Figure 3

Predicting pathway activity in tumours. A) Upper panel: predicted DART ERBB2 pathway activation scores in the basal (B) and HER2+ subtypes of ER-breast cancer and across six different breast cancer cohorts. P-values are from a one-tailed t-test since activity is predicted to be higher in the HER2+ subtype. The pruned network was only learned once, from the Wang set, and this same network was then used to compute pathway activity in the other 5 cohorts (Mainz,NCH,Frid,GH,Loi). Middle panel: predicted DART MYC activity scores across three breast cancer cohorts with combined expression and copy-number information (Wang, NCH, Loi). DART scores are plotted against segmented MYC copy number value (Wang,GH) or called copy-number state (NCH,1 = gain,0 = no-change or loss). One tailed P-value was estimated empirically from linear regression on permuted sample labels (Wang,GH) or from a one-tailed t-test (NCH). As in A), pruned network was learned in Wang, and same network used in NCH and GH. Lower panel: predicted DART TP53 activity scores in three lung cancer/normal data sets (Landi,Wachi,Su). Pruned network was learned in Landi, and this network tested in Wachi and Su. P-values are from a one-tailed t-test. We point out that DART learns the pruned network without using phenotypic sample information (i.e Basal/Her2, Copy-number, Normal/Cancer status), thus the results in Wang and Landi sets are not due to selection. B) For each of ERBB2, MYC and TP53 and for each data set we compare the significance of the associations (-log₁₀(P - value)) between the three methods (UPR-AV,PR-AV,DART). The green dashed line represents the line where P = 0.05 and values above it are declared significant.