Microsoft Research has introduced RetroChimera, a new framework for predicting retrosynthesis—the process of working backward from a target molecule to identify simpler precursor molecules. According to a paper published in the journal Nature, the framework addresses the complexity and high cost of chemical synthesis by proposing high-quality reaction routes automatically.
Product LaunchesMicrosoft ResearchRetroChimera
Microsoft Research Releases RetroChimera to Improve Small Molecule Synthesis Prediction
RetroChimera utilizes an ensembling strategy that combines two distinct models with complementary strengths: R-SMILES 2 and NeuralLoc. R-SMILES 2 is a Transformer-based de-novo model that predicts precursors directly from the input molecule, offering flexibility but potentially prone to hallucinations. In contrast, NeuralLoc is a graph neural network (GNN) based model that selects reaction templates based on existing patterns, providing more reliable and grounded outputs. By learning how to rank and combine these predictions, RetroChimera can leverage the specialized advantages of each model across various reaction classes.
In blind tests, expert chemists preferred the disconnections suggested by RetroChimera over those from preceding models and established approaches. The framework is available on GitHub under an MIT license and can be accessed via Microsoft Foundry. While the models are highly capable, the developers recommend that real-world applications be risk-assessed and verified by chemistry experts to mitigate potential hallucinations.
Sources
- Improving synthesis prediction of small molecules at scale with RetroChimera （Microsoft Research, 2026-09-21）
- RetroChimera