RetroChimera AI Improves Retrosynthesis, Solves 9 of 10 Targets
RetroChimera combines two AI models to improve retrosynthesis, winning chemists' preference and solving nine of 10 targets. Code and weights are now open.
Summary
RetroChimera, presented by Felix Pultar, John Gardner, Guoqing Liu and Marwin Segler on September 21, 2026 and detailed in Nature, automates retrosynthesis from target molecules to purchasable building blocks. It combines R-SMILES 2, a Transformer that directly generates precursors and handles large reaction changes but can hallucinate, with NeuralLoc, a graph neural network that applies training-derived templates reliably and excels on rare, localized reactions but cannot move beyond its template library. Learned, rank-dependent votes merge their proposals, reinforcing reactions predicted by both and approximately matching the stronger model across reaction classes.
Blind PhD-level chemists preferred RetroChimera's reaction predictions over its submodels, established approaches and recorded literature reactions. Experts accepted complete routes for nine of 10 challenging targets, versus five for the de novo model, four for the editing model and two for NeuralSym. Validation also covered rare-reaction recall, zero-shot transfer and fine-tuning on proprietary datasets. The system could accelerate design, make and test cycles for medicines, smart materials and sustainable agriculture, particularly alongside laboratory automation, although broader community testing must expose its shortcomings. The implementation and weights are MIT-licensed on GitHub, with checkpoint instructions there and access through Microsoft Foundry.
Positives
- Experts accepted RetroChimera routes for nine of 10 challenging targets, compared with five, four and two for the tested alternatives.
- Blind PhD-level chemists preferred its reaction predictions over constituent models, established systems and recorded literature reactions.
- Zero-shot transfer and fine-tuning succeeded on proprietary datasets, indicating adaptability beyond the original training distribution.
- R-SMILES 2 and NeuralLoc contribute complementary strengths across large, localized and low-precedence reaction changes.
- MIT-licensed implementation and weights on GitHub give researchers open access, while Microsoft Foundry provides another route to the model.
Risks & concerns
- R-SMILES 2 can hallucinate because its direct precursor generation is unconstrained.
- NeuralLoc cannot predict reactions absent from its training-derived template library.
- Chemical synthesis remains manual, time-consuming and costly, limiting large-scale exploration of computationally designed molecules.
- RetroChimera still needs wider community testing to identify shortcomings across varied chemical targets.
- Closed-loop planning and execution remain dependent on increasing laboratory automation rather than RetroChimera alone.