Assessing accuracy on truth data sets

Sample NA12878 and 22 samples from diverse L haplogroups were selected for in silico mixing experiments to create a large truth data set compared to the reference Chr M (totaling 1200 variants at 286 positions, including eight indels). For each L haplogroup sample, the number of mtDNA reads per sample was counted (SAMtools v1.8 idxstats [Li et al. 2009]), and then downsampling was performed (SAMtools v1.8) to create five BAM files containing a predefined ratio of reads from the L haplogroup sample and NA12878 (1%, 5%, 50%, 90%, 99%). For each mixture, total coverage was set to the L haplogroup sample's original coverage. GATK's HaplotypeCaller version 4.0.3.0 was used to call homoplasmic variants on the original BAMs before downsampling, with the ploidy argument set to 100. For each L haplogroup sample, a truth set was defined as variants present in the L haplogroup sample (allele count > 94/100) but absent in NA12878 (based on manual review using overlapping read pair data, with padding of 1 bp around each NA12878 variant). For each L haplogroup sample mixture, true- and false-positive calls were calculated against the sample-specific truth set and then summed across all 22 L haplogroup samples to create sensitivity and precision metrics.