete Atlas includes every target-method position and incomplete or unavailable outcomes.The printed views omit navigation controls; the offline Atlas permits inspection of structures and matched evidence at full size.Source: Figure S18 · pp. 59–60

Figure S19

  • On RF25 method-specific strictly closed routes, Rachel had the largest absolute supported-step count in forward-model agreement, but unmatched route populations preclude interpreting this as a higher support fraction.This diagnostic reports predictive agreement under the specified inputs and matching rules; it does not establish experimental correctness or feasibility.Source: Figure S19 · p. 61
Loading 2609.25118v1…