A new machine-learning based method to accurately assess copy number variants from whole genome sequencing data and its application on the analysis of the Hirschsprung disease genome

Abstract

Copy Number Variants (CNVs) are defined as DNA segments larger than 50 base-pairs with copy number changes. CNVs discovery is essential for uncovering genes/risk factors for a wide range of diseases, including Hirschsprung disease(HSCR; colon aganglionosis). Incidentally, gross deletions encompassing RETor EDNRBlead to the discovery of these two main HSCR genes. Importantly, our previous CNV analysis of HSCR patients identified the association of CNVs encompassing NRG3with HSCR, therefore vindicating a role for CNVs in HSCR. As CNV calling is still a challenge, we have devised and developed a new machine-learning based method CNV-JACG to accurately assess CNVs. Initially, in order to get the most comprehensive CNVs, we used four complementary CNV discovery methods namely CNVnator, Delly, Lumpy and Seeksv to call CNVs. We used 9 in-house trios to produce training dataset, 11 pairs of in-house duplicated samples for validation, and 2 trios from 1000 Genomes Project for evaluation. For deletions, after CNV-JACG assessment, the concordance between each pair of 11 duplicated samples increases from 63% to 88%, and the Mendelian inconsistent rate of the 2 trios decreases from 25% to 6%. In the benchmark sample NA12878, 84.7% of CNV-JACG predicted CNVs are consistent with previously published result. CNV-JACG has been used for the CNV analysis of whole genome sequencing data (pair-end 150bp, ~30X) generated from 443 HSCR patients and 493 matched controls. Here we will present the performance of our new program together with the genome-wide association results.