An efficient statistical methodology for peptide 3D structure clustering

The analysis of proteins and peptides conformations is of crucial interest to gain insights on their biological functions; it has therefore been an active research topic over the past decades. However, analyzing conformations of small size and highly flexible peptides remains a challenge due to their instability and their large number of different shapes. In this paper, an efficient methodology is proposed to analyze 3D structures of highly flexible elastin-derived peptides and to find out their principal conformations using a clustering algorithm. This methodology is based on a special representation of peptide structures, which has the great advantage to be neither affected by peptides' translations nor rotations, hence, avoiding the use of a complex superposition method. In addition, the proposed approach uses for the first time Kernel PCA to remove outlier structures that are not frequently present and do not resemble any other peptide structures. Outlier removal is very important in this context because, due to the instability of those peptides, a small portion of very different conformations, that seldom occur, can heavily affect the ensuing clustering results. Finally, the proposed approach latest step consists in hierarchical clustering, used as a non-supervised classification method to gather together similar structures. Experimental results, obtained using an existing database, show the relevance and the efficiency of the proposed method.