Figure 2

Performance of the structure prediction. Sensitivity (SN), specificity (SP) and Hamming distance (HD) of the structure prediction obtained in the simulations for different values of applied threshold. The dashed lines show the performance of SPACE before applying improvements and the continuous lines after applying Adapt algorithm and coherence correction. The Hamming distance error rate was calculated in the form of HD = (FP + FN) /N, i.e. the proportion of probes that are erroneously predicted. The sensitivity (or true positive rate, TPR) was calculated as SN = TP = (TP + FN), i.e. the proportion of probes that are predicted to hybridize against a transcript and in fact do hybridize. Finally, the specificity (or one minus the false positive rate, FPR) was calculated as SP = TN/(TN + FP), i.e. the proportion of probes that are predicted NOT to hybridize against a transcript and in fact do not hybridize. The x-axis shows the threshold in the estimated $\tilde{G}$ matrix to decide whether a probe hybridizes or not. The threshold that provides the minimum Hamming distance (overall error) is 0.5. A threshold near 0.8 provides the same sensitivity and specificity, both equal to 0.85; such value represents an error at least 25% lower than the crossing point for SPACE without improvements.