Given the obvious variability between calling algorithms, we and others13,47,48 have found that using multiple algorithms minimizes the number of false discoveries. Based on our experience this scheme allows for greater experimental validation by qPCR, which are typically >95% for variants >30 kb13. Because the algorithms use different strategies for CNV calling, their strengths can be leveraged to ensure maximum specificity. Nevertheless, we still observe that up to 50% of the calls detected by only one of two algorithms can be validated when compared to sample-level CNVs11 (Supplementary Fig. 15), indicating that CNVs may be missed in this stringent approach and that merging call sets from multiple methods could improve sensitivity. Our results also show that one single tool is not always best for each array, but that the optimal algorithm for a data set is dependent upon the noise specific to that experiment. As an example, iPattern was one of the best performing algorithms for high-quality data from Affymetrix 6.0, but would not work properly for noisier Affymetrix 6.0 data.
