Microsoft releases RetroChimera, an AI model for faster chemical synthesis planning

Finding efficient ways to synthesize new molecules remains one of the biggest bottlenecks in drug discovery and materials science. Microsoft Research has published a solution: RetroChimera, an AI framework that automatically designs synthesis pathways by working backwards from a target molecule to purchasable building blocks—and it outperforms existing approaches in the eyes of expert chemists.

How it works#

RetroChimera combines two models with complementary strengths. R-SMILES 2, a Transformer-based generator, predicts precursor molecules directly from an input compound, giving it flexibility to learn novel reaction patterns but also making it prone to generating unrealistic suggestions. NeuralLoc, a graph neural network, grounds its predictions in known reaction templates, making it more reliable for familiar chemistry but limited when encountering rare reactions.

Rather than choosing between them, RetroChimera uses a learned voting system that assigns rank-dependent weights to each model's proposals. When both models suggest the same reaction, their votes combine. This ensemble approach leverages what each model does best: R-SMILES 2 excels at reactions involving large molecular rearrangements, while NeuralLoc specializes in less common reactions and localized transformations.

Validation by human experts#

In blind evaluations, PhD-level chemists rated individual reaction predictions from RetroChimera as superior to those from either constituent model or established baseline systems. More tellingly, when chemists assessed complete synthesis routes, RetroChimera-generated paths achieved a 90% acceptance rate—compared to only 20–50% for routes from other methods. Since a single wrong step invalidates an entire route, this dramatic improvement reflects genuine practical value.

Release and applications#

RetroChimera is available on GitHub under an MIT license and can be accessed via Microsoft Foundry. The full results appear in Nature.

The tool could accelerate the design-make-test cycle in molecular science by helping researchers rapidly screen more candidate molecules and identify viable synthesis strategies. When paired with laboratory automation, such systems could eventually enable closed-loop discovery platforms that plan, execute, and learn from synthesis experiments autonomously.