Fig. 6

Test set performance of three methods to identify single-gene markers. The methods are evaluated by their ability to correctly classify more recently-tested cell lines as sensitive or resistant to the considered drug via the MCC on the test set. There is no overlap between test sets and those employed to identify all drug-gene associations (training sets). The three compared methods are those based on the chi-squared test (B), the MANOVA test (C) and their consensus (A; the association is significant if it is significant by both tests). We can see that the consensus method is the most predictive, followed by associations only significant with the chi-squared test (B) and those only significant by the MANOVA test (C). These results show that the overall predictive value of the markers revealed by the chi-squared test is higher than that arising from the MANOVA test and also that the consensus of both tests is more predictive than any of these two tests alone. While most of the markers provide better prediction than random classification (MCC = 0), their generally low test set MCC values regardless of the employed detection method highlight how hard is to identify predictive markers of drug sensitivity