Expression Quantitative Trait Locus (eQTL) mapping searches across the genome for markers associated with individual transcripts to identify loci containing regulatory elements [1, 2]. Although cis regulators are easily interpreted, assigning function to trans regulators is more difficult. At the transcriptional level, genes in trans are often co-regulated by transcription factors binding the same regulatory elements in the noncoding sequences of multiple genes [3]. When looking genome-wide, however, genes across many ontologies acting upstream of transcription factors participate in the co-regulation of genes [4, 5]. Because of the difficulty in predicting the trans-regulators, the majority of transcriptional regulatory proteins remain unknown.

Identification of genetic components of gene co-regulation is important because there is compelling evidence that aberrant gene co-regulation participates in human disease [6, 7]. Investigations into the genetic basis of gene co-regulation have used Bayesian networks to identify cis- and trans-acting factors controlling modules of co-regulated genes [8–10] or gene clusters. A key result here was the identification of associations that would have been missed when genes were tested individually, as in traditional eQTL. Biologically, this is very compelling because genes typically do not perform their functions in isolation but rather in coordinated groups. The existing Bayesian methods have focused primarily on the identification of yeast regulatory programs where other sources of information, such as sequence conservation, transcription factor binding site (TFBS) data, or protein interaction data are readily available and serve as prior information [8, 10]. Extension of these methods to human genetics with data from HapMap subjects has shown that sequence conservation and cis-regulatory information were the most useful prior data [8].

Studies in other human cohorts and mice have used directed Bayesian networks or undirected weighted gene coexpression networks to incorporate existing marker and phenotype data into models that have made biologically validated predictions [11–16]. These network-based methods and their alternatives (e.g. [17–21]) show great promise but often have high computational complexity, making them most practical for smaller datasets with limited numbers of traits. Furthermore, prior information is generally not readily available in humans. For example, most TF binding sites remain unknown and even within the same tissue, the vast majority of TFBS appear divergent between human and mouse [22]. This suggests that relying on sequence conservation or the presence of conserved TF binding motifs may miss some key associations and that agnostic, complementary methods should be developed.

Many algorithms that search for the determinants of gene co-regulation assign each gene to a single cluster (e.g. [4, 9, 23]), which is limiting, because genes can belong to different clusters under different biological conditions [24]. More recent network approaches overcome this problem by examining differential canonical correlation between multiple states, such as healthy and diseased, or with a reference [11]; these approaches, however, may rely on methods that are not robust to non-normal data to find correlated genes. This may be a problem for gene microarray expression data, which is often not normally distributed. Alternatively, robust, "mega-clustering" methods have been developed to provide improved estimates of co-regulation for microarray data [25]. One such algorithm-the 'Gene Recommender'-has successfully predicted previously unknown interactions that were verified experimentally in a multicellular organism [26]. A key property of the Gene Recommender is the categorization of genes into clusters under different conditions (i.e., allowing for "biclusters") where different samples' contributions to the given cluster may vary. Since inexpensive genotyping platforms can presently interrogate >1 million SNPs and we are rapidly shifting into the era of whole genome sequencing, existing genetics and systems biology methods would be nicely complemented by computationally feasible, agnostic approaches to the detection of trans-acting factors that regulate groups of genes.

Therefore, here we extend the Gene Recommender with an approach that can systematically identify trans loci controlling gene co-regulation. We broadly refer to the identification of gene co-expression trait loci as 'co-regulatory expression quantitative trait loci' (creQTL) mapping. Unlike trans-eQTL analysis, our method does not consider individual transcripts but rather focuses on multiple co-regulated transcripts. We provide a genome-wide implementation of the Gene Recommender and a statistical framework for association testing. The key steps of the creQTL approach are: gene clustering; calculation of each sample's similarity to each cluster; and statistical testing of how well genotype explains the similarity. We applied creQTL to a study of germline variants and tumor expression in endometrial cancer [27] and identified many loci significantly associated with gene co-regulation. These loci were commonly in noncoding regions closely in cis with genes encoding proteins required for transcription, signaling, cell adhesion, and development. Our results suggest that associating genetic variants with co-regulation via creQTL mapping provides an efficient and agnostic avenue for detecting biological factors important in the coordinate regulation of groups of genes.

The key steps in our creQTL approach and application to identify loci controlling co-regulation of genes are outlined in Figure 1. Because of the overall unreliability of low-expression measurements in gene microarray expression data and our interest only in genes with dynamic expression patterns, we prefiltered those data to exclude the bottom 50% of probes by variance, leaving 8,543 for inclusion in our analysis of gene clustering and QTL testing. 68,523 SNPs from the Affymetrix 100K chip met cutoffs for inclusion in our analysis.

From the expression data, we identified 282 gene clusters with at least 3 co-regulated genes. The distribution of gene cluster sizes is shown in Figure 2. Many genes present within the same clusters had overlapping, known biological functions, and genes within individual clusters often were syntenic, such as the IFNβ-induced genes IFI44, IFI44L, and IFIT1, -2, and -3, on chromosomal regions 1p31.1 and 10q23-26, contained within a cluster significantly associated with a SNP/locus upstream of RFC3 and STARD13.

We found 453 associations between genotype and Z2_E(j) with q-values < = 0.005 (Additional File 1 contains annotated associations) after 19,323,486 variance tests of 282 gene clusters and 68,523 SNPs. Note that a FDR q-value of 0.005 corresponded to an unadjusted p-value of approximately 1 × 10^-7. Table 1 and Figure 3 highlight two of the most significant creQTLs that were within 5 MB of a protein-coding gene. As shown in Figure 3, for statistically significant creQTLs, the SNP genotypes exhibit varying contributions to the clusters. Moreover, samples that contribute most to a given cluster have more extreme, more tightly-clustered expression values; samples that contribute little to a given cluster have expression values (and contributions) closer to zero with greater variability. Additional File 2 shows the q-value density for associations as a function of gene cluster size. As can be seen in the figure, there was no relationship between q-values and gene cluster size.

creQTL 1
SNP Annotation
Locus	dbSNP	Location (bp)	Position	Gene Symbol	Gene Title
chr13q12.3-q13	rs9315220	304602	upstream	RFC3	replication factor C (activator 1) 3, 38 kDa
chr13q12-q13	rs9315220	227702	upstream	STARD13	StAR-related lipid transfer (START) domain containing 13
Gene Cluster Annotation
Locus	Entrez ID	q creQTL	p eQTL	Gene Symbol	Gene Title
chr1p31.1	10561	1.13 × 10-4	0.21	IFI44	interferon-induced protein 44
chr1p31.1	10964		0.28	IFI44L	interferon-induced protein 44-like
chr10q23-q25	3433		0.21	IFIT2	interferon-induced protein with tetratricopeptide repeats 2
chr10q24	3437		0.49	IFIT3	interferon-induced protein with tetratricopeptide repeats 3
chr10q25-q26	3434		0.19	IFIT1	interferon-induced protein with tetratricopeptide repeats 1
chr12q24.1	4938		0.52	OAS1	2',5'-oligoadenylate synthetase 1, 40/46 kDa
chr21q22.3	4599		0.56	MX1	myxovirus (influenza virus) resistance 1, interferon-inducible protein p78 (mouse)
creQTL 2
SNP Annotation
Locus	dbSNP	Location (bp)	Position	Gene Symbol	Gene Title
chr1p31.1	rs2296697	0	intron	LPHN2	latrophilin 2
Gene Cluster Annotation
Locus	Entrez ID	q creQTL	p eQTL	Gene Symbol	Gene Title
chr2p25	6664	5.40 × 10-4	0.92	SOX11	SRY (sex determining region Y)-box 11
chr3p14.3	57408		0.38	LRTM1	leucine-rich repeats and transmembrane domains 1
chr11q22.3	4317		0.10	MMP8	matrix metallopeptidase 8 (neutrophil collagenase)
chr14q11.2	9362		0.38	CPNE6	copine VI (neuronal)
chr21q22.3	54058		0.64	C21orf58	chromosome 21 open reading frame 58
NA	NA		0.35	NA	NA

For each of the two creQTLs shown in this table (labeled 'creQTL 1' and 'creQTL 2'), the associated SNP (labeled 'SNP Annotation') is first shown with annotation for the nearest coding genes in cis. Beneath the SNP annotation is annotation for the significantly associated, co-regulated gene cluster (labeled 'Gene Cluster Annotation'). These two significant associations correspond to the two loci shown in Figure 3. 'Position' indicates the relative location of the SNP relative to the nearest coding gene; 'Location (bp)' indicates the physical distance, in base pairs, between the SNP and the nearest coding genes; p _eQTL indicates unadjusted significance of standard eQTL analysis.

For the association between rs9315220 and the cluster noted above, several other nearby SNPs (rs9315219, rs9315215, rs7981602, rs4943110) were also significantly associated with this cluster (see Additional File 1), reflecting the high linkage disequilibrium (LD) among them (the pairwise r² across all five SNPs ranged from 0.92 to 0.96 in these 52 samples). The smallest q-value was nearest to the STARD13 locus (q-value = 1.13 × 10^-4). Another strong association (q = 5.40 × 10^-4) shown in Figure 2 was between the intronic SNP rs2296697 in LPHN2 (latrophilin 2) and a cluster of genes including C21orf58, CPNE6, SOX11, MMP8, LRTM1, and one unannotated gene. The genes within this cluster have established roles in Ca²⁺-dependent signaling and cancer [37–41] (Discussion). Interestingly, a conventional trans-eQTL analysis between the SNPs and individual expression values of these co-regulated genes did not detect any associations (see Table 1); these results were typical of other clusters (data not shown).

The strongest regulatory loci for creQTL (q = 5.50 × 10^-4) are shown in Table 2. We calculated pairwise r² between all 453 SNPs from creQTLs with q < 0.005 and eliminated those with r² values above 0.5, leaving 338 creQTLs. Of the 338 SNPs remaining, 190 were in noncoding regions; 145 were intronic; two were in coding sequences; and one was in the 3' untranslated region. Interestingly, many of the associated creQTLs contained SNPs in noncoding sequences >500 kb from the nearest gene. After assigning the 68,523 SNPs to 8,398 protein-coding genes in cis (Methods), we plotted the resulting 8,398 q-values versus cis-location relative to the nearest gene in Figure 4. Consistent with the pattern for the highest-scoring hits at q < 0.005, associations appeared most enriched in the intronic regions, and within the non-intronic, noncoding regions, significant associations appeared most in regions >10 KB away from the coding sequence.

Locus	dbSNP	Location (bp)	Position	Gene Symbol	Gene Title	qcreQTL
chr6q16	rs959186	125994	downstream	RNGTT	RNA guanylyltransferase and 5'-phosphatase	1.13 × 10-4
chr13q12.3-q13	rs9315215	348455	upstream	RFC3	replication factor C (activator 1) 3, 38 kDa	1.13 × 10-4
chr13q12-q13	rs9315215	183849	upstream	STARD13	StAR-related lipid transfer (START) domain containing 13	1.13 × 10-4
chrXq21.33	rs4969656	695921	upstream	DIAPH2	diaphanous homolog 2 (Drosophila)	1.93 × 10-4
chrXq21.3-q22	rs4969656	2315179	upstream	NAP1L3	nucleosome assembly protein 1-like 3	1.93 × 10-4
chr3q13.1	rs4856121	2246645	upstream	ALCAM	activated leukocyte cell adhesion molecule	1.93 × 10-4
chr5q31-q32	rs319217	0	intron	PPP2R2B	protein phosphatase 2 (formerly 2A), regulatory subunit B, beta isoform	1.13 × 10-4
chr1p31.1	rs2296697	0	intron	LPHN2	latrophilin 2	5.40 × 10-4
chr2p21	rs222471	31099	downstream	KCNG3	potassium voltage-gated channel, subfamily G, member 3	4.91 × 10-4
chr2p21	rs222471	49701	upstream	COX7A2L	cytochrome c oxidase subunit VIIa polypeptide 2 like	4.91 × 10-4
chr4q21	rs1992489	59023	upstream	CXCL13	chemokine (C-X-C motif) ligand 13	5.19 × 10-4
chr4q21.1	rs1992489	282671	downstream	CCNG2	cyclin G2	5.19 × 10-4
chr14q13-q21	rs1951319	680177	downstream	RPL10L	ribosomal protein L10-like	5.48 × 10-4
chr14q21.2	rs1951319	717439	upstream	C14orf106	chromosome 14 open reading frame 106	5.48 × 10-4
chr8p23.2	rs1714757	0	intron	CSMD1	CUB and Sushi multiple domains 1	1.47 × 10-4
chr16p13.12	rs1159167	147599	upstream	ERCC4	excision repair cross-complementing rodent repair deficiency, complementation group 4	4.91 × 10-4
chr16p13.12	rs1159167	968670	upstream	CPPED1	calcineurin-like phosphoesterase domain containing 1	4.91 × 10-4
chr22q12.1	rs10483151	0	intron	TTC28	tetratricopeptide repeat domain 28	1.81 × 10-4

Associated gene clusters have been omitted for brevity and are listed in Additional File 1. 'Position' and 'Location (bp)' as for Table 1; q _creQTL ' indicates FDR-adjusted significance.

creQTL provides an agnostic framework for predicting trans regulators of clusters of genes. It does not require any one particular biclustering method or statistical test and can be extended to any species for which genetic markers (including those derived from linkage) and gene expression data can be obtained; for example, creQTL mapping in mice would allow the study of gene regulation in many disease models and in normal biological processes such as development, which are highly relevant to human disease. Like eQTLs, creQTLs may be incorporated into additional analyses. As recent studies have used eQTL data to assemble genetic loci into directed, Bayesian networks to predict causal relationships (e.g. [16, 42]), future studies could assemble creQTLs into directed or undirected gene networks. Because creQTL mapping groups co-regulated genes into a smaller number of modules, network construction using creQTL information in place of eQTL information should be computationally more efficient.

In our application to endometrial cancer, there was an abundance of creQTLs in noncoding regions and introns, and relatively very few in coding regions. Polymorphisms in coding regions, although observed less frequently, would likely generate more pronounced effects, and for signaling molecules, these could result in substantial downstream effects important for carcinogenesis due to a change in function of the protein. However, these alleles, which are likely to be under strong negative selection and very rare, would not be covered by the genotyping platform used in the present study. Unlike the results of eQTL studies, our strongest hits were not typically located in cis regions very close to the coding sequences of individual genes, but rather in more distant, noncoding regions located over 10 KB from the nearest gene. This suggests that regions that regulate clusters of genes, unlike those regulating individual genes, are located farther from coding sequences and may possibly function as general enhancers. However, without careful experimentation, such as promoter reporter gene assays that would specifically test the enhancer activity of these sequences, it is difficult to predict their actual roles. Future studies could address the relationship between location, allele frequency, and other empirical aspects of creQTLs in larger cohorts and healthy tissue. Collapsing SNPs into a dominant model would permit testing of less common SNPs, as requiring five observations in each genotype essentially restricted our analysis to more common variants.

Our results for creQTL from the Gene Ontology enrichment analysis indicated a regulatory hierarchy, consistent with data from wet-lab studies, where signal transduction molecules receive signals at the plasma membrane and transduce them through various intracellular intermediates to the nucleus to activate specific transcription factors, other DNA-binding proteins, and other transcriptional regulatory proteins. The relative abundances of DNA-binding proteins at the promoter region may control gene expression [43]. Because of the enrichment of cell adhesion and signaling ontologies at the cell junction and membrane at the highest significance levels, the major steps of gene regulation in endometrial cancer may occur not within the nucleus but rather at the cellular membrane. In addition, GO enrichment in genes required for telomere maintenance suggests additional target genes and loci for further investigation, as telomerase has long been considered an attractive target for therapy in endometrial carcinoma [44] and more recently, a mouse model of endometrial cancer showed that telomere length affected initiation of type II carcinogenesis [45].

Gene Ontology analysis is not without its own pitfalls. Besides the obvious effects of overlapping annotations for the majority of genes and annotation bias, the categories themselves often do not describe real biological phenomena. Given that transcription factors are very cell-type specific, a category that includes all of them is not particularly useful. In light of this, the q-values for transcription factor-related categories are likely to be overly conservative, and would probably be smaller if those TFs not expressed in this tissue were excluded. Future studies could address this problem using array and wet lab expression data.

For the association shown in Figure 2, the nearest coding sequence to rs9315220 is STARD13, a.k.a. DLC2 (deleted in liver cancer 2). All genes of this cluster but one (OAS1), including MX1, IFI44, IFIT2, IFIT3, IFI44L, and IFIT1, are known to be induced by IFNβ in humans[46]. IFNβ, besides its role in the treatment of multiple sclerosis (MS), is anti-tumorigenic [47–50]. Similarly, many genes from this cluster also have established or putative roles in cancer; IFI44 was part of a small set of genes independently validated for recurrence status in non-small cell lung carcinoma [37], is anti-proliferative in melanoma cell lines [38], and is upregulated in squamous cell carcinoma [39]. IFIT2 inhibits migration and proliferation in squamous cell carcinoma cultures [41]. IFIT1 was one of a small group of genes that predict carcinogenesis after training on DNA-damage response [40], while STARD13 is upregulated in human lymphocytes following gamma-irradiation [51], suggesting that IFIT1 may be a downstream target of STARD13 following DNA damage.

STARD13/DLC2 encodes a Rho GTPase activating protein and was identified from a region of chromosome 13 exhibiting a loss of heterozygosity in hepatocellular carcinoma [52]. It is a tumor suppressor that antagonizes Rho expression [53], and its downregulation or deletion has been reported in multiple cancers [54, 55]. Supporting its role as a regulator of IFNβ-induced genes, an intronic SNP in STARD13 was amongst eighteen that were replicated in a study of responders to IFNβ therapy in MS patients [56]. Further, SNPs in OAS1 have been associated with MS susceptibility [57]. Therefore, this particular gene cluster and its predicted regulatory partner, STARD13, may be interacting determinants of the response to various negative stimuli. This also proposes the hypothesis that the role of STARD13 as a tumor suppressor is from its induction in response to DNA damage, and that deletion of this region in at least ten different cancers is pathogenic possibly due to hyper-activation of Rho, which promotes migration, proliferation, and invasion [58]. In this case, beyond improvement of the total lack of significant association from the individual eQTLs, creQTL proposed a more complete picture of the biological pathway.

We detected a strong association of rs2296697, within an intron of LPHN2 (latrophilin 2), with a cluster of genes including C21orf58, CPNE6, SOX11, MMP8, and LRTM1. LPHN2 belongs to the latrophilin subfamily of G-protein coupled receptors, with known roles in cell adhesion and signal transduction [59]. Cell adhesion was among the most enriched categories from Gene Ontology analysis of creQTL. CPNE6 (copine VI), is a member of the ubiquitous copine family, which bind phospholipid in Ca²⁺-dependent manner and have roles in cellular division and growth [60]. Although very little is known regarding CPNE6, other members of the copine family have been shown to promote tumor cell migration [61] and repress NF-KB transcription [62]. SOX11 is a TF with important roles in brain development that is strongly upregulated in lymphoma [63] and malignant glioma, where it may affect tumorigenesis [64, 65]. SOX11 may also be a prognostic marker for recurrence-free survival in another uterine neoplasia, epithelial ovarian cancer [66]. MMP8 (human neutrophil collagenase) is a member of the matrix metallopeptidase family. Recent work has shown that a SNP in MMP8 may be a predictor of lung cancer risk [67], and that higher plasma levels of MMP8 may protect against lymph node metastasis in breast cancer [68]. Consistent with these findings, somatic mutations of MMP8 were common in melanomas, and mutated MMP8 failed to inhibit tumor formation in vivo [69], suggesting widespread roles for MMP8 in cancer progression. Three members of this gene cluster have no known roles, yet the functions of the other cluster members suggest they could contribute to cancer progression. In this case, the association of this cluster of genes with LPHN2 has suggested that LPHN2 may be a regulatory point for the modulation of genes with important roles in cancer, including MMP8 and SOX11.

eQTL analysis was not done in the original paper presenting this dataset [27], and the general lack of findings from the eQTL analysis may be from a lack of power, given that there were only 52 samples. This is not surprising: an eQTL study of 60 samples from the CEU cohort in the HapMap data identified only 10 cis-eQTLs and 94 in trans [70], although there may have been some technical issues here [71].

However, in the present study, creQTL appeared to identify interactions that were supported by recent work in the literature. A very compelling benefit of creQTL over trans-eQTL is greater computational efficiency and a reduced multiple testing burden, as there was an approximate 25-fold reduction in the number of statistical tests. By organizing genes into co-regulated clusters prior to statistical testing, the problem becomes more computationally feasible and the resulting output more biologically interpretable. The identification of more significant hits after adjustment was likely not simply due to a lighter multiple testing penalty; cis-eQTL required 1,468,327 tests, far less than the 19,323,486 tests for creQTL, and produced no statistically significant results even at the more liberal q-value cutoff of 0.05. Overall, these results suggest that individual SNP-gene interactions are more difficult to detect (as in eQTL) when compared to relative changes in the expression levels of tightly co-regulated genes, in agreement with previous work in yeast [9]. creQTL may be better-suited to the identification of the strongest trans drivers of gene expression because it better explains the data; i.e., genes do not work individually to exert their biological effects, but rather in tightly coordinated groups. However, our results regarding the location of these drivers of co-regulation relative to coding sequences suggests that they might be disjoint from cis drivers of individual gene expression.

Some trans-eQTL hits were significant and may merit further investigation based on previously established biological roles for associated loci, indicating complementary utility of eQTL and that the two methods may be most useful when applied side-by-side. Copy number analysis of this EC dataset by the original authors revealed two regions of gain that were predictive of survival - 3q26.32 and 12p12.1 [27]. Deletions were not considered. Because PIK3CA is located in this region, the authors proposed a role for the PI3-kinase pathway, and found supporting evidence for this theory based on indirect, bioinformatic analysis of existing, in vitro data. Located 200 kb from PIK3CA is ZNF639, a.k.a. ZASC1, which is often contained within the same amplicon in cancer [72, 73]. ZNF639 was originally identified in squamous cell carcinoma as a Kruppel-like transcription factor with mRNA expression levels prognostic for survival and metastasis [72, 73]. trans-eQTL identified a SNP less than 35 kb upstream of ZNF639 (SNP_A-1686963) driving expression of the TCF12 gene at 15q21. TCF12, a.k.a. HTF4 or ME1, is a basic helix-loop-helix TF with key roles in development [74]. In a mouse experiment, TCF12 was associated with obesity [14], a key prognostic factor for endometrial cancer [75]. Because the stoichiometric concentration of TFs within the nucleus may control gene expression [43], this association proposes a link between obesity and TCF12 through ZNF639. Because ZNF639 regulates anchoring of E-cadherin to the cytoskeleton through alpha N-catenin [76], given the importance of cell adhesion proteins in cancer [77], future studies of endometrial cancer might focus on ZNF639.

With the advent of low-cost SNP and expression arrays, human genetics has become a common tool for the dissection of gene regulation. Genetic approaches to the study of transcriptional regulation are compelling because they can detect regulatory molecules at all stages of gene regulation. Our approach expands upon previous methods and uses genetic variation to help identify transcriptional regulatory mechanisms, providing a biologically intuitive approach for detecting potential links between genotype and gene co-regulation.