RetroChimera tests synthesis routes against chemists' preferences
RetroChimera searches for routes from a target molecule to available starting materials by combining models with different inductive biases. The Nature paper describes tests on rare reactions and plausible-looking incorrect suggestions. A team including Microsoft Research, GSK, and Novartis compared the proposals with organic chemists' preferences; evaluation in real discovery settings remains the next step.
Artificial Intelligence··Morning
Working backward from a target molecule
RetroChimera proposes synthesis routes that work backward from a desired target molecule toward simpler, available starting materials. The task requires choosing which chemical reactions to apply and in what order. A team including researchers from Microsoft Research, GSK, and Novartis uses a learned ensemble to combine the results of two components with different inductive biases rather than relying on one model's proposal. The Nature paper describes the same method and goal.[1], [2]
Rare reactions form a central test
The study focuses on two weaknesses of earlier synthesis-planning models: reactions that are rare in training data but strategically important to chemists, and incorrect suggestions that appear plausible. RetroChimera was tested outside its training data, on reaction classes with few examples, and on pharmaceutical companies' internal datasets. The researchers therefore examined more than performance on a familiar distribution, asking whether useful proposals persisted when the system moved to other chemistry datasets.[1], [2]
Chemist preference does not replace laboratory use
Organic chemists evaluated routes proposed by RetroChimera against comparison-model routes and some previously published syntheses. Both sources report that experts more often preferred the system's proposals. That evaluation provides an important human check on whether routes make chemical sense. Practicality in a real discovery program remains untested. The researchers plan next to test the model in real discovery settings and see how its proposals fit laboratory workflows.[1], [2]
