ABSTRACT BACKGROUND Whole-genome sequencing (WGS) data may be used to identify copy number variations (CNVs). Existing CNV detection methods mostly rely on read depth or alignment characteristics (paired-end distance and split reads) to infer gains/losses, while neglecting allelic intensity ratios and cannot quantify copy numbers. Additionally, most CNV callers are not scalable to handle a large number of WGS samples. METHODS To facilitate large-scale and rapid CNV detection from WGS data, we developed a Dynamic Programming Imputation (DPI) based algorithm called HadoopCNV, which infers copy number changes through both allelic frequency and read depth information. Our implementation is built on the Hadoop framework, enabling multiple compute nodes to work in parallel. RESULTS Compared to two widely used tools – CNVnator and LUMPY, HadoopCNV has similar or better performance on both simulated data sets and real data on the NA12878 individual. Additionally, analysis on a 10-member pedigree showed that HadoopCNV has a Mendelian precision that is similar or better than other tools. Furthermore, HadoopCNV can accurately infer loss of heterozygosity (LOH), while other tools cannot. HadoopCNV requires only 1.6 hours for a human genome with 30X coverage, on a 32-node cluster, with a linear relationship between speed improvement and the number of nodes. We further developed a method to combine HadoopCNV and LUMPY result, and demonstrated that the combination resulted in better performance than any individual tools. CONCLUSIONS The combination of high-resolution, allele-specific read depth from WGS data and Hadoop framework can result in efficient and accurate detection of CNVs.

HadoopCNV: A dynamic programming imputation algorithm to detect copy number variants from sequencing data