ADME genes with liver eQTLs, which pass multiple correction thresholds in one of the four genome-wide studies, include COMT, CYP2D6, CYP3A4, CYP3A5, MTHFR, UGT1A1, and VKORC1 (Table 2). These genes belong to a set of ‘very important pharmacogenes’ which have been identified by the PharmGKB resource (http://www.pharmgkb.org/search/browseVip.action?browseKey=vipGenes). Furthermore, several of these eSNPs are markers of pharmacogenetic associations (Table 2). These findings show that liver eQTLs may help explain the associations of SNPs with clinical phenotypes and also provide a rationale for using liver eQTL data from relevant genes to select candidate SNPs for pharmacogenomic study. For example, the three eSNPs identified in the CYP2D6 gene (Table 2) could conceivably be used to test associations with pertinent clinical phenotypes of any of the 46 drugs that, according to PharmGKB curated information (http://www.pharmgkb.org/gene/PA128#tabview=tab6&subtab=21), are metabolized by the CYP2D6 enzyme.

Liver eQTL data can also be used to inform GWAS. GWAS have identified many SNPs which are associated with clinical traits in humans. Typically there is no mechanistic explanation for the association but eQTLs could help fill this gap in the knowledge. Amongst the liver eQTL studies, Greenawalt et al. identified 122 eSNPs which were significantly associated, after multiple test correction, with a clinical phenotype in a GWAS (Greenawalt et al., 2011) and we found 32 eSNPs that were significant GWAS SNPs or markers for a GWAS SNP through strong linkage disequilibrium (LD) (r² > 0.8) (Innocenti et al., 2011). In fact, there is a significant enrichment of eSNPs among SNPs associated with clinical phenotypes in GWAS (Nicolae et al., 2010). However, a critical piece of information which is lacking from these reports is the directionality of the eQTL association (i.e., which allele associates with higher/lower transcript levels) and, thus, hypotheses about the effect of a specific allele cannot be generated.

Another use for eQTLs is in determining which gene a SNP, associated with a clinical phenotype, may mediate its effect through as there is often uncertainty about GWAS SNPs located in intergenic regions. Schadt et al. provide examples of intergenic eSNPs that were found to associate with type 1 diabetes, coronary artery disease and LDL cholesterol levels in GWAS (Schadt et al., 2008). By identifying the gene whose expression the GWAS SNP associates with, evidence is provided to suggest that the gene may be a susceptibility factor for the disease.

There are many factors related to the liver eQTL studies which could act as confounders and affect eQTL reproducibility. Firstly, the liver eQTL studies use a variety of tissue sources from individuals with different ethnic backgrounds (Table 1). Tissue sources include resections of normal tissue from patients suffering from liver cancer (Schroder et al., 2011) or morbid obesity (Greenawalt et al., 2011). Resections of tissue from healthy and deceased liver donors, who may have undergone multiple pharmacotherapeutic treatments immediately prior to death, have also been used (Schadt et al., 2008; Innocenti et al., 2011). Secondly, different genotyping and gene expression platforms, and analysis methods were used to generate eQTLs (Table 1). Furthermore, these studies control for different covariates and have variable powering due to the number of samples, genotypes and transcripts interrogated (Table 1). Despite these issues, there is evidence that a substantial proportion of the cis-eQTL findings are reproducible. In our liver eQTL study, we attempted to replicate eQTL findings in two additional datasets, one of which was a subset of samples from the Schadt et al. study (Innocenti et al., 2011). 67% of the significant cis-eQTLs (Bayes factor (BF)>5) replicated in at least one of the additional datasets (p < 0.05 and concordant effect direction), but this replication rate dropped to 6% for trans-eQTLs (BF>5) (Innocenti et al., 2011). Greenawalt et al. examined the reproducibility of their liver cis-eQTLs (FDR<10%) in the Schadt et al. study, which had used the same microarray gene expression platform, and found a 66% overlap (Greenawalt et al., 2011). Using Bonferroni adjusted p-values, a much more stringent analysis, a comparison between the Schadt et al. and Schroder et al. studies showed that only 16% of cis-eQTLs replicated and no replication of trans-eQTLs (Schroder et al., 2011). Neither of the latter two studies state whether a concordant effect direction was a criterion for replication but it appears that liver cis-eQTLs are comparatively reproducible whilst trans-eQTLs are not. This observation may be partly explained by statistical power. Compared with cis-eQTLs, trans-eQTLs have a greater multiple test correction burden and, thus, a trans-eQTL is less likely to achieve statistical significance in two independent studies.

A salient finding in our study was that the eSNP with the highest BF was frequently located immediately upstream or downstream of the transcription start site (TSS) of the eQTL gene. Our group discovered that 75% of cis-eQTLs within 100 kb of the TSS were found to replicate, compared to 61% of those located more than 100 kb from the TSS (Innocenti et al., 2011). Furthermore, Schadt et al. noted in their study that 70% of the cis-eQTLs were found within 100 kb of the TSS (Schadt et al., 2008). Therefore, in general, it appears that reproducible liver cis-eQTL SNPs will be found close to gene TSSs. This suggests that these eSNPs may directly impact gene transcription by altering gene regulatory regions.

To examine tissue specificity of liver eQTLs, Schadt et al. compared their findings with eQTL analyses of human blood and adipose tissue (Emilsson et al., 2008). Approximately 30% of the cis-eQTLs identified in these tissues were also detected among the liver cis-eQTLs (FDR<10%) from their study (Schadt et al., 2008). Greenawalt et al., examined replication of their liver cis-eQTLs in adipose tissue from the same cohort of individuals used for liver eQTL analysis and found 46–48% of the liver cis-eQTLs replicated at p < 0.05 (Greenawalt et al., 2011). These observations suggest that some fraction of the genetic control of gene expression is specific to the liver. Indeed, Greenawalt et al. found that liver eQTLs, but not eQTLs identified from adipose tissue, were enriched for the KEGG pathway drug metabolism-cytochrome P450 genes. Furthermore, of the 11 ADME genes which had reproducible eQTLs between the Schroder et al. and Schadt et al. studies (Schadt et al., 2008; Schroder et al., 2011), CYP3A5, DHRS2 and SLC22A3 have been predicted to have liver-specific expression (Yu et al., 2006).

Gene transcript levels are a useful intermediate phenotype to study but measurements of protein levels or enzyme activity may provide a more informative phenotype for pharmacogenomics. One of the advantages of the liver eQTL model is that subcellular fractions can be generated from the liver, allowing proteomic and biochemistry studies to be integrated with genomic and transcriptomic data from the same tissue. For example, Yang et al. examined the enzyme activity of eight P450 proteins using liver microsomal samples derived from the same cohort as Schadt et al. (2008), and found that the activities of the CYP450s were positively correlated with mRNA transcript levels after adjusting transcript and activity levels for age, sex and study site (Yang et al., 2010). Transcript levels explained 10–30% of the variance in enzyme activity, thus, indicating that other biological factors, such as post-translational mechanisms, may also impact on activity levels. Fifty four enzyme activity QTLs (aQTLs) were identified for the eight P450 proteins: 24 SNPs were trans-aQTLs and 30 cis-aQTLs (FDR<10%) (Yang et al., 2010). The cis-aQTL SNPs were all associated with CYP2D6 enzyme activity and 29 of these were found to be also cis-eSNPs of CYP2D6 (FDR<10%).

The Gene Expression Omnibus (GEO) repository contains data from all the human liver eQTL studies (Table 1), although genotypes from the Schadt et al. and Greenawalt et al. studies are not available. The data cannot be easily interrogated and, thus, we plan to upload the liver eQTL findings of our study to the web-based SNP and CNV Association (SCAN) database (Gamazon et al., 2010). The results of the Schadt et al. study are already available through the Genotype-Tissue Expression (GTEx) eQTL browser (http://www.ncbi.nlm.nih.gov/gtex/GTEX2/gtex.cgi), and can be queried by gene or SNP and filtered by p-value or r². The GTEx browser provides further useful information by annotating corresponding eSNPs with significant findings from GWAS.

Single nucleotide base changes are the most common but not the only germline genetic variation present in the human genome. A study of genetic variation in eight individuals identified four million SNPs, ~800,000 insertion/deletions (>1 – 100 bp) and 1695 structural (>6 kb) variants, also known as copy number variants (CNVs) (Kidd et al., 2008). Although CNVs are much fewer in number than SNPs, their potential to impact on gene expression is greater due to their size and ability to increase gene copy number or ablate the transcriptional unit. Indeed, SNPs which are in strong LD (r² = 0.8) with CNVs are more likely to be LCL eQTLs and associated with the expression of more than one gene (Gamazon et al., 2011). Moreover, the importance of CNVs in pharmacogenetics is already well known. CNVs with functional effects on enzyme activity of proteins (e.g., CYP2D6, GSTM1 and GSTT1) have been observed to associate with drug efficacy and toxicity (He et al., 2011). It is likely that these CNVs may be eQTLs and, thus, there is a great need for liver CNV eQTL analyses to be carried out in the future.

Previous liver eQTL studies have interrogated genetic variation and gene expression at the genome level. However, due to the limitations of the oligonucleotide genotyping and gene expression microarray platforms, this corresponds only to a fraction of human genetic variation and does not encompass all transcript splice variants or gene isoforms.

Genotype imputation can be performed using genotype data from studies like HapMap or 1000 Genomes Project to greatly increase the number of genotypes interrogated through oligonucleotide genotyping platforms. Indeed, in our liver eQTL study we imputed genotypes from HapMap to generate more than two million genotypes (Innocenti et al., 2011). While still relatively expensive, next-generation sequencing has advanced to the point where it is becoming feasible to perform whole genome sequencing of DNA and mRNA to comprehensively characterize genetic variation and the entire transcriptome in a relatively large number of samples.

Although there is high correlation between transcript levels quantitated by RNA-seq and microarrays (Pickrell et al., 2010), RNA-seq provides a greater resolution which leads to detection of more eQTLs (Montgomery et al., 2010). The RNA-seq (next generation sequencing of mRNA) approach has other advantages over oligonucleotide expression arrays e.g., no saturation of intensity signal for high abundance transcripts and the ability to quantify allele-specific expression (Majewski and Pastinen, 2011). Additionally, RNA-seq can identify splicing QTLs which may not affect overall gene expression but alter transcript isoform levels (Pickrell et al., 2010).

eQTL analyses are informative but they cannot definitively identify the causative SNPs that affect gene expression. Due to LD, many eSNPs can represent the same causative genetic variant. Various approaches can be used to identify functional genetic variants which generate the eQTL signal. Reporter gene assays are useful in determining whether eSNPs in promoter, 5′UTR, 3′UTR and intronic regions have functional effects on gene expression (Innocenti et al., 2011). However, this technique cannot be easily applied in a high-throughput fashion and the effects of SNPs in coding regions are generally not amenable to testing.

Bioinformatic analyses may help to narrow the number of putatively causative SNPs, by identifying eSNPs which are located in functional genomic elements such as those defined by the Encyclopedia of DNA Elements (ENCODE) Project (Myers et al., 2011). Indeed, a study of LCL eQTLs found that 40% of the eSNPs were located in open chromatin regions and sites of DNaseI hypersensitivity or histone modification. Furthermore, SNPs in DNaseI hypersensitivity sites were four-fold more likely to associate with transcript levels compared with other SNPs (Gaffney et al., 2012). This epigenetic, and other regulatory, information can be combined with eQTL data to create a model which ranks eSNPs on their potential to affect gene expression (Gaffney et al., 2012). This approach has not yet been applied to a study of liver eQTLs but could prove extremely useful in identifying pharmacogenetic variants.

eQTL analyses generally examine only the effects of a single genetic locus on gene expression. However, it is possible that interactions of cis- and trans-eQTLs contribute to the regulation of gene expression. Indeed, 15% of transcripts were found to be regulated by a cis-trans eQTL interaction in a study of HapMap LCLs and the vast majority of the eSNPs were not significantly associated with gene expression in a single locus analysis (Becker et al., 2012). Therefore, important information may be lost by not considering these eQTL interactions and this approach could be used to detect epistatic pharmacogenetic variants which may not otherwise be identified.

The Yang et al. study (Yang et al., 2010) correlating liver eQTLs and P450 enzyme activity highlights the need for liver protein QTL analyses to characterize the genetic regulation of protein expression. High-throughput proteome-wide quantification of protein expression is much more problematic and less straightforward than that for transcriptome quantification. Liquid chromatography-tandem mass spectrometry (LC-MS/MS) techniques are providing a way forward but, currently, the best approach for sensitive and accurate protein quantitation may be to carry out targeted proteomic studies (Kline et al., 2009). For example, targeted LC-MS/MS analysis of several UGT1A isoforms has been successfully performed in human liver microsomes (Harbourt et al., 2012).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.