In this work, we advanced emerging descriptor-based screening and machine learning techniques to identify new families of catalysts to be synthesized. We present a computaional flow to evaluate the catalysis performance of single-atom catalysts for oxygen reduction reaction (ORR). The onset potentials of single-transitional metal on graphene defects computed by density funtional theory (DFT) were used as our input data, and electronegativity of metal atoms and surrounding atoms (C, N or B), the coordinate numbers of central metal atoms and neighbor atoms were set as the intrinsic features for predicting catalysis performance. We used random forest algorithm to train the machine learning model, and employed this moderl to generate prominent single-atom catalysts for ORR. The predictive model can predict the onset potential directly from the structural properties of single-atom-base systems, which is a great progress from the models which can only predict the adsorption energy of intermediates. Therefore, we applied various single-atom catalysts systems with structural information, and the model screened out the most promising ones among them. This work demonstrates that the collocation of traditional computational chemistry and data-driven science can rapidly expand computing power and solve the most challenging catalysis problems!