In this study, a novel protein submitochondia locations dataset was constructed which contained 1 293 proteins classified into four kinds of submitochondria locations. The GO information and homologous information was extracted to combine the feature vectors of proteins and the Supported Vector Machine algorithm was used to construct the classifier. As a result, by using the Jackknife Cross-Validation, an accuracy of 93.27% for four kinds of protein submitochondria locations and that of 94.73% for three kinds of protein submitochondria locations was obtained. Especially, the predictive accuracy for outer membrane of protein submitochondia locations was enhanced than previous methods. The data set of protein submitochondia locations constructed by ours has the intermembrane proteins compared to old ones. The intermembrane proteins have important functions in protein apoptosis. The integrity of data set and the improvement of prediction accuracy can help to understand the cell activity and internal biochemical process.