<p>This study presents a combinatorial machine learning approach for predicting the enantiomeric ratio ln(er) in cinchona-derived phase-transfer catalysts, based on 471 experimental ln(er) values and 47 molecular descriptors selected from a larger pool. These descriptors, primarily consisting of 2D autocorrelations and WHIM descriptors, mechanistically capture electronic and steric influences. Quantitative structure-selectivity relationship (QSSR) models were constructed using random forest and multilayer perceptron algorithms, and predictions from these models were combined through averaging to enhance predictive reliability. Robust performance was demonstrated by the ensemble model, achieving a determination coefficient (R²) of 0.778 and a root mean square error (RMSE) of 0.280 for the test set (94 ln(er)). Through mechanistic analysis, enantioselectivity was found to be enhanced by steric dissimilarity at intermediate bond distances, such as GATS5v and GATS2v, and uniform polarizability, as represented by MATS4p. In contrast, electronic irregularities, particularly long-range variations in ionization potential (GATS7i) and electronegativity (GATS7e, GATS4e), were observed to reduce ln(er). Fundamental insight into structural determinants of enantioselectivity is provided by this work, and a computational framework for catalyst design is established.</p> Graphical abstract <p></p>