Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning.

Active and geometric deep learning combined in a closed-loop workflow accurately predicts reaction outcomes and regioselectivity in drug discovery, achieving perfect pinpointing on challenging substrates.

• Why it matters: Accurate prediction of chemical reactions is crucial for efficient drug design, but current models are limited by scarce reference data and difficulty in predicting regioselectivity, hindering progress.
• What they did: Researchers developed an active learning strategy using a tree-based ensemble to guide laboratory experiments, creating a diverse C-H borylation benchmark dataset, and trained geometric graph neural networks with self-supervised tasks.
• The result: Augmenting symmetry-aware models with auxiliary tasks improved prediction performance, and prospective tests on complex N-heteroaryl substrates correctly identified borylation sites in all cases, demonstrating practical utility.