Data-Guided Descriptor Rebalancing for Monophosphine Ligand Training Set Development

Abstract Data-guided optimization strategies increasingly rely on well-designed training sets and meaningful chemical space representations, yet identifying appropriate descriptors and ensuring domain applicability remain persistent challenges. Here, we present an iterative workflow for refining monodentate phosphine-based training sets by coupling experimental reaction data with systematic feature evaluation. Using a pharmaceutically relevant cross-coupling reaction as a case study, we demonstrate how imbalances in steric and electronic descriptors of monodentate phosphines derived from the kraken database, along with limited feature overlap between training and validation sets, can challenge classification models and virtual screening performance. By rebalancing descriptor types, we generated a more chemically intuitive representation that improved predictive performance in a classification model. This approach enables the development of a redesigned monophosphine ligand screening set with an enhanced coverage of key structural and electronic dimensions. The resulting workflow provides a generalizable template for constructing more robust, diverse, and application-ready training sets for catalyst screening and reaction optimization within process chemistry.