Figure 2

The DART-CLQ algorithm.(A) Given an in vitro perturbation signature (depicted one of 40 genes) of up- and downregulated genes, one computes gene pairwise correlations in expression over the samples of a large training set. One would expect two genes that are commonly up- or downregulated in the in vitro signature to exhibit positive correlations, whereas two genes with one up- and the other downregulated, would be predicted to be anti-correlated (upper diagonal). If the perturbation signature has explanatory power in the training expression set, one would expect that observed correlations (lower diagonal) should agree, statistically, with the predicted ones, and if so, a consistency score can be derived. (B) If the consistency score is statistically significant, the correlation network is pruned to remove those observed correlations for which the directionality is inconsistent with the prior information, leaving only consistent and significant correlations. (C) The consistent and significant correlations define a correlation network with a maximally connected component as depicted. DART-CLQ infers all the largest cliques in this component and merges them (in practice, largest cliques exhibit very strong overlaps) to define an approximate clique gene module (CLQ-MOD). An example of a clique within this module is indicated by the square nodes (genes). (D) Given the approximate clique gene module, perturbation activity is now estimated by first z-normalizing each of a module gene’s expression profile (mean centering and unit variance scaling) over the samples and then constructing a weighted average, where we weight each gene according to its degree in the module (k _g) and whether it was predicted to be up- (σ _g=1) or downregulated (σ _g=−1). We note that the difference between DART-CLQ and DART is that DART estimates the activity score over the full maximally connected correlation network.