Measures for fuel auto-ignition behavior in form of research octane number (RON), motor octane number (MON), and octane sensitivity (OS) are heavily used for evaluation of fuel knock resistance. In order to allow their rapid assessment for novel fuel candidates, a number of quantitative structure-property relationship (QSPR) models have been proposed. While reasonable prediction accuracies have been achieved for predefined sets of fuel types and molecular groups as input features, the detailed effects of the structural groups and additional descriptors for the positions of side chains, double bonds, or oxygenated functionalities deserve further investigation. The tailored definition of such descriptors based on prior knowledge of the fuel ignition chemistry is expected to be particularly important due to the relative sparsity of available training data. The present work targets these aspects by systematically defining and incorporating such input features in QSPR models to predict RON, MON, and OS based on artificial neural networks. A set of experimental RON and MON data for 260 pure fuel components, including non-oxygenated and oxygenated hydrocarbons, serves as training basis. A backward feature elimination method is employed to reduce the numbers of predefined features of each model, finding that the OS model requires less input features for maximum prediction accuracy than the RON and MON models, while this direct modeling approach still yields lower prediction errors than retrieving the OS as difference of predicted RON and MON values. Finally, a sensitivity analysis on the input features provides insight into their effects on the predicted target quantities. [ABSTRACT FROM AUTHOR]