Figure 4

ROC curves for 7 different algorithms using the combined dataset of yeast transcriptional responses to various environmental changes [28] and the ChIP-chip data of Lee, 2002. We used the experiments with at least 5 time points or dose-response points for the total of 165 microarrays. KEGG pathways were used as the gold standard. The functional coherence of produced TMs was also compared to the functional coherence of TMs constructed by three other algorithms utilizing expression datasets containing the Gasch dataset used in our calculations. ROC points for TMs constructed by Naama Barkai's lab [11] utilizing only a very large expression dataset at 23 different cut-off levels are depicted by yellow spots. The seed module and extended module constructed by ReMoDiscovery [11] utilizing gene expression, ChIP-chip and binding sequence motif data are depicted by pink spots. B+C+, B-C+ and the C+ TMs constructed by combining B+C+ and B-C+ modules identified by COGRIM [20] utilizing gene expression, ChIP-chip and binding sequence motif data are depicted by dark blue spots. All three diagrams represent the same ROC curves/plots for different ranges of False Positive Rates (x-axis) ECIM results again dominated all other algorithms in terms of functional coherence.