Sparse Negative Binomial Signal Recovery for Genomic Variant Prediction in Diploid Species

Structural variants (SVs) - such as insertions, deletions, and duplications of an individual's genome - are associated with genetic diseases and promotion of genetic diversity. Detecting SVs of an unknown genome is a mathematically challenging problem since SVs are rare and prone to low-coverage noise. Common approaches to detect SVs in an unknown genome require sequencing fragments of the genome, comparing them to a high-quality reference genome, and predicting SVs based on identified discordant fragments. We developed a computational method which seeks to improve existing SV detection methods in three ways: First, we implement an optimization approach using a negative binomial log-likelihood objective function. Second, we use a block-coordinate descent approach to simultaneously predict if an SV is homozygous or heterozygous given genomic data of related individuals. Third, we model a biologically realistic scenario where variants in the child are either inherited or novel. We validate our framework with simulated data and demonstrate improvements in predicting SVs and detecting false positives.