Figure 4 shows the AUC distribution over 16 ovarian cancer transcriptomes (Section 2) for the best and worst cross-validation runs of SNVMix2 and SNVMix1 as well as the results from the Maq SNV caller with the two recommended settings of the r parameter (0.001 and 0.02). Both runs of SNVMix1 were statistically significantly better than the Maq runs [analysis of variance (ANOVA) test, P<0.0001], with mean AUC of 0.9557 ± 0.0100 and 0.9552 ± 0.0100, compared with 0.9290 ± 0.0120 and 0.9032 ± 0.0119 for the Maq runs. Furthermore, SNVMix2 without quality thresholds offers a slight performance improvement over SNVMix1. Although the improvement of SNVMix2 over SNVMix1 is not statistically significant, it is noteworthy that no thresholds of any kind were applied to the data and thus probabilistic weighting can eliminate the need for arbitrarily thresholding the data (see below).
