We genotyped 9380 samples from a prior Swedish schizophrenia association study (Bergen et al., 2012
) and 90 HapMap samples on the Illumina Infinium HumanExome BeadChip. As described in the Supplementary Methods
, we used Illumina’s GenCall algorithm to generate two sets of genotype calls and Illumina’s GenomeStudio software to normalize raw intensities.
First, we used linear regression (R Development Core Team, 2010
) to find the relationship between the means and standard deviations of the X
intensities from 10 538 autosomal, common sites using 4643 samples with genotype calls from a 9479 sample cluster file and found they were correlated (
= 0.027, P
< 2 × 10−16
) (Supplementary Table S1
). Next, we found that z
= 7 performed well to classify common variation (Supplementary Table S2
). Finally, new genotypes were assigned based on thresholds calculated from the linear regression model with z
= 7 and means and standard deviations calculated from 947 samples with GenCall genotype calls from a 90-sample cluster file.
To test how well zCall works for rare variation, we compared both GenCall and zCall exome chip genotype calls to whole-exome sequencing genotypes in 947 Swedish samples (Supplementary Methods
). To assess rare variant performance, we used the SNP-wise concordance (SWC), which is calculated by dividing passing sites by the total number of sites. A SNP is considered passing when the only error is a common homozygote being called a No Call. For example, when considering singletons (i.e. one heterozygote), the heterozygote is called correctly as a heterozygote and no common homozygotes are called as heterozygotes. For 10 075 autosomal, singleton sites, the SWC between GenCall and whole-exome sequencing was 92.49% and 96.84%, respectively, for zCall. The main error mode of GenCall was calling the singleton heterozygote as a No Call (6.61%) while the main error mode of zCall is calling common allele homozygotes as heterozygotes (2.32%). By restricting zCall to missing genotype calls, we observed a SWC improvement to 99.12% (Supplementary Fig. S1
We also tried using genotype calls from a larger cluster file (9479 samples), but the performance of zCall was slightly worse. Even though more singletons were called correctly when using a larger cluster file (SWC = 94.65% versus 92.49%), more genotype errors were made that cannot be recovered when only calling No Calls (SWC = 98.60% versus 99.12%). Therefore, zCall performs more effectively using a smaller cluster file.
We also assessed the effect of sample size on threshold definition. We used three sample sizes for threshold definition: 90, 947 and 9479. We found that 90 samples were insufficient to define the thresholds (SWC = 98.86%), but that there was no difference in SWC between 947 samples and 9479 samples (99.12% for both).
To provide an unbiased evaluation, we compared GenCall and zCall genotype calls to whole-exome sequencing data from 369 samples from the ARRA Autism Sequencing Consortium. For 10 712 singletons, we found the SWC using GenCall to be 93.12% and using zCall restricted to No Calls to be 99.27% (Supplementary Fig. S2
). We also compared our method with optiCall (Shah et al., 2012
) and found for 10 705 singletons, the SWC of optiCall was 98.21% versus 99.27% for zCall (Supplementary Fig. S3