Microsoft Research published RetroChimera in Nature. It is a new framework for retrosynthesis prediction. It combines two models with complementary strengths. It ranks their proposals to outperform either alone. PhD chemists preferred its predictions over prior models and literature. It achieved a 90% expert acceptance rate for complete routes. It solved nine of ten challenging targets in tests. The implementation and weights are open-sourced. It is available on GitHub and Microsoft Foundry. The team invites community experimentation and feedback.
What changed
Manual, time-consuming, and costly chemical synthesis planning dominated the field.
What it unlocks
Automatically proposing high-quality synthesis routes for new molecules at scale.
- 90% expert acceptance rate for RetroChimera routes
- 20-50% for sub-models or baseline
- 9 of 10 challenging targets solved
What you need to act on it
- access to GitHub or Microsoft Foundry
- MIT license compliance
Sources