Figure 2

Comparisons between performance of the M-measure, HMM-based CNA algorithms and the three-state Viterbi decoding based on the M-measure (3SMM) for binary mixtures. We simulated 200 datasets reflecting the mixture of a homogeneous tumor sample and a stromal component at different levels of the mixing coefficient. Panel A shows the average of the areas under the ROC curve (AUCROC) at the operative point for the correct classification of the three states (Eq.7) of GenoCNA (black), PennCNV (red), the M-measure (green) and a three-state Viterbi decoding based on the M-measure, 3SMM (blue). The dotted envelope around each curve represents two units of standard deviation centered at the mean performance. Overall, the M-measure performs equally well or better than the HMM-based algorithm. B. Inference of the mixing coefficient according to the genoCNA algorithm (black) or to the solution of the linear systems (Eq.3) in regions classified as loss based on the M-measure classification (green) or 3SMM (blue). The dotted envelope around each curve represents two units of standard deviation around the mean of the inferred coefficient. The value of the mixing coefficient, which is inferred from the classification using the M-measure and/or 3SMM, is not surprising as it reflects the deterministic nature of the underlying linear system. On the other hand, the low performance of genoCNA is attributed to the non-uniqueness of solutions in the problem of inferring exact copy numbers, particularly amplifications.