Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learning

Late-stage functionalization is a cost-effective method for optimizing drug candidates. This research presents a platform combining geometric deep learning and high-throughput reaction screening to address the challenges of late-stage diversification, focusing on borylation. The computational model accurately predicted reaction yields (4–5% mean absolute error) and classified reactivity of known and unknown substrates (92% and 67% balanced accuracy, respectively). Regioselectivity was also predicted with a classifier F-score of 67%. Applied to 23 diverse drugs, the platform identified numerous diversification opportunities. The influence of steric and electronic factors on model performance was quantified, and a user-friendly reaction format facilitated the integration of deep learning and high-throughput experimentation.

David F. Nippa, Kenneth Atz, Remo Hohler, Alex T. Müller, Andreas Marx, Christian Bartelmus, Georg Wuitschik, Irene Marzuoli, Vera Jost, Jens Wolfard, Martin Binder, Antonia F. Stepan, David B. Konrad, Uwe Grether, Rainer E. Martin, Gisbert Schneider