Protein surface properties using signal processing and statistical tools

(eng) This thesis is dedicated to the study of protein structure properties using 3D image processing tools. Proteins are macromolecules that rule almost every life processes by interacting with other molecules. The comprehension and the prediction of the interaction mechanisms, namely, protein docking, is of major interest for practical applications such as diagnosis and drug design. The applications related to protein docking have a high number of degrees of freedom and, moreover, there is a huge amount of available information about protein structures. Consequently, it is generally time consuming to proceed to protein docking or to screen data bases searching for a protein with some required properties. There is thus a need for fast and efficient algorithms for protein structure analysis . The contributions of this work mainly concern the localization of sites of interest on protein surfaces. To do so, surfaces were modeled as 3D polygonal meshes and algorithms were developed to extract features with a trade-off between short execution time and accuracy. The main improvements concern the generation of protein surface meshes, the approximation of geodesic distances (i.e. distances along the surface), and the computation of the travel depth (a descriptor of surface hollows depth). The locations of sites of interest were predicted by combining protein surface properties using machine learning tools. In this context, several classification and regression tools were compared on benchmark data sets. It was shown that for patch-based methods, regression is more appropriate than classification. The method was also validated for the particular case of antigens epitopes (i.e. the parts recognized by the antibodies) and provided better predictions than existing methods.