Comparison of kNN and k-means optimization methods of reference set selection for improved CNV callers performance

There are over 25 tools dedicated for the detection of Copy Number Variants (CNVs) using Whole Exome Sequencing (WES) data based on read depth analysis. The tools reported consist of several steps, including: (i) calculation of read depth for each sequencing target, (ii) normalization, (iii) segmentation and (iv) actual CNV calling. The essential aspect of the entire process is the normalization stage, in which systematic errors and biases are removed and the reference sample set is used to increase the signal-to-noise ratio. Although some CNV calling tools use dedicated algorithms to obtain the optimal reference sample set, most of the advanced CNV callers do not include this feature. To our knowledge, this work is the first attempt to assess the impact of reference sample set selection on CNV detection performance.

We used WES data from the 1000 Genomes project to evaluate the impact of various methods of reference sample set selection on CNV calling performance of three chosen state-of-the-art tools: CODEX, CNVkit and exomeCopy. Two naive solutions (all samples as reference set and random selection) as well as two clustering methods (k-means and k nearest neighbours (kNN) with a variable number of clusters or group sizes) have been evaluated to discover the best performing sample selection method.

The performed experiments have shown that the appropriate selection of the reference sample set may greatly improve the CNV detection rate. In particular, we found that smart reduction of reference sample size may significantly increase the algorithms' precision while having negligible negative effect on sensitivity. We observed that a complete CNV calling process with the k-means algorithm as the selection method has significantly better time complexity than kNN-based solution.