Protein Subcellular Classification using Machine Learning Approaches

Abstract

Subcellular localization of proteins is one of the most significant characteristics of living cells that may reveal plentiful information regarding the working of a cell. Subcellular localization property of proteins plays a key role in understanding numerous functions of proteins. The proteins, located in their respective compartments or localizations, are in- volved in their relevant cellular processes, which may include cell apoptosis, asymmetric cell division, cell cycle regulation, and spermatic morphogenesis. In fact, cells may not perform their regular operations well in case proteins are not found in their proper subcellular lo- cations. Improper localization of proteins may lead to primary human liver tumors, breast cancer, and Bartter syndrome. Protein sequencing has observed rapid expansion due to the advancement in genomic sequencing technologies. This led the research community to recognize the functionalities of different proteins. In this connection, microscopy imaging is providing protein images well in time with low cost compared to protein sequencing. However, automated systems are required for fast and reliable classification of these protein images. Comprehensive analysis of fluorescence microscopy images is required in order to develop efficient automated systems for accurate localization of various proteins. For this purpose, representation of microscopy images with discriminative numerical descriptors has always been a challenge. This thesis focuses on the identification of discriminative feature extraction strategies effective for protein subcellular localization, the recognition capability of the prediction sys- tems, and the reduction of classifier bias towards the majority class due to the imbalance present in data. The contributions of this thesis include (1) Analysis of different spatial and transform domain features, (2) Development of a novel idea for GLCM construction in DWT domain, (3) Analysis of SMOTE oversampling in the feature space, (4) Analysis of GLCM in the spatial domain for capturing discriminative information from fluorescence microscopy protein images along different orientations, (5) Exploitation of Texton images for their capability of extracting discriminative information along different orientations from fluorescence microscopy protein images, (6) Development of the web based prediction sys- tems that can be accessed freely by the academicians and researchers. Extensive simulations are performed in order to assess the efficiency of the proposed pre- dictions systems in discriminating different subcellular structures from various datasets.