Source-linked AI summary
An Agentic Retrobiosynthesis Framework with Learned Frontier Selection
Philippe Meyer, Guillaume Gricourt, Thomas Duigou, Joan Hérisson, Jean-Loup Faulon
TL;DR
The paper asks whether LLM search policy contributes independently of biochemical reaction generation in retrobiosynthesis. It fixes a deterministic reaction engine and trains or prompts Qwen2.5-7B policies to select frontier molecules, finding consistent fine-tuning gains across three benchmarks while noting dependence on the reaction-rule system and chassis sink.
Problem
Agentic control remains largely unexplored in retrobiosynthesis, leaving unclear how much learned frontier selection contributes when biochemical reaction generation is fixed.
Method
A deterministic biochemical engine generates validated transitions toward an E. coli terminal metabolite set, while prompted or LoRA-fine-tuned Qwen2.5-7B policies select frontier molecules from bounded observations.
Results
Across LASER, Golden, and BioNavi-NP, fine-tuned policies consistently outperform direct prompting and reach 78 ± 1%, 88 ± 3%, and 63 ± 2% solve rates at N = 200, versus MCTS rates of 75%, 80%, and 45%.
Takeaways & Limitations
Route supervision improves allocation of limited search budgets without replacing the underlying biochemical model and transfers beyond the MetaNetX-derived training routes.
Takeaways & Limitations
Success remains dependent on the reaction-rule system and chassis sink, while stereochemistry, thermodynamics, enzyme activity, yields, kinetics, and cellular physiology are not modelled.
Abstract
from arXiv · showhide
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 \emph{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 $65\pm1$\% solve rate at 10 expansions on LASER versus 59\% for MCTS, and at 200 expansions reaches $78\pm1$\% versus 75\% on LASER, $88\pm3$\% versus 80\% on the RetroPath RL Golden benchmark, and $63\pm2$\% 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.
1 Introduction
This study examines whether learned frontier selection can improve agentic retrobiosynthesis when biochemical reaction generation is fixed. It compares LLM policies with classical search methods in routes terminating at E. coli metabolites.
- Retrobiosynthesis recursively decomposes targets through enzyme-catalyzed transformations toward metabolites available to a host chassis.
- Comparable agentic control is largely unexplored in retrobiosynthesis.
- The framework holds biochemical reaction generation fixed and learns only which frontier molecule to expand next.
- Prompted and LoRA-fine-tuned Qwen2.5-7B policies are compared with BFS, DFS, greedy sink similarity, and MCTS using a strict choice-only interface.
2 Framework
The framework combines deterministic biochemical expansion with budgeted AND–OR search, while LLM policies choose among bounded frontier observations. Route replay supplies training decisions and supports construction of the candidate view.
- Deterministic biochemical expansion: The one-step engine screens 80,116 MetaNetX-derived biochemical reaction rules before retaining graph-valid outcomes.
- Budgeted AND–OR search: A molecule is solved by terminal-set membership or a solved child reaction, while a reaction is solved only when all precursor molecules are solved.
- Budgeted AND–OR search: The search policy selects one unsolved frontier molecule for deterministic expansion, with budget N defined as the number of molecule expansions.
- Agent observation: LLM policies receive at most 20 frontier candidates selected using pre-expansion ranking signals from the search state and reaction-rule applicability.
- Mined routes and corpus construction: Route mining produces training decisions after 12,251 biochemical routes are replayed and evaluation-target leakage is filtered.
- LLM frontier policies: The fine-tuned policy learns route-derived frontier choices, while direct prompting uses the same Qwen2.5-7B-Instruct model and observation-action interface.
3 Experiments
Experiments evaluate three retrobiosynthesis benchmarks under matched biochemical search settings. Performance is measured by solve rate across molecule-expansion budgets, with repeated seeds for LLM policies.
- The RetroPath RL Golden benchmark contains 20 curated experimental pathways and 70 reference one-step disconnections.
- LASER contains 141 metabolic-engineering targets without reference routes, while BioNavi-NP supplies an independent natural-product benchmark.
- BioNavi-NP evaluation samples 60 of 388 usable test targets using a fixed seed.
- All policies share reaction rules, ranking, terminal set, depth limit, graph updates, stopping conditions, and a 20-candidate LLM observation.
- Performance is reported as solve rate versus molecule expansions, with LLM results averaged over three seeds and deterministic policies evaluated once.
4 Results
Route-supervised frontier selection improves search performance across LASER, Golden, and BioNavi-NP, especially relative to direct prompting and MCTS at larger budgets. Candidate-view construction also affects which productive choices remain available to the LLM.
- Frontier observation: 84% coverage@20 is achieved by the four-signal portfolio, versus 80% for depth stratification and 79% for native-substrate similarity alone.The portfolio combines depth stratification, native-substrate similarity, reaction precedent, and molecular size.
- LASER: 65 ± 1% solve rate is reached by the fine-tuned policy at N = 10 on LASER, versus 59% for MCTS.At N = 200, the corresponding rates are 78 ± 1% and 75%.
- Golden: 68 ± 8% solve rate is reached by the fine-tuned policy at N = 10 on Golden, versus 45% for MCTS.At N = 200, the fine-tuned policy reaches 88 ± 3%, versus 80% for MCTS.
- BioNavi-NP: 63 ± 2% solve rate is reached by the fine-tuned policy at N = 200 on BioNavi-NP, versus 45% for MCTS and 39±2% for direct prompting.The policy reaches its first solution after 21.3 expansions on average, versus 38.6 for prompting and 25.0 for MCTS.
- Matched-observation control: Restricting classical policies to the same bounded observation leaves Golden BFS and MCTS solve rates unchanged at N = 200, while DFS and greedy similarity improve.This control does not explain the fine-tuned policy’s Golden advantage.
- Efficiency: At N = 200, fine-tuning requires 209, 168, and 417 seconds per obtained solution on LASER, Golden, and BioNavi-NP, respectively.Direct prompting requires 854, 972, and 3069 seconds on the same benchmarks; wall-clock values are hardware-dependent.
5 Discussion and Conclusion
Route supervision improves how a fixed biochemical search budget is allocated, while practical pathway value remains bounded by biological realism and model scope.
- Results: At N = 200, fine-tuned Qwen2.5 reaches 78 ± 1%, 88 ± 3%, and 63 ± 2% solve rates on LASER, Golden, and BioNavi-NP, versus 75%, 80%, and 45% for MCTS.The strongest reported gain is on BioNavi-NP, indicating transfer beyond MetaNetX-derived training routes.
- Results: Fine-tuning improves solve rate by 9–24 percentage points at N = 200 and reduces expansions needed to reach a first solution.All policies share the same observation and action space and biochemical reaction model, isolating route supervision within the LLM comparison.
- Implications: Efficient search can reduce the combinatorial space, while enzyme, thermodynamic, toxicity, flux, and physiological criteria remain necessary for pathway assessment.A solved route is a design hypothesis connecting a target to metabolites available in a host chassis, not by itself a validated production pathway.
- Implications: Agentic frontier selection could incorporate biological signals during search, making exploration more chassis- and constraint-dependent while retaining explicit control over transformations.The framework separates deterministic biochemical generation from learned search policy.
- Limitations: Performance remains dependent on the reaction-rule system and chassis sink, while stereochemistry, thermodynamics, enzyme activity, yields, kinetics, and cellular physiology are not modelled.The authors characterize solved routes as connectivity-level reachability to the selected E. coli chassis rather than experimentally validated production pathways.
- Conclusion: Route supervision improves limited-budget allocation without replacing the underlying biochemical model.The broader conclusion is that trained agents may learn domain-specific planning strategies from successful scientific trajectories.
- Scope: All molecular search operations use two-dimensional graph representations with stereochemical information removed.This representation is used for fingerprints, similarity computations, reaction-rule applications, and search operations.
B Reaction-center ECFP applicability filter
The reaction-center ECFP filter accelerates template screening by testing local fingerprint compatibility before graph-level validation, without changing the definition of valid retrosynthetic steps.
- Reaction preprocessing: MetaNetX reaction equations are instantiated in both orientations before atom mapping and decomposition into mono-substrate transformations.This preprocessing produces the reaction templates used by the one-step model.
- Template construction: Reaction templates encode substrate and product graph patterns around atoms whose local chemical signatures change between substrate and product.The signatures include charge, hydrogen count, valence, aromaticity, degree, ring membership, and neighboring bond orders.
- Applicability filtering: The applicability filter retains a reaction when its reaction-center ECFP is coordinate-wise compatible with the molecular system.This condition is necessary but not sufficient because ECFPs do not preserve complete molecular connectivity.
- Graph validation: All retained templates are subsequently applied to molecular graphs with RDKit, and only graph-valid products are returned.The vector-space test reduces expensive template applications without changing graph-level validity.
C Cofactors and their treatment in the search
Cofactors are treated as assumed-available auxiliary species rather than pathway precursors, while terminal metabolites are defined through an E. coli chassis sink.
- Cofactor set: Table 2 presents representative entries from the cofactor list, whose complete set contains 68 records.The table summarizes the auxiliary species used in the search rather than pathway-specific precursors.
- Cofactor set: The search uses a 68-entry biochemical cofactor set including ATP/ADP, NAD+/NADH, NADP+/NADPH, CoA, water, phosphate, and common metal ions.These species are auxiliary participants in biochemical transformations rather than pathway-specific precursors to synthesize.
- Search treatment: Generated cofactors are not added as unresolved frontier molecules, preventing common species from creating additional AND–OR graph branches.A solved route therefore assumes availability of the listed cofactors.
- Terminal set: A retrosynthetic branch is complete when all unresolved precursors belong to the sink or terminal building-block set.Sink membership represents an assumption about chassis availability rather than metabolite abundance under a particular growth condition.
- Terminal set: The E. coli sink is inherited from RetroPath RL and contains 753 unique terminal metabolites after processing the iML1515-derived collection.The source workflow uses genome-scale metabolic models, flux-variability analysis, database cross-references, and InChIKey-based deduplication.
E Frontier ranking diagnostics
The diagnostic evaluates how frontier-ordering signals preserve productive choices within a fixed 20-candidate observation. A four-signal portfolio achieves the strongest coverage, while MRR alone does not reliably indicate retention of productive options.
- Metrics: Coverage@k measures whether at least one productive choice remains visible after truncation, whereas MRR emphasizes the ranks of on-route choices.Coverage@20 is the primary criterion because the LLM selects only among displayed candidates.
- Metrics: 77–78% coverage@20 is achieved by the highest-MRR configurations, below the selected portfolio despite their stronger MRR.Higher MRR therefore does not necessarily retain more productive choices under top-20 truncation.
- Frontier ordering: 80% coverage@20 is reached by depth stratification alone.
F Search policies and agent interface
All methods select frontier molecules in a shared AND–OR search, but classical policies access the complete frontier while LLM policies choose from a bounded, ranked observation. The biochemical expansion engine and graph updates remain common across policies, isolating frontier selection as the policy decision.
- Shared interface: The shared engine constructs the frontier, expands the selected molecule, generates reactions, updates the graph, tests chassis membership, and propagates solved status identically across policies.The common search budget is the number of expanded molecules.
- Classical policies: BFS selects minimum-depth frontier molecules, whereas DFS selects maximum-depth frontier molecules.
- Classical policies: Greedy sink similarity expands the frontier molecule with maximum structural similarity to a terminal chassis metabolite.It does not account for the remaining explored route.
- MCTS: MCTS repeatedly descends from the root through molecule OR-nodes and reaction AND-nodes until reaching an eligible frontier molecule.Accumulated statistics use root-based paths because expanded molecule nodes leave the frontier.
- Agent interface: Table 4 contrasts complete-frontier access for classical policies with bounded-observation access for LLM policies.
- Agent interface: LLM policies return one displayed frontier-molecule index from a bounded portfolio, without generating reactions, selecting templates, or modifying the graph.The observation is constructed from depth, native-substrate similarity, biochemical precedent, and molecular size.
G LLM policy and supervised fine-tuning
The learned policy selects which frontier molecule to expand while a deterministic engine handles biochemical generation and graph updates. Route-replayed supervision trains this choice-only policy, with the frontier representation affecting how much supervision is recovered.
- Decision interface: The LLM returns only a frontier-molecule index; deterministic reaction generation, graph updates, and terminal tests remain outside the policy.The strict interface prevents the model from proposing reactions or directly evaluating their validity.
- Supervised training: 12,251 mined biochemical routes provide replayed search states paired with frontier choices that preserve attested routes.When multiple displayed molecules fit a route, a fixed convention selects the branch with the largest remaining route cost, making the label an imitation target rather than a unique optimum.
- Frontier construction: 36,703 replay decisions are recovered by the four-signal portfolio, 16% more than depth stratification before benchmark filtering.The portfolio loses 26% of otherwise available decisions to truncation, compared with 37% for depth stratification.
- Data split: After target-based filtering, 27,795 state–choice pairs remain, split into 25,150 training and 2,645 validation examples across 709 held-out targets.The split prevents decisions from the same route target from appearing in both sets.
- Training evaluation: Validation cross-entropy is 0.158 with mean token-level accuracy 0.946, measuring imitation of frontier-choice formatting rather than end-to-end search.The prompting baseline uses the same model interface without LoRA fine-tuning.
- Observation and budget: The LLM receives at most 20 portfolio-selected candidates, while the global search budget counts molecule expansions across the whole graph.The default candidate observation exposes SMILES and retrosynthetic graph depth, not engine-derived scores or reaction rules.
I Matched-observation control
The matched-observation control restricts classical policies to the same 20-candidate view available to the LLM while leaving the search setup otherwise unchanged. The restriction leaves strong policies unchanged but improves weaker ones, especially greedy sink similarity, and the interpretation is limited to Golden at N = 200.
- Control design: At N = 200 on Golden, classical baselines are compared on the complete frontier and the portfolio-defined 20-candidate view.Targets, expansion budget, reaction generation, graph updates, and policy definitions remain unchanged.
- Results: MCTS and BFS retain the same solve rates under the 20-candidate restriction, while DFS gains 5 percentage points and greedy sink similarity gains 20.The bounded view therefore affects weaker selection policies more than the two strongest classical baselines.
- Interpretation: The portfolio can filter candidates favored by poorly aligned criteria while leaving already effective policies unaffected.The large greedy-similarity gain is consistent with sink proximity being a weak guide to productive retrobiosynthetic search.
- Scope: The matched-observation control provides no evidence that complete-frontier access disadvantages classical baselines relative to the LLM.This control remains specific to Golden at N = 200.