Figure 3

ROC curves for training (A and C), and test (B and D) data, estimated when 0-10% of data annotations were randomly inverted in training. For each proportion of inverted annotations, 10 jacknife replicates (70%) of training data were generated. SVM parameters were independently optimized for each jacknife set and values reported are means. The lack of change of ROC curves for test data indicate the stability of the SVM to errors in training annotations with the Human/Mouse data.