Skip to main content
Fig. 5 | Genome Medicine

Fig. 5

From: Extracting a low-dimensional description of multiple gene expression datasets reveals a potential driver for tumor-associated stroma in ovarian cancer

Fig. 5

a For each of 90 INSPIRE modules (x-axis), the − log10 p from the Pearson’s correlation is shown (y-axis) for six different histological and clinical phenotypes. The p value threshold (shown by red dotted horizontal lines) is 5 × 10–3 for histological phenotypes and 5 × 10–2 for clinical phenotypes, which are harder to predict. We highlight modules 5, 6, 53, 54, 60, 78, and 81 that are significantly correlated with at least three of the six phenotypes in red. We also highlight module 30 in red since it is the only module that has a significant correlation with the vessel formation phenotype. Modules 5 and 6 achieve the first or second rank in terms of the significance of correlation with five of the six phenotypes. b For four different methods (the subnetwork markers, PCs, all genes, and INSPIRE latent variables), the prediction performance is compared for six prediction tasks in CV setting. In all bar charts except for the last one, a single accuracy (or concordance index for survival) is reported based on the predicted phenotype vector formed by pulling together the predictions for all folds. For the survival phenotype, an additional analysis is presented where the mean concordance index is reported across all 500 folds in 50 rounds of tenfold CV tests. For this bar chart (right), the p value from the Wilcoxon signed rank test and the 95 % confidence interval for the mean of a standard normal distribution fitted to the difference are reported to show the significance of the difference between INSPIRE and each of the alternative three methods (individual genes, PCs, and subnetworks). c For three different Pearson’s correlation p value thresholds (10–2, 10–4, and 10–6, respectively from left to right), the number of CNV levels that are significantly associated with the learned subtypes are shown for two published methods and INSPIRE. d The modules that differentiate the subtypes that are learned using INSPIRE features and the interactions among those modules as learned by INSPIRE. The modules are grouped and colored according to the subtypes they differentiate. Next to each one of the four module groups, there is the heat map of the features corresponding to the modules in this module group

Back to article page