A computational and experimental study of protein localisation determinants in the mammalian endomembrane system

The subcellular localisation of a protein, together with its sequence and structure, provide the first information about its function. Although several approaches for determining localisation exist, the most widely used experimental methods involve the overexpression of a GFP-tagged construct of the protein in a cultured cell or staining the endogenous protein with specific fluorescently-labelled antibodies. Computational sequence-based localisation predictors have been developed, and are of value, but they are still limited in their predictive power. Considering the now extensive use of imaging approaches, it is therefore unsurprising that much effort has been put into the development of automated image analysis methods to classify localisation. Image classifiers are typically trained on ground truth data, which introduces certain bias. The aim of modern algorithms is to eliminate human interaction and instead perform unsupervised classification of the images. The study presented in this thesis addresses three aspects of protein localisation methodology: sequence-based localisation prediction with short linear motifs (SLiMs), unsupervised image analysis with texture features and experimental determination of protein localisation.SLiMs are 3-12 amino acid long linear peptides enriched in disordered regions of proteins that interact with domains of other proteins. The aim of this study was to search for novel targeting SLiMs in a dataset of proteins for which their localisation had already been experimentally determined.