Dear Editor,

Eukaryotic transcriptional regulation networks are extremely complex. Usually, multiple transcription factors (TFs) bind to the promoter region of a gene and cooperate to control gene expression precisely. Identifying cooperative TFs remains a major challenge in modern biological research. Various types of data, including genomic sequences, expression profiles, ChIP-chip data and protein-protein interactions, have been used to identify mechanisms of cooperative transcriptional regulation. However, because of the noise inherent in these data and the fact that each data source only provides partial information about regulation, combining multiple types of data to improve their ability to infer cooperative TFs is advantageous 1, 2, 3.

In our previous work, we successfully integrated ChIP-chip data 4 and expression profiles with individual TF knockout strains 5 to unravel potential relations between TFs and their target genes 6. This combination of two independent and complementary sources of data improved the accuracy of our prediction. Here, we have extended the work to identify cooperativity between TFs in Saccharomyces cerevisiae. We achieved high prediction performance by identifying the most statistically significant overlap of target genes regulated by two TFs in ChIP-chip data and TF knockout data. In addition, we attempted to find the appropriate point to which extent the threshold should be relaxed by looking at the increasing number of cooperative TFs identified within different threshold ranges. Finally, identified TF pairs were ranked using Fisher's combined probability test 7 by combining two independent P-values calculated from the ChIP-chip and knockout data (METHODS, Supplementary information, Data S1).

This analysis identified 186 cooperative TFs. The identified cooperative TFs, the P-value calculated from the ChIP-chip data and knockout data, the combined P-value and any previous experimental and computational evidence are listed in Supplementary information, Data S1-Table S1. Figure 1 shows the cooperative network of TFs, which are colored and clustered according to their functions. This network suggests that different biological processes, such as the cell cycle, stress response pathways and metabolism, are closely connected to each other. We were pleased to find that many previously characterized cooperative TFs showed highly significant cooperativity measures in our results. Of the top 20 predicted pairs with characterized TFs, 16 pairs have been reported in the literature (Supplementary information, Data S1-Table S1) and 9 of these have been experimentally validated (Supplementary information, Data S1-Table S1). For example, SWI4-SWI6, ACE2-SWI5 and MBP1-SWI4 are known cooperative TFs that control the cell cycle. DAL81 facilitates the binding of STP1 to SPS sensor-regulated promoters 8. The galactose-activated transcription of GAL genes occurs when GAL3 binds GAL80 9. Another seven pairs, TEC1-TYE7, HIR3-YOX1, SPT23-YOX1, GAT3-RAP1, ACE2-MBP1, GAT3-RGM1 and RAP1-YAP5, have not been experimentally validated; however, they are supported by numerous computational studies (Supplementary information, Data S1-Table S1). For the remaining four pairs, GTS1-RIM101, RPN4-STB2, CHA4-GAT3 and STP4-TEC1, the potential for cooperativity can be inferred from the literature. For example, RPN4-STB2 together with another two pairs, STB2-YRR1 and RPN4-YRR1 (not in the top 20 but listed in Supplementary information, Data S1-Table S1), form a cooperative triad. Researchers have shown the coordinated action of RPN4, PDR3 and YRR1 on the transcriptional activation of FLR1 when adapting yeast to mancozeb 10. Both PDR3 and RPD3 control PDR5 expression 11, indicating their coordinated action, and STB2 has been detected in the protein complex containing RPD3 12. Therefore, it is highly probable that RPN4-STB2 and STB2-YRR1 are cooperative. In addition, many predicted cooperative TFs not ranked in the top 20 list have also been experimentally validated; for example, HAP2-HAP4, RPN4-YRR1, STP1-STP2 and YHP1-YOX1 (Supplementary information, Data S1-Table S1). All of these examples suggest that our predicted cooperative TFs are promising and interesting subjects for future experiments.

Figure 1
figure 1

Cooperative network of TFs. This network is built from the 186 cooperative TFs identified in this study. TFs are colored according to their functions annotated in SGD and CYGD. Blue node: cell cycle; red node: stress response; green node: metabolism; pink node: multiple function; yellow node: uncharacterized. Edges between TFs represent the types of evidence that support the cooperativity. Thick black edge: supported by experimental evidence; thin black edge: supported by computational evidence; thin gray edge: currently unsupported by any evidence.

Full size image

We further compared the power of our method with three existing methods developed by Banerjee and Zhang 1, Nagamine et al. 2 and Yu et al. 3. The overlaps between these predictions are low, which may be due to the different sources of data used in each study. We compiled 27 TF pairs from the MIPS transcription complex catalog as our benchmark data set for TF cooperativity (Supplementary information, Data S1-Table S2), which is the only high-quality data set of TF cooperativity currently available. We compared the significance of the overlap of different predictions with this data set using Fisher's exact test. The results showed that our predictions had a more significant overlap with the standard data set than the other three sets of predictions (Supplementary information, Data S1-Table S3), suggesting that the combination of binding and functional data helps improve prediction accuracy.

Using the identified cooperativity between TFs, we predicted functions for 12 uncharacterized TFs: STP4, SNT2, EDS1, STB4, YDR049W, YDR266C, YER130C, YPR196W, YFL052W, YPR022C, YFL278C and YML081W. We assumed that a given TF has a high probability of functioning in the same processes as its cooperative TF partners (Supplementary information, Data S1).

We attribute the reliability of our method to two features. First, ChIP-chip and knockout data are complementary and independent. ChIP-chip data contain information about the binding between a TF and its target(s), whereas TF knockout data provide information about the functional relationship between a TF and the genes it regulates. Thus, by combining the binding and functional data, we can identify TF pairs that both bind to target genes and work as a complex. Second, we used an optimization procedure to calculate the most significant overlap of target genes regulated by two TFs by the stepwise relaxation of P-value thresholds. When a stringent P-value threshold (0.001) was used, only 20 cooperative TF pairs were identified (Supplementary information, Data S1-Table S4), of which 4 pairs contained uncharacterized TFs. Of the remaining 16 TFs, 13 pairs were supported by literature and only 6 of these pairs were experimentally validated. In comparison, 14 pairs had supporting evidence and 9 of these pairs were experimentally validated out of the top 16 pairs with characterized TFs in our results. Many well-known cooperative TF pairs were missed when using the stringent threshold, including SWI4-SWI6, GAL3-GAL80 and MCM1-YOX1. When we relaxed the threshold to 0.005 and omitted the optimization, 117 pairs were discovered, of which 44 pairs had evidence, as compared with 68 out of 186 pairs when the optimization was included. Our method also achieved higher Jaccard similarity scores than the method without the optimization (Supplementary information, Data S1-Figure S3). These results suggest that selecting a suitable but not too stringent P-value threshold is a feasible way to uncover more interactions and achieve a low false-positive rate. The optimization principle makes sense not only in statistics but also in biology because TFs are independent. Setting the same threshold for each TF does not take this independence into account and thus could exclude some significant cooperative TFs.

In conclusion, our work provides an initial step toward identifying cooperative TFs by integrating binding and functional information in a robust manner with few arbitrary thresholds. We successfully identified many cooperative TFs that had previously been experimentally confirmed. In addition, we identified many novel potentially cooperative TFs that could lead directly to new hypotheses for future experiments. The cooperative TF networks we constructed suggest that intensive cross talk occurs between cell cycle, metabolism, protein synthesis and filamentous growth pathways at the level of transcriptional regulation. If the appropriate knockout expression profiles and genome-wide location data were available, our method could identify cooperative TFs under different conditions in yeast or other species. Although we focused on cooperativity between TFs, our method would work equally well to detect cooperativity between other regulatory factors whose binding sites can be identified (for example, microRNAs) or between TFs and other regulatory factors. Our program is available on request.