Microsoft Research published the RetroChimera model for predicting molecular retrosynthesis

The research division Microsoft Research published the RetroChimera model in the journal Nature, a framework for predicting retrosynthesis — a process in which a target molecule is broken down backwards into simpler starting materials so that it can actually be made in a laboratory. The model combines two different approaches: R-SMILES 2, a transformer model that generates molecular precursors directly from data (flexible but prone to hallucinations), and NeuralLoc, a model based on a graph neural network that selects and applies reaction templates trained on data (more accurate but limited to known templates). RetroChimera combines the ranked predictions from both models using a learned voting mechanism that weights their proposals at individual positions in the ranking according to their reliability.

According to Microsoft, chemists with PhDs in blind tests preferred predictions for individual reactions from RetroChimera over predictions from both component models, over older established methods and over recorded reactions from the scientific literature. According to the source, the model performs well on both common and rare reaction types.

RetroChimera is available on GitHub under the MIT license and is also accessible through the Microsoft Foundry service; the company provides instructions for obtaining the trained model (checkpoint) in the repository on GitHub. Microsoft Research states that the aim is to accelerate the cycle of designing, making and testing molecules across fields such as drug development or smart materials design, and, in combination with laboratory automation, it expects progress towards self-improving systems for planning and carrying out synthesis.