UCLA team develops machine learning tool to predict drug-building chemical reactions and speed discovery

A research team from the University of California, Los Angeles, and the University of Utah has developed a machine learning system to predict outcomes of asymmetric cross-coupling reactions used in drug synthesis. The effort was co-led The model was trained solely on results from four academic papers on nickel-based catalysts with varying ligands. It converts reaction components into numerical data and screens tens of thousands of structures to forecast which handed form of a molecule will form. The workflow supports predictions for reactions outside the original training data. Gallarati said physics-based computational tools are too expensive for large-scale predictions, so the team built statistical models that remain accurate while using less data and computation. Bucci said the approach reduces lab tests from 50 or 60 reactions to five or 10, saving weeks or months and lowering costs for materials that must be purchased or synthesized. Doyle said the workflow is transparent and can provide chemical insights even when predictions miss the mark. Sigman said the tool could help pharmaceutical companies optimize known reactions for new compound targets during clinical development. The work received support from the Swiss National Science Foundation, the U.S. National Science Foundation and the National Institutes of Health.