Robust out-of-distribution prediction of Buchwald-Hartwig reactions.

A new framework improves out-of-distribution prediction accuracy for Buchwald-Hartwig reactions, achieving reliable results on unseen substrates and conditions.

• Why it matters: Accurate predictive models are essential for advancing pharmaceutical synthesis, but current datasets are noisy, fragmented, and limited in scope, hindering model generalization.
• What they did: The authors integrated multiple reaction datasets into a standardized, high-quality database and used active learning to expand chemical space, training a model with enhanced predictive capabilities.
• The result: The resulting model successfully predicted novel reactivity and was validated through experimental reagent recommendations, paving the way for accelerated drug discovery and robust machine learning applications.