Paper deep dive

Strategy-first synthesis planning for complex natural products

Daniel Armstrong, Xuan-Vu Nguyen, Octavian Susanu, Gabriel Gibberd, Théo A. Neukomm, Taddäus Strunden, Dan Forster, Morgane Delattre, Shawn Teh, Clément Rols, John Federice, Hayden Leatherwood, M. Lavelle Barnes, Maarten R. Dobbelaere, Peter Wipf, Jon T. Njardarson, Jieping Zhu, Philippe Schwaller

Year: 2026Venue: arXiv preprintArea: cs.MAType: PreprintEmbeddings: 118

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:The total synthesis of a complex molecule is among the most demanding intellectual and experimental feats in chemistry: a chemist must plan many steps ahead for how to assemble simple building blocks into an intricate target, devise backup strategies, and anticipate procedural challenges. It is also a profoundly creative activity. For half a century, efforts to automate the retrosynthetic design of natural products and other complex molecules have drawn on catalogued reactions, and the resulting tools now report near-complete success on benchmarks built from that same source. But these tools were shaped to fit benchmarked chemistry, and they falter on many natural products, the frontier of the field, whose densely functionalized, polycyclic architectures demand precisely the inventive chemistry the record contains least. Whether a machine could reasonably design such syntheses like an expert chemist does has remained unclear. Here, we show that SynthEx, an agentic framework built on large language models, plans routes to complex natural products that lie beyond the reach of conventional design algorithms. SynthEx proposes competing strategies, assembles a sequence of routine and key steps into a cohesive route, and critiques and improves its own design; the chemistry it favours is more convergent than existing tools produce, and spans a region of reaction space that catalogue-based tools cannot match. Most notably, in blinded assessments, expert chemists judged its key steps comparable to those of published human syntheses and engaged with them as genuine synthesis plans, a response algorithmic route prediction has not previously accomplished. We release routes to more than a thousand natural products as SynthAtlas, an open, interactive database, and anticipate it will become a shared resource for a collection of complex target molecules that lack existing literature routes.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

117,088 characters extracted from source content.

Expand or collapse full text

Strategy-first synthesis planning for complex natural products Daniel Armstrong 1† , Xuan-Vu Nguyen 1† , Octavian Susanu 1 , Gabriel Gibberd 1 , Th ́eo A. Neukomm 1 , Tadd ̈aus Strunden 1 , Dan Forster 2 , Morgane Delattre 2 , Shawn Teh 2 , Cl ́ement Rols 2,3 , John Federice 5 , Hayden Leatherwood 5 , M. Lavelle Barnes 6 , Maarten R. Dobbelaere 1,4 , Peter Wipf 6 , Jon T. Njardarson 5 , Jieping Zhu 2,3 , Philippe Schwaller 1,3* 1 Laboratory of Artificial Chemical Intelligence (LIAC), ́ Ecole Polytechnique F ́ed ́erale de Lausanne (EPFL), Lausanne, Switzerland. 2 Laboratory of Synthesis and Natural Products (LSPN), ́ Ecole Polytechnique F ́ed ́erale de Lausanne (EPFL), Lausanne, Switzerland. 3 National Centre of Competence in Research (NCCR) Catalysis, Lausanne, Switzerland. 4 Laboratory for Chemical Technology, Ghent University, Zwijnaarde, Belgium. 5 Njardarson Laboratory, University of Arizona, Tucson, United States. 6 Wipf Group, University of Pittsburgh, Pittsburgh, United States. *Corresponding author(s). E-mail(s): philippe.schwaller@epfl.ch; Contributing authors: daniel.armstrong@epfl.ch; nguyen.nguyen@epfl.ch; pwipf@pitt.edu; njardars@arizona.edu; jieping.zhu@epfl.ch; † These authors contributed equally to this work. Abstract The total synthesis of a complex molecule is among the most demanding intellec- tual and experimental feats in chemistry: a chemist must plan many steps ahead for how to assemble simple building blocks into an intricate target, devise backup strategies, and anticipate procedural challenges. It is also a profoundly creative activity. For half a century, efforts to automate the retrosynthetic design of nat- ural products and other complex molecules have drawn on catalogued reactions, 1 arXiv:2608.07454v1 [cs.MA] 7 Aug 2026 and the resulting tools now report near-complete success on benchmarks built from that same source. But these tools were shaped to fit benchmarked chem- istry, and they falter on many natural products, the frontier of the field, whose densely functionalized, polycyclic architectures demand precisely the inventive chemistry the record contains least. Whether a machine could reasonably design such syntheses like an expert chemist does has remained unclear. Here, we show that SynthEx, an agentic framework built on large language models, plans routes to complex natural products that lie beyond the reach of conventional design algorithms. Freed from the fixed reaction libraries that confine those planners, SynthEx proposes competing strategies, assembles a sequence of routine and key steps into a cohesive route, and critiques and improves its own design; the chemistry it favors is more convergent and elegant than existing tools pro- duce, and spans a region of reaction space that catalogue-based tools cannot match. Most notably, in blinded assessments, expert chemists judged its key steps comparable to those of published human syntheses and engaged with them as genuine synthesis plans, a response algorithmic route prediction has not previ- ously accomplished. We release routes to more than a thousand natural products as SynthAtlas, an open, interactive roadmap, and anticipate it will become a shared resource for a collection of complex target molecules that lack relevant current literature designs. Keywords: Synthesis Planning, LLMs, Agentic Scientific Discovery, Total Synthesis, Multiagent systems 1 Main The total synthesis of natural products is among the most challenging and creative endeavors in chemistry. Faced with an intricate molecular architecture and complex stereochemical relationships, a chemist develops a reaction-by-reaction recipe (syn- thetic route), deciding which bonds to forge and in what order, when to introduce and remove protective groups, and how to control the three-dimensional outcome of every step to create the target molecule. Corey’s retrosynthetic analysis gave this reasoning a formal language, reducing a formidable target to stepwise disconnections, resulting in a sequence of established, feasible transformations [1–4]. The discipline has repaid the effort many times over: total synthesis is where new reactions prove their worth and many, now-standard, transformations trace their origins to the assembly of a nat- ural product [5–7]. Yet designing a route to a demanding target remains an art that only a small community of experts practice fluently [8], and it is that expertise, not the chemistry it draws on, that proves hardest to articulate and imitate. For half a century, chemists have sought to automate synthesis planning. Computer-Assisted Synthesis Planning (CASP) began with Corey and Wipke in the 60s [9–11], with manually coded reaction logic, before shifting to reaction tem- plates—deterministic rules mined at scale from literature and patents [12–14]. Trained retrosynthesis models can now propose multistep routes in minutes, and on standard 2 benchmarks drawn from these same corpora, they succeed impressively [15–24]. How- ever, this success reflects the shared distribution of the benchmarks and training data, which are dominated by the recurring, routine transformations of streamlined medic- inal and process chemistry. The highly context-specific disconnections that complex targets demand are precisely what these tools struggle to capture [25–27]. This limitation has become visible in two ways. First, the benchmarks themselves have saturated: on the patent-derived test sets that the field has long used, state- of-the-art planners now report success approaching completeness. Like early coding benchmarks that saturated before the field shifted to agentic software engineering tasks [28, 29], existing synthesis benchmarks no longer probe the axes on which model capa- bilities are improving [24, 30–33]. Second, and more tellingly, synthetic planning tools falter the moment they leave that chemical space [34–36]. Complex natural products, with their densely fused, stereochemically rich architectures, are out-of-distribution for these tools. This weakness is likely structural to the reaction rule approach itself. Rules, whether as explicit templates, or implicit in neural network weights, rely on applying memorized patterns from frequently occurring reactions. Natural product synthesis, however, routinely requires bespoke, inventive transformations—such as complex ring- closing cascades or intricate structural rearrangements—that depend entirely on the specific, global context of the molecule. Because these reactions occur too rarely to form robust templates [25], or are too rare to be recalled by a neural network [24, 26], a planner confined to a historical reaction library fundamentally struggles to navi- gate the unique chemical environments of natural products. The field has accordingly moved past judging a planner on whether it can successfully retrieve a pathway, to whether a chemist would actually use the resulting route [37]. Large language models (LLMs) suggest a way past this limitation. Having absorbed, through pretraining, a vast body of chemical knowledge and context that extends well beyond any single reaction database, they can act less as retrieval engines than as chemical reasoning engines [38–44]. Bran et al. showed that an LLM can judge a synthesis, aligning a chemist’s stated strategy with candidate routes and rank- ing them in agreement with human experts [8]. Frameworks such as LARC [45] and MMORF [46] use LLMs as evaluators, selecting and scoring nodes against user-defined criteria while a traditional planner expands the search space. Synthelite [47] demon- strates that an LLM can directly plan, casting the model as a language-based policy prior that proposes each disconnection in natural language and so steers the tree search directly, in place of a trained reaction policy network; these proposals, how- ever, are still grounded by retrieval against a fixed template library. Each advances the role an LLM can play, yet all share one ceiling: whether the model ranks syn- thetic strategy or tactics or proposes the next step, it selects within a reaction space a conventional planner has already fixed, and has no ability to instantiate reactions outside that space. This is what makes complex-molecule synthesis the most relevant available test of chemical reasoning: because the required disconnections are, by con- struction, decoupled from any reaction database, success is not limited to reaction retrieval, and a system that plans these routes must reason about reactivity, strategy, and stereocontrol rather than recall them through a template. Planning a synthesis in 3 this way requires the planner to adapt chemistry principles to any given reaction and is a necessary precondition for the automation of chemical reasoning. Here we introduce SynthEx, an agentic synthesis planner that removes this con- straint at its source (Fig. 1). Rather than selecting disconnections from a fixed template library, SynthEx expresses each as an ordered list of atom-level graph edits, a format we call ReactionJSON. By letting the language model write these edits directly, Syn- thEx transitions from a critic that ranks known templates to a policy that generates novel expansions. Around this representation, SynthEx builds the iterative reasoning of a human chemist: it proposes competing high-level strategies, expands each into a complete pathway, and critiques and repairs its own chemistry step by step (Fig. 1a). Applied to a collection of over a thousand documented natural products, Syn- thEx opens a region of synthesis space that a state-of-the-art search does not reach. It designs routes for the large majority of targets for which a leading template-based planner solves only a small fraction, and its advantage widens, rather than narrows, as targets grow more complex: the additional rings, connectivities, and functional groups that cause a conventional search to fail have a lower impact on SynthEx. Its bond con- structions are also markedly more convergent than those existing tools propose, most of them uniting two independent fragments, and they occupy a distinct reaction space that state-of-the-art single-step models rarely propose [12, 24, 25, 48]. This is not the same chemistry produced in new contexts but different chemistry altogether: SynthEx favors constructive, bond-forming steps, ring constructions and cross-couplings, over the functional group manipulations that dominate template-based output, a qualita- tive signature of SynthEx’s strategic preferences. None of the benchmark targets are flagged as synthesized in NPAtlas, so the routes SynthEx proposes are not retrievals of existing literature. In blinded review, expert chemists rated SynthEx’s key steps on par with published human syntheses and, additionally, singled out individual discon- nections, such as Grob fragmentation, [3+2] dipolar cycloaddition, or intramolecular cascade Michael additions, as genuinely elegant (Fig. 1d). We release these routes as SynthAtlas (Fig. 1c), an open, interactive resource comprising 1,098 natural-product targets, 3,243 routes with corresponding synthetic strategies, and 33,145 reactions. The focus of this work is on reach, convergence, distinctiveness, and strategy of proposed syntheses. We regard wet-lab feasibility as the next frontier in CASP, with the field moving from graph-based solve rate toward establishing proven chemical validity. To contribute to the automation of scientific discovery, every route is released in SynthAtlas, an open, interactive platform on which chemists can explore, compare, and comment on them. We envision that this atlas of predicted synthetic routes has the potential to spark ideas and discussions, and radically alter the way synthetic chemistry is approached, in analogy to how predicted protein structures transformed how biologists reason about molecules they may never crystallize [49]. 2 Results 2.1 SynthEx: a multi-agent system for synthesis planning SynthEx plans a synthesis the way a chemist does: it settles on a strategy before it commits to a route, and it revises the route without abandoning the strategy. Five 4 HNHN O - ON + O O Br N O O - ON O O Br O O Br O N O N + O - O N O Br O O N O - O N O O N O N + O - O N O O O N + O - O N Intramolecular cascade aza-Michael/Michael additionIntramolecular cascade aza-Michael/Michael addition O N O O Br O N O N O O Br O NHO Chiral [3+2] Dipolar Cycloaddition N O N O Stereocontrolled multicomponent [3+2] azomethine ylide cycloaddition to form the second pyrrolizidine core. Stereo controled by chiral-bisoxazoline catalyst Reductive ketyl-olefin radical cyclization to form the central fused cyclopentanes before a Grob fragmentation to form the final cyclooctene ring system Route analysis: Excellent step economy for the skeletal rearrangement O HO HO HO O Ketyl-olefin radical cyclization and Grob fragmentation Elegant multistep strategic planning Mechanistically complex asymmetric catalysis One-pot cascade reactions 1. STRATEGY GENERATOR a. SynthEx agentic synthesis planning pipeline Proposing key disconnections STRATEGY GENERATOR Tandem aza-Cope rearr. / Mannich Fischer indolisation Intramolecular Diels-Alder Expert chemist’s strategy (optional) 2. ROUTE BUILDER Building the full route for every strategy ROUTE BUILDER ReactionJSON, template-free reaction representation N H N H H Hydrogenation Heck coupling N H N H H N H HN HCHO + N H N H I 3. VERIFY · ITERATE 3. ROUTE CRITIC - EDITOR Iterative refinement b. LLM-based route analytics ANALYST Classifying route quality Analysis agent 12345678910 0 200 400 600 800 1000 Number of r outes Route feasibility analysis Feasibility Score c. Public database of synthesis routes d. Representiative targets and key steps SynthAtlas Open interactive resource VERIFIED ROUTE → ANALYSIS Key step Strategy-based step-wise route Supporting N H N H H Route target Melonine Abstract conditions Potential risks CRITIC flag problematic reactions EDITOR implement changes R1 R2 R3 Reorder R1 and R2 Add a protecting step P4 R2 R1R3 R2 R1R3 P4 R2 R1R3 P4 SynthAtlas HN O N O O N O O H N O Intramolecular cascade aza-Michael/Michael addition to construct the core indolizidine and cyclopentane rings in a single reaction, forming multiple bonds and stereocenters simultaneously. Abstract conditions: Polar aprotic solvent and mild base at ambient to elevated temperatures Dibohemamine A Traversiadiene Cyclopiamine B 1,098 natural products 3,243 synthetic strategies 33,145 reaction steps Up to three unique routes for each target Text descriptions and condition annotation HO O O OH Molecule Problematic step Improved / added step Tandem aza-Cope rearr. / Mannich Fig. 1: The SynthEx agentic synthesis-planning pipeline and the SynthAtlas resource. a, The planning pipeline has three stages with corresponding subagents. First, the Strategy Generator proposes several high-level strategies for the target molecule, each cen- tered on a key disconnection. In the second stage, the Route Builder expands each strategy into a pathway, expressing disconnections in ReactionJSON, a template-free representation in which reactions are specified as graph-edit operations; key and supporting steps are dis- tinguished. Next, the Critic flags problematic steps (pink) and an Editor applies edits to the route (orange), in an iterative loop that leaves the overall strategy intact. b, Verified routes are passed to an LLM-based analysis stage that classifies route quality, flagging potential risks and key reaction steps, and scoring feasibility; the histogram shows the distribution of route feasibility scores across the corpus. c, The routes are released as SynthAtlas, an open, interactive resource built over NP-Atlas targets. d, Representative key steps produced by SynthEx for the natural products Traversadiene, with a two-step cyclization followed by Grob fragmentation, Dibohemamine A, an asymmetric dipolar cycloaddition, and Cyclopi- amine B, a cascade aza-Michael/Michael addition to close two rings in a single step. 5 stages, each driven by a dedicated language-model agent, implement this (Fig. 1a). The Strategy Generator agent proposes diverse high-level strategies for the target, each anchored on a key disconnection. A Route Builder expands every strategy into a complete pathway, expressing each disconnection in ReactionJSON, a template- free representation that we introduce here in which a reaction is an ordered list of atom-level graph edits. A Critic then simulates each reaction in the forward direction and flags those that are chemically not feasible, and an Editor repairs them through surgical edits that leave the strategy intact. Finally, an Analyst scores the finished route for feasibility and identifies its key steps and risks (Fig. 1b). The architecture descends from Synthelite [47], which treats a language model as the search policy over a fixed template library while delegating single-step feasibility to that same library. The first stage emphasizes strategic planning before reaction searching. Rather than commit to one disconnection, the Strategy Generator returns a diverse set of competing strategic hypotheses, three per target by default, each built around a key step that the downstream stages then test and prune. This mirrors the way a chemist entertains several plausible routes before investing in one, and it frames synthesis planning as hypothesis generation followed by exploration rather than as a single-shot search. Strategies can also be steered by constraints expressed in natural language, a required starting material, or a free-text instruction from a chemist, although for this work we leave the choice to the model. For the second stage, we design ReactionJSON, a reaction representation the model authors itself. Earlier LLM planners choose among entries in a template library, which serves complex reactions in unfamiliar structural contexts poorly; the obvious alter- native, having the model generate reaction SMILES directly, is unreliable [8, 50]. ReactionJSON avoids both bottlenecks: the model neither names a transformation nor redraws a molecule, but writes the ordered graph edits that convert a product into its precursors; applying those edits to the mapped product yields the precursors deterministically (Methods 4.2). This converts the language model from a subroutine that selects from a limited number of available steps to a designer, and in doing so lifts the search off the reaction frequency and similarity precedence that limits every template-based and corpus-trained planner alike: SynthEx can propose chemistry that is common in the named-reaction and total-synthesis knowledge a model has absorbed but rare in any reaction catalogue, and therefore invisible to tools trained on one. A practical artifact of ReactionJSON is that a route can be rendered as an editable object. Because a ReactionJSON route is a text object anchored on atom maps, it can be revised in place: the Critic and Editor can critique and repair the chemistry sur- gically without re-running the search that produced it: reordering reactions, inserting protections, or replacing a disconnection (Section 2.5) without re-running the search that produced it (Section 2.5). A repair loop of this kind would be impractically expensive if each edit required a full tree search, and unreliable if SMILES needed regenerating by an LLM. This flexibility is what allows the system to genuinely correct its own chemistry rather than merely rank it. For every target SynthEx returns several distinct strategies, each carrying step- and route-level reasoning. Before quantifying reach across the full benchmark, we examine three demanding molecules in detail. 6 2.2 Strategic synthesis planning and reasoning A solve rate cannot convey the essence of what a route actually proposes. Before turning to benchmark-scale reach (Section 2.4), we examine the chemistry itself on three targets, ordered by how much external validation is available against which to check the proposals.. For Okaramine M a route was published after the model’s training cut-off (Methods 4.2), and SynthEx recovers it. Melonine presents a harder test: a total synthesis was likewise published after the cut-off, but here SynthEx does not converge on it, proposing instead an alternative disconnection that one of us (J.Z.), whose group has worked on this target, judges feasible and worthy of experimental validation. For the conversion of Chanoclavine to Lysergol, we are aware of no reported route across this specific gap: it is an open problem, posed here by an author working on it, on which the only available check is expert judgment of a novel proposal. In all three, what matters is not merely whether a route is found but whether the chemistry withstands a synthetic chemist’s scrutiny. 2.2.1 Okaramine M: recovering an expert route published after the training cut-off Okaramine M (Fig. 2a) is proposed to be a key intermediate in the biosynthesis of the Amauromine class of natural products [51], whose members exhibit, among oth- ers, vasodilating activity [52] and anticancer activity [53]. Okaramine M and other Amauromine-class natural products have been synthesized from the TIPS-protected derivative of Okaramine M [51], demonstrating the synthetic value of this intermedi- ate, which we selected as our target. This example demonstrates that SynthEx can reproduce a route conceived by expert organic chemists and experimentally validated. The first part of SynthEx’s strategy forms the TIPS-monoprotected derivative of pre-Okamauromine through standard protection, condensation and deprotection reactions. The key step then involves the tandem prenylation of the pre-Okamauromine derivative and formation of the hexahydropyrroloindole core: a prenyl cation adds at the C3 position of the unprotected indole through electrophilic addition, followed by trapping of the resulting iminium cation by a nitrogen atom of the diketopiperazine moiety. In proposing this step, SynthEx displays both a capacity for elegant chemistry and a grasp of subtle mechanistic considerations: it correctly identifies the C3 position of the indole as more nucleophilic than C2, and further recognizes that TIPS protection at the indole nitrogen lowers the nucleophilicity of its C3, so that electrophilic addition is directed selectively to the C3 of the unprotected indole. The synthesis is completed by N-acylation of the indole. Compared with the reference route [51], SynthEx follows it closely: it captures the key tandem prenylation and iminium-trapping step, and it forms pre-Okamauromine through the stepwise condensation of two distinct tryptophan precursors rather than a direct dimerization of tryptophan, which would give the diketopiperazine derivative in low yield [51]. In one respect SynthEx departs from the published sequence, and the departure may be an improvement. The literature route installs the TIPS group on pre- Okamauromine, whereas SynthEx protects from the first step, which would avoid 7 a. A literature-matched route for TIPS-protected L-exo/endo-Okaramine M b. A creative approach for (±) Melonine c. Lysergol semisynthesis from Chanoclavine SynthEx independently recovers a route published after its training cut-off. L-exo/endo-Okaramine M TIPS-protected L-exo/endo-Okaramine M TIPS-protected pre-Okamauromine Input (target) Input (target) (±) Melonine Tandem aza-Cope / Pictet-Spengler Prenylation / Iminium cation trapping Input (target) Lysergol Input (starting material) Chanoclavine SynthEx’s intermediate Proposed biomimetic intermediate SynthEx converges on a disconnection an expert group attempted, by a different path they judge more likely to work. Given only a starting material and a target, SynthEx designs the elegant Hofmann-Löffler-Freytag disconnection. SynthEx-generated strategy: “Utilize a Hofmann-Löffler-Freytag reaction... Arrive at Chanoclavine C/C(=C\[C@H]1[C@@H] (C2=CNC3=C=C1=C23)NC)/CO as a starting material.” + Hoffmann-Löffler-Freytag Tandem aza-Cope / Pictet-Spengler Tandem prenylation / Iminium trapping Fig. 2: Three case studies in strategic reasoning. a, Okaramine M, with the mechanism of the tandem prenylation and iminium trapping inset. b, Melonine, with the mechanism of the tandem aza-Cope / Pictet–Spengler key step. The gray inset compares the biomimetic pre-Mannich intermediate proposed in the literature with SynthEx’s. The hydrogens causing potential steric clash with the piperidine ring are highlighted in red. c, Chanoclavine to Lysergol; the two structures shown were the entire input to the planner, which proposed the Hofmann–L ̈offler–Freytag sequence unprompted. Colored atoms and bonds highlight the reaction center in the key step of each route. 8 the mixture of mono- and bi-protected pre-Okamauromine that late-stage protection invites. We have not tested this, so it stands as a proposal rather than a result; but it is a substantive variation on the expert route, not a paraphrase of it. That SynthEx recovers the strategic logic of an experimentally validated synthesis published online in June 2025, well after the model’s training cut-off, and offers a defensible refinement of it, illustrates a capacity for retrosynthetic reasoning rather than recall. Two limits of this comparison should be stated. The target is the TIPS-protected intermediate rather than Okaramine M itself, chosen because the Amauromine-class syntheses proceed through it; and recovery of the strategy was judged by inspection of the two routes, not by execution. What the case shows is that SynthEx reconstructs expert strategic logic it cannot have seen, not that its route would perform as written. 2.2.2 Melonine: converging on the same key disconnection Melonine is a pentacyclic monoterpene indole alkaloid [54]: an indoline fused to a quinolizidine, carrying the 2,2,3-trisubstituted motif characteristic of the schizozygane alkaloids [55] and a two-carbon bridge between the C2 and C20 positions. Its con- gested, bridged architecture has made Melonine a challenging target, and two total syntheses have been reported to date [56, 57]. Both postdate the model’s training cut-off, Yokoshima’s route appearing online in February 2025 and Zhu’s route in 2026. SynthEx’s strategy is displayed in Figure 2b. It begins by iodination of Boc- protected tryptamine followed by deprotection. The next steps involve a reductive amination and an intramolecular Heck coupling. The key step is a highly elegant tandem reaction: the condensation of the secondary amine in the substrate and formaldehyde enables an aza-Cope rearrangement to take place; the newly formed iminium undergoes a Mannich reaction with the nucleophilic C3 position of the indole, followed by a migration of an alkyl chain from the C3 to the C2 position, these last two steps mimicking a Pictet–Spengler reaction; the cascade ends with a reduction, unlike a classical Pictet–Spengler that would restore aromaticity. It is to note that a cyclization through Mannich reaction is also proposed in the biosynthesis of Melonine, starting from a reduced form of the intermediate that SynthEx employs [54]. Finally, SynthEx reaches the target Melonine by selective hydrogenation. SynthEx’s strategy differs from Melonine’s two published total syntheses: Yokoshima and co-workers have as key steps an oxidative intramolecular aziridination followed by nucleophilic attack of an arylamine [56] while Zhu and co-workers employ as key steps a bis-cyclisative diamination followed by a lactamization [57]. Zhu and co-workers moreover attempted a route along the proposed biosynthetic pathway, cen- tered on exactly the Mannich cyclization that SynthEx selects as its key step. That attempt failed, and the reason is conformational: in their substrate the conformation required for cyclization suffers a severe steric clash between the piperidine ring and the C–H bonds of the CH 2 CH 2 linker, so the iminium cannot adopt a geometry from which the indole can reach it. Full experimental detail is in the Supporting Information of the work [57]. One of us (J.Z.) led those experiments, which allows the comparison with SynthEx’s route to be made directly rather than inferred. The distinction is the linker. Because SynthEx generates its iminium through an aza-Cope rearrangement rather than by 9 condensation, the CH 2 –CH 2 single bond of the failed substrate is replaced by an HC=CH double bond. That change removes the clash, and the resulting intermediate should reach a reactive conformation considerably more easily. A similar cyclization on a related substrate with similar reactive conformation has been realized by the same group [58]. On present evidence, SynthEx’s route is therefore a live experimental proposal rather than a repeat of a known failure, and it is one we intend to test. The convergence is worth separating from the outcome. Given only the structure, and with no published route to Melonine available to it, SynthEx identified the same disconnection that an expert group judged worth committing laboratory effort to. Agreement with an idea experts chose to test is a demanding standard because it is independent of whether the idea worked. That SynthEx then reached the same disconnection by a route which, on the assessment of the group that ran the original experiments, may succeed where theirs did not, is a stronger result than convergence alone. 2.2.3 Chanoclavine to Lysergol: planning from an advanced intermediate A synthetic campaign does not always begin with simple commercial materials. A group may hold a hard-won advanced intermediate and be unable to close the remain- ing gap to the target, in which case the useful question is not how to make the molecule but how to cross the finishing line from what is already accomplished. SynthEx accepts a required starting material as a constraint on strategy generation (Section 2.1, Meth- ods 4.2), which turns it on that problem directly. We tested this on the unprecedented conversion of Chanoclavine, a tricyclic ergot alkaloid that has been produced on scale through a combination of synthesis and bioengineering [59], into Lysergol, which requires forming the D ring of the ergoline skeleton (Fig. 2c). Given only the two structures of Chanoclavine and Lysergol, the Strategy Gener- ator returned a strategy centered on a Hofmann–L ̈offler–Freytag reaction: chlorinate the secondary amine, photolyze to generate the nitrogen radical, abstract a hydrogen through a 1,6-hydrogen atom transfer (1,6-HAT) to functionalize an allylic methyl group, and use the resulting alkyl chloride as the electrophile for a base-promoted intramolecular N-alkylation that closes the D ring, forming Elymoclavine, which upon a double-bond migration, is converted into Lysergol. Even though a 1,6-HAT is far less common than a 1,5-HAT, such a reaction has literature precedence [60]. More- over, the rigidity of the substrate, the absence of any H atom correctly aligned for a 1,5-HAT, and the stability of the allylic radical that would be formed through 1,6- HAT are all factors that support the feasibility of this step in SynthEx’s strategy. It is also noteworthy to mention that SynthEx’s strategy is a unique way to reach Lysergol from Chanoclavine, differing from the proposed biosynthetic pathway whose key steps involve the oxidation of Chanoclavine to Chanoclavine-aldehyde followed by condensation to reach an iminium cation [61]. Where the conventional disconnection for this ring closes onto a carbon that already bears a functional handle, SynthEx instead installs the handle where the chem- istry needs it, through remote functionalization of an unactivated C–H bond. This is precisely the kind of low-frequency synthetic tactics that Section 2.3 shows to be 10 scarce in the patent record, and SynthEx proposed it unprompted to bridge a specific two-compound gap. This is the mode of use we expect to matter most in practice. A campaign stalled a few steps from its target has already heavily invested in the formation of its advanced intermediates, and the value of a planner there is not a route from commercial mate- rial but alternative ways to succeed in the last few steps. Because strategies can be conditioned on an intermediate, SynthEx can be asked that question directly. The three targets were selected for their utility to showcase the capability of the pipeline, but we note that this level of chemical sophistication is not isolated. Experts engaged with the proposed steps as chemistry to be reasoned about rather than as output to be scored, a response that algorithmic route prediction has not previously drawn. We now move on to quantitative evidence to examine whether this qualitative distinctiveness holds systematically. 2.3 Analysis of SynthEx’s reaction space Because the model writes disconnections rather than choosing them from a library (Section 2.1), SynthEx is not bound by the fixed templates and patent-frequency priors of conventional planners. We hypothesized that this would push it into regions of reaction space that patent-trained tools systematically under-represent; that the resulting chemistry would be practically useful rather than merely unusual; and that much of it would be absent from the proposals of even a state-of-the-art single-step model. We tested each prediction in turn, and then asked which chemistry accounts for the difference. We first characterize the SynthAtlas reaction corpus of 33,145 steps produced by running SynthEx across the more than 1,000 natural-product targets (Fig. 1c, Section 4.1), by how readily existing tools recognize it. We use three classification tools, one of them in two modes: NameRXN, Rxn-INSIGHT and ReactionClassifier [62–64]. ReactionClassifier is a hierarchical neuro-symbolic namer in which a fingerprint-based Multi Layer Perceptron proposes a class within a corpus-derived taxonomy and the label is returned only if one of the retrosynthetic templates belonging to that class reproduces the recorded product, so that recognition requires an explicit template match. The decisive comparison is NameRXN, an expert-curated dictionary tied to no training corpus, which is essentially at parity between SynthEx and USPTO (Fig. 3b): the reactions are therefore not ill-defined or unnameable. What separates them is their origin rather than their validity. The USPTO-derived ReactionClassifier recognizes 15–25 percentage points fewer SynthEx reactions than USPTO reactions, confirming that this nameable chemistry is scarce in the patent record. The two corpora also separate geometrically, a principal-component projection of ReactionClassifier’s neural network’s output layer places them in largely distinct regions rather than dispersing them through one another (Fig. 3a); we show this as a visualization rather than as independent evidence, since that classifier is itself trained on patent reactions and some separation follows from that alone. For a template-based tool, non-recognition directly implies non-reproduction, as a reaction absent from the library cannot be proposed. Models that construct discon- nections rather than retrieving them from a library are not bound in this way. Because 11 they predict precursors directly, either by editing the product graph or by generat- ing them de novo, they can in principle express any transformation, yet they remain anchored to the distribution of its training reactions, which is again the patent corpus; representational freedom is not distributional freedom. We therefore tested reacha- bility against RetroChimera [24], a state-of-the-art single-step model that ensembles a graph-editing component and a de novo Transformer with complementary induc- tive biases, and which its authors report to remain robust outside its training data. Presenting the product of each SynthEx reaction to RetroChimera (trained on Pis- tachio), the SynthEx disconnection appears in its top-1 prediction for only 13.5% of steps and its top-5 for 31.4% (Fig. 3c); therefore, more than two-thirds of SynthEx’s transformations are absent from the top-5 of a leading corpus-trained model. This gap is most pronounced in SynthEx’s signature chemistry—ring-forming disconnec- tions—where top-1 recovery falls to 2.3% and top-5 to 10.9%. Deeper prediction lists do not lift the ceiling, RetroChimera recovering only 52.0% of disconnections overall and 25.8% of the ring-forming ones even at top-50. Top-50 is in any case a generous regime: a multi-step search at the depths these targets demand can expand only a handful of candidates per node, so a disconnection buried far down a ranked list is rarely reached in practice. The chemistry SynthEx proposes is thus not a faster path to the same routes, but a region that a template library cannot reproduce and that a state-of-the-art corpus-trained model rarely proposes. Finally, we examine what chemistry drives this separation. The single clearest sig- nature is ring construction (Fig. 3d): SynthEx forms a ring in 16.0% of its steps, against 9.9% for USPTO, 8.7% for recent academic reactions, and only 2.8% for RetroChimera’s own top-1 predictions. Ring formation is precisely the low-prior, high-entropy chemistry described above: a large, heterogeneous family of individually rare transforms, poorly served by rigid traditional retrosynthesis models. Whereas NameRXN recognizes two-thirds of SynthEx’s ring-forming steps, USPTO-derived methods identify fewer than one in ten (Table 1), indicating how poorly such reactions are served by USPTO-derived tooling. This ring-building bias is one facet of a broader constructive signature. C–C bond formation is the single largest class of SynthEx’s named steps, at 22.5%, more than double its share of RetroChimera’s predictions on the same targets (9.7%); and these constructions are predominantly convergent, with 63.5% uniting two independent fragments through aldol addition, Horner–Wadsworth–Emmons and Julia–Kocienski olefination, Suzuki and Stille coupling, or olefin metathesis, and the remaining 36.5% closing rings intramolecularly. The corpus-trained model instead defaults to conser- vative functional-group editing: protecting-group manipulations account for 40.0% of RetroChimera’s top-1 disconnections but only 27.0% of SynthEx’s steps, and RetroChimera strongly over-proposes reduction and oxidation reactions and protection adjustments while under-proposing organometallic coupling. Furthermore, ReactionJSON lets the language model express many different modes of ring construction and reorganization. Figure 3e illustrates this range with three examples: an intramolecular spiroketalization, a pinacol rearrangement of a polycyclic skeleton, and a domino aza-Michael that closes several rings in one operation. 12 USPTO a. Reaction space of SynthEx vs. USPTO Ordered SMIRKS Hybrid Rxn- INSIGHT NameRXN 0 20 40 60 80 100 SynthExUSPTO b. Coverage by reaction classification tools c. SynthEx's reaction recovered by RetroChimera 1 10 r eactions r ecognised (%) 203040 50 Top-k 0 10 20 30 40 50 52% 26% 57% all ring-forming non-ring SynthEx recovery (%) SynthEx USPTO CRD Chimera 0 5 10 15 20 d. Share of ring forming reactions e. Flexible reaction output via verbalized intention and ReactionJSON ring-fo r ming r eactions (%) ReactionJSON Target / Intermediate Expanding intermediate Retro disconnection Verbalized Intention Partial Synthesis Tree Rearrangement Intramolecular reaction Domino sequence Pinacol rearrangement Spiroketalization Domino aza-Michael Retro Disconnections LLM Expansion Policy "op": "break_bond", "map_a": 11, "map_b": 13 "op": “change_bond_order", "map_a": 8, "map_b": 16, "delta": 1 The reaction is an oxidative pinacol rearrangement of a polycyclic indole precursor to form a spiro-oxindole, proceeding via oxidation of the indole C2-C3 double bond followed by a 1,2- alkyl shift. ... N OTBSO N NO 2 H H O N TBSO O N NO 2 H H N O N O Br O H O 2 N H N O HN O Br O H O 2 N H O O O OTBS OH HO O TBSO O SynthEx Fig. 3: SynthEx occupies a distinct reaction space that template-based tools can- not reproduce, enriched in ring construction. a, PCA of the output layer of a neural classifier co-embedding of SynthEx reactions (blue) with USPTO reactions (pink); SynthEx reactions form a largely contiguous territory distinct from the bulk of the patent-derived taxonomy. b, Fraction of reactions recognized by four classifier configurations for Syn- thEx versus a random USPTO sample; the two corpus-derived methods (ReactionClassifier, Ordered and Hybrid) recognize 15–25 percentage points fewer SynthEx reactions, whereas the corpus-independent NameRXN dictionary is at parity. c, Fraction of SynthEx ground- truth disconnections recovered within RetroChimera’s top-k predictions (k = 1–50) for all steps, ring-forming steps and non-ring-forming steps; recovery of ring-forming disconnections saturates far below completeness. d, Ring-forming reactions as a fraction of each corpus; SynthEx (16.0%) far exceeds USPTO, recent academic reactions (CRD), and RetroChimera’s top-1 predictions (2.8%). e, The Route Builder ’s natural-language reasoning, its ReactionJ- SON graph-edit output, and representative ring constructions. The chemistry demonstrated in the Okaramine, Melonine and Lysergol routes, ring- forming, convergent, and drawn from named-reaction rather than patent-frequency space, is not a property of those three targets; it is representative of the key steps found in the SynthAtlas corpus. 13 Table 1: Recognition of SynthEx’s ring-forming steps. Fraction of the 5,318 ring-forming and 27,827 non-ring-forming SynthEx steps named by each classifier. NameRXN, an expert-curated dictionary inde- pendent of any training corpus, names two-thirds of the ring-forming steps; every USPTO-frequency-derived method recognizes almost none. ReactionClassifier (Ordered) and ReactionClassifier (Hybrid) are the two variants of our template-based classifier (Section 2.3). The ratio is ring-forming over non-ring-forming recognition; a value near unity indi- cates no ring-specific deficit. ClassifierRing-formingNon-ring-formingRatio NameRXN66.8%79.1%0.84 Rxn-INSIGHT3.2%47.8%0.07 ReactionClassifier (Ordered)6.9%48.4%0.14 ReactionClassifier (Hybrid)2.2%35.6%0.06 2.4 Reach across the benchmark and expert assessment of the chemistry The above analysis has confirmed why SynthEx succeeds where traditional tools fail, but does not establish how often SynthEx succeeds — on what fraction of complex targets it returns a complete route to purchasable building blocks. We quantify that reach across the benchmark next. Existing multistep retrosynthesis benchmarks based on the patent literature, such as USPTO-190 [20], PaRoutes [30] and Pistachio-Hard [65], are by and large saturated, with state-of-the-art tools suggesting reasonable routes within a sensible compute budget [20, 23, 24, 66]. However, such benchmarks are largely in distribution for existing models, and do not permit claims about perfor- mance on the frontier of organic chemistry, the total synthesis of natural products. We therefore benchmark SynthEx on 1,098 structurally complex natural products drawn from NP-Atlas (Methods 4.1). A target is scored as solved only when a complete route to purchasable building blocks is found. For a conventional planner, these targets are effectively out of reach even under a generous compute budget. We use AiZynthFinder as the baseline because it is a widely adopted open-source template planner and, being inexpensive per expansion, can be run near-exhaustively; its failures therefore reflect the limits of the patented reaction space rather than an exhausted search budget. Run near-exhaustively over our building-block stock, without a practical expansion cap, to a maximum search depth of 25 and a wall-clock limit of 30 minutes per target, AiZynthFinder expands a median of roughly 2.9× 10 4 nodes yet solves only 13.8% of the benchmark (151/1,098; Methods 4.4). The barrier is therefore not search budget but reaction space: the dis- connections these targets require lie outside the template library, so no amount of additional expansion within it reaches them. The benchmark subsets make the same point in a way that also rules out a misconfigured baseline (Fig. 4a). On the control subset, which was assembled from structurally undemanding targets, AiZynthFinder solves 80%, close to the performance 14 Table 2: Reach on the natural-product benchmark (n = 1,098). AiZynthFinder is run near-exhaustively (no expansion cap, maximum depth 25, 30 minutes per target); SynthEx’s strategic layer is the language-model decomposition alone; the stitched result adds a short-budget (maximum depth 6) AiZynthFinder completion of the remaining unsolved leaves. The same template engine that solves only 13.8% as a stand- alone planner completes SynthEx’s simplified leaves to a 63.9% target-level solve rate. MethodSolvedSolve rate AiZynthFinder (exhaustive baseline)15113.8% SynthEx, strategic layer only27525.0% SynthEx, stitched (strategy + completion)70263.9% such tools report on the patent-derived benchmarks they were built for; SynthEx solves 95%. On the more complex targets, the performance of AiZynthFinder drops, to 12% on the complexity-dense set and only 4% on the large complex set, while SynthEx retains most of its reach. The failure we report is specific to structural complexity rather than a property of how the baseline was configured. The control subset sharpens this into something close to a controlled comparison, because of how it was selected. Those targets were chosen precisely because a short, resource-limited template search had failed on them (Methods 4.1); given the generous budget used here, the same planner solves 80% of them. Their original failure was therefore a budget failure, and more compute repairs it. The complex subsets were run under exactly the same generous budget and are not repaired: the planner still returns nothing for 88–96% of them. The same intervention that rescues one group leaves the other untouched, which is the distinction between a search that is under-resourced and one that is looking in the wrong space. SynthEx overcomes this barrier by widening the reaction space the search can draw on, not by searching the existing one harder. By guiding toward articulated key steps, its language-model strategist proposes complexity-reducing disconnections that the template library does not contain, simplifying each target toward precursors a short template search can finish; unsolved leaves are then completed by an AiZyn- thFinder call under a deliberately small budget (maximum depth 6), which succeeds precisely because the leaves handed to it are simple. The effect is cumulative (Table 2): the strategic layer alone, without any leaf completion, already solves 25.0% of tar- gets (275/1,098), exceeding the exhaustive template baseline; adding the short-budget completion of its leaves raises the target-level solve rate to 63.9% (702/1,098). The same tool that fails as a stand-alone planner succeeds as a completion engine once SynthEx has done the strategic work of substantially decomplexifying the target prod- uct through elegant, constructed steps, a division of labor whose advantage widens sharply with molecular complexity (Fig. 4b). Solve rate alone is an inadequate measure of a route’s quality, with the feasibility of the individual reactions mattering at least as much [17, 31]. We therefore asked 15 expert chemists a narrower and more answerable question than “is this a good route”: holding the overall strategy fixed, how do the reactions SynthEx selects compare with those a chemist selected for the same target? Conditioning on a shared strategic frame is what makes the comparison inter- pretable, since it isolates the quality of the chemistry from differences in overall plan. We assembled 70 natural-product targets, each with a total synthesis published after the model’s training cut-off and for which SynthEx also returned a route to the same target, and retained the 47 on which SynthEx independently arrived at a strategy congruent with the published one. Key steps were drawn from both routes, rendered identically and stripped of any cue to their origin, and rated by ten synthetic chemists from total-synthesis groups on four axes: feasibility, strategic value, elegance and overall quality. This yielded 1,040 ratings over 148 unique key steps (Methods 4.10). Individual raters are not independent replicates; each rates many items, and the items themselves are rated by an uneven number of raters. Hence we treat the rater, not the rating, as the unit of resampling: all intervals below come from a cluster bootstrap over the ten raters (resampling raters with replacement and recomputing item means from whichever ratings survive), which is the appropriate correction and is applied throughout rather than chosen post hoc. On this basis, SynthEx key steps were statistically indistinguishable from the published expert steps on feasibility (δ = −0.01, 95% CI [−0.09, +0.08]), elegance (δ = −0.09, [−0.19, +0.03]) and overall quality (δ =−0.09, [−0.14, +0.01]), where δ is Cliff’s delta over per-item mean scores and positive values favor SynthEx (Fig. 4c). Strategic value is the one axis with a detectable gap (δ = −0.14, [−0.21, −0.03]), which remains on the literature-favoring side even under Holm correction for testing four axes (98.75% CI [−0.24, −0.01]). Equivalently, the largest true difference consistent with the data is small on every axis: |δ|≤ 0.08 for feasibility, 0.20 for strategic value, 0.18 for elegance and 0.13 for overall. This one detectable gap is smaller than the disagreement among the raters them- selves. Recomputing δ from each rater’s own scores alone, the standard deviation across the ten individual estimates exceeds the pooled effect on every axis, including strate- gic value (ratio 1.47; elegance 1.96; overall 1.10; Fig. 4d), and it is not shared evenly: one participating group shows a clear literature preference on this axis while the other two show essentially none. We directly tested whether the four-axis rating profile can identify a step’s source: a logistic classifier on the per-item mean scores, evaluated by leave-one-item-out cross-validation, gives an area under the ROC curve of 0.48 (95% CI [0.38, 0.57]), indistinguishable from a permutation null (Fig. 4e); the panel’s own scores carry no detectable signal about which steps are SynthEx’s. Strategic value is therefore the one axis worth reporting as a real, if small and rater-dependent, differ- ence; with strategy controlled, whatever separates expert from machine on that axis lies in higher-order planning rather than in the soundness of the steps themselves, and it is not a difference the expert panel could reliably act on step-by-step. That is also the one stage at which SynthEx is built to accept human direction. The Strategy Generator takes chemist-supplied strategies and constraints in natural language (Section 2.1), and we withheld that input throughout in order to measure the system unaided. However, whether such steering would narrow the gap is a question these data cannot answer. 16 f. Comparison of route lengths per target e. Classification of reaction source via human ratings a. Solve rate by benchmark subset b. Solve rate by target molecular weight SynthEx shorter Literature shorter 0 5 10 15 20 25 30 SynthEx r oute length SynthEx / Literature Effect sizes by rating axis -0.4-0.20.0 0.20.4 Cliff's (SynthEx vs Literature) Feasibility Strategic value Elegance Overall Lit.SynthEx c. Expert rating differences for key steps -0.6-0.4-0.2 0.00.20.40.6 Cliff's (SynthEx vs Literature) R02 R10 R07 R05 R01 R06 R03 R08 R04 Njardarson Group LSPN Wipf Group d. Inter-rater variability for strategic value 0.0 0.2 0.4 0.6 0.8 1.0 False positive rate T rue positive rate AUC = 0.48 [0.38, 0.57] Complex, large (n=852) Complexity- dense (n=123) Control (n=123) 0 20 40 60 80 100 solve rate (%) 4 12 80 56 87 95 204 - 282 282 - 360 360 - 438 438 - 516 516 - 593 593 - 671 671 - 749 749 - 827 827 - 904 904 - 982 molar mass (g/mol) SynthEx AiZynthFinder SynthEx AiZynthFinder 051015202530 Literature route length 0.00.20.40.60.81.0 Pooled Chance Njardarson Group LSPN Wipf Group Fig. 4: General results on the SynthEx benchmark. a, Comparative solve rates of SynthEx and AiZynthFinder on several benchmark subsets, detailed in Supplementary Infor- mation: Benchmark Construction. b, Performance degradation relative to target molecular weight in Daltons. c, Differences in expert ratings for key steps (SynthEx vs. literature) across feasibility, strategic value, elegance, and overall axes. d, Inter-rater variability (heterogene- ity) on the strategic value axis. e, Predictive performance of human ratings, displaying the AUC of a logistic classifier used to predict whether a key step originates from the literature or SynthEx. f, Target-wise comparison of route lengths between academic literature and Syn- thEx routes. On targets where both a literature synthesis and a SynthEx route exist, SynthEx routes are frequently the longer of the two (Fig. 4f, right). This is expected rather than damaging. A published total synthesis is the endpoint of an optimization campaign in which steps were telescoped, protecting groups designed out and sequences shortened over months of laboratory work, whereas a SynthEx route is a first proposal that has never met an experiment. The comparison worth drawing is with what an automated planner produces, not with the distilled result of a human campaign; brevity relative to the literature is a target for the field, not a claim we make here. In the SI D.2 we provide a similar figure comparing route lengths where both AiZynthFinder and SynthEx find a route, with SynthEx consistently finding shorter pathways for a given target. 17 2.5 Iterative and granular synthetic route improvement Iterative feedback loops have proven effective in coding agents [67, 68], and the case for one is stronger still in retrosynthesis, where a proposed step can fail in more ways than a line of code can. The disanalogy matters as much as the parallel. A coding agent is corrected by a compiler and a test suite, which are ground truth; synthesis planning has no such oracle short of the laboratory. We therefore use a language-model critic as a stand-in, and are explicit below about what that does and does not establish. Inspired by how contemporary coding agents interact with source files through sur- gical editing and exact text replacement [29], we translate each synthetic route into a RouteJSON document, a linear sequence of ReactionJSON entries (Fig. 5a). Reaction- JSON anchors its graph-edit operations on atom maps, thereby avoiding branching [34, 69] and allowing the agent to focus on the transformation itself rather than rewrit- ing whole molecules [70–72]. This enables flexible manipulation of the synthetic tree through the textual modality, extending beyond the insertion of protections and depro- tections [73]: the agent can reorder reactions, insert or delete reaction steps, and alter reaction conditions or functional groups. Before surfacing a route to human users, the third phase of SynthEx runs an iter- ative action–feedback loop between a Critic and an Editor (Fig. 5a). The Critic acts as a reaction harness, simulating each reaction in the forward direction and flagging chemically infeasible steps, termed blocking reactions [8]. The Editor is then tasked with resolving all blocking reactions through surgical route edits while keeping the core strategy of the route intact. The resulting route is then re-evaluated by the Critic agent, and the loop continues until either all blocking reactions are resolved or a maximum number of iterations is reached. The blocking-reaction rate, computed per route as the number of blocking reactions divided by the total number of reactions, falls steadily with iteration, from approxi- mately 0.27 before any repair to approximately 0.06 after six iterations (Fig. 5b). The feasibility distribution shifts accordingly when the finished routes are re-scored by the Analyst in phase 4, with fewer infeasible and poor routes and more good and excellent ones (Fig. 5c). We note that this feasibility criterion rests on the knowledge of the language models themselves. While they can flag many obvious problems, certain blind spots may be shared across the agents, since they share the same LLM backbone. Establishing true feasibility requires experimental validation. Figure 5d illustrates this pipeline on a concrete SynthAtlas target, Monascus- pirolide A, whose route quality improves markedly after the improvement loop. The original route proposed by the Route Builder in phase 2 is flagged with several fea- sibility issues. First, the acid-labile spiroketal ring is installed mid-route, preventing subsequent steps from being run under acidic conditions; indeed, as the Critic agent notes, the late-stage Friedel–Crafts reaction, typically conducted with a Brønsted or Lewis acid, could decompose the ring. To resolve this blocking step, the Editor relocates the spiroketalization to the final step, after the Friedel–Crafts reaction, min- imizing the ring’s exposure across the synthesis. Second, in the Claisen condensation step, the Critic agent observes that the acetophenone substrate carries an α-keto ester group that is more electrophilic than the ester of the external substrate, potentially 18 step_id: R1 condition: ... reactionJSON: - op: break_bond, map_a: 2, map_b: 3 - op: change_bond_order, map_a: 2, map_b: 4, delta: 1 Friedel-Crafts Friedel-Crafts HWE olefination Claisen condensation Claisen condensation Spiroketalization Spiroketalization step_id: P1 condition: ... reactionJSON: - op: break_bond, map_a: 6, map_b: 7 - op: add_group, map_idx: 6, fragment_smiles: *OCC[O:8] - op: add_bond, map_a: 6, map_b: 8 step_id: R1 condition: ... reactionJSON - op: break_bond, map_a: 2, map_b: 3 - op: change_bond_order, map_a: 2, map_b: 4, delta: 1 - op: add_group, map_idx: 3, fragment_smiles: *MgBr step_id: P2 condition: ... reactionJSON: - op: break_bond, map_a: 6, map_b: 8 - op: break_bond, map_a: 6, map_b: 11 - op: add_group, map_idx: 6, fragment_smiles: *O, order: 2 P1 P2 R1 O OH O O 1 2 4 5 6 7 O O 1 2 4 5 6 7 CH 4 O OH O O O O CH 3 MgBr 1 3 3 + + + 2 4 5 6 7 O OH 1 3 2 4 5 6 7 1 3 2 4 5 6 8 11 8 9 10 9 10 RouteJSONImproved RouteJSON Added reactions/functional groups step_id: R1 - The nucleophiles (CH4) is missing MgBr. - The aldehyle is more favorable as the electrophile compared with the ketone. Route critic Problematic routeRoute at iteration 0 Route at iteration 6 Friedel-Crafts Requires strong acids that could break the acid- sensitive spiroketal Improvement agent Critic agent 11 1 3 2 4 5 6 8 9 1011 OH HO R1 O O O HO O O HO O O O O O O O O O O O O O O O O O O OH O O O HO O O HO O OHO O O O OH O OHO O O O OH O OH O O O O O OHO OOOH O O OHO OOOH O O O + OTBDMSO t-BuO O OTBDMS O O Iterate Claisen condensation . . . The alpha-keto ester is more electrophilic than the external ester. Feedback by Critic Dihydroxylation Dihydroxylation Changes made by Editor Deprotections . . . . . . t-BuO O OTBDMS O O + OTBDMSO O Reorder the Friedel-Crafts and the Spiroketalization steps Move HWE olefination to early stage to remove the alpha-keto ester before Claisen condenstaion. Insert protections / deprotections to avoid side reactions of Claisen condensation. Monascuspirolide A Monascuspirolide A a. Route improving - critic loop b. Blocking rate vs. Improvement iterations c. Feasibility before vs after improvement d. Case study of route improvement for the synthesis pathway of Monascuspirolide A Fig. 5: SynthEx’s iterative route refinement. a, The iterative action–feedback loop between the Critic and the Editor. b, The average per-route blocking-reaction rate decreases as the number of iterations increases. c, Overall route feasibility across SynthAtlas before and after the improvement loop, showing fewer infeasible and poor routes and more good and excellent ones. d, A worked example of the pipeline applied to a synthetic route for Monascuspirolide A. leading to polymerization. The Editor resolves this by moving the Horner–Wadsworth– Emmons (HWE) olefination to an earlier stage, eliminating the ketone at the carbon 19 α to the ester group of the acetophenone. Finally, the acid introduced after HWE ole- fination is protected as its tert -butyl ester and the alcohols are protected as TBDMS ethers, shielding these sites from the strong base used in the Claisen condensation. 2.6 SynthAtlas: an open resource over natural-product synthesis As a result of our benchmark, we release SynthAtlas, an open resource of SynthEx- designed syntheses of natural products (Fig. 1c). The release comprises 1,098 natural-product targets, 3,243 routes with synthetic strategies (a mean of 2.95 per target) and 33,145 fully specified reaction steps (mean 10.22 steps per route). Some strategies and corresponding key reactions are illustrated in Fig. 1d. Every reaction carries a complete atom-mapping, produced by construction from the graph-edit rep- resentation rather than by a post hoc mapper, together with structural descriptors including ring-formation flags and reactant count. Because the mapping falls out of the representation, the corpus is internally consistent in a way that post hoc mapped datasets are not. Routes are browsable through an interactive platform, on which each strategy can be inspected step-by-step alongside the agent’s reasoning, compared against the alternative strategies proposed for the same target, and commented on by other chemists. The resource is intended to be useful in two ways. Because it captures the con- vergent, ring-building chemistry that current retrosynthesis models under-represent (Fig. 3 and Table 1), and because its individual steps were rated comparably with pub- lished expert steps (Section 2.4), it offers training and evaluation material of a kind the patent corpora do not supply. And because every target was selected as having no reported total synthesis at the time of curation, each released route is a dated, public prediction about a molecule nobody has yet made. We intend to report concordance as syntheses of these targets appear. 3 Discussion The multistep benchmarks the field has long used are now highly saturated. Patent- derived planners report success approaching completeness on them, so differences in score no longer resolve differences in capability, much as early coding benchmarks stopped separating models before the field moved from isolated scripts to repository- scale software engineering tasks [28, 29]. Solve rate has not stopped being a meaningful quantity; it has stopped varying on those targets. We therefore evaluated on the regime where the open problem lies, a set enriched for the structural complexity that defeats standard planning tools and against which a near-exhaustively resourced template planner solves only 13.8% [31, 32, 37]. On this set SynthEx returns complete routes for 63.9% of targets, close to a five-fold increase in reach, and its margin widens rather than narrows as the molecules grow more complex. Reach matters most exactly where existing tools reach least: a planner that returns a route for one target in eight leaves the great majority of the frontier untouched, however sound its chemistry may be for the remainder. 20 What makes this possible is a change in the chemistry available to the planner, not merely in the efficiency of the search. SynthEx plans in a region of reaction space that catalogue-based planners cannot reproduce and corpus-trained models rarely pro- pose, favoring the convergent, ring-forming, bond-constructive steps that characterize expert total synthesis and that the patent record systematically under-represents. On individual targets, this chemistry withstands inspection. For Okaramine M it recovers the strategic logic of an expert route published after the cut-off and offers a defensible refinement of the protecting-group sequence; for Melonine it converges, unaided, on the same difficult disconnection an expert group judged worth attempting, and reaches it by a variant that the group who ran those experiments assess as more likely to suc- ceed than their own. In blinded assessment, the differences between its key steps and published expert steps were small on all four axes, with strategic value the remaining edge held by literature chemistry. We would go further and propose complex natural-product synthesis as the setting in which planners should now be measured. It probes chemical reasoning more directly than any patent-derived set, since the disconnections it demands are by construction absent from the catalogues, and its difficulty scales continuously with structural com- plexity rather than saturating. Capabilities of this kind tend to look discontinuous from the outside, a task on which every system scores near zero admitting one that scores ten per cent and ceasing to separate systems at all not long after; a benchmark is most useful before that happens, and natural-product planning is still early on that curve. The routes are the second output. SynthAtlas releases 1,098 targets, 3,243 routes with strategies and 33,145 atom-mapped steps, a body of convergent, ring-forming chemistry of a kind the patent corpora do not contain and against which the next gen- eration of planners can be trained and evaluated. Because every target was selected as having no reported total synthesis, the release is also a set of dated, public predictions about molecules nobody has yet made, and we will report concordance as syntheses of them appear. What we report is reach and per-step quality, not experimental feasibility. We have not verified stereochemical outcomes, and expert review surfaced occasional selectivity and feasibility errors; the expert comparison is conditional on a shared strategic frame, so it speaks to the chemistry SynthEx proposes once a viable strategy is found rather than to how reliably it finds one; the improvement loop is scored by the same class of model that performs the repairs, so it measures internal convergence rather than experimental validity; and the language-model backbone remains costly relative to a template search. A route a chemist judges sound on paper is a hypothetical evaluation, not a result. The axis on which expert judgment still holds an advantage is also the most amenable to user provision through natural-language strategy descriptions. The most productive near-term arrangement may therefore be neither autonomous planning nor unaided human design, but a chemist supplying the strategy and the agent working it out. Beyond that, the integration of reaction conditions and, ultimately, closed-loop validation in the laboratory remains the next frontier, and the planners that follow 21 should be judged against experimental proof, on targets where the question is still open. 4 Methods 4.1 Curation of natural-product targets The benchmark was drawn from the NPAtlas database (release 202409), retain- ing only natural products with no reported total synthesis according to NPAtlas. From these we selected structurally complex molecules, using a Bertz complexity of 900–2200, 24–65 heavy atoms, 4–12 stereocenters, 2–8 rings and at most 14 rotatable bonds. We then clustered them by Morgan-fingerprint similarity with BitBirch [74], keeping the medoid of each cluster and retaining sparse clusters so as to span the chemical space of the collection. This representative core forms the complex, large subset (n = 852). We further added two smaller subsets chosen to probe specific failure modes. The complexity-dense subset (n = 123) comprises molecules for which AiZynthFinder’s route is unexpectedly long for their molecular complexity. These are targets whose difficulty lies in producing a concise route, rather than finding one at all. The control subset (n = 123) comprises structurally simple molecules that a short, resource-limited AiZynthFinder run nonetheless failed to solve. Because low structural complexity would normally predict an easy synthesis, these unsolved cases are informative, as their failure turns out to be an artifact of the limited search budget rather than a property of the molecules, and under the generous budget used in our benchmark runs most of them are solved. They therefore serve as a control, separating failures caused by an under-resourced search from failures caused by the target lying outside the reachable reaction space. Together these selections yield 1,098 targets, the benchmark used throughout. 4.2 The SynthEx multi-agent system SynthEx is an agentic, large-language-model (LLM) retrosynthesis planner imple- mented as a subclass of AiZynthFinder’s Monte-Carlo tree search [66], in which the neural template expansion policy is replaced by an LLM-guided policy. Strategy generation. For each target a Strategy Generator first proposes several independent one-sentence synthetic strategies (default three per target). The prompt directs the model through a fixed four-point analysis (scaffold motif, key bond-forming reaction, functional-group conflicts and protection, and stereocenters) and requests strategies each achievable in one to two reaction steps; the model returns structured JSON. Chemical constraints (a required starting material, or free-text human guidance) are optional and are otherwise left to the model’s judgment. Each strategy seeds the subsequent search as a steering query. 22 Reaction representation and graph edits. Disconnections are represented not as reaction SMILES but as an ordered list of atom- level graph-edit operations (the operations field of an EditBasedRetroReaction). Each operation is a JSON object keyed by atom-map number; the opera- tion set comprises ten primitives: breakbond, addbond, changebondorder, changeatom, setexplicith, addgroup, removegroup, invertstereocenter, clearstereocenter and setbondstereo. As a worked example, a retro-Diels– Alder step is encoded as two changebondorder operations (restoring the diene and dienophile double bonds) followed by two breakbond operations (cleaving the two new sigma bonds). Applying the operation list to the mapped product yields the mapped precursors deterministically. LLM configuration. The backbone is Gemini (model identifier gemini-3.1-pro-preview) accessed through the google-genai SDK. Google’s published API documentation states a knowledge cut-off of January 2025 for all Gemini 3 models, including gemini-3.1-pro-preview [75]; we take the cut-off from that documentation rather than from the model’s own report, which is not a reliable source for it. The cut-off precedes the model’s February 2026 release by thirteen months, leaving a wide win- dow of post-cut-off literature against which the routes can be compared. The same documentation directs users to the Search Grounding tool for information after the cut-off; that tool was never enabled here, as set out below. Two caveats attach to any cut-off argument. A documented cut-off describes the pretraining corpus, whereas post-training data is often more recent and is not generally disclosed. We therefore treat the cut-off as reasonable evidence against retrieval, rather than as proof of it. Separately, the target molecules and their scaffolds are likely present in the pretraining corpus as isolated natural products, together with discussion of their biosynthesis; our claim is not that the targets are novel to the model but that no synthetic route to them was available before the cut-off. Presenting the target structure confers little information on route design, as a chemist handed the same structure must still plan the disconnections. No tools were available to the language model at any point. Search grounding was never enabled and the model had no web access, so it could draw on no information beyond its training data at inference time. The only lookups in the pipeline are against the fixed local template library and the building-block stock, both of which predate the training cut-off. The strategy planner runs at temperature 0.1 and the next-step generator at 0.3; no top p, topk or sampling seed is set, so generations are not determin- istic. Each call is retried up to three times with a 600 s timeout. The Phase-1 guided search is bounded by a step limit of 25. The exact run configuration is syntheliteconfig/configs/synthelite.gemini31.codeops.yml. 23 4.3 Route criticism and improvement Each completed route is serialized as a RouteJSON document, a linear sequence of ReactionJSON entries, so that edits can be applied to individual steps without regenerating the route. Three agents then operate on it. The Critic agent traverses the route and simulates each reaction in the forward direction, labeling a step blocking when it judges the transformation chemically infea- sible as written, for example because a functional group elsewhere in the substrate is incompatible with the stated conditions. The per-route blocking rate reported in Fig. 5b is the number of blocking steps divided by the number of steps in that route. The Editor receives the route together with the Critic’s annotations and resolves the blocking steps by editing the RouteJSON in place. The permitted operations are reordering steps, inserting or deleting steps, and altering reaction conditions or functional groups; the agent is instructed to preserve the key disconnection and the overall strategy. The edited route returns to the Critic, and the loop continues until no blocking steps remain or an iteration cap is reached. The Analyst scores the finished route for overall feasibility on a five-point scale (infeasible, poor, acceptable, good, excellent ) and identifies its key steps and principal risks (Fig. 1b). All three agents use the same backbone as the planner. The loop is an internal consistency procedure, not an external validation: the agent that scores the improve- ment shares a backbone with the agents that produce it, so a declining blocking rate demonstrates convergence against the Critic’s criterion and not experimental feasibility. 4.4 Search and solve criteria A building block is deemed purchasable if its full InChIKey is present in the com- bined ZINC and eMolecules stock (39,684,411 InChIKeys), matched exactly on the full InChIKey. A target is scored as solved only when a complete route is returned in which every leaf is purchasable. This logic is inherited from AiZynthFinder [66]. We report three conditions, all evaluated against the same 1,098-target benchmark and the same ZINC plus eMolecules stock. (i) AiZynthFinder exhaustive baseline: AiZynthFinder run directly on each intact target with the USPTO expansion, ring- breaker and filter policies, under a deliberately generous budget (maximum 25 transforms, 1,500 iterations, 1,800 s wall-clock, returning the first solution); no per- node expansion cap was set, so the library defaults apply (at most 50 templates per node, cumulative probability 0.995; exploration constant 1.4). (i) SynthEx strategic layer only : a target counts as solved if any strategy reaches an all-in-stock precur- sor set using the LLM-guided layer alone, without template completion. (i) SynthEx stitched : any precursor not purchasable after the strategic layer is completed by a short template search (AiZynthFinder, maximum 6 transforms, 500 iterations, 1,200 s), and the strategic and template fragments are stitched into a single route. This is a comparison of reach, not of compute cost. The LLM planner is substan- tially more expensive per target than a template search; the baseline is therefore run under a generous budget so that its failures reflect the reach of the USPTO reaction 24 space rather than an exhausted budget. For reference, the baseline expends a median of approximately 2.9× 10 4 expansion-policy calls per target (median ≈ 1, 500 MCTS iterations), whereas the short leaf-completion searches use a median of 163 expan- sion calls (median 74 iterations); a target with no returned result is counted as an incomplete, hence unsolved, search. 4.5 Reaction classification and recognition analysis Reactions were classified using three recognition tools, two external and one our own. The first, NameRXN (NextMove Software, filbert2 v3.7.0), deems a reaction rec- ognized when it is assigned any named class other than “Unrecognized.” The second, Rxn-INSIGHT (rxn-insight v0.1.3), scores a reaction as recognized when it returns a classification other than “OtherReaction.” Finally, ReactionClassifier was evaluated in two operating modes: an Ordered mode, which applies a hierarchical library of SMIRKS templates and accepts the first matching product, and a Hybrid mode, which initially predicts a reaction class using a DRFP fingerprint classifier (2048 bits, radius 2) before applying strict SMIRKS matching restricted to that class’s tier-three subset. ReactionClassifier serves as a purpose-built diagnostic for evaluating distribu- tional distance from the USPTO corpus; by design, it recognizes approximately 57% of USPTO reactions (57.3% in Ordered mode). Consequently, any reduced coverage observed on SynthEx chemistry measures the extent to which that chemical space diverges from the underlying USPTO distribution, rather than indicating a technical limitation of the classifier itself. To establish a baseline for this recognition compar- ison, a USPTO reference set was constructed by sampling 100,000 mapped USPTO reactions (random seed 42), yielding 92,857 parseable, single-product reactions after filtering to compare against the 33,145 SynthEx steps. 4.6 Reaction-space embedding The reaction-space map (Fig. 3a) is a principal-component projection of the output- layer activations of a neural reaction classifier: each SynthEx and USPTO reaction is passed through ReactionClassifier, its output-layer representation is taken as a reaction embedding, and the SynthEx and USPTO embeddings are co-projected by PCA into two dimensions [64]. Because that classifier is itself trained on patent reactions, a degree of separation between the two corpora is expected on those grounds alone; we therefore present the map as a visualization of the recognition results rather than as independent evidence of distinctness. 4.7 Single-step reachability Single-step reachability was assessed with RetroChimera, an ensemble single-step ret- rosynthesis model (a SMILES-transformer component and a template-localization component) using the Pistachio-trained checkpoint. For each SynthEx step, the model was queried on the largest product fragment and asked for its top 50 predictions. A step counts as recovered at rank k if, within the top k predictions, some predic- tion’s set of reactant fragments matches the ground-truth reactant set. Matching is on canonical SMILES with fragments canonicalized and sorted; stereochemistry is 25 retained (isomericSmiles=True). Over the full corpus (n = 33,145) recovery is 13.5%, 31.4% and 52.0% at top-1, top-5 and top-50 respectively; on the ring-forming subset (n = 5,318) it falls to 2.3%, 10.9% and 25.8%. 4.8 Structural descriptors and ring formation Structural descriptors were computed per reaction after atom-mapping. A reactant is treated as a true reactant only if it contributes more than three mapped heavy atoms to the product, which removes reagents, solvents and catalysts; only single-product reactions are analyzed, and malformed reactions are excluded. A step is ring-forming when the product contains more SSSR rings than the true reactants combined (RDKit CalcNumRings). Of the 33,145 parseable SynthEx steps, 5,318 (16.0%) are ring-forming and 27,827 are not. Carbon–carbon bond formations are those with NameRXN level- one code 3; among these, a step is classified as intermolecular when it has two or more true reactants and intramolecular when it has one. Protecting-group manipula- tions were counted as NameRXN level-one codes 5 or 6 (protection and deprotection combined); on this definition they account for 27.0% of named SynthEx steps versus 40.0% of named RetroChimera disconnections. 4.9 Atom-mapping and the SynthAtlas release Atom mapping is produced by construction: because each precursor is generated by applying explicit graph edits to the mapped product, product-derived atoms retain their parent map numbers and newly introduced atoms are assigned fresh numbers in sequence. No external atom-mapping model (for example RXNMapper) is used at any point [76]. The released SynthAtlas comprises 1,098 targets with a releasable route, 3,243 strategies (mean 2.95 per target) and 33,145 reaction steps with valid atom mapping (mean 10.22 steps per route). 4.10 Expert key-step evaluation Individual key steps were rated by ten expert chemists from academic total synthe- sis groups in Switzerland and the United States in a blinded web application. Each item was a single key step; the raters saw the target and reaction images and a copy- able SMILES string, rendered for both sources through a single RDKit engine with source-neutral filenames and re-canonicalized SMILES, so that the source (SynthEx or literature) never reached the client; the source mapping was held only in an offline key. The comparison was conditioned on a shared strategic frame: from the 70 tar- gets for which both a published literature route and a SynthEx route to the same target were available, we retained the 47 on which SynthEx independently arrived at a strategy congruent with the published one (congruence judged by an LLM), and drew key steps from both routes for these targets. Key steps were deduplicated by exact canonical reaction SMILES so that no rater scored two identical steps. Scoring was independent, not a paired forced choice: each key step was rated on its own on four axes (feasibility, strategic value, elegance and overall), on a five-point scale with anchors at 1, 3 and 5; the objective axes were presented first and elegance last to limit halo bias, and “overall” was asked as a separate fourth question rather than a 26 composite. In total 1,040 ratings were collected over 148 rated items from ten raters (per-axis counts, SynthEx/literature: feasibility 672/362, strategic 670/359, elegance 669/358, overall 672/359). Agreement between sources on each axis is summarized by Cliff’s δ, defined as P (SynthEx > literature)− P (SynthEx < literature) over per-item mean scores, so that a positive value indicates SynthEx rated higher. Because each rater scores many items and each item is scored by an uneven number of raters, ratings are not indepen- dent replicates; we therefore treat the rater, not the rating, as the unit of resampling, and report 95% confidence intervals from a cluster bootstrap over the ten raters (2,000 resamples, seed 7), pre-specified as the primary analysis rather than chosen after inspecting the data. Testing across the four axes is corrected by a Holm step-down procedure; strategic value is the only axis whose Holm-corrected interval excludes zero (Section 2.4). On conventional bands for Cliff’s δ, all four effects are negligible to small. As a robustness check we recompute the comparison rater-by-rater (Section 2.4), with a leave-one-rater-out sensitivity analysis, and with Krippendorff’s ordinal α for inter-rater reliability4. The literature key steps were drawn from a set of recent total syntheses published after the model’s training cut-off. From 145 curated routes, 138 yielded a usable key step; running SynthEx on these targets produced a route to the same target for 70 of them (the pool above), of which 47 were strategically congru- ent. Key steps were extracted by a combination of computational and LLM analysis: routes were reconstructed and atom-mapped programmatically from the scraped reac- tions, and an LLM (Claude Opus 4.8, structured output) then produced a strategy description and identified the key-step reactions; all routes were human-confirmed. 4.11 Statistics, software and reproducibility Benchmark construction and the labeling-set build used a fixed random seed (42); the rating-study bootstrap uses seed 7 with 2,000 resamples. Analyses used RDKit, AiZynthFinder [66], NameRXN (filbert2 3.7.0), Rxn-INSIGHT (gen-rxn-insight 0.1.3) and ReactionClassifier [62–64]. The language-model planner is not deterministic: no sampling seed is set, and temperatures of 0.1 and 0.3 are used for strategy generation and next-step gen- eration respectively (Methods 4.2). Re-running SynthEx on a target will therefore not reproduce a route exactly. All reported figures come from a single run over the benchmark. Acknowledgements. The authors thank collaborators in EPFL Laboratory of Artificial Chemical Intelligence (LIAC) for helpful discussions. Funding. This work was supported by the Swiss National Science Foundation through the National Centre of Competence in Research (NCCR) Catalysis (225147) and through the grant (214915). TAN acknowledges support from Intel and Merck KGaA via the AWASES programme. MD acknowledges financial support from the Research Foundation – Flanders (FWO Vlaanderen) through postdoctoral fellowship grant 1266226N and travel grant V414426N. XVN acknowledges the support from the AiChemist project via MSCA Doctoral Network. GB acknowledges the support from the LowDataML project via MSCA Doctoral Network. This research was conducted with support from Google.org and the Google Cloud Research Credits program for 27 the Gemini Academic Program. PS is part of the Reaxys R&D collaboration network. NTJ, JF and HL acknowledge support from the US National Science Foundation (US NSF) grant CHE-2449261. Competing interests. The authors declare no competing interests. Data availability. SynthAtlas, comprising 1,098 natural-product targets, 3,243 strategic routes and 33,145 fully specified and atom-mapped reaction steps, is released at https: //synthatlas.epfl.ch and archived at Zenodo (DOI to be assigned). The benchmark target list, the AiZynthFinder baseline results, the reaction-classification outputs underlying Fig. 3 and Table 1, and the anonymized expert key-step ratings under- lying Fig. 4c are included in the same database. The building-block stock is derived from ZINC and eMolecules and is subject to those providers’ terms. NPAtlas (release 202409) is publicly available. Pistachio and NameRXN are commercial and cannot be redistributed. Code availability. The SynthEx framework is released under an open-source license at the project repository, https://github.com/schwallergroup/SynthEx.git. Deterministic ReactionClassifier is released under an open-source license at the project repository, https://github.com/schwallergroup/ReactionClassifier.git. References [1] Corey, E., Ohno, M., Mitra, R.B., Vatakencherry, P.A.: Total synthesis of longifolene. Journal of the American Chemical Society 86(3), 478–485 (1964) [2] Corey, E.J., Cheng, X.-M.: The Logic of Chemical Synthesis. John Wiley & Sons, Weinheim (1995) [3] Nicolaou, K.C., Sorensen, E.J.: Classics in total synthesis: targets, strategies, methods (1996) [4] Nicolaou, K.C., Vourloumis, D., Winssinger, N., Baran, P.S.: The art and science of total synthesis at the dawn of the twenty-first century. Angewandte Chemie International Edition 39, 44–122 (2000) [5] Trost, B.M., Fleming, I.: Comprehensive Organic Synthesis: Selectivity, Strategy, and Efficiency in Modern Organic Chemistry vol. 8. Elsevier, Oxford (1991) [6] Nicolaou, K., H ̈arter, M.W., Gunzner, J.L., Nadin, A.: The wittig and related reactions in natural product synthesis. Liebigs Annalen 1997(7), 1283–1301 (1997) [7] Heravi, M.M., Zadsirjan, V.: Recent Applications of Selected Name Reactions in the Total Synthesis of Alkaloids. Elsevier, Amsterdam (2021) [8] Bran, A.M., Neukomm, T.A., Armstrong, D., Jonˇcev, Z., Schwaller, P.: Chem- ical reasoning in llms unlocks strategy-aware synthesis planning and reaction 28 mechanism elucidation. Matter 9(5) (2026) [9] Corey, E.J., Wipke, W.T.: Computer-assisted design of complex organic syn- theses: Pathways for molecular synthesis can be devised with a computer and equipment for graphical communication. Science 166(3902), 178–192 (1969) [10] Corey, E., Wipke, W.T., Cramer I, R.D., Howe, W.J.: Computer-assisted syn- thetic analysis. facile man-machine communication of chemical structure by interactive computer graphics. J. Am. Chem. Soc. 94, 421–430 (1972) [11] Corey, E.J., Long, A.K., Rubenstein, S.D.: Computer-assisted analysis in organic synthesis. Science 228, 408–418 (1985) [12] Lowe, D.M., Corbett, P.T., Murray-Rust, P., Glen, R.C.: Chemical name to structure: OPSIN, an open source solution. ACS Publications (2011) [13] Lowe, D.M.: Extraction of chemical structures and reactions from the literature. PhD thesis, University of Cambridge (2012) [14] Lowe, D.: Chemical reactions from US patents (1976-Sep2016) http://doi.org/10. 6084/m9.figshare.5104873.v1 (2017) [15] Segler, M.H., Waller, M.P.: Neural-symbolic machine learning for retrosynthesis and reaction prediction. Chem. Eur. J. 23, 5966–5971 (2017) [16] Coley, C.W., Rogers, L., Green, W.H., Jensen, K.F.: Computer-assisted retrosyn- thesis based on molecular similarity. ACS Cent. Sci. 3, 1237–1245 (2017) [17] Segler, M.H., Preuss, M., Waller, M.P.: Planning chemical syntheses with deep neural networks and symbolic ai. Nature 555(7698), 604–610 (2018) [18] Schwaller, P., Gaudin, T., Lanyi, D., Bekas, C., Laino, T.: “Found in Transla- tion”: predicting outcomes of complex organic chemistry reactions using neural sequence-to-sequence models. Chem. Sci. 9, 6091–6098 (2018) [19] Jin, W., Coley, C., Barzilay, R., Jaakkola, T.: Predicting organic reaction out- comes with weisfeiler-lehman network. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neu- ral Information Processing Systems 30, p. 2607–2616. Curran Associates, Inc., ??? (2017) [20] Chen, B., Li, C., Dai, H., Song, L.: Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search (2020). https://arxiv.org/abs/2006.15820 [21] Sacha, M., Blaz, M., Byrski, P., Dabrowski-Tumanski, P., Chrominski, M., Loska, R., Wlodarczyk-Pruszynski, P., Jastrzebski, S.: Molecule edit graph attention network: Modeling chemical reactions as sequences of graph edits. J. Chem. Inf. Model. 61, 3273–3284 (2021) https://doi.org/10.1021/acs.jcim.1c00537. PMID: 29 34251814 [22] Schwaller, P., Petraglia, R., Zullo, V., Nair, V.H., Haeuselmann, R.A., Pisoni, R., Bekas, C., Iuliano, A., Laino, T.: Predicting retrosynthetic pathways using transformer-based models and a hyper-graph exploration strategy. Chem. Sci. 11, 3316–3325 (2020) [23] Liu, G., Xue, D., Xie, S., Xia, Y., Tripp, A., Maziarz, K., Segler, M., Qin, T., Zhang, Z., Liu, T.-Y.: Retrosynthetic planning with dual value networks. In: International Conference on Machine Learning, p. 22266–22276 (2023). PMLR [24] Maziarz, K., Liu, G., Misztela, H., Kornev, A., Gai ́nski, P., Hoefling, H., Fortu- nato, M., Gupta, R., Segler, M.: Chimera: Accurate retrosynthesis prediction by ensembling models with diverse inductive biases. arXiv preprint arXiv:2412.05269 (2024) [25] Genheden, S., Norrby, P.-O., Engkvist, O.: Aizynthtrain: robust, reproducible, and extensible pipelines for training synthesis prediction models. Journal of Chemical Information and Modeling 63(7), 1841–1846 (2023) [26] Schwaller, P., Laino, T., Gaudin, T., Bolgar, P., Hunter, C.A., Bekas, C., Lee, A.A.: Molecular transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS Cent. Sci. 5(9), 1572–1583 (2019) [27] Tanovic, S., Wieczorek, E., Duarte, F.: An exploration of dataset bias in single- step retrosynthesis prediction. Digital Discovery 5(2), 793–802 (2026) [28] Jimenez, C.E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., Narasimhan, K.: Swe-bench: Can language models resolve real-world github issues? In: Inter- national Conference on Learning Representations, vol. 2024, p. 54107–54157 (2024) [29] Yang, J., Jimenez, C., Wettig, A., Lieret, K., Yao, S., Narasimhan, K., Press, O.: Swe-agent: Agent-computer interfaces enable automated software engineering. Advances in Neural Information Processing Systems 37, 50528–50652 (2024) [30] Genheden, S., Bjerrum, E.: Paroutes: a framework for benchmarking retrosyn- thesis route predictions. ChemRxiv (2022) [31] Maziarz, K., Tripp, A., Liu, G., Stanley, M., Xie, S., Gai ́nski, P., Seidl, P., Segler, M.: Re-evaluating retrosynthesis algorithms with syntheseus. arXiv preprint arXiv:2310.19796 (2023) [32] Tripp, A., Maziarz, K., Lewis, S., Segler, M., Hern ́andez Lobato, J.M.: Retro- fallback: retrosynthetic planning in an uncertain world. In: International Confer- ence on Learning Representations, vol. 2024, p. 30687–30744 (2024) 30 [33] Morgunov, A., Batista, V.S.: Procrustean bed for ai-driven retrosynthesis: A unified framework for reproducible evaluation. arXiv preprint arXiv:2512.07079 (2025) [34] Xuan-Vu, N., Armstrong, D., Joncev, Z., Schwaller, P.: Tempre: Template generation for single and direct multi-step retrosynthesis. arXiv preprint arXiv:2507.21762 (2025) [35] Hassen, A.K., Lai, H., Genheden, S., Preuss, M., Clevert, D.-A.: Synthe- sis planning in reaction space: a study on success, robustness and diversity. Digital Discovery 5(4), 1623–1634 (2026) https://doi.org/10.1039/d5d00280j https://pubs.rsc.org/d/article-pdf/5/4/1623/12756083/d5d00280j.pdf [36] Tran, S.B., Roh, J., Coley, C.W.: Quantifying the failure modes of current one- step retrosynthesis models. Chemical Science (2026) [37] Morgunov, A., Shee, Y., Soudackov, A.V., Batista, V.S.: The syntax of matter: Synthesis planning as the foundation of generative chemistry. ChemRxiv (2026) https://doi.org/10.26434/chemrxiv.15001278 [38] Jablonka, K.M., Ai, Q., Al-Feghali, A., Badhwar, S., Bocarsly, J.D., Bran, A.M., Bringuier, S., Brinson, L.C., Choudhary, K., Circi, D., et al.: 14 examples of how llms can transform materials science and chemistry: a reflection on a large language model hackathon. Digital discovery 2(5), 1233–1250 (2023) [39] Jablonka, K.M., Schwaller, P., Ortega-Guerrero, A., Smit, B.: Leveraging large language models for predictive chemistry. Nat. Mach. Intell. 6(2), 161–169 (2024) https://doi.org/10.1038/s42256-023-00788-1 [40] Alampara, N., Schilling-Wilhelmi, M., R ́ıos-Garc ́ıa, M., Mandal, I., Khetarpal, P., Grover, H.S., Krishnan, N., Jablonka, K.M.: Probing the limitations of mul- timodal language models for chemistry and materials research. arXiv preprint arXiv:2411.16955 (2024) [41] Bran, A., Cox, S., Schilter, O., Baldassari, C., White, A.D., Schwaller, P.: Aug- menting large language models with chemistry tools. Nature Machine Intelligence, 1–11 (2024) [42] Boiko, D.A., MacKnight, R., Kline, B., Gomes, G.: Autonomous chemical research with large language models. Nature 624(7992), 570–578 (2023) [43] Mirza, A., Alampara, N., Kunchapu, S., R ́ıos-Garc ́ıa, M., Emoekabu, B., Krish- nan, A., Gupta, T., Schilling-Wilhelmi, M., Okereke, M., Aneesh, A., et al.: A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists. Nature Chemistry, 1–8 (2025) [44] Armstrong, D., Jonˇcev, Z., Bran, A.M., Schwaller, P.: Synthstrategy: Extracting 31 and formalizing latent strategic insights from llms in organic chemistry. arXiv preprint arXiv:2512.01507 (2025) [45] Baker, F.N., Adu-Ampratwum, D., Averly, R., Yu, B., Sun, H., Ning, X.: Larc: Towards human-level constrained retrosynthesis planning through an agentic framework. arXiv preprint arXiv:2508.11860 (2025) [46] Baker, F.N., Nguyen, T., Averly, R., Yu, B., Adu-Ampratwum, D., Sun, H., Ning, X.: Mmorf: A multi-agent framework for designing multi-objective retrosynthesis planning systems. arXiv preprint arXiv:2604.05075 (2026) [47] Xuan-Vu, N., Armstrong, D., Wehrbach, M., Bran, A.M., Jonˇcev, Z., Schwaller, P.: Synthelite: Chemist-aligned and feasibility-aware synthesis planning with llms. arXiv preprint arXiv:2512.16424 (2025) [48] Reaxys database. (Accessed Jul 29, 2021) (2024). https://w.reaxys.com [49] Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., ˇ Z ́ıdek, A., Potapenko, A., et al.: Highly accurate protein structure prediction with alphafold. Nature 596(7873), 583–589 (2021) [50] Runcie, N.T., Imrie, F., Deane, C.M.: Molecular representations for large language models. arXiv preprint arXiv:2605.01822 (2026) [51] Schmelzer, D., Stark, C.B.: Synthesis of Okaramine M, Its Conversion to Amauromines, and Concise Bidirectional and Biomimetic Total Synthesis of Amauromines. Organic Letters 27(27), 7367–7371 (2025) [52] Takase, S., Kawai, Y., Uchida, I., Tanaka, H., Aoki, H.: Structure of amau- romine, a new alkaloid with vasodilating activity produced by amauroascus sp. Tetrahedron Letters 25(41), 4673–4676 (1984) [53] Ishikawa, K., Hosoe, T., Itabashi, T., Wakana, D., Takizawa, K., Yaguchi, T., Kawai, K.-i.: Novoamauromine and ent-Cycloechinulin: two new diketopiper- azine derivatives from Aspergillus novofumigatus. Chemical and Pharmaceutical Bulletin 58(5), 717–719 (2010) [54] Kouam ́e, T., Bernadat, G., Turpin, V., Litaudon, M., Okpekon, A.T., Gallard, J.-F., Leblanc, K., Rharrabti, S., Champy, P., Poupon, E., et al.: Structure Reas- signment of Melonine and Quantum-Chemical Calculations-Based Assessment of Biosynthetic Scenarios Leading to Its Revised and Original Structures. Organic Letters 23(15), 5964–5968 (2021) [55] Zhang, X.: Vallesamidine and schizozygane alkaloids: rearranged monoterpene indole alkaloids and synthetic endeavours. Natural Product Reports 41(5), 784– 812 (2024) 32 [56] Matsuyuki, Y., Umekubo, N., Yokoshima, S.: Total Synthesis of Melonine. Organic Letters 27(9), 2065–2068 (2025) [57] Go ̈elo, V., Wang, Q., Zhu, J.: Total Synthesis of (+)-Melonine and (+)-N 4 - Oxy Melonine Enabled by an Intramolecular Alkene Diamination Reaction. Angewandte Chemie International Edition, 8101956 (2026) [58] Delayre, B., Piemontesi, C., Wang, Q., Zhu, J.: TiCl 3 -Mediated Synthesis of 2,3,3- Trisubstituted Indolenines: Total Synthesis of (+)-1,2-Dehydroaspidospermidine, (+)-Condyfoline, and (−)-Tubifoline. Angewandte Chemie International Edition 59(33), 13990–13997 (2020) https://doi.org/10.1002/anie.202005380 https://onlinelibrary.wiley.com/doi/pdf/10.1002/anie.202005380 [59] Ma, Y., Yan, J., Yang, L., Yao, Y., Wang, L., Gao, S.-S., Cui, C.: A hybrid system for the overproduction of complex ergot alkaloid chanoclavine. Frontiers in Bioengineering and Biotechnology 10, 1095464 (2022) [60] Dupeyre, R.-M., Rassat, A.: Application de la reaction de Hofmann-L ̈offler- Freytag synthese de derives diaza-2, 6 adamantane. Tetrahedron Letters 14(29), 2699–2701 (1973) [61] Jakubczyk, D., Cheng, J.Z., O’Connor, S.E.: Biosynthesis of the ergot alkaloids. Natural Product Reports 31(10), 1328–1338 (2014) [62] Nextmove Software NameRXN http://w.nextmovesoftware.com/namerxn. html. (Accessed March 29, 2022). http://w.nextmovesoftware.com/namerxn. html [63] Dobbelaere, M.R., Lengyel, I., Stevens, C.V., Van Geem, K.M.: Rxn-insight: fast chemical reaction analysis using bond-electron matrices. Journal of Cheminfor- matics 16(1), 37 (2024) [64] Armstrong, D., Dobbelaere, M., Olikauskas, V., Avila, H., Susanu, O., Waser, J., Schwaller, P.: Agentic generation of verifiable rules for deterministic, self- expanding reaction classification. arXiv preprint arXiv:2607.01061 (2026) [65] Yu, K., Roh, J., Li, Z., Gao, W., Wang, R., Coley, C.W.: Double-ended synthesis planning with goal-constrained bidirectional search. arXiv preprint arXiv:2407.06334 (2024) [66] Genheden, S., Thakkar, A., Chadimov ́a, V., Reymond, J.-L., Engkvist, O., Bjer- rum, E.: AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. J. Cheminf. 12, 1–9 (2020) [67] Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Lan- guage agents with verbal reinforcement learning. Advances in neural information processing systems 36, 8634–8652 (2023) 33 [68] Takerngsaksiri, W., Pasuksmit, J., Thongtanunam, P., Tantithamthavorn, C., Zhang, R., Jiang, F., Li, J., Cook, E., Chen, K., Wu, M.: Human-in-the-loop software development agents. In: 2025 IEEE/ACM 47th International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP), p. 342– 352 (2025). IEEE [69] Wang, H., Guo, J., Kong, L., Ramprasad, R., Schwaller, P., Du, Y., Zhang, C.: Llm-augmented chemical synthesis and design decision programs. arXiv preprint arXiv:2505.07027 (2025) [70] Edwards, C., Lai, T., Ros, K., Honke, G., Cho, K., Ji, H.: Translation between molecules and natural language. Proc. Conf. Empirical Methods Nat. Lang. Process., 375–413 (2022) [71] Walters,P.:SillyThingsLargeLanguageModelsDoWith Molecules.http://practicalcheminformatics.blogspot.com/2024/10/ silly-things-large-language-models-do.html Accessed 2025-02-03 [72] Jang, H., Jang, Y., Kim, J., Ahn, S.: Can LLMs Generate Diverse Molecules? Towards Alignment with Structural Diversity (2024) [73] Westerlund, A.M., Sigmund, L.M., Mijangos, M.V., Kannas, C., Genheden, S., Kabeshov, M.: Toward lab-ready ai synthesis plans with protection strategies and route scoring. Journal of Chemical Information and Modeling 66(11), 6361–6375 (2026) [74] P ́erez, K.L., Jung, V., Chen, L., Huddleston, K., Miranda-Quintana, R.A.: Bit- birch: efficient clustering of large molecular libraries. Digital Discovery 4(4), 1042–1051 (2025) [75] Google: Gemini 3 models. Gemini API documentation. https://ai.google.dev/ gemini-api/docs/gemini-3. Knowledge cutoff January 2025. Accessed 25 July 2026 (2026) [76] Schwaller, P., Probst, D., Vaucher, A.C., Nair, V.H., Kreutter, D., Laino, T., Reymond, J.-L.: Mapping the space of chemical reactions using attention-based neural networks. Nat. Mach. Intell. 3, 144–152 (2021) A Assembling the blind chemist evaluation set We asked expert chemists to score individual synthetic key steps drawn from two sources: the key steps of recently published academic total syntheses, and the key steps of the routes proposed by our tool for the same targets. Each chemist saw one key step at a time and scored it on four scales. The labelers saw the target structure and the reaction diagram. Steps from the two sources were mixed, blinded and shown 34 in a scrambled order, so raters could not tell which had come from the literature and which from our tool. B Extracting the key steps The two sources are matched by target, but the key step of each route was identified differently by design. For the literature we identified it independently, whereas for our tool we used the step the tool itself nominated; the chemists then scored both on the same scales, so the comparison is between the steps each source treats as pivotal rather than one assessor’s view of both. For the published syntheses we first reconstructed each route from the scraped reaction records: reactions were grouped by DOI, atom-mapped, and joined into a synthesis tree by matching the products of one reaction to the reactants of the next, allowing for small stereochemical mismatches during matching. We numbered the reac- tions of each route in synthetic order, from starting materials to final target, and gave the numbered route to a large language model (Claude Opus 4.8), which returned a short description of the strategy and the number of the single reaction that defines it (occasionally two, for a route that genuinely hinges on a convergent union followed by a cascade cyclization). That nominated reaction is the literature key step. Of 145 targets, 138 yielded a key step; the remainder failed to reconstruct cleanly and were dropped. C Removing routine chemistry A central claim of the tool is that it proposes non-obvious disconnections, so we removed steps that are routine and carry little discriminating signal—protecting-group manipulations, standard functional-group interconversions, amide and ester formation, and the like. To do this consistently we classified every candidate step with a published reaction classifier, which assigns a step to a named class only when a reaction template reproduces its product exactly. Sixty of the 358 steps were classified in this way; the remaining 83% were left untouched. We did not, however, discard every named reaction, because some of the most strategic steps carry familiar names. We therefore kept classified steps whose class was a skeleton-forming or olefination reaction, alkynylations, sp 2 –sp 2 couplings, aldol and Horner–Wadsworth–Emmons and Wittig olefinations, organolithium additions, olefin metathesis and Claisen rearrangements and removed the rest, which were protecting- group cleavages, oxidations and reductions, ether, amide, ester, silyl, thioether and N - alkyl bond formations, arene halogenations and Buchwald–Hartwig aminations. This removed 8 literature steps, leaving 156 steps. In addition, we deduplicated reactions across the corpus, removing another 8 steps, for a final count of 148 steps. D Scoring and blinding Chemists scored each step on four independent 1–5 scales: feasibility (would the reac- tion work as drawn, considering reactivity and selectivity); strategic value (how much 35 the step builds the target, through the bonds and stereocenters it forms); elegance (a holistic judgment of the synthetic design); and an overall “would I run this” score. A prompt on every item asked raters, when judging feasibility, to consider whether con- ditions exist that could achieve the required chemo-, regio- and stereoselectivity. Each item also allowed a free-text comment and a flag to mark a step for later exclusion. Steps were presented without any indication of their source. The mapping from item to source was held in a separate key that the interface never accessed. Literature and tool steps were paired and then scrambled within and across pairs so that neither position nor ordering revealed the source, using a fixed random seed so that the whole set can be reproduced exactly. D.1 Efficiency of the expansion policies The two planners arrive at comparable coverage through very different amounts of search. Run on the intact target, AiZynthFinder expands tens of thousands of nodes per molecule—a mean of 24,218 calls to its expansion policy, median 28,802 (Fig. 6). Most of that effort goes into failures: a target it cannot solve simply runs out the 1500-iteration budget, whereas the ones it does solve usually finish in a few hundred calls. SynthEx’s LLM policy works under a hard budget of 75 calls per target (three strategies, 25 expansions each) and reaches a slightly higher raw solve rate (25.0% vs. 13.8%), roughly two to three orders of magnitude fewer policy invocations for the same job. Completing SynthEx’s unsolved leaves with AiZynthFinder is where the template planner is at its best: the leaves are small, near-terminal fragments, and when one is solvable AiZynthFinder finds a route almost immediately (median 6 expansion calls). What a target actually costs at this stage is set by the leaves AiZynthFinder fails on— each of those runs out its budget at around 1,650 calls—so, summed over a target’s handful of leaves, the stitch adds a mean of ∼3,400 template calls (median ∼3,100). Even so, the full pipeline sits an order of magnitude below the cost of running AiZyn- thFinder on the intact molecule, and solves far more of them. One caveat when reading Fig. 6: a Gemini expansion and an AiZynthFinder template call are not the same unit of work—one is a single large-model inference, the other a small template-network pass—so the bars count how often each policy is queried, not the underlying compute. D.2 Route length on jointly solved targets Where both methods reach a target, SynthEx’s routes are shorter. Restricting to the n = 134 targets for which both SynthEx and AiZynthFinder return a complete route to purchasable building blocks, SynthEx’s median full route length (its own strategic disconnections plus any AiZynthFinder-completed leaves) is 5 steps (mean 5.7) against AiZynthFinder’s median 11 steps (mean 12.2) (Fig. 7). SynthEx reaches the target in fewer steps on 105 of the 134 targets (78%). 36 10 2 10 3 10 4 10 5 expansion / model calls per target AiZynthFinder (13.8% solved) SynthEx, LLM only (25.0% solved) SynthEx (63.9% solved) 24,218 75 3,499 Mean 10 2 10 3 10 4 10 5 expansion / model calls per target 28,802 75 3,176 Median SynthEx Gemini LLM callsAiZynthFinder template calls Fig. 6: Expansion/model calls per target, shown as the mean (left) and median (right) over the full corpus (n = 1098); solve rate annotated beside each method. AiZynthFinder run on the intact target is compared with SynthEx’s LLM policy alone and the full pipeline, in which AiZynthFinder completes SynthEx’s unsolved leaves. LLM and template calls are different units of work; the bars count invocations, not compute. 0510152025 AiZynthFinder route length 0 5 10 15 20 25 SynthEx route length y = x a 0510152025 Route length (number of reactions) 0 5 10 15 20 25 Routes at given length b SynthEx AiZynthFinder Full route lengths to stock, SynthEx : AiZynthFinder Fig. 7: SynthEx plans shorter routes. Full route length (to purchasable building blocks) for SynthEx versus AiZynthFinder on the n = 134 targets both methods solve. Left: per- target comparison; most points fall below y = x (105/134). Right: route-length distributions (dotted lines: medians, SynthEx 5 vs AiZynthFinder 11). E Robustness of the expert key-step comparison The main text (Fig. 4c) reports the four-axis comparison between SynthEx and liter- ature key steps as a cluster bootstrap over ten raters, the appropriate resampling unit since each rater scores many items and each item is scored by an uneven number of raters. This section reports that comparison decomposed by participating group, its dependence on any single rater, and the reliability of the rating instrument itself. 37 Figure 8 decomposes the comparison by participating group. The pattern under- lying the pooled strategic-value effect is uneven across groups rather than uniform: LSPN’s own ratings alone give δ = −0.20 (95% CI [−0.29, −0.10]), the only per- group interval on any axis that excludes zero, while the Njardarson Group (δ = +0.01, [−0.08, +0.12]) and the Wipf Group (δ = +0.02, [−0.19, +0.24]) show no detectable difference on the same axis. No group shows a detectable difference on feasibility, elegance or overall quality. This unevenness is not an artifact of the groups rating different material. Figure 9 shows the number of key steps each pair of groups rated in common: the Njardarson Group and LSPN rated the full item set (148 of 148 items each, entirely overlapping), and the Wipf Group’s 95 items are a strict subset rated identically by both other groups. The three groups’ effect sizes are therefore directly comparable and are not confounded by item difficulty. Figure 10 shows each group’s raw score distribution by axis. Groups differ notice- ably in overall scale use (some are more lenient across both sources than others), which is why every effect size in this section and the main text is computed within rater or within group rather than by comparing raw means across raters directly. Figure 11 extends the rater-level heterogeneity view (main text Fig. 4d) to all four axes. The spread of individual raters’ own estimates exceeds the pooled effect on every axis (main text), and this figure shows the pattern is not specific to strategic value: on every axis, at least one rater’s own comparison sits on the opposite side of zero from the pooled estimate. Table 3 reports the pooled effect size recomputed with each rater removed in turn, using the same cluster bootstrap as the primary analysis. The maximum shift from dropping any single rater is 0.068 (feasibility), 0.033 (strategic value), 0.065 (elegance) and 0.037 (overall); on strategic value, the effect remains negative under every possible single-rater removal (δ ∈ [−0.168,−0.109]). No axis is materially dependent on any one rater. Table 4 reports inter-rater reliability (Krippendorff’s α, ordinal metric, on the 148 items with three or more raters) for each axis: 0.23 for feasibility, 0.24 for strategic value, 0.13 for elegance and 0.17 for overall. Values below 0.4 are conventionally con- sidered poor agreement, indicating that individual raters agree with each other only weakly on the absolute quality of a given step, independent of source. This is the same phenomenon quantified directly in the main text (rater heterogeneity exceeding the pooled source effect on every axis) and is the reason the comparison is reported at the level of a bootstrap over raters rather than a single pooled estimate treated as if it had no clustering structure. 38 0.60.40.20.00.20.40.6 Cliff's Feasibility Strategic value Elegance Overall Njardarson Group 0.60.40.20.00.20.40.6 Cliff's LSPN 0.60.40.20.00.20.40.6 Cliff's Wipf Group S1. Per-lab effect sizes (rating-level bootstrap, n=500) Fig. 8: Per-group effect sizes. Cliff’s δ (SynthEx vs. literature) by rating axis, computed separately within each participating group (rating-level bootstrap, 500 resamples). LSPN is the only group whose interval excludes zero on any axis (strategic value); the Njardarson Group and the Wipf Group show no detectable difference on any of the four axes. Njardarson Group LSPN Wipf Group Njardarson Group LSPN Wipf Group 14814895 14814895 959595 S2. Item x lab overlap (common rated items) 0 20 40 60 80 100 120 140 items Fig. 9: Item-by-group overlap. Number of key steps rated in common by each pair of groups. The Njardarson Group and LSPN rated the full 148-item set each; the Wipf Group’s 95 items are a strict subset rated identically by both other groups, so the per-group effect sizes in Fig. 8 are not confounded by item difficulty. 39 Njardarson Group LSPN Wipf Group 1 2 3 4 5 Feasibility Njardarson Group LSPN Wipf Group Strategic value Njardarson Group LSPN Wipf Group Elegance Njardarson Group LSPN Wipf Group Overall S3. Per-lab score distributions (leniency differences) Fig. 10: Per-group score distributions. Raw score distributions (both sources pooled) by rating axis and participating group, showing that groups differ in overall scale use. This is why every effect size in this section and the main text is computed within rater or within group, never from raw means compared across raters directly. 0.60.40.20.00.20.40.6 R08 R01 R05 R03 R04 R06 R10 R02 R07 Feasibility 0.60.40.20.00.20.40.6 R02 R10 R07 R05 R01 R06 R03 R08 R04 Strategic value 0.60.40.20.00.20.40.6 R10 R02 R08 R03 R05 R04 R07 R01 R06 Elegance 0.60.40.20.00.20.40.6 R08 R05 R03 R02 R01 R06 R04 R10 R07 Overall S4. Rater-level heterogeneity, all four axes Fig. 11: Rater-level heterogeneity, all four axes. Individual raters’ own Cliff’s δ (Syn- thEx vs. literature, computed from each rater’s own scores only), extending main text Fig. 4d to feasibility, elegance and overall quality. Marker shape indicates participating group (squares: Njardarson Group; circles: LSPN; triangles: Wipf Group); color indicates the direc- tion of that rater’s own estimate (blue: SynthEx-favoring; pink: literature-favoring). The dashed line is the pooled estimate for that axis. On every axis, at least one rater’s own com- parison falls on the opposite side of zero from the pooled estimate. 40 Table 3: Leave-one-rater-out. Pooled Cliff’s δ (SynthEx vs. literature; positive favors SynthEx) recomputed with each rater removed in turn, using the same cluster-bootstrap point estimate as the primary analysis. None is the full-panel estimate (main text Fig. 4c). No axis is materially dependent on any one rater; strate- gic value remains negative under every single-rater removal. Rater droppedFeasibilityStrategic valueEleganceOverall R01−0.011−0.137−0.128−0.082 R02−0.011−0.109−0.079−0.069 R03−0.007−0.164−0.089−0.088 R04−0.003−0.168−0.108−0.102 R05−0.012−0.128−0.112−0.083 R06−0.012−0.144−0.123−0.099 R07−0.034−0.136−0.109−0.095 R08+0.056−0.146−0.026−0.053 R09−0.010−0.149−0.098−0.087 R10−0.025−0.125−0.064−0.106 None−0.011−0.142−0.091−0.091 Table 4: Reliability. Ordinal Krippen- dorff’s α per rating axis, computed over the 148 items with three or more raters. Values below 0.4 are conventionally rated poor (Krippendorff, 2004): raters agree only weakly with each other on the absolute qual- ity of a given step, independent of source, which is why the main text reports effect sizes with a cluster bootstrap over raters rather than treating the panel as a single low- variance measurement. AxisKrippendorff’s α n items Feasibility0.233148 Strategic value0.238148 Elegance0.131148 Overall0.174148 41

Instant research discovery