Prediction of Subnuclear Location for Nuclear Protein

摘要

To play a biomolecular function, a protein must be transported to a specific location of cell. Also in a nucleus, a nuclear protein has its own location to fulfil its role. In this study, subnuclear location of nuclear protein was predicted from protein sequence by using deep learning algorithm. As a dataset for experiments, 319 non-homologous protein sequences with class labels corresponding to 13 classes of subcellular localization (e.g. "Nuclear envelope") were selected from public databases. In order to achieve better performance, various combinations of feature generation methods, classification algorithms, parameter tuning, and feature selection were tested. Among 17 methods for generating features of protein sequences, Composition/Transition/Distribution (CTD) generated the most effective features. They were further selected by randomForest package for R. Using the selected features, quite high accuracy (99.91%) was achieved by a deep neural network with seven hidden layers, maxout activation function, and RMSprop optimization algorithm.