Study of underlying repeats in genomic DNA sequences with neural networks

Abstract

Hidden periodicities play a very important role in the regulatory and structural functioning of genomic DNA strands. Primarily, it concerns the fundamental three-periodicity inherent to protein coding regions in all taxonomic groups, two-periodicity in introns of eukaryots as well as periodicities related to helix and chromatin pitches, while the other periodicities appear to be species specific. Rather roughly (and without sharp boundary) the underlying periodicities may be divided by two groups. In the first case the periodicities are due to particular nucleotides (or very short oligomers) quasi-regularly positioned in a seemingly random background. This type of regularity can be identified via either standard frequency analysis or more elaborate Fourier methods. For the second group a periodicity is related to the quasi-random replacements in initially complete repeating motifs (situation typical, e.g., for modifications of satellites). In the last case the statistical reconstruction of underlying repeats is a much less trivial task. The authors show that this problem can successfully be solved with multi-symbol extension of energy-minimizing neural networks (EMNN). The reconstruction of underlying motifs may shed additional light on the evolutionary and functional modifications in various genomes.