Figure 3

Performance comparisons between HMM-based CNA algorithms and M-measure for mixtures of three states. We simulated 200 datasets by mixing a stromal component with relative concentration of α, with two independent tumor components with relative concentrations of 2(1- α)/3 and (1- α)/3, respectively. To assess performance, the output of these algorithms is first classified into three states (gain, loss and normal), and the average between the AUCROCs at the operative points (Eq.7) is computed for GenoCNA (black) and the M-measure (green). The correct state is defined as the state of the aberrant signal with the largest coefficient. The dotted envelope around each curve represents two units of standard deviation, and the central solid line represents the mean performance. Overall, the M-measure (green) exhibits a high degree of robustness to increasing levels of stromal contamination. Real samples will generally have more than two components, including the stromal fraction. Therefore, although genoCNA is not designed to handle heterogeneous samples with more than two components, its performance has been added to the plot to show the way in which state-of-the-art HMM-based methods perform.