Our strategy to identify transcriptional regulatory networks combines two independent, but complementary methods: TF binding site identification and analysis of gene expression profiling data (Figure 1). TF binding site analysis is used to identify genes containing overrepresented binding sites of a TF, whereas analysis of gene expression profiling data results in one or more genes that are coherently expressed with a TF. Our hypothesis is that genes identified using both methods are more likely to be regulated by a TF than those genes identified by either of the methods alone. Thus, to predict regulatory targets of a TF by sequence analysis, all the TF binding sites that are conserved in human and mouse were identified, and binding probability scores of each TF binding each gene were calculated using a regulatory sequence analysis pipeline 29 (Figure 1a). To identify genes with a statistically significant binding score, the P value for observing a binding score by chance was calculated by randomizing all the identified TF binding sites in the genome (Figure 1b). In parallel, gene expression profiles were obtained from cultured bone marrow cells that are stimulated with the granulocyte colony-stimulating factor (G-CSF) growth factor to simulate the in vivo myeloid differentiation program 30. This system allowed us to identify important genes that are upregulated during myeloid differentiation. Template matching was used to identify genes whose expression is similar to a specific expression pattern or 'template'. This step identified six coexpressed gene clusters (Figure 1c). The regulatory interactions identified by sequence analysis were used to construct regulatory networks for each cluster, which were then consolidated to establish the regulatory network underlying myeloid differentiation (Figure 1d). Each of these steps is described in greater detail in the following sections.

To identify regulatory targets of TFs using genomic sequence information, all of the evolutionarily conserved TF binding sites in the human genome were identified using 596 known TF binding profiles curated in the TRANSFAC 31 and JASPAR 32 databases. Using TF binding sites found in the non-coding sequence of a gene, binding probability scores 29, which assess the likelihood of a TF regulating a gene, were calculated for each TF-gene pair (Figure 1a). The P value for observing a binding score for a TF-gene pair by chance was then calculated by permutation of all the binding sites in the genome (Figure 1b). By applying a P value cutoff, genes that have statistically significant binding scores for a TF were identified as putative targets of that TF. The appropriate P value cutoff was determined empirically to be 0.005 by using the total number of transcriptional regulatory interactions estimated in a previous study 15. As a result, 106,997 TF-target gene pairs (that is, a TF regulating a target gene), including 6,474 TF-target TF pairs (that is, a TF regulating another TF gene) were identified. Using human-mouse ortholog gene pairs calculated using the HomoloGene database (see Materials and methods), these transcriptional regulatory interactions predicted in human were mapped to orthologs in mouse and thus generated 102,346 TF-target gene pairs. To determine if these regulatory relationships were supported by other computational prediction methods, these results were compared to the data curated in PReMod, a database of genome-wide cis-regulatory module predictions 33. As a result, 40.3% of these TF-target pairs were also predicted in PReMod.

To elucidate the transcriptional regulatory networks underlying myeloid development, expression profiling data were utilized from a previous study that employed a well characterized model of in vitro myeloid differentiation 30. In this model, G-CSF is used to stimulate the maturation of enriched myeloid progenitors. During the seven-day time course, the predominant cells in culture at days 2 and 3, at days 4 and 5, and at days 6 and 7 are promyelocytes, mid-myeloid cells, and terminally differentiated myeloid cells, respectively (Figure 2a). Using these data, we identified coherently expressed genes during myeloid differentiation. Because myeloid development is a unidirectional, progressive event, it was hypothesized that genes regulated during this process have relatively simple expression patterns (that is, up-regulated or down-regulated at one or more points during myeloid development). In fact, comparing gene expression profiles during the in vitro system to an exhaustive list of temporal patterns revealed that the majority of genes that were triggered at some point in the seven-day time course were up-regulated either on just one day or over two consecutive days (Additional data file 1). Therefore, we focused on these two types of expression patterns. Expression patterns that were upregulated on just one day (day 0 to day 7) or over two consecutive days (days 0 and 1 to days 6 and 7) during the myeloid development were defined as 'templates'. The Pearson's correlation coefficient was calculated for each gene expression profile and template. To focus on the transcriptional regulation of the most coherently expressed genes, a correlation coefficient cutoff of 0.9 was used to identify genes whose expression profiles match each template. Using this method, six coexpressed gene clusters were identified, including genes upregulated on day 0, 1, 2, or 7, and genes upregulated on days 0 and 1 or days 6 and 7 (Figure 2b; Additional data file 2). The clusters that were upregulated on day 0, on days 0 and 1, and on days 6 and 7 include the most genes (267, 138, and 118 genes, respectively). These clusters contain many well known genes that are associated with myeloid differentiation, including those encoding myeloid differentiation antigens (for example, Cd2, Cd3d, Cd5), and terminal myeloid differentiation genes (for example, Mmp9, Fpr1, and Itgam).

The results of genomic sequence analysis and expression profiling analysis were integrated to construct the genetic network associated with each coexpressed gene cluster. The regulatory targets of myeloid TFs were identified by the intersection of genes found in the same cluster with a TF (that is, those having a similar expression pattern to that of the TF) and genes having statistically significant binding scores with that TF. Using this approach, 96 and 25 TF-target gene pairs were identified for the gene clusters upregulated on day 0 and days 0 and 1, respectively (Additional data file 3). Note that there were also genes that were down-regulated on days 0 and 1 and days 6 and 7 (Additional data file 4), but none of the TFs that were up-regulated at those time points and that had a known binding profile were predicted as a regulator of these genes. Thus, these genes may be regulated by other myeloid TFs whose binding profiles are not yet available. Using the entire set of predicted TF-target gene pairs, a genetic network was constructed for each gene cluster (Additional data file 5). In these networks, TFs and their target genes are represented by nodes, and a directed edge is drawn from a TF to a gene if the TF regulates the gene. The identified genetic networks allowed for the identification of previously unknown TFs that regulate myeloid differentiation as well as regulatory target genes of known myeloid regulators. For example, Egr1 was shown to be a candidate myeloid regulator by previous studies 343536. However, what myeloid genes are directly regulated by Egr1 is unclear. Using our results, seven genes were found to be potentially regulated by Egr1, including three genes encoding TFs (Figure 3a). Among these genes, Dusp5 and Egr1 are both strongly upregulated after interleukin-5 treatment in eosinophils 37. Lmna modulates cellular responses to the transforming growth factor-beta 1 (Tgfb1) signaling pathway 38, and Tgfb1 is regulated by Egr1 39. Further computational analysis showed that all of these genes have evolutionarily conserved Egr1 binding sites in their proximal promoter region (Figure 3b). These results suggest that the regulatory networks constructed by our method identified potential target genes of Egr1.

The previous analysis identified myeloid TFs that are coherently expressed with their target genes during myeloid differentiation. However, there may be additional myeloid TFs that may not share a similar expression profile with their target genes (for example, myeloid TFs that are constantly expressed). The TF binding site analysis described above identifies TFs that regulate individual genes in the coexpressed gene clusters. Thus, PAP 29 was utilized to identify additional myeloid TFs that regulate a set of genes in myeloid gene clusters. PAP scores each TF and predicts TFs that regulate a set of coexpressed genes using a statistical model that is based on TF binding sites and that is used to calculate a P value to assess the statistical significance of this binding (see Materials and methods). Using a P value cutoff of 0.05, up to five additional TFs were identified for each of the coherently expressed gene clusters (Table 1). A majority of these additional TFs are known myeloid regulators or are involved in leukemia pathophysiology, including AML1, PU.1, and C/EBPα. These TFs were added into the genetic network for each gene cluster as new nodes (Additional data file 6), and connections from these TFs to other genes in each individual network were made based on the TF-target gene pairs predicted by genomic sequence analysis (Additional data file 7).

To systematically study the transcriptional regulatory mechanisms underlying myeloid differentiation, focus was given to TF genes in each coexpressed gene cluster. These TFs presumably modulate myeloid development by regulating the genes in the same gene cluster (that is, genes that were coexpressed with the TFs). Indeed, 12 of the 17 TFs found in these gene clusters were previously associated with myeloid differentiation or myeloid disorders (Table 2). Therefore, the regulatory networks of TFs were extracted from each genetic network identified for each gene cluster (Additional data file 8). Because genes in each coexpressed cluster are upregulated at different time points during myeloid development, each individual network represents a 'sub-network' of the entire transcriptional regulatory network for myeloid differentiation. Therefore, a comprehensive transcriptional regulatory network was constructed by combining each individual network identified in each coexpressed gene cluster. Namely, each individual network was joined by the common TFs to build a combined network (Figure 4a).

The identified genetic network for myeloid differentiation has several interesting features. First, the integration of individual networks into the complete myeloid development network was consistent with the time at which each individual gene cluster was upregulated (that is, genes upregulated earlier occupied the upper part of the network and genes upregulated later occupied the lower part). Second, multiple regulatory pathways converge on a single TF, Rora. These characteristics highlight the genetic circuitry that may be operating in the myeloid differentiation.

To identify transcriptional regulatory mechanisms that are dysregulated in acute promyelocytic leukemia, TFs that are differentially expressed in APL compared to normal cells were identified as follows. Gene expression profiles in APL were collected from a mCG-PML-RARα knock-in mouse model 30. PML-RARα is a fusion protein observed in most APL patients, and the majority of PML-RARα knock-in mice eventually develop APL 30. Because APL is characterized by an arrest of the normal myeloid differentiation program in the promyelocytic stage, and promyelocytes are the predominant cells at day 2 and day 3 of the in vitro myeloid maturation program, gene expression data measured at these two days were compared to those measured in APL mice. The software program SAM (Significance Analysis of Microarrays) 40 was used to identify 602 differentially expressed genes in APL. Among these genes, 472 were overexpressed in APL, and 130 were underexpressed. These differentially expressed genes include seven encoding TFs (Stat5b, Fos, Atf1, Arid5b, Rxra, Mybl2, Nfyc) that have characterized binding profiles, termed the APL dysregulome (Table 3). Four of the seven dysregulated TFs, Atf1, Fos, Rxra and Stat5b, have been previously associated with acute myeloid leukemia 41424344, and ninety-six genes that were differentially expressed in APL were identified as targets of these TFs (Additional data file 9).

To test if these seven TFs in the APL dysregulome participate in a common regulatory pathway (that is, their abnormal expression is the cause or result of a single regulatory cascade), transcriptional regulatory interactions between these TFs were identified using regulatory sequence analysis (see above and Materials and methods). Namely, TF-A and TF-B are connected to form a regulatory pathway if TF-B has over-represented binding sites of TF-A. As a result, six of these seven TFs were shown to form a common regulatory pathway (Figure 4b). The last TF, Nfyc, neither regulates nor is regulated by any of the other six TFs, but it regulates a gene (Actl6a) in common with Atf1. Two of these seven TFs, Nfyc and Mybl2, are expressed at lower levels in APL than in normal promyelocytes. It is interesting to note that while Atf1 regulates Fos and Actl6a, Fos is up-regulated and Actl6a is down-regulated in APL. This suggests that Atf1 may act as both a transcriptional activator and a repressor, possibly depending on different cooperative factors. This hypothesis is supported by a previous study of Atf1 45. Thus, to identify cooperative TFs of Atf1, TFs that regulate Fos or Actl6a and have similar expression profiles to Atf1 in APL were identified. This analysis identified Egr2 and Nfyb as cofactors of Atf1 in the regulation of Fos and Actl6a, respectively.

To study how the APL dysregulome perturbs the genetic network of normal myeloid differentiation, the regulatory cascade of the seven dysregulated TFs were joined with the normal myeloid genetic network (Figure 4c). All the predicted regulatory interactions between any two TFs within the normal genetic network were maintained, and the TFs that were predicted as targets of at least one dysregulated TF were identified (Table 4, Figure 4c). Interestingly, these two networks could be simply combined through a common TF, Fos. This result suggests that the genetic network of normal myeloid differentiation is perturbed, and this dysregulation is mediated through Fos. Furthermore, these results predict a change in the genetic circuitry wherein the normal cascade is regulated by Rora while the pathophysiology observed in APL is mediated by Rxra. Bona fide downstream targets of Rora and Rxra need to be identified, and their functions in normal myeloid development or APL need to be elucidated to further validate the role of Rora and Rxra in normal or leukemic biology.

While the proposed genetic network predicted that the APL pathway converged on Rxra, the relationship between PML-RARα and the APL dysregulome was not uncovered. Thus, to test whether the APL dysregulome (Figure 4b) is caused either directly or indirectly by PML-RARα, PML-RARα TF targets were identified using RARα binding profiles. Because it has been shown that PML-RARα binds to a much broader range of binding site architectures than the normal RARα, eight binding profiles of PML-RARα with various orientations and spacings (DR2, DR3, DR4, DR5, DR6, DR12, IR0, and ER8) were created based on a previous experimental study 46 in addition to the RARα binding profiles in TRANSFAC. Using the same binding site permutation algorithm, none of the TFs in the APL dysregulome were predicted as direct targets of PML-RARα, suggesting that dysregulation of these TFs was mediated by other TFs.

Thus, to identify direct targets of PML-RARα, gene expression profiles collected from cultured bone marrow cells derived from young mCG-PML-RARα knock-in mice were used (see Materials and methods). First, to identify genes dysregulated in preleukemic promyelocytes, expression data from days 0, 2, and 7 in normal and preleukemic promyelocytes were analyzed using SAM. No genes were differentially expressed at day 0, whereas 73 and 1,028 genes were differentially expressed at day 2 and day 7, respectively. Using the eight binding profiles of PML-RARα and the RARα binding profiles in TRANSFAC, six TFs differentially expressed at day 7 were predicted as direct PML-RARα targets (Table 5). Interestingly, one of the TFs in the normal network, Egr1, is also predicted as a PML-RARα target; however, Egr1 is expressed at a normal level in young, preleukemic mice. Therefore, it is possible that Egr1 may not be a direct mediator of PML-RARα in leukemogenesis. Collectively, these results suggest a model of APL pathogenesis in which PML-RARα regulates the APL dysregulome through six mediator TFs. This circuitry ultimately converges to create the APL dysregulome, hallmarked by activation of Rxra, which then triggers downstream events (Figure 5).

In this report, we propose a novel approach to genetic network identification that combines two independent types of information, gene expression profiling data and computational identification of TF binding sites. Using gene expression data, genes coherently expressed with TFs were first identified. In parallel, direct regulatory targets of TFs were predicted by a computational model that calculates binding scores for each coexpressed gene and assesses statistical significance using binding site permutation. These two types of information were then integrated to construct the genetic network for each coexpressed gene cluster, which were subsequently consolidated into a comprehensive network. We used this approach to identify the genetic network in normal myeloid differentiation and to determine how this network is perturbed in APL. This approach is general and may be applied to delineate genetic networks operating in other complex human diseases.

A major challenge in expression data-based genetic network construction is the ability to distinguish direct regulatory targets of TFs from indirectly regulated downstream genes. To predict direct targets accurately, coexpressed genes were scored using the identified TF binding sites, and the statistical significance of each score was determined. This method is different from other existing approaches in several ways. First, it does not compare the frequency of TF binding sites to a specific set of reference sequences. Instead, the binding scores of genes in the entire genome are calculated, and the P value for observing a score is determined by permutation of all the binding sites. Secondly, unlike most previous methods where TF binding sites are only identified in the proximal promoter region, our model considers binding sites located in evolutionarily conserved sequences in the entire gene locus. This includes a significant number of additional, highly conserved sites found in introns and distant genomic regions. These modifications and improvements make our predictions more accurate due to a better scoring model and more comprehensive due to a more complete set of evolutionarily conserved TF binding sites 47.

In this study, we used an in vitro, G-SCF driven myeloid differentiation system to model normal myeloid maturation and compared its gene expression profile to that of an in vivo APL mouse model. Although the in vitro GCS-F driven myeloid differentiation is not equivalent to in vivo differentiation, it is a validated surrogate that mimics human myeloid differentiation 30. Moreover, this system is currently the only practical and technically feasible platform for the study of normal murine myeloid development. Therefore, we believe the data provided by this system can be used to infer hematopoietic gene regulation.

The genetic network of TFs identified by our computational method provides several new insights into the normal and aberrant regulatory pathways that may drive myeloid differentiation and in acute promyelocytic leukemia, respectively. These predictions are compared to several previous findings.

First, we observed that the normal and disease regulatory cascades converge on Rora and Rxra, respectively. Rora and Rxra belong to the same family of nuclear receptors and are related to Rara, and Rxra forms a heterodimer with Rara. Although the specific functions of Rora in myeloid differentiation are still unknown, all three nuclear receptors bind to very similar DNA sequences ((A/G)GGTCA) 4648, implying that they may regulate the same genes. In addition, the human RORA gene is located within a highly conserved region on chromosome 15 near the human PML gene 49. Therefore, our results suggest a novel role for Rora in myeloid development. Furthermore, the dysregulated pathway in APL leads to an overexpression of Rxra, and the absence of Rxra in the normal network implies that Rxra is a potential effector protein in APL but not in normal myeloid development. In fact, recent studies have shown that although Rxra is not required in normal myelopoiesis 50, it is an essential component for the PML-RARα complex to initiate APL in mice 5152. Therefore, the prediction of Rxra as a potential key participant in APL pathogenesis is supported by in vivo studies.

Second, the normal and the disease pathways did not share many common TFs. Instead, only one TF in the normal network, Fos, which modulates myeloid cell survival and differentiation 42, was dysregulated in APL. Our analysis identified that the TFs dysregulated in APL may mediate the perturbation of the normal myeloid genetic network through Fos. Indeed, previous studies have demonstrated that PML-RARα promotes cell growth by activating Fos 53, and that PML-RARα directly functions as a co-repressor of Fos in the absence of retinoic acid 54. Our data suggest that over-expression of Fos in APL may have an aberrant activation function as well. This over-expression of Fos may be induced by PML-RARα and may, in turn, cause the dysregulation of other TFs involved in APL.

Third, we identified a set of six TFs as direct targets of PML-RARα. Among these six TFs, five of them were overexpressed in APL, and one TF was underexpressed, suggesting PML-RARα may act as both a repressor and an activator. In agreement with this observation, PML-RARα was first demonstrated as an enhanced repressor of retinoic acid target genes by its stronger binding to corepressors than wild-type RARα 5556, but it is becoming clear that PML-RARα may also function as an activator 5457. While a version of PML-RARα that can only repress gene expression is still leukemogenic 58, whether PML-RARα can trigger APL pathogenesis by gene activation is unknown. Our analysis suggests that PML-RARα may indeed be an activator and provides a potential mechanism through which this may occur.

Finally, our data are consistent with a previous study that showed that relative levels of Sfpi1 regulate macrophage versus neutrophil differentiation 59. Importantly, at low Sfpi1 levels (neutrophil differentiation), we did not find that Egr1/2 are targets of Sfpi1 whereas Laslo et al. 59 found that Egr1/2 are targets of Sfpi1 during macrophage differentiation where Sfpi1 is expressed at high levels. Therefore, there may be weaker binding sites for Sfpi1 in Egr1 or Egr2 promoters, and thus the activation of Egr1 and Egr2 may require a higher expression level of Sfpi1.

The computational approach used in this study is based on the binding profiles of TFs in TRANSFAC and JASPAR. Therefore, the quality of our prediction and the accuracy of our conclusions may be dependent upon the quality of the binding models in these databases. Some TFs require other co-factors to accomplish their regulatory functions, and so their binding specificities or preferences may also be dependent on their binding partners. Such information may not have been included in the TRANSFAC or JASPAR models. For example, the TF RXRA studied in this work typically forms a heterodimer with various co-factors, including RAR, VDR, TR, or PPAR at various spacings. The four binding profiles we used were created using different binding partners of RXRA, including PPAR (M00518), RAR and TR (M00963), and VDR (M00966). Thus, our prediction did consider the binding site of RXRA with different binding partners. However, our prediction may not include all the allowable spacings between the two sites. Therefore, our results may be further improved when more precise and complete binding profiles are available.

While the direct targets of PML-RARα in APL were computationally identified, the direct targets of these mediator TFs could not be determined. Therefore, the genetic network in APL was not fully elucidated (Figure 5). In order to identify targets of these mediator TFs, knowledge of their binding profiles is required, information that is currently not available. In fact, while there are more than 2,000 TFs predicted in the human genome 60, only a quarter of them have known binding profiles. Thus, the predicted genetic network may be further improved and expanded when more TF binding profiles become available. Additional TF binding profiles may be generated by traditional protein-DNA binding assays 61 or by computational approaches that utilize evolutionary conservation of functional sequences. This latter approach is used to predict DNA binding profiles on a genome-wide scale. For example, DNA binding patterns may be identified by calculating the conservation rate of a given oligonucleotide across the genome 62, or by clustering genes that share common conserved sequences 63. Thus, the employment of such methods is a rational next step toward the refinement of genetic networks. Integration of this computational component would not only increase our understanding of the molecular mechanisms underlying APL but would also facilitate the construction of more comprehensive regulatory networks driving other complex diseases.

The genomic sequences of human, mouse, and rat were acquired from the NCBI's Genome Assembly Project 64. Genome build 35 was used for human, build 33 for mouse and build 3 for rat. The genomic sequence of a gene locus was defined as the sequence between the end of the upstream gene and the end of the gene itself. Within this range, protein coding sequences were masked and excluded from the search of TF binding sites. Repetitive elements were also masked by the program RepeatMasker 65 using slow and sensitive mode (the -s flag). Human, mouse and rat ortholog gene groups (13,194 in total) were identified using the annotation of NCBI's HomoloGene database as previously described 29. Genomic sequences of the genes in the same ortholog group were then aligned using the program TBA 66.

To identify TF binding sites, 596 vertebrate TF binding profiles were collected from the TRANSFAC (version 9.1) and JASPAR databases. The program PATSER 67 was used to search for matches of these profiles in the genome using default cutoff scores (the -li option). This cutoff score is calculated as follows: for each position, PATSER scores the subsequence and calculatesthe P value for observing the same score or higher at thatposition 68. A P value cutoff is calculated for each binding profile using its information content. The score corresponding to that P value cutoff is then chosen to be the cutoff score. After all the TF binding sites were identified, binding probability scores 29 for each TF-gene pair were then calculated using evolutionarily conserved TF binding sites found in the non-coding sequence of a gene. For mammals, functional regulatory elements have been found in distant upstream regions 6970 as well as intronic sequences 7172. However, searching for TF binding sites in the entire intergenic sequence and in the entire gene locus frequently results in a high false discovery rate. Therefore, to overcome this problem, we considered only evolutionarily conserved TF binding sites in the 'proximal promoter region' of a gene and in the most conserved sequence regions within a gene locus defined by multi-species conserved sequences (MCSs) 73.

The proximal promoter region of a gene was defined as the 10 kb upstream sequence and the 5 kb downstream sequence from the transcription start site, regardless of the presence of an upstream gene or a downstream gene. MCSs are defined as the top 5% conserved sequences in the human genome when compared to another 11 vertebrate genomes 73. The human MCSs were first downloaded, and the MCSs in mouse and rat were defined by mapping the human MCSs to the mouse or rat genomes using multiple sequence alignments generated by TBA. These multiple sequence alignments were also used to identify evolutionarily conserved TF binding sites, which were defined as sites located within conserved sequence regions, present in all the species and aligned in the multiple sequence alignment 74.

To predict regulatory targets of TFs, genes that have statistically significant binding scores 29 for a TF were identified. The statistical significance of a given binding score was evaluated by the P value for observing an equal or higher score by chance. This P value was calculated by permutation of all the TF binding sites in the genome. In this algorithm, each individual binding site of a TF was randomly assigned to genes in the genome based on a precalculated probability distribution calculated as follows: the probability for a gene to acquire a given TF site is the length of the TF binding site search range of that gene (that is, the proximal promoter region and the MCSs; see above) divided by the sum of the TF binding site search ranges of all the genes in the genome. This permutation of TF binding sites was performed for 10,000 iterations. After each iteration, a new binding probability score was calculated for each gene using the TF sites randomly assigned to that gene. The P value for observing a score for a gene was then calculated by the number of iterations where a binding score equal to or higher than the true score was obtained for that gene, divided by the total number of iterations.

Gene expression profiles in normal myeloid differentiation were obtained from a G-CSF stimulated in vitro myeloid differentiation model 30. Expression data were collected from cultured bone marrow cells in two independent experiments during a seven-day time course. Gene expression data in APL were collected from bone marrow cells of six adult PML-RARα knock-in mice based on a previously developed murine APL model 30. Expression profiles of the cultured bone marrow cells were also collected from two sets of young PML-RARα knock-in mice. For the expression data collected from cultured cells, only probesets that were present on at least one day in at least one experiment were considered.

The coexpressed gene clusters during myeloid differentiation were identified using the template matching clustering tool in the software suite FunctionExpress 75. The templates for each upregulated expression pattern were created manually, including genes that were upregulated on just one day or over two consecutive days during the seven day time course. The Pearson's correlation coefficient was used to quantify the similarity between the template and the expression profile of each probeset. A cutoff of 0.9 was applied to the correlation coefficient to identify the coexpressed gene clusters.

PAP 29 was used to identify additional TFs that may regulate genes in each coexpressed gene cluster. For each gene cluster, PAP ranked all the TFs by their R-scores, which were calculated based on overrepresentation of their binding sites in the coexpressed genes. To find TFs that have a statistically significant R-score, a P value for each R-score was calculated using 10,000 randomly selected gene clusters of the same size. For each random set, the R-scores for each TF were calculated, and the P value for a TF was calculated as the number of gene clusters that had an equal or higher score than that of the original gene cluster, divided by 10,000, the total number of random sets.