The discovery of new transformations in organic synthesis is driven by chemical intuition, although accurate prediction of reaction outcomes using human expertise remains a major challenge. Predictive models learnt from large experimental datasets have the potential to shift this paradigm, with diverse applications ranging from catalyst design to drug discovery.1,2 However, data for a reaction of interest is often limited in scope and unreliable. Moreover, the de novo generation of the large quantities of high quality data required to fit multi-parameter machine learning models using traditional synthetic methods is currently unfeasible. High-throughput experimentation shows great promise to address the problem of large scale data generation; thousands of reactions can be carried out on a nanomolar scale utilising automated liquid handling platforms and rapid LCMS analysis.3 Here we use high-throughput approaches to generate significant quantities of yield data for the simple, yet powerful, SNAr reaction. We combine this data with both structural and physico-chemical descriptors to train a series of models that are able to accurately predict reaction yield for compounds that are not included in the training set. We carry out prospective experiments to validate the model and determine the regions of chemical space that it can accurately capture. Iterations of experiment and modelling in an active learning approach are used to determine how data collection capacity can be best leveraged to expand the model’s trust region in chemical space.

Beyond chemical intuition: Combining high-throughput experimentation with machine learning to build accurate models

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.