The experiments were performed in triplicate for each sample, allowing us to analyze the reproducibility in CNV calling between replicates. A CNV call was considered replicated if there was a reciprocal overlap of at least 80% between CNVs in a pair-wise comparison. Reproducibility was measured by calculating the sample level Jaccard similarity coefficient, defined as the number of replicated calls divided by the total number of nonredundant CNV calls. We first investigated these parameters across the full size range, including all CNVs >1 kb and with a minimum of five probes, representing cut-offs typically used in research projects. The summary data of call reproducibility and the number of calls for each platform and algorithm combination are shown in Figures 2a,b for high-resolution platforms and Supplementary Figure 8a for lower resolution platforms. The reproducibility is found to be <70% for most platform and algorithm combinations. In general, the most recently released arrays perform better, resulting in more CNV calls reproducibly obtained between replicates. Of the CGH arrays, the Agilent (2 × 244K) platform produced the largest number of CNV calls (an
