Improving synthesis prediction of small molecules at scale with RetroChimera
Microsoft Felix Pultar, John Gardner, Guoqing Liu, Marwin Segler
Microsoft Research just open-sourced RetroChimera, a model that plans chemical synthesis routes. It beat older systems and even chemists’ expectations in blind tests.
Based on reporting by Microsoft, Felix Pultar, John Gardner, Guoqing Liu, Marwin Segler — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft Research has published RetroChimera, a retrosynthesis model that tries to do for molecule-making what search tools did for information retrieval: cut through a messy problem and surface good options faster. The work is in Nature, and the team has also released the implementation and weights on GitHub, with access through Microsoft Foundry as well.
The core idea is an ensemble. RetroChimera combines two models with different strengths: R-SMILES 2, a Transformer-based de-novo system that predicts precursors directly, and NeuralLoc, a graph neural network that works from reaction templates. One is more flexible but can hallucinate. The other stays closer to known chemistry but can get boxed in when a reaction is outside its template library.
That mismatch is the point. The two systems do not just vote; RetroChimera learns how much to trust each one at different ranks, then re-ranks their proposals into a single prediction. Microsoft says the blend helps it handle both common and rare reaction classes, and that it can cope better with reactions involving bigger structural changes as well as those with more localized changes.
The validation is the part that will get chemists’ attention. In blind tests, PhD-level chemists preferred RetroChimera’s individual reaction predictions over earlier models and recorded literature reactions. In expert review of ten challenging multistep targets, RetroChimera succeeded on nine, compared with five for the de novo model, four for the editing model, and two for NeuralSym, which Microsoft describes as a strong baseline.
The broader pitch is straightforward: if synthesis planning gets faster and less manual, researchers can test more candidate molecules in drug discovery, smart materials, and other molecular science work. Microsoft is also hinting at a future where this kind of planner sits inside more automated, closed-loop lab systems. That’s the real story here: not just another chemistry model, but a push to make synthesis planning less artisanal and more scalable.
My take — AI-written commentary, not fact-checked reporting
The interesting bit is not that AI can suggest chemistry routes; plenty of models can make a plausible guess and call it science. It’s that the better approach here is an ensemble with a learned referee, which is a quiet admission that chemistry still punishes overconfident one-model swagger. Open-sourcing the weights is the right move too, because in chemistry, closed boxes tend to age badly and get adopted slowly.
Read more about this at: Microsoft