Prediction of Protein Subcellular Location using Residue Exposure

This is the authors' abstract. We don't add key points for this paper.

The proteins perform their functions in associated cellular locations. Therefore, subcellular location is a key-feature in the functional characterization of proteins. The experimental methods of determining a protein's subcellular location are costly, time consuming, error prone and can not cope with exponentially growing genomic and proteomic data. Therefore, computational prediction of protein subcellular location is a major effort in bioinformatics research. Subcellular location of a protein can be predicted either from its sequence by identifying the targeting peptide and motifs, or by homology to proteins of known location. Another approach, which is complementary, exploits the differences in amino acid composition of proteins associated to different cellular locations. This is an especially useful approach if motif and homology information are missing. In this study, we expand this approach taking into account amino acid composition at different levels of amino acid exposure. Through careful selection and data integration we created a high quality dataset of proteins with known structure and location. The members of three subcellular location categories were considered: nuclear, cytoplasmic and extracellular, plus the extra category nucleocytoplasmic, accounting for the fact that a large number of proteins shuttle between nucleus and cytoplasm. We explored the relationship between residue exposure and protein subcellular location. The analysis demonstrated that amino acids at different levels of exposure have signal about the location of proteins. For the classification purpose we applied a novel approach of two stage classification. At stage one, multiple Support Vector Machines (SVMs) were trained to score eukaryotic protein sequences for membership to each location class. In stage two, an artificial neural network (ANN) was used to propose a category from the scores assigned to the four locations in stage one. The method reaches an accuracy of 68% when using as input 3D-derived values of amino acid exposure. Calibration of the method using predicted values of amino acid exposure allows classifying proteins without 3D-information with an accuracy of 62%. The algorithm is implemented as the web server 'NYCE'. We compared the performance of NYCE against other state-of-the-art subcellular location prediction tools. The comparison revealed the fact that 'NYCE' performs reasonably well compared to other tools, though using a limited set of information. A major challenge of protein subcellular location prediction methods based on homology is that there are very similar proteins that act in different subcellular locations. Using pairs of paralog proteins experimentally known to be in different locations, we demonstrated that our algorithm can evaluate proteins independently of their homology. NYCE can discern proteins in different locations even if they share high levels of identity whereas other tools fail to do so.