Key Takeaway
A multi-agent AI system named SynthEx can design complete synthetic routes for 63.9% of challenging natural products (vs 13.8% for a leading template-based planner) by generating atom-level reaction edits and planning strategy-first.
ON THIS PAGE
Core Insights
Planning strategy before individual steps and letting a language model write atom-by-atom reaction edits lets the system propose chemistry outside standard reaction libraries. SynthEx finds many more routes for structurally complex natural products, favors convergent and ring-forming bond constructions, and produces shorter, more human-like routes. Expert chemists rated its key steps comparable to published human syntheses, and the authors released 3,243 predicted routes for 1,098 targets as an open resource called SynthAtlas for Multi-Agent Scientific Research.
Data Highlights
1SynthEx returns complete routes for 63.9% of a 1,098-target complex natural-product benchmark versus 13.8% for a leading template-based planner (about a five-fold increase).
2SynthAtlas release: 1,098 targets, 3,243 routes, and 33,145 atom-mapped reaction steps available for inspection and reuse.
3Ring-forming steps are much more common in SynthEx output (16.0% of reactions) than in patent-derived or template-based outputs (around 2.8%).
What This Means
AI engineers building multi-agent systems and planning pipelines can adopt the strategy-first, editable-reaction representation to break out of template limits. Synthetic chemists and pharma R&D teams benefit by getting creative, human-like route ideas and convergent disconnections they might not think of, accelerating ideation and experimental planning Semantic Capability Matching Pattern.
Explore evaluation patternsSee how to apply these findings
Key Figures
Fig 1: Figure 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 centered 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 distinguished. 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 Cyclopiamine B, a cascade aza-Michael/Michael addition to close two rings in a single step.
Fig 2: Figure 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–Löffler–Freytag sequence unprompted. Colored atoms and bonds highlight the reaction center in the key step of each route.
Fig 3: Figure 3 : SynthEx occupies a distinct reaction space that template-based tools cannot 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 SynthEx 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 k predictions ( k = 1 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 ReactionJSON graph-edit output, and representative ring constructions.
Fig 4: Figure 4 : General results on the SynthEx benchmark. a , Comparative solve rates of SynthEx and AiZynthFinder on several benchmark subsets, detailed in Supplementary Information: 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 (heterogeneity) 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 SynthEx routes.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Routes are hypothetical and not experimentally validated: stereochemical outcomes and reaction conditions were not verified in the lab. The system can still produce selectivity or feasibility issues that require expert review. Running the language-model backbone is computationally costly compared with template lookups, and expert guidance (e.g., supplying strategy constraints) remains valuable Hallucination Propagation.
Methodology & More
SynthEx is a multi-agent synthesis planner that flips the usual search order: it first generates multiple high-level strategies for how to break the target into parts, then expands each strategy into a full route. Instead of choosing from a fixed library of named reactions, the language model writes reactions as ordered atom-level graph edits (a format called ReactionJSON ReAct Pattern (Reason + Act)). A Critic simulates forward chemistry to flag impossible steps and an Editor applies surgical repairs while keeping the overarching strategy intact; an Analyst then scores and annotates the final routes. This editable, text-centered representation lets the system propose and iteratively refine novel, context-sensitive chemistry without re-running expensive searches.
Benchmarked on 1,098 structurally complex natural products (no known total synthesis in the source database), SynthEx produced complete routes for 63.9% of targets—close to five times the reach of a state-of-the-art template planner. Its proposals are qualitatively different: more convergent, shorter (median route length 5 vs 11 in a comparison subset), and richer in ring-forming and bond-constructive steps. In blinded expert review, key disconnections from SynthEx were rated on par with literature examples. The authors released all results as SynthAtlas (3,243 routes, 33,145 atom-mapped steps) to serve as a public corpus for idea generation and future training/testing. The next steps for the field are integrating reaction conditions, stereochemical validation, and experimental closed-loop testing to turn promising paper routes into proven lab syntheses. Handoff Pattern
Need expert guidance?We can help implement this
Credibility Assessment:
Multiple authors with modest h-indices (several between ~5–11) and some recognizable names in chemistry/ML, but venue is arXiv and no top‑tier institutional/major lab affiliations listed — solid but not top-tier credibility.