Fig. 2

The approach implemented in PathCORE-T to construct a pathway co-occurrence network from an expression compendium. a A user-selected feature extraction method is applied to expression data. Such methods assign each gene a weight, according to some distribution, that represents the gene’s contribution to the feature. The set of genes that are considered highly representative of a feature’s function is referred to as a feature’s gene signature. The gene signature is user-defined and should be based on the weight distribution produced by the unsupervised method of choice. In the event that the weight distribution contains both positive and negative values, a user can specify criteria for both a positive and negative gene signature. A test of pathway enrichment is applied to identify corresponding sets of pathways from the gene signature(s) in a feature. We consider pathways significantly overrepresented in the same feature to co-occur. Pairwise co-occurrence relationships are used to build a network, where each edge is weighted by the number of features containing both pathways. b N permuted networks are generated to assess the statistical significance of a co-occurrence relation in the graph. Here, we show the construction of one such permuted network. Two invariants are maintained during a permutation: (1) pathway side-specificity (if applicable, e.g. positive and negative gene signatures) and (2) the number of distinct pathways in a feature’s gene signature. c For each edge observed in the co-occurrence network, we compare its weight against the weight distribution generated from N (default: 10,000) permutations of the network to determine each edge’s p-value. After correcting the p-value by the number of edges observed in the graph using the Benjamini—Hochberg procedure, only an edge with an adjusted p-value below alpha (default: 0.05) is kept in the final co-occurrence network.