Machine learning-assisted molecular design for efficient 19F hyperpolarization
TL;DR
A machine-learning-assisted strategy for identifying and ranking substrate-specific structural and electronic factors that contribute to efficient 19F signal amplification by reversible exchange.
Abstract
Efficient 19F signal amplification by reversible exchange (SABRE) remains difficult because enhancement depends on multiple substrate-specific structural and electronic factors. Here we report a machine-learning-assisted strategy for identifying and ranking high-performance 19F SABRE substrates. A chemical space of more than 180 000 fluorinated N-heterocycles was mapped using refined molecular descriptors, and 33 representative fluorinated pyridine substrates were selected for experimental SABRE screening. Random forest classification identified N_Sum_of_connectivity as a key descriptor separating high- and low-enhancement substrates. For substrates classified in the high-enhancement regime, forward stepwise multivariate linear regression with leave-one-out cross-validation produced an interpretable five-descriptor model for quantitative signal-enhancement-factor prediction. External validation with six additional substrates showed good agreement between predicted and experimental values for high-enhancement candidates. Notably, 3-amino-4-fluoropyridine gave an experimental enhancement of 7133-fold, close to the predicted value of 6637-fold, corresponding to approximately 3.2% polarization at 1.4 T without co-ligands. The selected substrates further enabled single-scan simultaneous hyperpolarization of 15 fluorinated pyridines. This workflow provides an interpretable framework for data-driven optimization of 19F SABRE substrates.