How AiZynthFinder results compare to the bench

You have an AiZynthFinder result. Here is what it supports.

• ⊘ "Solved" is not "makeable." Solved means a route was found whose leaves are in the configured stock. That is a statement about the template library and the stock set as much as about your molecule. A solved route can be unexecutable, and an unsolved molecule can be trivially makeable. The error runs in both directions, and both have been measured: against 44 compounds with published syntheses, AiZynthFinder at default settings found routes for 20 — sensitivity 0.45 — while against real AstraZeneca synthesis records, template models over-predicted makeability for novel combinatorial designs.
• ⊘ An unsolved result is not a verdict that your molecule cannot be made. It is a verdict that no route exists on this template library, to this stock, within this budget. Re-running one search tool with different hyperparameters reclassified 10% of a set from makeable to not.
• Quote the solve rate that matches this configuration, not the headline. The widely cited 80–86% was measured on a proprietary Reaxys + Pistachio + AstraZeneca-ELN template model and AstraZeneca's internal stock. Neither runs here. This platform runs the public USPTO policy against a ZINC-only stock at 100 iterations, and the closest independently measured anchor for that setup is 76.4% on drug-sized ChEMBL molecules — falling to 9.5% on heavy molecules of 500–800 Da.
• ⊘ Do not rank routes by the score. AiZynthFinder's own authors state the score "was designed to support the tree search and is rather indiscriminate with regard to the quality of the route." It is the search's reward — short routes with a large in-stock fraction — not a probability that the chemistry works.
• ⊘ On this platform, is_solved is weaker than the library's own verdict. The deployed Modal app sets the top-level is_solved to number_of_routes > 0, so it means "the search returned a route tree," not "every leaf is in stock." The real per-route verdict is present inside each route record. See known defects.
• ★ A synthetic-accessibility score is not a substitute for this search, which is why GenMol and REINVENT send you here — but not for the reason you might expect. Scores predict a search's verdict reasonably well (SAscore reaches AUC 0.85–0.90 against AiZynthFinder solvedness). What they cannot give you is a usable cutoff — there is "no threshold value at which the SA, SC, and SYBA scores can be set that clearly separates compounds that can and cannot be solved by AiZynthFinder" — and they rank within a chemical series but not across one, which is exactly what a generated set is. See score versus search.
• Chemists have graded these routes, and the two studies that did it disagree. Blinded, with 17 raters on a deliberately hard set, roughly 44% of expert votes on individual steps flagged a problem. Unblinded, single-rater, on a random test set, 45 of 50 top-1 routes were called synthetically feasible. The blinded figure is the better evidence. No survey has ever asked chemists directly whether they find CASP route output usable.
• Confirm makeability with a chemist, then by running the chemistry. A route is a literature-shaped hypothesis about disconnections; it carries no conditions, no yields and no selectivity analysis. Every published route that reached a flask — for this tool and for every other — was modified by a chemist first.

Jump to agreement with wet lab or against other in silico methods. The general framing for every tool is in reading an in silico result against the bench.

How in silico is used currently

Retrosynthesis planning is used to generate and triage route hypotheses, not to decide them. In practice that is three things in a live program:

• Makeability triage on a design set. You have more structures than you can make — from a generator, a virtual library, or a screening shortlist — and you want to deprioritize the ones no software can disconnect. This is the mode the platform's GenMol and REINVENT pages point at, and the one where the per-molecule bar is low and the set is what matters.
• Route scouting on a single committed target. You have decided to make a compound and want alternative disconnection strategies to compare against the one you drew, including ones a chemist would not have reached for.
• Idea generation inside a human-run route design. This is the tool's own developers' framing. Writing about AiZynthFinder 4.0, the AstraZeneca Molecular AI group list "accurate scoring of routes" as an open challenge, "which is essential for the software to recommend routes rather than serving as an ideation tool."

What retrosynthesis software is not used for, by anyone who reports their validation honestly, is concluding that a molecule can be made. Every published route that reached a flask had a chemist between the search and the bench.

Against other in silico methods

Every comparison in this section is confounded by basis. No two retrosynthesis tools in the published literature have been compared on the same template library and the same stock set, so the differences below mix the search algorithm, the reaction data and the building-block catalogue together.

Against ASKCOS

The one published head-to-head is in AiZynthFinder's own paper, author-reported, on 100 compounds drawn at random from ChEMBL. ASKCOS ran on the public web server with a max depth of 6 and a 120-second expansion time; AiZynthFinder ran the CLI with the ZINC stock and the USPTO policy.

⊘ Those two columns differ in their templates and in their stock, so the 7-compound gap is not attributable to either tool. The authors say so themselves: "We have found that route finding capability depends on the stock that is used as stop criteria in both tools," and then quantify it on their own arm — "If we include the readily available Enamine building blocks in the stock, we could find routes for an additional 10 compounds." Adding one vendor catalogue moved AiZynthFinder from 55 to 65, past ASKCOS, without changing the software at all.

The authors also flag their own test set: "these 100 compounds are not necessarily a representative part of the chemical space that might be relevant in a drug design project. Thus, the test set should be viewed as an illustration of the capacity of the software rather than a go-to benchmarking set."

Against Retro*

On a benchmark of 189 USPTO targets deliberately hardened by removing molecules that a heuristic breadth-first planner could already solve, with building blocks from a 231-million-molecule eMolecules catalogue and a budget of 500 one-step model calls, the Retro* authors report:

⚠︎ Read that MCTS row carefully. It is the Retro* authors' own reimplementation of Segler-style MCTS with PUCT, not AiZynthFinder, and the budget is counted in one-step model calls rather than iterations. It is evidence that best-first A*-style search dominates MCTS on this benchmark; it is not a measurement of the software that runs here. The benchmark also has a built-in overlap guarantee in the opposite direction: the test set keeps "only the routes whose reactions are all covered by the top-50 predictions by the one-step model," so every target is known in advance to be reachable by the model's own templates.

Against Segler-style 3N-MCTS

AiZynthFinder is an open reimplementation of the Segler neural-MCTS approach, and the original reports a far higher solve rate — 95.24% ± 0.09 on 497 molecules, in 13.0 s per molecule. ⊘ The basis is the entire explanation, and almost none of it is available here:

The iteration gap is three orders of magnitude, and it is explicitly the lever: the same study solved ">80% of the test set with a time limit of 5 s per target molecule" and 92% at 60 s, while its heuristic baseline reached 4% at 60 s. Independent work on AiZynthFinder makes the same point from the other side, noting that "3N-MCTS utilizes 100,000 iterations per molecule, in contrast to 100 used by AiZynthFinder." ⚠︎ A 95% solve rate on a Reaxys-derived rule set with a 423,731-molecule stock and a 100,000-iteration budget tells you nothing about what this platform will return.

The Segler rules' own coverage is also stated and is a useful ceiling: the expansion and rollout rule sets "cover 52% and 79% of all chemical reactions from 2015 and after," and the expansion policy's top-1 accuracy was 31%, rising to 72.5% at top-50.

Against other search algorithms

An independent group at Université Paris Dauphine reported that Nested Monte Carlo Search and Greedy Best First Search both outperform AiZynthFinder's base MCTS on a 60-molecule benchmark — 40 drugs from PubChem plus "20 SMILES representation of molecules selected by Bayer's chemists" — and concluded that "the search algorithms are bounded by the quality of the policy network." ⚠︎ That study graded route quality by solve rate alone and states no stock set, so it is a direction, not a figure. And the Nottingham group's modified UCT variants solved between 600 and 900 more molecules than unmodified MCTS across 50,000 ChEMBL targets. This platform runs base MCTS only; the alternative algorithms AiZynthFinder supports are not exposed.

Chen B, Li C, Dai H, Song L. Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search. Proceedings of the 37th International Conference on Machine Learning. 2020. doi: 10.48550/arXiv.2006.15820

Segler MHS, Preuss M, Waller MP. Planning chemical syntheses with deep neural networks and symbolic AI. Nature. 2018;555(7698):604-610. doi: 10.1038/nature25978

Roucairol M, Cazenave T. Comparing search algorithms on the retrosynthesis problem. Molecular Informatics. 2024;43(7):e202300259. doi: 10.1002/minf.202300259

A score is not a search

★ This is the comparison the generative pages need. GenMol and REINVENT both tell you to settle makeability here rather than with a synthetic-accessibility score. That advice is correct, but not because the scores are bad at predicting what a search will say — they are reasonably good at it. It is correct for four narrower reasons, each measured.

First: what a score actually is. SAscore is a fragment-frequency heuristic, not a route search. It sums ECFC_4 fragment contributions derived from 934,046 PubChem molecules and adds a penalty for ring complexity, stereocentres, macrocycles and size. Its authors chose that design deliberately — "a decision not to rely on comprehensive databases of reactions and reagents… clearly favored implementation of a method based on molecular complexity" — and they validated it, after the algorithm was frozen, against 40 molecules rated by 9 Novartis chemists (r² = 0.890 against the mean chemist rank; inter-chemist r² averaged 0.718 over all pairs, range 0.450–0.892). They also stated its ceiling themselves: the SAscore "cannot compete with more sophisticated approaches for estimation of synthetic accessibility which reconstruct the full synthetic path, in cases when the results are critical." SYBA's authors say the same of theirs, adding that a complexity score "cannot… incorporat[e] other factors such as the availability of starting materials, reaction yields or a price aspect."

Second: a score predicts the search well enough, and that is the trap. On a 49-compound curated set, SAscore predicted AiZynthFinder's own solved/unsolved verdict at AUC 0.90 / accuracy 0.81 (RAscore 0.85/0.85, SCScore 0.67/0.69, SYBA 0.66/0.67), and that study's conclusion is explicitly pro-score. On ~20,000 held-out ChEMBL compounds, SAscore reached AUC 0.85 against AiZynthFinder solvedness, SYBA 0.74 and SCScore 0.61 — against 0.93–0.95 for ECFP6-based classifiers trained on the same labels. On ASKCOS solvedness, SA Score reached AUC 0.87 — where, tellingly, raw SMILES string length reached 0.69 and SCScore only 0.61.

Third: ⊘ there is no threshold. This is the finding that justifies running the search. From the AstraZeneca group who built RAscore on AiZynthFinder's own labels:

"we have found that there is no threshold value at which the SA, SC, and SYBA scores can be set that clearly separates compounds that can and cannot be solved by AiZynthFinder, as shown by the overlapping histograms… Thus, there is potential for them to be misused when filtering large virtual libraries."

That gap between ranking and deciding has been quantified since: against Retro* labels on 1,800 structurally complex molecules, SAscore scored ROC-AUC 0.772 but F-score 0.333 — it orders those molecules acceptably while being close to useless at any fixed cutoff. A score orders candidates; it does not decide them. And the ordering degrades exactly where generative design lives: every score tested fell monotonically as molecules grew complex (SAScore ROC-AUC 0.980 → 0.818, SYBA 0.968 → 0.784, across three 1,800-molecule sets).

Fourth: a score ranks within a chemical series but not across series. Seven chemists blind-labelled 30 generated molecules feasible or not, 12 of 30 feasible, with the chemists "all agree[ing] with each other except two chemists on two molecules." SA_Score's AUC against those labels was 1, 0.33 and 0.96 inside the three single-reference generations but only 0.63 across all 30 pooled. The authors' explanation:

"the SA Score… performed well for each separate generation, i.e., it was able to assess the relative complexity of molecules similar to each other, but it failed to discriminate the synthetic accessibility of molecules with very different chemical structures… Complexity and synthetic accessibility are two different notions, and although they are usually correlated, complex molecules can be feasible and simple molecules can be unfeasible… the SA score is better suited to assess molecular complexity."

So a score is usable to rank a scaffold series and not usable to compare a diverse generated set — which is precisely what a generator produces.

Two corollaries worth carrying.

• A score trained on a search's labels is a model of that search, not of chemistry. RAscore was trained on 400,000 AiZynthFinder outcomes, and the Warsaw group put the consequence cleanly: "RAscore's good result is not surprising, because it was trained on the outcomes of the AiZynthFinder algorithm. This, combined with the low sensitivity of AiZynthFinder, allows us to claim that RAscore is a precise heuristic of AiZynthFinder outcomes, but not necessarily a synthetic accessibility score in general." ⚠︎ RAscore's labels were also generated against "the ACD catalogue, Enamine building block set, and AstraZeneca internal database" — a stock set that is not public and is not this platform's.
• ⊘ Optimizing a generator against a score is reward hacking. Running REINVENT with a single score as the only reward, "SAScore steers the model towards a minimized number of rings and linearly fused scaffolds between aromatic rings, whereas SCScore favours simple long-chain structures — both strategies artificially inflate the respective reward but sacrifice the accessible chemical space." The scales are also not interchangeable: on one unconditioned generated set, SAscore called 99.63% of molecules easy to synthesize and SCScore called 40.87% easy — a 58.8-point disagreement between two cheap scores on identical molecules.

What the measured workflow actually is. Not score-or-search, but score inside generation and search on the survivors. On hard GuacaMol objectives, 30.2% of the top-100 molecules had an ASKCOS route unbiased, rising to 80.2% when generation was biased by SA_Score (trivial objectives: 60.1% → 91.0%). For reference on the same pipeline: MOSES 89.8%, ChEMBL 68.3%, ZINC 60.8%, GDB17 3.5% synthesizable. But a score constraint buys fewer usable molecules than a real retrosynthesis constraint: on a Pi3K/mTOR lead-optimization task, an SA-score constraint produced 311 molecules satisfying all objectives against 850 for a retrosynthesis-based constraint — while chemists rated the top 10 from each about equally (SAscore 0.63, retrosynthesis 0.63, no constraint 0.00). ⚠︎ So the score's weakness there is yield of usable candidates, not per-molecule quality. Be precise about that.

Skoraczyński G, Kitlas M, Miasojedow B, Gambin A. Critical assessment of synthetic accessibility scores in computer-assisted synthesis planning. Journal of Cheminformatics. 2023;15(1):6. doi: 10.1186/s13321-023-00678-z

Thakkar A, Chadimová V, Bjerrum EJ, Engkvist O, Reymond JL. Retrosynthetic accessibility score (RAscore) — rapid machine learned synthesizability classification from AI driven retrosynthetic planning. Chemical Science. 2021;12(9):3339-3349. doi: 10.1039/d0sc05401a

Ertl P, Schuffenhauer A. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of Cheminformatics. 2009;1(1):8. doi: 10.1186/1758-2946-1-8

Voršilák M, Kolář M, Čmelo I, Svozil D. SYBA: Bayesian estimation of synthetic accessibility of organic compounds. Journal of Cheminformatics. 2020;12(1):35. doi: 10.1186/s13321-020-00439-2

Parrot M, Tajmouati H, da Silva VBR, Atwood BR, Fourcade R, Gaston-Mathé Y, Do Huu N, Perron Q. Integrating synthetic accessibility with AI-based generative drug design. Journal of Cheminformatics. 2023;15(1):83. doi: 10.1186/s13321-023-00742-8

Gao W, Coley CW. The Synthesizability of Molecules Proposed by Generative Models. Journal of Chemical Information and Modeling. 2020;60(12):5714-5723. doi: 10.1021/acs.jcim.0c00174

Chen S, Jung Y. Estimating the synthetic accessibility of molecules with building block and reaction-aware SAScore. Journal of Cheminformatics. 2024;16(1):83. doi: 10.1186/s13321-024-00879-0

Hastedt F, Hellgardt K, Yaliraki S, Zhang D, del Río Chanona EA. MolPrice: assessing synthetic accessibility of molecules based on market value. Journal of Cheminformatics. 2025;17(1):150. doi: 10.1186/s13321-025-01076-3

Gao S, Zhou X, Liang L, Lin J. RetroScore: graph edit distance-guided retrosynthesis for accessibility scoring with route metrics. Journal of Cheminformatics. 2025;18(1):10. doi: 10.1186/s13321-025-01138-6

In silico vs. wet lab

The standard wet-lab approach for the question AiZynthFinder answers is a chemist designing a route and then executing it at the bench — or, short of execution, an expert chemist's route assessment. That is a process and a judgement, not a single measurement technique, and it changes how the table reads.

There is no RMSD or AUC here. A route is not a prediction about one measurable quantity. What the literature measures instead is one of three different things, and they are not interchangeable:

1. Solvability — does the search find a route whose leaves are all in the configured stock? This is the only one the tool reports, and it is a property of the stock set and the template library as much as of the molecule.
2. Route acceptability — would an expert chemist accept the proposed route?
3. Execution — was the route run, and did it work?

Read each cell against its own question.

Read the accuracy column by cell, never down the column. The three entries are not the same measurement and the two in silico rows were not measured on the same template library or the same stock set, which is the central point of this page.

On that reading, the honest position for this chain is: seconds instead of days to find out whether a route exists on a public template library and a fragment-sized ZINC stock, at a cost of one credit, with nothing measured about whether the route would work. The chain's value is triage throughput. The bar it has to clear is a chemist's judgement.

⚠︎ Two asymmetries matter. A chemist can make a compound the search cannot disconnect, because the template library is a record of published reactions and the chemist is not. And the bench reports what happened, while the search reports what the templates find precedented — those are not interchangeable at any score.

⚠︎ The platform's own index page and quickstart say a typical run takes 5–30 minutes. That is inconsistent with the configured 120-second search limit and with the 139-second padded runtime that job pricing is derived from. Treat the configured limit as authoritative.

Agreement with wet lab

The agreement question splits in two and the two halves must be kept apart: would a chemist accept the proposed route, and was the route run, and did it work. Both have partial answers for this tool, and neither has a success rate. Each subsection below states which question it answers.

Solvability, with the stock set and template library that produced each number

⊘ A solve rate without its stock set and template library is unusable. Every figure below carries both. The spread between them — 9.5% to 99.1% for the same software — is the most useful thing on this page.

Author-reported, AiZynthFinder 4.0, on four target sets. Note that the two sets with the high figures used a proprietary template model and AstraZeneca's internal stock, neither of which runs on this platform:

Those four rows are the source of the "80–86%" and "71% on ChEMBL" claims that circulate — including in this platform's own index page. The 80–86% pair does not describe this platform. The row whose configuration resembles ours is the ChEMBL row at 71%, and even that one combined E-molecules with ZINC. In the same paper, the authors measured what each half contributed: "the E-Molecules stock set is most useful as it covers on average 80% of the starting material, whereas ZINC stock only covers 60% on average." This platform has the 60% half.

The GDB row is the useful extreme. GDB MedChem is enumerated chemical space, so its compounds have no literature precedent at all, and the authors read the 10% as "highlighting a disconnection of the current template-based model trained on historical reaction data, with the chemistry needed to find synthesis routes for the enumerated, and therefore potentially non-synthesizable, GDB compounds."

Independently measured, and the figures that match this configuration most closely. A University of Nottingham group ran unmodified AiZynthFinder MCTS as the control arm of an algorithm study, varying only the stock representation and the iteration budget:

⊘ Same software, same version, same molecules — 76.4% to 99.1% on the stock representation and the iteration budget alone. And 9.5% when only the molecule set changes. This is why a bare solve rate is not a number you can act on. The same authors note that "3N-MCTS utilizes 100,000 iterations per molecule, in contrast to 100 used by AiZynthFinder," and that "increasing the iteration count is beneficial in solving more molecules." This platform defaults to an iteration ceiling of 100 and a 120-second wall clock, and chat runs cannot change either; a direct job request can raise both, within an iteration ceiling of 5000 and a wall clock of 1800 seconds.

The same group also replicated the published baseline: "The literature benchmark for the 50K ChEMBL data set was 73% using the default parameters. We were able to replicate this with AiZynthFinder 3.7." A third measurement, from the AstraZeneca group, puts default settings at 73% on ChEMBL and 85% on their in-house set. Three independent-ish measurements landing at 71–76% on ChEMBL is the most reliable thing on this page.

Route acceptability — would a chemist accept the route?

Chemists have graded AiZynthFinder routes directly, twice, and the two studies disagree. Report both; they used different instruments on different sets.

The blinded, multi-rater study — roughly 44% of expert votes on individual steps flagged a problem. Seventeen chemists, "most… from AstraZeneca, with some external experts," graded 50 routes and 204 reactions generated by AiZynthFinder v4.3.1 at default settings, taking the top-1 route per target. Each expert graded 20 randomly sub-selected routes, about 4 experts per route. Among expert votes on individual reactions, the majority category was "Reaction feasible" at 55.6%, with "Selectivity issues" 11.5%, "Functional group issues" 8.4% and "Unnecessary step" 8.0%. ⊘ Read that the other way: on roughly 44% of expert votes, a chemist flagged something wrong with the step.

⚠︎ Three things bound that figure. The 50 compounds were drawn from the Journal of Medicinal Chemistry "with the help of the experts to identify drug-like molecules challenging for retrosynthesis analysis" — a set deliberately curated to be hard, not a representative sample. Both solved and partially-solved routes were included. And one of the four flagged categories was partly an instrument artefact rather than chemistry: on "Unnecessary step," the authors found that "for the majority of the cases AiZynthFinder generated compounds that were supposed to be in stock but were tagged as not in stock in the stock solutions file."

★ That study also wrote down the scale this question actually needs — "Route feasible as it is / Route feasible with few modifications / Route feasible with significant modifications / Route unfeasible / Route was not solved to building blocks." ⊘ The distribution of chemists' votes across those five options is in that paper's Supplementary Information and could not be retrieved. So the headline figure a reader most wants — what share of AiZynthFinder routes a chemist calls runnable as-is — is not quoted on this page, because it was not read. The underlying response data is public if someone wants to derive it.

The unblinded, single-rater study — 90% of top-1 routes called synthetically feasible. A separate AstraZeneca-affiliated preprint had expert chemists analyse the top-1 AiZynthFinder route for 50 random test molecules (USPTO expansion and filter models, eMolecules stock snapshot of January 2023): "for 30 out of 50 molecules, the top-scored route was observable from literature experimental studies. Out of the other 20 molecules, for 15 of them, the top-1 scored route was deemed to be a good, often shorter, alternative… this results in 45 of 50 synthetically feasible routes, or 90%." Four needed step modifications; one failed outright.

⊘ That 90% is much weaker evidence than the blinded 44% figure, and the reasons matter. The paper never states how many chemists took part, reports no blinding and no inter-rater agreement — each route appears to have been rated once. The experts were explicitly not blind to the model: they analysed "the route quality suggested according to the model's score predictions." It is a preprint, not peer reviewed. And the paper's own abstract misdescribes the number as "90% correspondence to the expert synthesis evaluation," which reads as a model–expert agreement rate; it is not, it is an expert acceptance rate. Treat 90% as an upper bound produced by an unblinded, single-rater, developer-run assessment of a random test set, and ~56%-of-steps-clean as the blinded multi-rater reading of a deliberately hard set.

One independent assessment is harsher still. A commercial group that had PhD-level chemists label over 4,500 reactions, then graded top-1 routes from four systems on a four-tier scale ending in "Nonsense," reports that "AiZynthFinder demonstrates significant limitations, failing to identify viable pathways for significant number of the target molecules while generating a substantial proportion of unreliable routes classified as Nonsense and Rather Not." ⊘ Their per-system percentages exist only inside a stacked bar chart and are not quoted here. Their framing of the field is worth carrying: "To date, none of these approaches have been extensively validated through human expert evaluation."

And the sharpest quantified gap comes from AstraZeneca's own group. Running AiZynthFinder's production template model and MCTS over 1,748 targets from PaRoutes set-n1 with an eMolecules stock and default hyperparameters, they compared plain solvability against solvability plus satisfaction of bond constraints — explicit statements of which bonds a chemist wants disconnected and which preserved:

"Although the standard search solved most targets (84.91%) in a short time, it only solved and satisfied the bond constraints for 54.80% of the targets."

⊘ Roughly 30 points of that solve rate are routes that a chemist's stated disconnection preferences would reject. The authors' framing of the gap is that contemporary tools "generate solved routes for the majority of target molecules, but do not consider the prior knowledge of the chemist, including specific bonds that should disconnect or remain connected throughout the routes." Adding a disconnection-aware model and a multi-objective search raised constraint satisfaction to 75.57%; neither is available on this platform, which runs single-objective MCTS on the state score.

⚠︎ That measurement is on the AstraZeneca production template model and an eMolecules stock, not this platform's public USPTO policy and ZINC stock. Read the 30-point gap as the finding, not the absolute figures.

A second, qualitative finding from the same group names the specific reason a solved route can be unexecutable. Evaluating AiZynthFinder-generated reaction trees in 2026, they report that such tools "frequently produce reaction trees with selectivity issues, manifested by competing reactive sites. Therefore, manual postprocessing is often required to formulate appropriate protection strategies, which prevents full automation." A solved route with competing reactive sites is not a route you can hand to a chemist unchanged.

Blind comparison against literature routes has been done — for other tools, and it came out favourably for the machine. Two studies, neither on AiZynthFinder:

• A double-blind A/B test, 3N-MCTS against literature. Chemists chose between two step-count-matched routes to the same molecule "on the basis of personal preference and synthetic plausibility," with neither raters nor conductors told which was which. Over 128 submitted judgements the split was literature 43.0%, machine 57.0%, P = 0.26, and the authors concluded that "organic chemists should consider the quality of retrosynthetic routes generated by a machine to be on par with reported routes for molecules of practical relevance." ⚠︎ Three caveats the paper does not state. The 45 raters were "postgraduate students who had specialized in organic chemistry," not practising medicinal or process chemists. P = 0.26 is a failure to reject, not a demonstration of equivalence — no equivalence test and no power analysis was run, and the non-significant direction actually favours the machine. And 45 raters across 9 route pairs would allow 405 judgements, so each pair rests on about 14 judgements, not 45.
• A "synthesis Turing test," Chematica against literature. Twenty machine-designed and twenty literature pathways were presented in no particular order to 18 synthesis experts — academic total-synthesis principal investigators, the opposite end of the seniority scale. They could not tell them apart: AUC 0.53 ± 0.03, "close, to within error, to random guessing," and seniority did not help (AUC against H-index, R² = 0.000267). Machine routes scored marginally higher on the human-versus-machine scale (5.17 against 4.58) and on expert-rated elegance (4.75 against 4.55). ⚠︎ It was single-blind, not double-blind — the conductors held the key — and the two arms' reaction conditions came from different provenance.

⊘ Neither is a measurement of AiZynthFinder, and neither asks the question a reader actually has. Both measure detectability and preference between two routes, not "would you run this." No published survey, poll or interview study has ever asked medicinal or synthetic chemists whether they find CASP route output acceptable, trustworthy or usable. Every acceptance datum in this literature sits inside a model paper as a validation arm, written by the model's own team.

⚠︎ Finally, calibrate the referee. Chemist judgement is itself a noisy instrument, and nowhere near a gold standard. Asked to accept or reject fragments, 19 doctorate-level medicinal chemists reached 75% consensus on only 8% of about 3,685 compounds, with individual acceptance rates ranging from 7% to 97% (mean 45%) and self-consistency on 227 deliberately repeated compounds averaging 0.52. In a separate pairwise-preference study at Novartis, inter-chemist agreement over 220 pairs was Fleiss κ 0.40 then 0.32 across two pilot rounds of 9 and 14 chemists, within-chemist agreement Cohen κ 0.60 and 0.59, and per-chemist κ ranged from 0.16 to 1.00 with four values not significant at p = 0.1 — after which the authors "forwent the pair repetition requirement in the production runs and considered all participating chemists as a single labelling oracle," measuring no agreement at all thereafter. ⚠︎ Both are molecule-level, not route-level. The point stands either way: a figure for how often chemists accept a route is bounded by how often chemists agree with each other.

Against molecules known to have been synthesized

The cleanest wet-lab anchor available is also the most uncomfortable. An independent group at the University of Warsaw assembled 49 compounds — "The majority of these compounds are drugs or plant metabolites, of which 44 have documented synthesis" — and ran AiZynthFinder at default parameters:

"AiZynthFinder found a synthetic route for 22 compounds. For all found synthetic routes, 20 of them are known (precision 0.91), and for all known synthetic routes, 20 of them are found (sensitivity 0.45)."

⊘ Read the sensitivity figure. Of 44 molecules that demonstrably can be made — somebody published the synthesis — the search found a route for 20. When it did produce a route, that route was a known one 91% of the time; but it missed more than half of the molecules whose makeability is a matter of record. An unsolved result on this platform is not evidence that a molecule cannot be made. That is the single most important sentence on this page for anyone deciding what to drop from a design set.

⚠︎ That study states only "default parameters" for the AiZynthFinder arm — no template library, no stock set, no budget — so this figure is reported without its mandatory basis and the absolute values should not be transferred to another configuration. The direction is what carries.

The same asymmetry shows up elsewhere, and in both instruments at once. Of 18 structurally complex molecules that all have literature syntheses, Retro* "could be successfully planned… within 10 steps" for only five, while a normalized SAScore "assigns a high score (> 0.5) to 17 of the tested molecules." The search said not makeable for 13 molecules that had been made; the score said makeable for essentially all of them. Both instruments fail on real complex chemistry, and they fail in opposite directions.

⚠︎ And the search label is not stable against its own settings. Re-labelling one 900-molecule easy-to-synthesize set with the same search tool under different hyperparameters flipped 90 of 900 (10%) from easy to hard. As that study puts it, a CASP tool "is highly dependent on its hyperparameters (such as iteration budget, tree width) as well as the database it was trained and the building block dataset. Thus, running the HS dataset with different hyperparameters/building blocks will classify some compounds as ES, and vice versa." Treat a solved/unsolved verdict as a reading of one configuration, not a property of the molecule.

Against real synthesis records — the measurement that shows "solved" is not "makeable"

The closest thing in this literature to a wet-lab concordance measurement is AstraZeneca's dataset study, which had the one thing a benchmark normally lacks: records of which compounds were sent for synthesis and which were actually synthesized. Template models were run on 1,731 compounds from 41 AstraZeneca virtual libraries designed between October 2017 and January 2019, and on the top 125 small-molecule therapies of 2018 by sales.

The finding, in the authors' words: "The models, regardless of reaction dataset, consistently over-estimate the number of compounds that can be synthesized in the case of the virtual libraries, and under-estimate with regards to the top 125 small molecule therapies."

Broken down by library type, the model "consistently over-predicted the number of compounds that could be synthesized for the 'combinatorial' library," while "the number of compounds that could be synthesized for the 'bespoke' library was consistently under-predicted." The authors give the mechanism: over-prediction happens "because some of the predicted steps may not translate to the wet lab," and "the conditions required to carry out the reaction in the forward direction are not predicted." Their conclusion is the sentence this page is built around: "only considering the number of compounds for which routes can be predicted does not afford enough granularity for the assessment of synthetic routes, and CASP tools."

⚠︎ The per-dataset percentages behind that finding are presented only in the paper's figures, so this page quotes the direction and the mechanism, not the magnitudes.

Execution — were the routes run, and did they work?

AiZynthFinder routes have reached the bench, and n is small, selected, and always chemist-modified. No prospective campaign with a denominator and a success rate has been published. Three sources report real compounds:

Seven AstraZeneca case studies, no denominator. AiZynthFinder runs automatically on every designed compound at AstraZeneca — "Over 5000 compounds per month are analysed," and "Since 2020… it has generated retrosyntheses for over 130 000 molecules." Of that volume, seven cases are written up, and in every one of the seven the proposal was altered before or during execution: a one-step cyclisation became three steps from a cheaper starting material; a reductive amination was run as oxime formation plus Zn reduction; a one-pot three-component coupling "was split into two steps" to minimise side reactivity; a sequence needed "the addition of TBS as an orthogonal protecting group strategy"; a cyclisation turned out to require Boc protection first; and a proposed decarboxylative bromination became an iodination because "37 was found to sublime during isolation." Two of the seven targets were made and found inactive. One worked at scale — a decarboxylative iodination at "92% yield on >12 g (60 mmol) scale… the largest scale instance of an AI-proposed reaction run at AstraZeneca to date" — and the step after it "suffered from irreproducibility and did not prove to be scalable."

★ The authors' own characterisation is the most useful sentence written about this tool:

"The first molecule that most new users test is one that they have personally synthesised recently, and AiZynthFinder rarely replicates their route exactly. Due in part to our self-imposed requirement to run fast searches, AiZynthFinder often gets close to a good route. Thus, experienced users seek inspiration from AiZynth rather than perfection."

And their named failure modes: "Common problems include proposals that would lead to undesired regioselectivity, functional group incompatibility, or overgeneralisation of precedented reactions to an inappropriate context." They also insist on the human gate — "these results are only useful if viewed and evaluated by expert synthetic chemists" — and decline the strong claim themselves: "We have little doubt that many of these solutions could have been found by 'traditional' means."

⚠︎ Seven selected positive cases, written by the tool's own users, with no statement of how many AiZynth-informed syntheses were attempted. No success rate can be derived, and none should be quoted.

Three compounds made, one active. The one fully auditable funnel in this literature used DrugEx v3 with an in-house synthesizability classifier trained on AiZynthFinder's own route-found labels, running the public USPTO NeuralSym policy and MCTS at 900 s, 1,000 iterations and depth 8. Its funnel: 100,000 molecules generated → 32,907 passing activity and synthesizability at probability 0.8 → 20,055 with an in-house route → 4,675 with a route of five steps or fewer → 950 passing Lipinski → 609 after a novelty cut → three selected by domain experts on diversity, "chemical eye" and a 1–2 step route. All three were made "using the suggested synthesis routes" and assayed against monoglyceride lipase: one at IC50 1 µM, two at around 100 µM. The authors apply the field's own bar themselves: "a stricter boundary of ≤10 μM, generally used for hit finding, only leaves one candidate that can be classified as active." They also note the compounds' novelty was limited — "2 and 3 are variations of the closest ligand."

★ That study also measured the stock-set effect directly, and it is the closest published analogue to this platform's situation. It ran the same molecules against two building-block sets: the 17,422,831 ZINC building blocks provided by AiZynthFinder — the same stock this platform uses — and 5,955 in-house Leiden blocks. "Solvability rates for Caspyrus centroids are around 60%" with the in-house set against "around 70%" with ZINC, and "the shortest synthesis route found with in-house building blocks is, on average, two reaction steps longer." Shrinking the catalogue costs about 12 points of solve rate and lengthens every route.

One academic route executed, with a substitution. Of 1,139 enumerated bicyclic diamines, AiZynthFinder "suggested that 573 molecules (50%) were accessible from commercial precursors." For the one target it solved, "the synthesis proposed by AiZynthfinder for the (5,7)-diamine 2 succeeded" — but "due to the high cost of the commercial precursor 8 identified by AiZynthfinder, we started from pyrrolocyclohexanone 5" instead, reaching the target in 7 steps and 16% overall yield. ⊘ And in the other direction: "although AiZynthfinder did not propose a synthesis for the unknown (6,7)- and (7,7)-diamines 3 and 4, the Beckmann rearrangement approach proved suitable in both cases." Two molecules the search called unsolvable were made by chemists who generalised the motif themselves — the sensitivity-0.45 finding above, playing out in one lab.

What the wider CASP field has executed, for calibration

These are other tools and are reported so the AiZynthFinder evidence has a scale. Do not read any of them as AiZynthFinder figures.

• 8 targets planned by Chematica, 8 executed, 8 reported successful. The regime matters more than the count: "No alterations in the proposed routes were allowed save for straightforward adjustments in reactions conditions (e.g., temperature, solvent, specific base, catalyst, etc.) for the sake of optimization." Disconnections frozen; conditions human-optimised. Route selection also applied two human filters that demoted the top-ranked route in 3 of the 8 cases, and six of eight targets were supplied precisely because previous in-house attempts "were very low yielding, not scalable, or altogether failed" — so the baseline was selected to be weak. Execution ran against pre-registered industrial deliverables: each step attempted no more than three to five times, within 8 weeks and 70 hours of bench work.
• 3 natural products planned by Chematica, 3 executed. Same regime: "no alterations to the proposed disconnections were allowed," with conditions adjustable — and in practice one proposed reagent was swapped for a more reactive iodoacetal. An overall yield is reported for only one of the three.
• 15 substances made on a robotic flow platform from 8 ASKCOS routes. Conditions were human-adapted for every target: "Adjustments to the AI-proposed synthetic route required for compatibility with continuous flow are recorded in these reusable recipes," and "we rely on expert input from chemist users, minimally including residence times, equivalence ratios, and concentrations, to translate recommendations into practice." At least one planned target never reached the robot: bezafibrate "was planned by the software, but after expert evaluation and screening was found to be a poor candidate for translation into continuous flow." The paper states no total count of routes planned, so 15 is a count, not a success rate. The authors' own summary: "routes still require human input and process development."
• One API route from Synthia, executed, with named failures. A helicase-primase inhibitor was taken from four CASP tools; Synthia was chosen because it delivers "detailed reaction conditions and recipes." Its suggested conditions then failed repeatedly in the lab: "no product was recovered, and a complex mixture was observed in the NMR spectra"; a copper coupling run without ligand because "Synthia suggested that no ligands can be used" gave "no product formation… the highest yield observed was only 13%"; and suggested coupling agents "both failed to produce the desired product, necessitating the screening of a series of coupling agents." The published route improved overall yield from 8% to 26% across 7 steps, and to 35% once a human removed a step.

⊘ Across every one of these, a chemist stood between the search and the flask. None of them is a measurement of what happens if you act on a route unread.

Shields JD, Howells R, Lamont G, Leilei Y, Madin A, Reimann CE, Rezaei H, Reuillon T, Smith B, Thomson C, Zheng Y, Ziegler RE. AiZynth impact on medicinal chemistry practice at AstraZeneca. RSC Medicinal Chemistry. 2024;15(4):1085-1095. doi: 10.1039/d3md00651d

Hassen AK, Šícho M, van Aalst YJ, Huizenga MCW, Reynolds DNR, Luukkonen S, Bernatavicius A, Clevert DA, Janssen APA, van Westen GJP, Preuss M. Generate what you can make: achieving in-house synthesizability with readily available resources in de novo drug design. Journal of Cheminformatics. 2025;17(1):41. doi: 10.1186/s13321-024-00910-4

Carrel A, Yiannakas A, Roukens JJ, Reynoso-Moreno I, Orsi M, Thakkar A, Arus-Pous J, Pellegata D, Gertsch J, Reymond JL. Exploring simple drug scaffolds from the generated database chemical space reveals a chiral bicyclic azepane with potent neuropharmacology. Journal of Medicinal Chemistry. 2025;68(9):9176-9201. doi: 10.1021/acs.jmedchem.4c02549

Klucznik T, Mikulak-Klucznik B, McCormack MP, Lima H, Szymkuć S, Bhowmick M, Molga K, Zhou Y, Rickershauser L, Gajewska EP, Toutchkine A, Dittwald P, Startek MP, Kirkovits GJ, Roszak R, Adamski A, Sieredzińska B, Mrksich M, Trice SLJ, Grzybowski BA. Efficient syntheses of diverse, medicinally relevant targets planned by computer and executed in the laboratory. Chem. 2018;4:522-532. doi: 10.1016/j.chempr.2018.02.002

Mikulak-Klucznik B, Gołębiowska P, Bayly AA, Popik O, Klucznik T, Szymkuć S, Gajewska EP, Dittwald P, Staszewska-Krajewska O, Beker W, Badowski T, Scheidt KA, Molga K, Mlynarski J, Mrksich M, Grzybowski BA. Computational planning of the synthesis of complex natural products. Nature. 2020;588(7836):83-88. doi: 10.1038/s41586-020-2855-y

Coley CW, Thomas DA 3rd, Lummiss JAM, Jaworski JN, Breen CP, Schultz V, Hart T, Fishman JS, Rogers L, Gao H, Hicklin RW, Plehiers PP, Byington J, Piotti JS, Green WH, Hart AJ, Jamison TF, Jensen KF. A robotic platform for flow synthesis of organic compounds informed by AI planning. Science. 2019;365(6453):eaax1566. doi: 10.1126/science.aax1566

Teixeira RI, Andresini M, Luisi R, Benyahia B. Computer-aided retrosynthesis for greener and optimal total synthesis of a helicase-primase inhibitor active pharmaceutical ingredient. JACS Au. 2024;4(11):4263-4272. doi: 10.1021/jacsau.4c00624

Voinarovska V, Mercado R, Kabeshov M, Genheden S. Do humans and large language models agree on the quality of synthesis plans? Digital Discovery. 2026;5(9):3648-3660. doi: 10.1039/d6dd00266h

Guo Y, Kabeshov M, Le THD, Genheden S, Bergonzini G, Engkvist O, Kaski S. An Expert-Augmented Deep Learning Approach for Synthesis Route Evaluation. Research Square. 2025. doi: 10.21203/rs.3.rs-6106987/v1

Segler MHS, Preuss M, Waller MP. Planning chemical syntheses with deep neural networks and symbolic AI. Nature. 2018;555(7698):604-610. doi: 10.1038/nature25978

Kutchukian PS, Lou D, Shakhnovich EI, et al. Inside the mind of a medicinal chemist: the role of human bias in compound prioritization during drug discovery. PLoS One. 2012;7(11):e48476. doi: 10.1371/journal.pone.0048476

Choung OH, Vianello R, Segler M, Stiefl N, Jiménez-Luna J. Extracting medicinal chemistry intuition via preference machine learning. Nature Communications. 2023;14(1):6651. doi: 10.1038/s41467-023-42242-1

Sadowski M, et al. Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers. 2025. doi: 10.48550/arXiv.2510.10645

Genheden S, Thakkar A, Chadimová V, Reymond JL, Engkvist O, Bjerrum E. AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. Journal of Cheminformatics. 2020;12:70. doi: 10.1186/s13321-020-00472-1

Saigiridharan L, Hassen AK, Lai H, Torren-Peraire P, Engkvist O, Genheden S. AiZynthFinder 4.0: developments based on learnings from 3 years of industrial application. Journal of Cheminformatics. 2024;16:57. doi: 10.1186/s13321-024-00860-x

Thakkar A, Kogej T, Reymond JL, Engkvist O, Bjerrum EJ. Datasets and their influence on the development of computer assisted synthesis planning tools in the pharmaceutical domain. Chemical Science. 2020;11(1):154-168. doi: 10.1039/c9sc04944d

Blackshaw TM, Davies JC, Spoerer KT, Hirst JD. Enhancing Monte Carlo Tree Search for Retrosynthesis. Journal of Chemical Information and Modeling. 2025;65(13):6537-6546. doi: 10.1021/acs.jcim.5c00417

Westerlund AM, Barge B, Mervin L, Genheden S. Data-driven approaches for identifying hyperparameters in multi-step retrosynthesis. Molecular Informatics. 2023;42(11):e202300128. doi: 10.1002/minf.202300128

Shields JD, Howells R, Lamont G, Leilei Y, Madin A, Reimann CE, Rezaei H, Reuillon T, Smith B, Thomson C, Zheng Y, Ziegler RE. AiZynth impact on medicinal chemistry practice at AstraZeneca. RSC Medicinal Chemistry. 2024;15(4):1085-1095. doi: 10.1039/d3md00651d

Westerlund AM, Saigiridharan L, Genheden S. Human-guided synthesis planning via prompting. Chemical Science. 2025;16(32):14655-14667. doi: 10.1039/d5sc00927h

Westerlund AM, Sigmund LM, Mijangos MV, Kannas C, Genheden S, Kabeshov M. Toward Lab-Ready AI Synthesis Plans with Protection Strategies and Route Scoring. Journal of Chemical Information and Modeling. 2026;66(11):6361-6375. doi: 10.1021/acs.jcim.6c01147

What the figure doesn't cover

• ⊘ "Solved" is defined by the configuration, and this configuration is the weak end of the published range. One expansion policy (uspto), one stock (zinc), 100 iterations, a 120-second ceiling. AstraZeneca's production setup, which produced the 80–86% figures, runs "three expansion models in parallel: the general template-based model, the RingBreaker, and a reaction look-up function" against an internal stock. Neither the extra policies nor the internal stock exists here, and chat runs cannot raise either limit, though a direct job request can.
• ⊘ The stock is a fragment set, not a general catalogue. The ZINC stock that ships with AiZynthFinder's public data was built by downloading "tranches including fragment compounds (molecular weight up to 250 D and log P up to 3.5) that had reactivity labeled as 'standard' or 'reactive', resulting in 17,422,831 compounds" — and an independent group using the public bundle in 2025 confirms the same figure, "17,422,831 Zinc building blocks provided by AiZynthfinder." A perfectly orderable advanced intermediate outside that molecular-weight window counts as not in stock, which turns a makeable molecule into an unsolved one. The AiZynthFinder 4.0 authors measured the cost of that narrowness directly: eMolecules "covers on average 80% of the starting material, whereas ZINC stock only covers 60% on average."
• Training overlap is measurable here, and it is large for famous molecules and zero for novel ones. Comparing target sets against the products in each reaction dataset, the top 125 pharmaceuticals by 2018 sales overlapped USPTO 1976–2016 at 47%, Pistachio at 58%, Reaxys at 70% and Reaxys + patents at 78% — while AstraZeneca's virtual-library designs overlapped every one of them at 0%. A solve rate measured on known drugs is partly a lookup. If your molecule is novel, the GDB row (10%) and the heavy-molecule row (9.5%) describe your situation better than the ChEMBL row (71–76%).
• ⊘ Single-step accuracy and route-finding ability move in opposite directions. On the same models and the same target sets, "the top-1 accuracy on the test set is not reflective of overall model performance and increases with template specificity. In contrast, the overall performance of the model decreases with increased template specificity." So a reported top-k retrosynthesis accuracy is not evidence about solve rate, and the literature-reaction test sets those top-k numbers come from overlap the training templates by construction.
• The template library is almost entirely a function of which reaction dataset you chose. Across USPTO, Pistachio, Reaxys and an AstraZeneca ELN subset, "only 2% of the extracted templates are common between all datasets." Separately, of the top-50 templates the policy proposes — the exact cutoff configured here — "only ca. 34% of the top 50 templates are applicable on average in the best case."
• ⊘ The score is a search reward, not a feasibility estimate. AiZynthFinder's authors: the score "reflects the fraction of solved precursors and the number of reactions required," and "was designed to support the tree search and is rather indiscriminate with regard to the quality of the route (i.e. if it's a good route or not) and should be interpreted with care." Treat the platform's own outputs guidance that a score above 0.8 marks a high-confidence route as a statement about search progress only. In any case the score is not populated in a platform result — see known defects.
• ⊘ No conditions, no reagents, no yields, no selectivity. A route here is a sequence of disconnections. "The conditions required to carry out the reaction in the forward direction are not predicted." That is a difference between tools, not a universal limit: a 2024 API campaign chose Synthia over three competitors specifically "due to its ability to deliver more comprehensive reaction pathways, accompanied by detailed reaction conditions and recipes." AiZynthFinder does not emit those. ⚠︎ And conditions are no guarantee either — in that same campaign, Synthia's suggested conditions gave "no product… and a complex mixture was observed in the NMR spectra" on one step and "no product formation was observed" on another.
• ⊘ Speed was traded for depth, deliberately, and this platform takes the fast end of that trade. AstraZeneca's chemists say it directly: "AiZynthFinder trades depth for speed and the results sometimes suffer accordingly," and they attribute the tool's failure to reproduce a chemist's own route "in part to our self-imposed requirement to run fast searches." Their production median is under three minutes. This platform is capped tighter — 100 iterations and 120 seconds — and neither limit is reachable from chat.
• ⊘ A solved route is not a route a chemist would accept. Running AiZynthFinder's production template model and MCTS against 1,748 targets from PaRoutes set-n1 with an eMolecules stock, AstraZeneca found that while "the standard search solved most targets (84.91%)... it only solved and satisfied the bond constraints for 54.80% of the targets" — where the bond constraints encode which bonds a chemist wants disconnected or preserved. Roughly 30 points of that solve rate are routes that a chemist's stated disconnection preferences would reject.
• The version here is ahead of every published figure. The platform pins AiZynthFinder 4.4.1. The author-reported figures above are for 4.0, the independent ones for 3.7 and 4.3, and the ASKCOS comparison for the 2020 release. No figure on this page was measured on 4.4.1.
• Route comparison is an open problem, by the developers' own account. "It is still inherently difficult to compare output from retrosynthesis experiments on a large scale, and improvements are often judged on a case-by-case basis."
• Nothing on this platform has been validated for AiZynthFinder. This page is literature only. No figure here was measured on a platform run.

Known issues on this tool

These are known issues, and each one changes how you must read a result.

⊘ "Solved" means a route came back, not that you can buy its starting materials. The top-level flag reports only that the search returned at least one route. AiZynthFinder's own per-route verdict — the one that actually asks whether every leaf is in stock — is carried inside each entry of the routes array, so read it there. A fix for the top-level flag is pending.

The consequence: the stock-coverage numbers that would let you check a solved claim do not render on the result card — in-stock precursor counts, route score and step count come back empty. So "solved" is a claim you currently cannot verify from the result alone; open the route tree and read the leaves.

⚠︎ And the stock list itself can be wrong, which presents as bad chemistry. In the 17-chemist grading study, 8.0% of expert votes flagged an "Unnecessary step," and on inspection "for the majority of the cases AiZynthFinder generated compounds that were supposed to be in stock but were tagged as not in stock in the stock solutions file" — the search kept pulling apart a molecule it should have stopped at. If a route takes an obviously orderable intermediate one step too far, suspect the stock list before the templates.

Go / no-go

Go — act on the result when:

• You are triaging a design set and want to deprioritize the molecules no public template library can disconnect. A solved/unsolved split over a set is the use this tool supports best, and it is why GenMol and REINVENT point here.
• You want disconnection ideas for a target you have already committed to making, and a chemist is going to read them. This is the developers' own framing of what the tool does well today.
• Your target is drug-sized and resembles published medicinal chemistry, so the 71–76% ChEMBL regime applies rather than the 9.5% heavy-molecule regime.
• You read the per-route is_solved inside the routes array rather than the top-level flag, and you open the route trees to see which leaves the search actually reached.
• You treat an unsolved result as "no route on this template library and this fragment-sized ZINC stock," which is a statement about the configuration — not as "this molecule cannot be made."

No-go — do not act on the result when:

• You are reading is_solved as a makeability verdict. Against real synthesis records, template models over-predict makeability for novel combinatorial designs and under-predict it for known drugs.
• You are ranking routes by score. The score is the search's own reward, its authors call it "rather indiscriminate with regard to the quality of the route," and on this platform it is not populated at all.
• You are quoting 80–86% for this platform. That figure belongs to a proprietary template model and an internal stock that do not run here.
• You are handing a route to a CRO or ordering reagents without a chemist reading it. The route carries no conditions, no yields and no selectivity analysis, and AiZynthFinder trees are documented to contain competing reactive sites that need a protection strategy the search did not plan.
• You need a rate to plan against for whether a solved route will actually run. AiZynthFinder routes have reached the bench, but only as selected case studies with no denominator — and in every published case a chemist altered the proposal first. No prospective campaign with a success rate exists, so the honest answer is that nobody has measured it.
• You are comparing two runs made with different stocks, policies or budgets. Those numbers are not comparable, and the spread from configuration alone is larger than the spread between tools.

Caveats on reading this page: every figure above is quoted from the cited source with its original metric and sample, and each is labelled author-reported or independent, and body-text or abstract-only. Figures from different studies are not comparable, because template libraries, stock sets, search budgets and target sets differ between them — which on this page is the finding rather than a caveat. Where a figure for AiZynthFinder does not exist, this page says so rather than substituting one from another tool or version.