Our results will be highly dependent on the accuracy of alignments as well as the consistency and accuracy of mapping qualities reported by the aligner. Results in Table 2 showed that the combined approach of stringent thresholding on mapping quality and modeling the uncertainty of the remaining reads gave the highest accuracy. Given that the most gain was obtained in precision, it suggests that false positive predictions may indeed be reduced with more accurate alignments. As read lengths increase with technology development and mapping algorithms improve, we expect that the input to SNVMix will be of higher quality, which should yield more accurate results. Moreover, alignment using Maq presents a drawback with regard to SNVMix2's model. When a short read can be aligned to more than one position in the genome with the same mapping quality, this read is dropped, being assigned a mapping quality of zero. SNVMix2's design would be able to leverage the read's quality amongst the distinct coordinates and still use the information it conveys to predict SNVs. The performance of this will be evaluated in future work.
