Another aspect of CNV accuracy is how well the reported start and end coordinates correlate with the actual breakpoints, and how reproducible the breakpoint estimates are. To analyze breakpoints, we first investigated reproducibility in the triplicate experiments. For every CNV call set generated, we took the regions called in at least two of the three replicate experiments for each sample and calculated the distance between the breakpoints on the left and right side of the CNV call, respectively. The distribution of differences in breakpoint estimates between replicate experiments reflects, in part, the resolution of the platform and the reproducibility of the data (Fig. 3). To normalize for variable probe density between platforms, we performed the same analysis based on the number of probes that differed between replicate CNV calls, and the results are very similar (data not shown). One observation from these analyses is that there are clear differences between analytic tools when applied to the same raw data. Algorithms with predefined variant sets (e.g., Birdsuite42) and algorithms searching for clustering across samples (such as iPattern13) show substantially better reproducibility in breakpoint estimation for common CNVs than do algorithms treating each sample as an independent analysis.
