Effective normalization for copy number variation detection from whole genome sequencing

Background: Whole genome sequencing enables a high resolution view of the human genome and provides unique insights into genome structure at an unprecedented scale. There have been a number of tools to infer copy number variation in the genome. These tools, while validated, also include a number of parameters that are configurable to genome data being analyzed. These algorithms allow for normalization to account for individual and population-specific effects on individual genome CNV estimates but the impact of these changes on the estimated CNVs is not well characterized. We evaluate in detail the effect of normalization methodologies in two CNV algorithms FREEC and CNV-seq using whole genome sequencing data from 8 individuals spanning four populations. Methods: We apply FREEC and CNV-seq to a sequencing data set consisting of 8 genomes. We use multiple configurations corresponding to different read-count normalization methodologies in FREEC, and statistically characterize the concordance of the CNV calls between FREEC configurations and the analogous output from CNV-seq. The normalization methodologies evaluated in FREEC are: GC content, mappability and control genome. We further stratify the concordance analysis within genic, non-genic, and a collection of validated variant regions. Results: The GC content normalization methodology generates the highest number of altered copy number regions. Both mappability and control genome normalization reduce the total number and length of copy number regions. Mappability normalization yields Jaccard indices in the 0.07 - 0.3 range, whereas using a control genome normalization yields Jaccard index values around 0.4 with normalization based on GC content. The most critical impact of using mappability as a normalization factor is substantial reduction of deletion CNV calls. The output of another method based on control genome normalization, CNV-seq, resulted in comparable CNV call profiles, and substantial agreement in variable gene and CNV region calls. Conclusions: Choice of read-count normalization methodology has a substantial effect on CNV calls and the use of genomic mappability or an appropriately chosen control genome can optimize the output of CNV analysis.