Classification of Local Protein Structural Motifs by Kohonen Networks

Abstract

Kohonen networks were used for automatic classification of local structural elements derived from a set of 136 non-homologousproteins. A reduced representation of protein backbones based on dihedral phi and psi angles was employedfor construction of training patterns. Segments of nine residues were transformed into a 16-dimensional description by their angular values. Kohonen-mapping yielded a two-dimensional topological map of representative structural motifs. Helical structures and sheets are located on opposite sides of the feature map, and several intermediate forms are found. The map wasusedto trace protein backbones of two proteins, cytochrome bs and y-IV crystallin, leading to characteristic trajectories in the feature map.