Data set heterogeneities and their effects on the derivation of contact potentials

Abstract

A new class of protein structure prediction methods bases on threading a sequence through a known 3D structure. These techniques offer a means of recognizing similarity in cases of distant evolutionary relationship, where the fold has been conserved to a greater extent than its sequence. The scoring table(s) which are used to evaluate different sequence to structure alignments are derived from a particular database of known structures. The distribution and the reliability of the threading scores is effected by parameters of this database, such as length and composition of the included chains, and the number of alternative alignments. Weinvestigated several compositional aspects of two databases which were used in protein fold recognition studies. Among them are the numberof chains from different folding classes, the amino acid distribution within the classes, and the sequence lengths.