Computing Genomic Signatures Using de Bruijn Chains

Abstract

Genomic DNA sequences have both deterministic and random aspects and exhibit features at numerous scales, from codons to regions of conserved or divergent gene order. Genomic signatures work by capturing one or more such features efficiently into a compact mathematical structure. We examine the unique manner in which oligonucleotides constitute a genome, within a graph-theoretic setting. A de Bruijn chain (DBC) is a kind of de Bruijn graph that includes a finite Markov chain. By representing a DNA sequence as a walk over a DBC and retaining specific information at nodes and edges, we obtain the de Bruijn chain genomic signature θdbc, based on graph structure and the stationary distribution of the DBC. We demonstrate that the θdbc signature is information-rich, efficient, sufficiently representative of the sequence from which it is derived, and superior to existing genomic signatures such as the dinucleotide odds ratio and word frequency based signatures. We develop a mathematical framework to elucidate the power of the θdbc signature to distinguish between sequences hypothesized to be generated by DBCs of distinct parameters. We study the effect of order of the θdbc signature, genome size, and variation within a genome on accuracy. We illustrate its superior performance over existing genomic signatures in predicting the origin of short DNA sequences.