Microsoft Releases RetroChimera AI for Chemical Synthesis
Microsoft has open-sourced RetroChimera, an AI framework published in Nature that combines two distinct models to predict chemical synthesis routes and accelerate drug discovery.
Microsoft Research has open-sourced RetroChimera, a new machine learning framework designed to automate retrosynthesis, the process of planning how to create complex molecules from simpler building blocks. Published in the journal Nature, the system is now available under an MIT license on GitHub and through Microsoft Foundry. RetroChimera addresses a major bottleneck in medicine and materials science by predicting high-quality chemical synthesis pathways, helping researchers bypass manual and costly planning phases.
The framework achieves its high accuracy by combining two distinct deep learning architectures with complementary strengths. The first, R-SMILES 2, is a Transformer-based de-novo model that predicts precursor molecules directly from the input, excelling at reactions that involve large molecular changes but occasionally suffering from hallucinations. The second, NeuralLoc, is a graph neural network that matches target molecules to a library of reaction templates, making it highly reliable for localized or low-precedence reactions but limited by its template library. RetroChimera uses a learned ensembling strategy to rank and merge the predictions of both models, leveraging the strengths of each.
In validation studies, RetroChimera significantly outperformed existing baselines. When tested on ten highly challenging target molecules, RetroChimera successfully proposed viable multistep synthesis routes for nine of them. In comparison, the standalone de-novo model succeeded on five, an editing model succeeded on four, and NeuralSym, a strong baseline model, succeeded on only two. Furthermore, blind tests revealed that PhD-level chemists preferred RetroChimera's individual reaction predictions over those of preceding models and even recorded literature reactions.
For chemical and pharmaceutical practitioners, this development provides a highly reliable tool to scale up the design-make-test cycle. By combining the flexible generation of Transformers with the grounded constraints of graph neural networks, RetroChimera allows researchers to confidently explore novel, rare, and complex chemical reactions. This reduces the specialized expertise required for synthesis planning, paving the way for closed-loop, automated laboratory systems.
This is our own summary of reporting by Microsoft Research Blog
