Previous studies have employed stringent thresholding for removing poor quality bases and/or reads (Ley et al., 2008; Morin et al., 2008). We propose that this may throw out informative data, and we extend SNVMix1 to explicitly encode base and mapping qualities by using them to probabilistically weight the contribution of each nucleotide to the posterior probability of a SNV call. In addition, we explore how to optimally combine thresholding and probabilistic weighting in order to obtain more accurate results. We show (Section 3) how this extended model, which we call SNVMix2, confers an increased specificity in our predictions.
