Mining human mitochondrial and mitochondrial associated proteins based on SVM and neural network

Human mitochondrial proteins are involved in fundamental biological process including apoptosis, energy production and many metabolic pathways, prediction of mitochondrial proteins is a major challenge in genome annotation. In this study, we implemented a machine learning approach and developed reliable neural network and SVM based methods to classify human mitochondria proteins with high confidence. We used experimentally characterized human mitochondria proteins as positive training datasets and human proteins localized in other organelles as negative training datasets for neural network, support vector machine, naive bayes and bayes network classification. In addition, we constructed simple amino acid composition model, a hybrid model of simple amino acid composition combining amino acid chemical-physical properties, and dipeptide amino acid composition model. With 5 fold cross-validations, the results demonstrate that multiple perceptrone neural network performs better than SVM in all three training models. We concluded that our classification approach utilizing empirically characterized human mitochondria protein sequences is a valuable tool for classifying human mitochondria proteins.