We observed that the variability in CNV calls was larger when using different CNV calling algorithms on the same raw data, compared to when the same algorithm is used on the data from different laboratories (Supplementary Figs. 8b,c). We find that results originating from different laboratories tend to cluster together, indicating that the site where the experiment was performed has less effect on resulting data than the choice of platform or algorithm. Interlaboratory variability correlates with reproducibility, and platforms exhibiting high reproducibility in replicates also seem more robust to interlaboratory variability. The exceptions to this are the Affymetrix arrays, where CNV calls are highly dependent on the reference data set used for analysis. We observe that the sample-level concordance of CNV calls between any combinations of two algorithms is typically 25–50% within a platform, and even lower for comparisons across platforms (Supplementary Fig. 9a). Larger CNVs would be expected to show higher concordance and we therefore divided the data into CNVs of 1–50 kb and variants >50 kb. Although we see improvement, the degree of concordance between platforms rarely exceeds 75% (Supplementary Fig. 9b,c).
