A series of ball–disk contact friction tests were carried out using a microtribometer to study the tribological characteristics of steel/steel rubbing pairs immersed in 47 different organic compounds as lubricant base oils. The structures and their friction data were included in a back-propagation neural network (BPNN) quantitative structure tribo-ability relationship (QSTR) model. Following leave-one-out (LOO) cross-validation, the BPNN model shows good predictability and accuracy for the friction parameter (R2 = 0.994, R2(LOO) = 0.849, and q2 = 0.935). Connectivity indices (CHI) show the large positive contribution to friction, which imply that friction performance has a strong correlation with molecular structure. The BPNN–QSTR models can flexibly and easily estimate the friction properties of lubricant base oils.