Assignment of endogenous retrovirus integration sites using a mixture model

TL;DR

This paper introduces a new statistical mixture model to accurately detect endogenous retrovirus integration sites in genomes, improving upon existing methods by accounting for read count variability, and demonstrates its application in studying mule deer genetics.

Contribution

A novel two-component mixture of negative binomial distributions for assigning ERV presence, outperforming traditional threshold-based methods.

Findings

Effective in identifying ERV sites in mule deer without genomic resources

Revealed patterns of shared ERV sites related to animal relatedness

Improved accuracy over existing detection methods

Abstract

Structural variation occurs in the genomes of individuals because of the different positions occupied by repetitive genome elements like endogenous retroviruses, or ERVs. The presence or absence of ERVs can be determined by identifying the junction with the host genome using high-throughput sequence technology and a clustering algorithm. The resulting data give the number of sequence reads assigned to each ERV-host junction sequence for each sampled individual. Variability in the number of reads from an individual integration site makes it difficult to determine whether a site is present for low read counts. We present a novel two-component mixture of negative binomial distributions to model these counts and assign a probability that a given ERV is present in a given individual. We explain how our approach is superior to existing alternatives, including another form of two-component mixture model and the much more common approach of selecting a threshold count for declaring the presence of an ERV. We apply our method to a data set of ERV integrations in mule deer [Odocoileus hemionus], a species for which no genomic resources are available, and demonstrate that the discovered patterns of shared integration sites contain information about animal relatedness.