Description of the method to infer co-regulation networks by phylogenetic footprint discovery. (A) Over-represented dyads in promoters of LYS1 (left) and LYS12 (right) orthologs in Saccharomycetes. Sig: binomial significance. Bold brown and green italics characters highlight the dyads matching the motifs bound by Lys14p (WWWTCCRnYGGAWWW) and Gcn4 (TGAGTCA) (38). (B) Sequence logos built from the over-represented dyads matching the Gcn4p and the Lys14p motifs compared to the reference motif (TRANSFAC). (C) Feature maps showing the location of the over-represented dyads in the promoters of orthologs for LYS1 (left) and LYS12 (right). Each box indicates an exact match for a dyad, with a height proportional to the binomial significance. (D) Subset of the dyad significance matrix selected for illustrating the correspondences between dyads and genes. The full matrix comprises 3585 rows (genes) and 33 276 columns (dyads). Values indicate the binomial significance sig_gj of each dyad (j) in the phylogenetic footprints of each gene (g). Negative significance values are denoted by a dot, and are replaced by a zero value for further computation. Boxes highlight groups of dyads that match the motif annotated for a given regulator (Upc2p, Gcn4p, Leu3p and Lys14p). (E) Detail of computation of the DPbits score. (F) Co-regulation network built by tracing an edge between each pair of genes having a strictly positive DPbits score. Edge thickness and color gradient (yellow–red) are proportional to the DPbits.

Comparison between scoring metrics. The ability of different scoring metrics to predict co-regulation is evaluated by comparing their performances on receiver operating characteristic (ROC) curves (A, C, E) and precision-recall curves (B, D, F), using the different reference networks: annotated regulons (A, B), STRING co-expression (C, D) or ChIP-on-chip co-binding (E, F), respectively. Each curve represents one of the metrics used for comparing dyad profiles. Dotted lines: negative controls performed by permuting the cells of the dyad significance matrix (50 permutations per curve). Orange curves indicate the correspondences between the STRING co-expression network and the annotated regulons (A, B) or ChIP-on-chip co-binding network (D, E), respectively. The STRING network interactions are weighted according to their co-expression score derived from Pearson's correlation coefficient (L. J. Jensen, personal communication). For the STRING co-expression network, the negative control was performed by permuting the edges of the original network (50 repetitions).

Co-regulation network inferred by phylogenetic footprint discovery. (A) High-scoring layer (DPbits ≥ 5) of the co-regulation network inferred between all yeast genes. Edge thickness reflects the prediction score. Insets highlight areas of the network corresponding to groups of functionally related genes. Each color corresponds to one annotated regulon. (B) Subset of the inferred network restricted to the 612 genes of the reference data set (80 regulons, DPbits ≥ 1). Green edges: co-regulations already supported by annotations; non-supported predictions, combining novel predictions and false positives.

Correspondences between inferred co-regulation network and three ‘reference’ networks: annotated regulons, microarray co-expression (STRING) and co-binding (ChIP-on-chip), respectively. (A) Impact of the score threshold on the size of the co-regulation network (restricted to the subset of interactions between genes found in at least one annotated regulon) and on its correspondence with annotated regulons. Note that some of the ‘false positives’ may correspond to actual co-regulations not yet documented in the database. (B) Sn and PPV curves as a function of the DPbits score of the predicted interactions. Inferred co-regulation versus annotated regulons. (C) Inferred co-regulation versus STRING co-expression. (D) Inferred co-regulation versus ChIP-on-chip. (E) STRING co-expression versus annotated regulons. (F) STRING co-expression versus ChIP-on-chip. Dotted lines: negative controls performed by permuting the cells of the dyad significance matrix (50 permutations per curve) (A–D) or by permuting the edges of the original network (50 repetitions) (E–F).

Impact of regulon size on sensitivity. (A) Gene–gene cliques generated by specific (GAL4, MET4) and global (GCN4, ZAP1) TFs. Correctly predicted co-regulations (true positives) are highlighted in green, with edge thickness proportional to the DPbits score. Orange edges indicate false negatives (FN), i.e. existing co-regulations missing from the predicted co-regulation network. Edge-wise sensitivity (E_sn) is the fraction of correctly predicted edges. Node-wise sensitivity (N_sn) is the fraction of target genes linked by at least one prediction. (B, C) Impact of regulon size (abscissa) on edge-wise (b) and gene-wise (c) sensitivity, respectively. All numerical values are provided in Table 4.

Experimental validation of selected inferred regulations by qRT–PCR and ChIP analyses. Each panel shows a cluster of neighbor genes in the inferred network, together with the experimental measurement of the transcriptional responses. (A) Effect of Lovastatin and UPC2 deletion on RNA levels of four potentially Upc2 regulated genes, using ERG2 as control. Lovastatin (40 µg/ml) was added to the cultures for 16 h before cell harvesting. RNA levels were quantified by qRT–PCR and data were expressed versus the actin gene (ACT1) used a non-regulated reference. (B) Effect of the RPN4 deletion on two potentially Rpn4 regulated genes using RPN2 as a positive control. Cells were grown on the YPD medium; experiments were performed as in (A). (C) Derepressing effect of the triple ure2 gzf3 dal80 mutations on the RNA levels of four genes potentially subject to nitrogen catabolite repression. Cells were grown on glutamine as the sole nitrogen source. Experiments were performed as in (A). (D) ChIP PCR analysis of four potentially Mbp1 regulated genes. Mbp1-HA expressing cells were grown on YPD and the promoter occupancy by Mbp1 was monitored by ChIP for RNR1 (a known target gene of Mbp1) and the four indicated genes. Data were expressed versus the GAL1 gene used as a reference. Color code for the network clusters: genes known to be submitted to a common regulation are highlighted in yellow. Cyan-shaded octagonal boxes indicate genes submitted for experimental validation.