RetroChimera

A single-step retrosynthesis model from Microsoft Research AI for Science in Cambridge, UK. Given a target molecule, it proposes reactant sets that could make it, the step that multi-step synthesis-planning search calls recursively. RetroChimera ensembles two sub-models introduced with it. R-SMILES 2 is a Transformer that generates precursors de novo; it does well on reactions that change the molecule a lot but can hallucinate. NeuralLoc is a graph neural network that selects a reaction template and predicts where on the target to apply it; its outputs are more reliable and it learns a template from a few examples, but it cannot go beyond its template library. A learned, rank-dependent vote merges their ranked outputs (votes add when both models propose the same reactants), so the ensemble roughly matches whichever sub-model is stronger on each reaction class. The date marks the first public release of the code and weights (v1.0.0); the arXiv preprint was first posted on December 6, 2024, and the peer-reviewed Nature paper ("Chemist-aligned retrosynthesis by ensembling diverse inductive bias models", 26 authors, received August 2025) appeared with Microsoft Research's announcement on September 21, 2026.

The main checkpoint is trained on the proprietary Pistachio reaction database, which the preprint describes as 3.5 times the size of USPTO-FULL, and tested on a time split of reactions added in 2024. In the preprint, RetroChimera set a new state of the art on USPTO-50K and USPTO-FULL beyond top-1, raising top-10 accuracy by 1.7% and 1.6%. In blind comparisons, PhD-level organic chemists preferred its predictions over the published reference reactions and over other AI models. In an expert assessment of complete multi-step routes for ten challenging targets, it succeeded on nine, against five for the de novo sub-model, four for the editing sub-model and two for the NeuralSym baseline. The paper also reports zero-shot transfer and fine-tuning on internal reaction data from two pharmaceutical companies; the preprint's zero-shot test used Novartis data, and the Nature author list adds GSK chemists alongside researchers from Cambridge, Jagiellonian University and Wuppertal. Corresponding authors are Krzysztof Maziarz, Guoqing Liu and Marwin Segler; Christopher Bishop is a co-author.

Code and checkpoints are MIT-licensed on GitHub and PyPI (package retrochimera), built on Microsoft's syntheseus library, and the model is also offered through Microsoft Foundry. The code and the Pistachio, USPTO-50K and USPTO-FULL checkpoints (on figshare) have been public since v1.0.0 on November 30, 2025; a forward reaction-prediction checkpoint was added in July 2026, and v1.3.0, released with the paper, added fine-tuning and a consensus ensembling mode. The README warns that the checkpoint covers reaction data only through 2023, that lower-ranked outputs are increasingly likely to be hallucinations, and that chemistry experts must verify predictions before use. The repository had 74 stars at filing.