We ranked the detected candidate mutated positions according to P-values. We did not use the nominal P-values to call the significant mutations. Instead, we relied on the calibration sample to derive the thresholds. Since there is an imbalance between the number of true positives (calibration SNPs) and true negatives (invariant bases), we used Matthew's correlation coefficient (MCC) [29] to determine the best threshold that maximizes MCC phi coefficient. This non-parametric procedure used in machine learning is optimal to classify binary information such as distribution between true negative and true positive.
