Infinite string block matching features for DNA classification

Abstract

Automatic classification of DNA can be performed in a number of ways using a variety of features. This study introduces a novel technique for generating global features for DNA classification. Based on a new technique, the “do what's possible” representation, infinite string generators are evolved to produce strings with a maximized collection of matching blocks above a critical length in the target DNA. Most global DNA features, such as GC-content or those in spectrum string kernels, capture diffuse statistical information about the target DNA. Infinite string matching is based on multiple loci, and thus finds a different type of global feature than most techniques now in use. It is discovered that the block-matching score for evolved infinite string generators is able to cleanly separate high-entropy synthetic DNA data sets using a single feature threshold classifier. Preliminary evaluation on human endogenous retrovirus sequences shows that evolved infinite string generators locate promising features on biological data as well.