Machine learning models supporting synthesis planning applications are largely limited to the chemistry seen in training, and the accuracy and diversity of their predictions are often diminished in sparsely populated chemical subspaces. By measuring how different datasets affect the performance of trained models, we can make stronger assertions regarding the expected coverage and novelty of synthesis planning solutions, and design datasets that will open up previously difficult areas of science.
In this study, scientists at Bayer demonstrate the significant impact that scientist-curated reactions from the CAS Content Collection have on the predictive power of a synthesis planning model. Accuracy in prediction of outcomes in rare reaction classes increased significantly – a boost of 32 percentage points – expanding understanding into new, useful chemistry.
