Chiral atom-centered scalar triple products for predicting enantiomer elution order in HPLC.

The assignment of the stereochemical configuration of enantiomers is crucial for structure elucidation in drug discovery and molecular design. Machine learning models trained on experimental data can predict chiral observable properties and assist in configuration assignment by comparison with experimental results. We implemented scalar triple product (STP) descriptors with atomic properties from RDKit libraries to create chiral-atom-centered variants (cSTP) that more effectively encode molecular chirality. Using four high-resolution liquid chromatography datasets with different chiral stationary phases (Chiralpak AD-H, CROWNPAK CR(+), CROWNPAK CR-I(+), and Lux cellulose-1), we trained Random Forest models to predict enantiomer elution order. The cSTP descriptors, particularly those calculated within a single-bond sphere around chiral centers, performed closely to conventional Morgan fingerprints-superior to the latter in generalizability and with specific datasets. The best-performing cSTP_1 descriptors achieved 77-100% accuracy on unseen test sets and superior performance on molecules structurally dissimilar to the training set. These results demonstrate that chirality-sensitive molecular representations centered on stereogenic centers are highly effective for predicting enantioselectivity in chromatography and suggest potential applications in automated configuration assignment and quality control of chiral compounds.