Rare CNVs were further filtered by three metrics: (1) Confidence score (CS), (2) segmental duplication (SD) content, and (3) overlap with validated common copy-number loci. CS was calculated using MeZOD, an outlier-detection-based method published previously (McCarthy et al., 2009), where the rare CNV call is assigned a p value based on the distribution of probe ratios across the reference population. Thresholds for CS were then adjusted within each size class of CNV to achieve a 5% rate of mendelian inconsistency across all size classes (Figure S1). Second, we removed rare CNVs which had > 70% overlap with known SDs from the UCSC hg18 Human Genome browser annotations. A SD filter is helpful because it eliminates regions where the exact location, boundaries, and patterns of inheritance of the CNV calls are often too difficult to determine from array CGH data due to the complexity of the local genomic architecture. This final rare CNV call sets consisted of 3,856 CNVs in 788 offspring including BD, SCZ, and controls and in 45 ASD subjects (Table S1).
