Strategy-first synthesis planning for complex natural products

The paper addresses the long-standing challenge of automating retrosynthetic design for complex natural products. The authors note that while computer-assisted synthesis planning (CASP) has made significant strides, existing tools are fundamentally limited: "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."

The authors argue that this limitation is structural: "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."

The authors introduce SynthEx, an agentic synthesis planner that removes this constraint at its source. Rather than selecting disconnections from a fixed template library, SynthEx expresses each disconnection as an ordered list of atom-level graph edits in a format called ReactionJSON. This allows the system to transition from a critic that ranks known templates to a policy that generates novel expansions.

The architecture consists of five stages, each driven by a dedicated language-model agent:

1. Strategy Generator: proposes diverse high-level strategies for the target, each anchored on a key disconnection

2. Route Builder: expands every strategy into a complete pathway, expressing each disconnection in ReactionJSON

3. Critic: simulates each reaction in the forward direction and flags those that are chemically not feasible

4. Editor: repairs them through surgical edits that leave the strategy intact

5. Analyst: scores the finished route for feasibility and identifies its key steps and risks

The ReactionJSON representation is central: "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."

The paper examines three demanding targets to demonstrate strategic reasoning:

Okaramine M: SynthEx recovers a route published after the model's training cut-off (Gemini 3.1, with documented knowledge cut-off of January 2025). SynthEx independently recovers a route published after its training cut-off, including the key tandem prenylation and iminium-trapping step. The authors note SynthEx even offers a defensible refinement: The literature route installs the TIPS group on pre-Okamauromine, whereas SynthEx protects from the first step, which would avoid the mixture of mono- and bi-protected pre-Okamauromine that late-stage protection invites.

Melonine: SynthEx converges on the same key disconnection as an expert group, but through a different path that the group judges more likely to succeed. The paper explains: "Because SynthEx generates its iminium through an aza-Cope rearrangement rather than by condensation, the CH2–CH2 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."

Chanoclavine to Lysergol: Given only a starting material and target, SynthEx proposed a Hofmann–Löffler–Freytag reaction sequence: "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."

The authors demonstrate that SynthEx occupies a distinct reaction space from template-based tools. Key findings:

• NameRXN, an expert-curated dictionary independent of training corpora, recognizes SynthEx reactions at parity with USPTO reactions

• USPTO-derived classifiers recognize 15–25 percentage points fewer SynthEx reactions

• RetroChimera, a state-of-the-art single-step model, recovers only 13.5% of SynthEx disconnections at top-1 and 31.4% at top-5

• Ring-forming disconnections are particularly poorly recovered: top-1 recovery falls to 2.3% and top-5 to 10.9%

The clearest signature is ring construction: 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.

On a benchmark of 1,098 natural products from NP-Atlas, SynthEx dramatically outperforms a conventional planner: AiZynthFinder expands a median of roughly 2.9 × 104 nodes yet solves only 13.8% of the benchmark (151/1,098). By contrast, SynthEx returns complete routes for 63.9% of targets (702/1,098), close to a five-fold increase in reach.

The advantage widens with complexity: on the complex, large subset, AiZynthFinder solves only 4% while SynthEx retains most of its reach. The control subset shows this is a reaction-space problem, not a search-budget problem: 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.

The most notable result is the blinded expert assessment: In blinded assessments, expert chemists judged its key steps comparable to those of published human syntheses. Ten synthetic chemists from total-synthesis groups rated 148 unique key steps on four axes (feasibility, strategic value, elegance, overall quality), yielding 1,040 ratings. The results: "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])."

Strategic value was the only axis with a detectable gap (δ = −0.14, [−0.21, −0.03]), though this one detectable gap is smaller than the disagreement among the raters themselves. Crucially, 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 — the expert panel could not reliably distinguish SynthEx's steps from literature steps.

The paper describes an iterative critique-and-repair loop: "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 approximately 0.27 before any repair to approximately 0.06 after six iterations."

The Monascuspirolide A case study illustrates the pipeline, where the Editor reorders steps, moves reactions to earlier stages, and inserts protections to resolve flagged issues.

The authors release SynthAtlas, an open, interactive resource comprising 1,098 natural-product targets, 3,243 routes with corresponding synthetic strategies, and 33,145 reactions. Every reaction carries a complete atom-mapping, produced by construction from the graph-edit representation rather than by a post hoc mapper.

The authors propose complex natural-product synthesis as the setting in which planners should now be measured, arguing it probes chemical reasoning more directly than any patent-derived set. They acknowledge important limitations: 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 improvement loop is an internal consistency procedure, not an external validation: the agent that scores the improvement shares a backbone with the agents that produce it.

The authors suggest 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, with the integration of reaction conditions and, ultimately, closed-loop validation in the laboratory remaining the next frontier.

Based on this paper, I would make the following specific improvements to AI systems for chemical synthesis planning (and similar strategic-planning domains):

1. Replace template-based disconnection with generated, atom-level graph-edit policies. Instead of ranking known reaction templates, the AI should express every disconnection as an ordered list of atom-mapped graph edits (like ReactionJSON). This lets the AI propose transformations never seen in training data.

The improved system can invent bespoke ring-closing cascades, rearrangements, and tandem reactions for complex natural products, rather than failing when no memorized pattern applies.

2. Add a strategy-first hierarchical planner. The AI should first propose diverse high-level synthetic strategies, each anchored on a key disconnection, then independently expand each strategy into a complete route.

The improved system can explore multiple strategic directions and commit to those with global coherence, avoiding the myopia of single-step retrosynthetic searches.

3. Build a self-correcting loop with forward-simulation critique and surgical repair. Generated reactions should be simulated forward to detect infeasibility; the AI should then propose minimal edits that restore chemistry without changing the overall strategy. Run this loop iteratively.

The improved system can reduce blocking reactions (from 27% to 6% in the paper) and produce routes that are internally consistent and chemically plausible.

4. Explicitly optimize for strategic value and elegance, not just feasibility. The paper shows expert chemists value strategic bond-breaking and aesthetic efficiency; current systems undervalue this. The AI should be trained or prompted to identify key disconnections and avoid redundant functional-group shuffling.

The improved system can generate synthetic routes whose key steps are statistically indistinguishable from published human syntheses, even under blinded expert review.

5. Reward ring-forming reactions more heavily. Template-derived training data heavily under-represents ring construction (USPTO 9.9%, AI predictions 2.8%, vs. SynthEx 16%). The AI should be explicitly encouraged to propose ring-forming disconnections.

The improved system can design routes for polycyclic natural products that stump conventional planners, increasing solve rates from ~14% to ~64% on a benchmark of 1,098 natural products.

6. Use structured, atom-resolved reaction output rather than SMILES or arrow-pushing text. Every reaction should be specified as a complete atom mapping produced by construction from graph edits.

The improved system can output unambiguous, machine-executable synthesis routes that can be validated by forward simulation, checked for atom conservation, and directly used to generate experimental plans.

7. Adopt the natural-product frontier as the primary evaluation environment. The AI should be benchmarked on complex, non-patent-derived targets (e.g., NP-Atlas) where rule-based systems fail—not on reaction databases that the system was trained on.

The improved system can be meaningfully compared to human total synthesis and will improve chemical reasoning rather than database memorization.

8. Enable human-AI strategy collaboration. The system should accept a human-specified high-level strategy and expand it into a full, refined, and scored route.

The improved system can act as an interactive synthesis assistant: a chemist proposes the key disconnection, and the AI works out the complete pathway, flags risks, repairs infeasible steps, and suggests improvements—sharply accelerating complex synthesis design.

9. Use a multi-agent architecture with role-separated agents (strategy generator, route builder, critic, editor, analyst) communicating through a shared structured format. Each agent should be fine-tuned for its specific task, and the critic/editor loop should be trained on simulated forward-reaction feedback.

The improved system gains modularity, interpretability, and reliability: each route decision is attributable to a specific agent, and failures can be traced and repaired without restarting the entire search.

10. Integrate uncertainty-aware scoring to flag experimental risk. The analyst agent should score each route not only for reachability but also for stereochemical uncertainty, selectivity concerns, and feasibility of conditions, and should explicitly list key steps and risks.

The improved system can surface warning flags on potentially problematic steps, allowing chemists to prioritize which steps need experimental validation—moving toward closed-loop lab integration.

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.