Active and geometric deep learning advances chemical reaction prediction in data-scarce drug discovery
Mason Minot, Yannick Stenzhorn, Jens Wolfard, Sebastian G. Strobel, Philippe Jablonski, Daniel Zimmerli, Martin Binder, Uwe Grether, Rainer E. Martin, Alex T. Müller, David F. Nippa, Kenneth Atz, Gisbert Schneider · ChemRxiv · 2025
Late-stage functionalization (LSF) is a critical approach for optimizing complex molecules during the advanced stages of drug discovery. Although computational models offer a pivotal opportunity to enhance LSF by predicting reaction outcome, their current impact is constrained by insufficient reference data and the challenge of accurately predicting reaction regioselectivity. To address these critical limitations, we present an integrated closed-loop workflow that combines efficient reaction data acquisition with high-accuracy deep learning. We developed an active learning strategy, driven by a computationally efficient tree-based ensemble, to guide the acquisition of the most informative laboratory experiments. This experimental campaign produced new reactions on novel and complex substrates, yielding the most diverse benchmark yet reported for C–H borylation. The profoundly expanded bespoke dataset allowed us to train and evaluate geometric graph neural networks. We demonstrate that augmenting these symmetry-aware models with self-supervised auxiliary tasks consistently improves performance for both reaction outcome prediction and regioselectivity prediction. Prospective tests on unseen substrates, featuring challenging N-heteroaryl motifs, pinpointed the correct borylation positions in all cases, thereby closing the loop between model-guided experimentation and deep learning, and illustrating the practical utility of our workflow for efficient drug optimization.