ChemDraw-Guided Design of Descriptors for Quantitative Statistical Models

High Resolution Image Download MS PowerPoint Slide Quantitative models in transition-metal catalysis are typically restricted to narrow catalyst families, largely because no universal featurization strategy exists for systems that differ in ligand architecture, metal, and/or coordination geometry. On the example of regioselectivity in propene polymerization, we introduce a simple approach allowing the design of custom ligand-independent 3D-steric descriptors guided by ChemDraw depictions of competing transition states. A simple multiple linear regression model trained on a rigorously curated high-throughput dataset (86 catalysts, including octahedral and tetrahedral geometries, 2 metals, 3 families, and 8 classes) achieves high accuracy ( R 2 = 0.88) with only four 3D-steric descriptors and one electronic descriptor. Importantly, the model is fully interpretable: descriptors directly map onto specific transition-state interactions (and the ChemDraw depictions), enabling clear connections between ligand modifications and catalytic outcomes. This approach demonstrates how experimental intuition, encoded through ChemDraw depictions, can be translated into robust quantitative predictions, portable across catalyst classes.