We then studied the errors owing to under-calling of true germline events in the matched normal with the same approach but instead using the ~1 million germline variant loci in the same territory (Fig. 3d–f). In classifying an event as germline or somatic, MuTect uses different prior probabilities at sites of common germline variation versus the rest of the genome, and therefore we report the false positive rates separately for these two scenarios (Fig. 3d) along with the power to have classified such events (Fig. 3e–f). We observe that with ≤ 7 reads in the normal at novel germline sites (Fig. 3e) or with ≤ 18 reads at sites of known germline variation (Fig. 3f), there is insufficient data to classify a variant as being somatic or germline, and hence we keep such sites as ‘variant’ and never make false positive somatic calls in these cases. Once there is sufficient data to make a classification, the error rate drops rapidly from 2.4×10−3 at 8x in the normal to below 0.2×10−3 at 12x, which corresponds to less than one misclassified germline in the entire exome (~30mb in the exome × 50 novel germline variants/mb × 0.2×10−3 error rate).
