Figure 4

A schema of the LINK and STAR models: (A) the LINK model is designed to find a subset of conditions in which the expression profiles of a TF-target gene pair (link highlighted in pink in the upper left panel) are positively or negatively correlated. In addition it finds other known or putative target genes whose expression profiles are correlated with the TF or its target under this subset of conditions. The regulating TF and its target gene (Gene₁) are the core set of input genes inserted into the USA. The blue and yellow nodes in the known local network represent the input gene set employed by the USA. This input set is also indicated in a genome-wide vector that has only two nonzero elements representing the TF (+1) and its target (+1 for activation and -1 for suppression). The dataset we insert into the USA is the yeast compendium data described in Fig. 1 and schematized in the miniature matrix consisting of 11 genes and 15 experimental conditions shown between the left and right panels. The red, blue and yellow entries correspond to up-regulation, down-regulation and intermediate expression levels respectively. The algorithm finds the conditions in which TF and Gene1 are correlated. It also finds the additional genes denoted in red (Gene₂, Gene₄, and Gene₅), whose expression profiles across the subset of experimental conditions correlate positively or negatively with the TF or Gene1. Altogether, these genes and conditions constitute the UTM shown in the left middle panel. We predict that links of the known regulatory interactions TF→Gene₂ and TF⊣Gene₄ are functional under the united transcriptional module (UTM) conditions in which the TF, Gene₁ and Gene₂ are up regulated (red pixels in the UTM matrix) and Gene₄ is down regulated (blue pixels). For the predicted link TF→Gene₅, we further compute a MATCH score between the TF position-specific weight matrix (PWM) and the promoter region of Gene₅. Links with a score higher than a threshold value of 0.94 (i.e. PWM match) are reported along with the experimental conditions supported by their corresponding UTM as illustrated in the lower left panel. (B) the STAR model enables us to find an alternative subset of experimental conditions in which Gene1 is directly regulated by TF (illustrated by the link highlighted in pink in the right upper panel). It searches for conditions in which the expression profiles of Gene1 is positively or negatively correlated with some of the genes regulated by TF (highlighted in light blue). In addition, it is designed to identify new genes that are positively or negatively correlated with the input core of target genes under the same subset of experimental conditions, and whose promoters contain sequences similar to the TF binding site. In the STAR model, we apply the USA (right center panel) to the core set of input genes consisting of all the TF target genes (Gene_1,2,3,4) excluding the TF itself. As in the LINK model, the input set of genes is indicated by the nonzero elements of a genome-wide vector (highlighted in green in the upper right panel), which denotes the regulatory relationship between the TF and its target genes. In addition to links from the original local (STAR) network, two new links, i.e. TF⊣Gene₆ and TF→Gene₇, are predicted based on their co-expression with the core genes, and the match between their promoter region and the TF PWM. Each target is predicted to be affected by the TF under the experimental conditions of the respective UTM.