were validated using Sanger amplicon sequencing (Section 2). These consisted of 305 true SNVs (variants seen in the Sanger traces) and 192 that could not be confirmed in the Sanger traces. Table 2 shows the sensitivity, precision and F-measure results of SNVMix2, SNVMix1 and SNVMix2 combined with base and mapping quality thresholding at both 10× and 40× coverage at a p(SNV) (Section 2) threshold determined using a FPR ≤ 0.01. We did not include a comparison to Maq at these 497 positions since the results would be biased toward the SNVMix1 model that led us to identify them in the first place. SNVMix2 and SNVMix1 showed similar F-measure at both 10× and 40× reinforcing that the probabilistic weighting confers equal accuracy without having to select arbitrary quality thresholds. In addition, both SNVMix2 and SNVMix1 had higher accuracy at 40× and 10× (Table 2). Interestingly, all the models had increased false negative rates in the 40× genome in comparison with the 10× genome. Upon further review of the SNVMix2 positions predicted at 10×, but not at 40×, we examined that the majority (9 out of 13) were marginally below threshold and significant probability mass was indeed on the P(ab) state (>0.99)
