Fig. 4
From: Unearthing new genomic markers of drug response by improved measurement of discriminative power
Potential false-positive marker of the MANOVA test incorrectly rejected by the chi-squared test. (Left) The scatter plot for the drug-gene association (GW441756-FLT3) with the largest -logPMANOVA among those not significant according to the chi-squared test. Hence, mutated-FLT3 is a marker of sensitivity to the experimental drug GW441756 according to the MANOVA test, but not according to the chi-squared test. In the plotted training set, this marker offers practically no discriminative power as further evidenced by a φ of just 0.05 and similar drug response (logIC50) distributions of mutated and WT cell lines. However, this marker provides an MCC of 0.10 on the test and hence this is a false negative of the chi-squared test. (Right) Conversely, to assess the consistency of the MANOVA test, we searched for the drug-gene association with largest -logPχ2 among those with a similar -logPMANOVA to that of GW441756-FLT3, which is Dasatinib-BCR_ABL. Whereas the p-value for Dasatinib-BCR_ABL is of the same magnitude as that for GW441756-FLT3 using the MANOVA test (PMANOVA~ 10− 10), the p-values for the same associations using the chi-squared test differ is almost 27 orders of magnitude. Thus, unlike the chi-squared test, the MANOVA test is unable to detect the extreme difference in discriminative power offered by these two drug-gene associations. Indeed, the BCR_ABL translocation is a highly discriminative marker of Dasatinib sensitivity (φ = 0.65), as also evidenced by the barely overlapping drug response distributions from each set of cell lines. This is confirmed in the test set, where the Dasatinib-BCR_ABL marker obtains an MCC of 0.21