How RetroChimera predicts molecular synthesis routes

Microsoft Research reports that experts accepted complete proposed routes for 9 of 10 challenging targets in an assessment of RetroChimera. The count measures expert judgments of proposed routes; it is not a tally of laboratory syntheses. RetroChimera predicts how a target molecule might be broken into simpler precursors, combining two different prediction models to rank possible reaction steps.

What RetroChimera predicts

Retrosynthesis works backward from a target molecule: it proposes reactions that could break that molecule into simpler precursors, potentially continuing toward purchasable building blocks. RetroChimera ranks possible reaction predictions, and a search can connect individual steps into proposed multistep routes. Microsoft Foundry describes the model’s input as a molecule represented in SMILES, a text format for encoding chemical structures. (Microsoft Research; Microsoft Foundry Labs)

How RetroChimera combines two models

RetroChimera combines R-SMILES 2, a Transformer-based model that generates precursor molecules de novo, with NeuralLoc, a graph-based model that applies reaction templates. A reaction template represents a known pattern for transforming one structure into another. R-SMILES 2 can generate proposals flexibly; NeuralLoc draws on its template library. A learned, rank-dependent voting method combines and reorders their predictions, and a reaction proposed by both models can receive votes from each. (Microsoft Research)

What experts accepted in the 10-target assessment

Microsoft Research reported expert acceptance of complete routes for 10 challenging targets. The counts for RetroChimera and three comparison systems were:

The figures describe accepted proposed routes in this assessment. They offer a comparison of the systems on the same target set, rather than a general success rate across molecules or a measure of completed synthesis. (Microsoft Research)

A separate chemist-preference study

The research paper describes a different evaluation: nine PhD-level organic chemists made 599 pairwise comparisons, and preferred RetroChimera’s top prediction over the dataset’s ground-truth reaction at a mean rate of about 64%. This measures chemists’ preferences in those comparisons, not the proportion of complete routes accepted in the 10-target assessment. (Research paper)

What the company-data evaluations measured

The paper reports a zero-shot evaluation on 10,444 internal Novartis reactions, without additional fine-tuning for that test. It also describes evaluations using the public USPTO-50K and USPTO-FULL datasets and the proprietary Pistachio dataset. Separately, Microsoft Source reported that the pretrained model was adapted to GSK’s internal chemical data; that adaptation is distinct from the Novartis zero-shot test. (Research paper; Microsoft Source)

Nature published the study on September 21, 2026. The company-data evaluations and expert assessments describe computational and expert review of predictions; the paper’s route illustrations also leave out key experimental details. (Nature)

Why a predicted route is not a lab protocol

The paper’s illustrated routes omit reagents, solvents and reaction conditions, so they do not provide a complete experimental procedure. The authors also warn that models can hallucinate, particularly for reactions farther from their training distribution, and identify systematic errors in training data as another limitation. (Research paper)

Microsoft Research says RetroChimera’s implementation and weights are available on GitHub under the MIT license, and identifies Microsoft Foundry as another access route. The Microsoft Foundry Labs listing labels the model “Experimental.” (Microsoft Research; Microsoft Foundry Labs)