Microsoft open-sources RetroChimera, a retrosynthesis model for small-molecule synthesis planning

Microsoft Research announced RetroChimera, a retrosynthesis model published in the journal Nature that proposes synthesis routes for small molecules by working backward from a target molecule to purchasable building blocks.

The model combines two sub-models with complementary strengths: R-SMILES 2, a Transformer-based de novo model that predicts precursors directly from the input molecule, and NeuralLoc, a graph neural network model that selects reaction templates and applies them to the target molecule.

RetroChimera uses a learned ensembling strategy to rank and combine predictions from both sub-models, leveraging their complementary strengths across reaction types, including rare reactions.

In blind tests, PhD-level chemists preferred RetroChimera’s individual reaction predictions over preceding models and recorded literature reactions, and the model succeeded on nine out of ten challenging targets in expert assessments of multistep synthesis routes.

The team reports successful zero-shot transfer and fine-tuning on proprietary datasets, and they open-sourced RetroChimera’s implementation and weights to enable researchers to accelerate development of medicinally relevant molecules and advanced materials.