Use of Cell Lines in the Investigation of Pharmacogenetic Loci

Abstract

Drug response and toxicity, complex traits that are often highly varied among individuals, likely involve multiple genetic and non-genetic factors. Pharmacogenomic research aims to individualize therapy in an effort to maximize efficacy and minimize toxicity for each patient. Cell lines can be used as a model system for cellular pharmacologic effects, which include, but are not limited to, drug-induced cytotoxicity or apoptosis, biochemical effects and enzymatic reactions. Because severe toxicities may be associated with drugs such as chemotherapeutics, cell lines derived from healthy individuals or patients provide a convenient model to study how human genetic variation alters response to these drugs that would be unsafe or unethical to administer to human volunteers. In addition to the traditional candidate gene approaches that focus on well-understood candidate genes and pathways, the availability of extensive genotypic and phenotypic data on some cell line models has begun to allow genome-wide association (GWA) studies to simultaneously test the entire human genome for associations with drug response and toxicity. Though with some important limitations, the use of these cell lines in pharmacogenomic discovery demonstrates the promise of constructing a more comprehensive model that may ultimately integrate both genetic and non-genetic factors to predict individual response and toxicity to anticancer drugs.

INTRODUCTION

Personalized medicine has the ambitious goal of both maximizing effective therapy and avoiding adverse effects by tailoring medical care based on one’s genetic profile. Drug response is often not consistent across patients and ranges from beneficial effects to lack of efficacy to the extreme of fatal adverse drug reactions (ADRs). ADRs are responsible for 5–7% of annual hospital admissions in the Western countries [1] and estimated to be between the 4^th and 6^th leading cause of death in the United States [2]. Of particular concern are chemotherapeutic drugs because of their narrow therapeutic index. However, the genetic variants that influence clinical responses to these anticancer drugs remain largely unknown. Hence, identifying patients at risk for severe ADRs to chemotherapeutic agents prior to treatment is an important area of research and will improve current oncology practice. We will focus this review on the discovery of pharmacogenetic loci that may predict toxicities to drugs using EBV (Epstein-Barr Virus) immortalized lymphoblastoid cell lines (LCLs), with particular emphasis on the use of the International HapMap Project [3, 4] samples.

1. Phamacogenetic and pharmacogenomics

Traditionally, pharmacogenetic studies have focused on a handful of well-recognized pharmacology-related genes (e.g. drug metabolizing enzymes, transporters) to search for genetic variants such as single nucleotide polymorphisms (SNPs) that determine the variability in drug response and toxicity. For example, UGT1A1*28 (UDP glucuronosyltransferase 1 family, polypeptide A cluster) polymorphism was identified as the major predictive pharmacogenetic marker for severe hematological toxicity (neutropenia) to irinotecan [5], a commonly prescribed anticancer agent both as a single agent or in combination therapy. More recently, with the availability of the whole-genome DNA sequences [6, 7] and the data from several large scale human genetic variation projects (e.g. the International HapMap Project [3, 4]), the field of pharmacogenomics now aims to examine variation in the entire human genome to identify variants that control susceptibility to therapy-induced toxicity [8, 9] as well as to provide drugs to patients who may respond well.

2. Human genetic variation

Human genetic variation is reflected in the observed differences in DNA sequences between the genomes of individuals or different human populations. The most common form of genetic variation is SNPs, which have been implicated in complex traits since the very beginning of the genomic era, marked by the launch of the Human Genome Project [6, 7]. To date, common SNPs have been found to be associated with various complex phenotypes such as the risks of common diseases (e.g. breast cancer [10, 11], diabetes [12, 13], obesity [12, 14], coronary artery disease [15] and Alzheimer’s disease [16]), adult height [17] and hair color [18]. It is estimated that the average proportion of nucleotide differences between a randomly chosen pair of humans lie between 1 in 1,000 and 1 in 1,500 base pairs [19, 20]. Variation can be measured at both the individual level and at the population level. Of the ~0.1% of DNA that varies among individuals, approximately 85–90% of genetic variation is found within three continental groups (Asians, Europeans and Africans), and only an additional 10–15% of variation is found between any two populations [20–22].

In addition to SNPs, there are other forms of genetic variations in the human genome. For example, copy number variants (CNVs), which can be deletions or amplifications of a genomic region, also account for a substantial fraction of natural genetic variation [23, 24]. Since many CNVs result in differential levels of gene expression, these structural variants may contribute to a significant proportion of phenotypic variation [25]. Like SNPs, CNVs in the human genome have also been associated with the susceptibility to various human diseases [26]. Because of the complexity of the human genome, to identify the genomic contributors to the common disorders and complex traits, it will be necessary to systematically explore both SNPs and other genetic variations (e.g. CNVs) in association studies.

3. Drug response and toxicity is a heritable, complex trait

Individual response to therapeutic treatment has been shown to be a heritable trait that may be partially attributed to genetic diversity. For example, it was estimated that the heritability for susceptibility to cisplatin-induced cytotoxicity was 0.47, indicating sensitivity to the cytotoxic effects of cisplatin is under appreciable genetic influence [27]. Furthermore, there is evidence that susceptibility to cisplatin-induced cytotoxicity is likely due to multiple loci, with low locus-specific heritability contributing to the trait. At the same time, the variability of individual drug response has been observed by clinical practitioners among different human populations, age groups and genders [28–30]. Therefore, drug response and toxicity are likely complex traits, which may be determined by multiple genes and possibly other non-genetic (e.g. environment, age, diet) factors.

ADVANTAGES OF CELL LINE-BASED MODELS

1. Challenges of studying toxic drugs

The majority of drugs currently on the market to treat cancer cause adverse side effects, precluding the use of normal, healthy volunteers for genetic studies. For example, cisplatin, a platinating agent commonly used to treat many human cancers, has been associated with platelet activation in vitro [31], and there have been reports of serious vascular toxicity [32] as well as nephrotoxicity, neurotoxicity and ototoxicity [33, 34] associatedwith its use in the clinic. Furthermore, many anticancer drugs present a narrow therapeutic index, indicating that small changes in dosage could cause unacceptable toxic response which in extreme cases could be fatal. Moreover, toxicity is observed earlier than the therapeutic effect, therefore, toxic effects represent a major endpoint for pharmacodynamic studies of anticancer agents [35]. Clinically, patients taking these drugs require constant monitoring so that the dosage can be adjusted to assure uniform and safe results.

2. Advantages of cell line-based models

Use of cell line-based models controls for certain non-genetic factors such as administration time and diet, because cells can be grown and treated with drugs under the same condition. Different dosages of a drug (even at a concentration that may cause potential cell death) can be administered to cell lines to observe their response. Compared to clinical trials on human subjects, it is also much easier to repeat experiments using cell lines. In addition, extensive genotypic (e.g. SNP genotypes) data on some commercially available cell lines such as the International HapMap LCL samples have been made public. Particularly, the commercial availability of the HapMap cell lines through the Coriell Institute for Medical Research (Camden, NJ) (http://www.coriell.org) has allowed researchers in both human genetics and pharmacogenomics to accumulate various phenotypic data including gene expression and drug response [36] as well as other useful annotations (e.g. positions under recent positive selection [37]) on the same cell lines.

LYMPHOBLASTOID CELL LINE RESOURCES

1. Available genetic variation datasets for LCLs

SNP genotypes for ~1.8 million SNP markers (May, 2003) for LCLs derived from 48 CEPH (Centre d’Etude du Polymorphisme Humain) families [38] can be downloaded from the SNP Consortium database of the SNP Consortium Linkage Map Project (http://snpdata.cshl.edu) [39]. Most SNPs in this database are clustered in very closely linked sets (two or three SNPs within 100 kb), with an average intercluster distance of approximately 3 Mb [40]. Samples from some of these CEPH families (e.g. CEPH 1362, 1408) were included in the International HapMap Project [3, 4]. Therefore, more extensive genetic variation data are available for those LCLs.

The International HapMap Project [3, 4] aims to develop a human haplotype map that catalogues the common patterns of human DNA sequence variation. The HapMap Phase 1/2 Project released a human haplotype map of over 3.1 million SNPs [41] on a panel of 270 EBV-transformed LCLs derived from apparently-healthy individuals of three major continental populations (CEU: 30 CEPH parents-offspring trios from Utah, USA; YRI: 30 Yoruba parents-offspring trios from Ibadan, Nigeria; 45 unrelated Han Chinese from Beijing, China and 45 unrelated Japanese from Tokyo, Japan). In addition to the frequencies of SNP alleles, raw genotypes in each population and other information (e.g. phased haplotypes, linkage disequilibrium information), which can be viewed and downloaded from the HapMap Project website (http://www.hapmap.org) [42]. During the past a few years, researchers working on these samples also generated comprehensive CNV data. The CNV data that involve gains or losses of several kilobases to hundreds of kilobases of genomic DNA among phenotypically normal individuals are publicly available through the Database of Genomic Variants (DGV) (http://projects.tcag.ca/variation/) [43, 44]. The current DGV can be used to browse 1,447 CNV regions (~12% of the human genome) of the 270 HapMap samples screened using the Affymetrix Human Mapping 500K EA Array as well as clone-based comparative genome hybridization with a Whole Genome TilePath (WGTP) array [24]. The sample level data can be accessed from the Copy Number Variation Project of the Sanger Institute (http://www.sanger.ac.uk/humgen/cnv/data/) [24]. Other genetic variation data such as short-tandem-repeat polymorphism (STRP) data [45] have also been generated from these samples and are expected to be publicly available for analyses in the near future. Table 1 lists a summary of some whole-genome resources of genetic variations in LCLs.

Table 1.

Some publicly available whole-genome resources of LCLs.

Date Type		Technology Platform	Samplesa	Website & Reference
Genotype	SNPs	PCR and sequencing by Celera or Motorola	48 CEPH families	http://snpdata.cshl.edu [39]
	SNPs	Affymetrix Human Mapping 100K and 500K Arrays; Perlegen	90 CEU, 90 YRI, 45 CHB and 45 JPT lines	http://www.hapmap.org [41]
	CNVs	Affymetrix Human Mapping 500K EA Array; Whole Genome TilePath Array	90 CEU, 90 YRI, 45 CHB and 45 JPT lines	http://www.sanger.ac.uk/humgen/cnv/ [24, 49, 117] http://projects.tcag.ca/variation/ [43]
mRNA Expression	Gene-level	Affymetrix Human Focus Array	233 CEPH lines	http://www.ncbi.nlm.nih.gov/geo/ [GSE1485] [40]
	Gene-level	Affymetrix Human Focus Array	60 CEU, 41 CHB and 41 JPT lines	http://www.ncbi.nlm.nih.gov/geo/ [GSE5859] [46]
	Gene-level	Affymetrix Human Focus Array	60 CEU, 60 YRI and 82 AA lines	http://www.ncbi.nlm.nih.gov/geo/ [GSE10824] [47]
	Gene-level	Affymetrix Human Focus Array	8 CEU and 8 YRI lines	http://www.ncbi.nlm.nih.gov/geo/ [GSE7236] [48]
	Gene-level	Affymetrix Human Exon 1.0 ST Array	87 CEU and 89 YRI lines	http://www.ncbi.nlm.nih.gov/geo/ [GSE7851] [51]
	Exon-level	Affymetrix Human Exon 1.0 ST Array	87 CEU and 89 YRI lines	http://www.ncbi.nlm.nih.gov/geo/ [GSE9703] [50]
	Gene-level	Illumina Sentrix Human-6 Expression BeadChip Version 1	90 CEU, 90 YRI, 45 CHB and 45 JPT lines	http://www.ncbi.nlm.nih.gov/geo/ [GSE6536] [49]

^aCEPH, Centre d’Etude du Polymorphisme Humain; CEU, CEPH samples from Utah, USA; YRI, Yoruba people form Ibadan, Nigeria; CHB, Han Chinese from Beijing, China; JPT, Japanese from Tokyo, Japan; AA, Coriell Human Variation Panel (African American).

Besides the above whole-genome genetic variation data, some ongoing deep resequencing projects (e.g. the SeattleSNPs Project at http://pga.gs.washington.edu) have released genotypic data of some gene targets on some HapMap samples. These will be introduced more in the section of “Current Developments and Outlook”.

2. Available expression and other phenotype datasets for LCLs

Using the Affymetrix Human Focus Array (focus array), 233 LCLs from 14 CEPH families [38] (CEPH 1333, 1340, 1341, 1345, 1346, 1347, 1362, 1408, 1416, 1418, 1421, 1423, 1424 and 1454) were phenotyped for transcriptional expression [40]. During the past a couple of years, gene expression data on the HapMap samples have been generated using different microarray platforms. Spielman et al. profiled 60 unrelated CEU, 41 CHB and 41 JPT samples using the focus array [46]. The same group recently generated expression data on 60 unrelated YRI, 60 unrelated CEU and 82 non-HapMap Coriell Human Variation Panel AA (African American) samples using the focus array [47]. At a smaller scale (8 CEU and 8 YRI samples), Storey et al., generated expression data using the focus array [48]. In contrast, Stranger et al used the Illumina BeadChips to obtain gene expression data on the entire panel of 270 HapMap samples [49]. Furthermore, our lab used the Affymetrix Human Exon Array (exon array) to obtain both exon-level (probeset) [50] and gene-level (transcript) [51] expression profiles of 87 CEU and 89 YRI samples. Unlike the focus array (a conventional 3′ oligonucleotide array) and the Illumina platform, which can only measure gene-level expression, the exon array has the flexibility to measure the expression of the entire human exonome. Though lacking more comprehensive comparisons, there is evidence that the exon array may provide more accurate measurements of gene expression than the conventional 3′ arrays (e.g. the focus array) [52]. The above whole-genome expression data are available through the NCBI Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/) website (Table 1). In addition, published pharmacology-related phenotypes such as cell growth inhibition, apoptosis after drug treatment and drug metabolizing enzyme activity on both HapMap and some non-HapMap LCLs are available through the PharmGKB (Pharmacogenetics and Pharmacogenomics Knowledge Base) website (http://www.pharmgkb.org) [53]. Together with the genetic variation data, these resources are of tremendous value for human genetic and pharmacogenomic studies [36].

3. Human genetic studies using LCLs

Because gene expression is an intermediate phenotype between DNA sequence variation and more complicated cellular or whole-body phenotypes such as risks of common diseases and drug response, studying the genetics of gene expression regulation can provide insights into the human genome and complex traits/diseases. During the past few years, progress has been made in dissecting the genetic architecture of gene expression using LCLs. Gene expression variation has been observed within a population (e.g. the CEPH cell lines [40, 54–57]) or between continental populations (e.g. the HapMap LCLs [46, 48, 49, 51]). In addition, gene expression differences between males and females were reported using both the CEPH [58] and HapMap [59] cell lines. Furthermore, both cis-acting (local) and trans-acting (distant) eQTLs (expression quantitative trait loci) have been mapped in the human genome [60, 61]. Recently, the genetics of transcript isoform variation (alternative splicing) within the CEU samples [62, 63] and between the CEU and YRI samples [50] have also been interrogated using the exon array. Because different populations and genders may have different risks of common diseases and response to therapeutic treatments (e.g. some anticancer drugs [64]), these LCLs, particularly the HapMap samples, provide a convenient model for studying the health disparities (e.g. drug response variation) between major human populations as well as between males and females. Interestingly, population differences in gene expression were observed in some of the important pharmacogenes using the HapMap LCLs [65], as maintained by the PharmGKB [53] VIP (Very Important Pharmacogenes) database. In addition, we set up a searchable database, SCAN (SNP and CNV Annotation Database, http://www.scandb.org), which can be useful for both gene-centric and SNP-centric queries for eQTLs using the exon array data (gene-level) on the CEU and YRI samples [51, 60]. We expect to update the SCAN database to accommodate the exon-level expression data [50] and CNV data [24] in the future.

USE OF CELL LINES FOR PHARMACOGENOMIC DISCOVERY

LCLs derived from targeted populations have been utilized to evaluate genetic factors responsible for radiation sensitivity [66], sensitivity to mitomycin C in patients with Wilms’ tumor [67], camptothecin-induced apoptosis [68], bleomycin-induced chromatid breaks and oxidant-induced apoptosis [69] as well as response to ionizing radiation [70]. For example, studies using LCLs indicated that cellular radiosensitivity was a general property of human cells and the patterns of cellular radiosensitivity suggested autosomal dominant inheritance [66]. In another study, LCLs were used as a control to illustrate the difference in cell death between Werner Syndrome (WRN) and WRN-deficient individuals after treatment with the DNA-topoisomerase-I-trapping drug camptothecin [68]. Generally, there are two main categories of approaches (Fig 1) to identifying genetic determinants responsible for drug response and toxicity. Each has its own advantages and limitations. Though the full strength of the LCL model could be reflected more in the whole-genome approaches (including GWA and linkage analysis), combining both the candidate and whole-genome approaches at different stages of pharmacogenomic and pharmacogenetic studies may be necessary for discovering and validating truly functional genetic variants. We will discuss these approaches and present some examples using them in this section. Fig. (1) shows an overview of these two general approaches.

Fig. 1.

General approaches to pharmacogenetic and pharmcogenomic studies. (A) Candidate gene approach; (B) Whole-genome approach. Though SNPs are shown, other genetic variation data can also be used in these approaches.

1. Traditional Candidate gene approach

When the function of a gene is well-established, researchers may focus on that gene to search for functional or associated genetic variants for drug response and toxicity. Similarly, if a particular genetic variant (e.g. SNP) has been associated with drug response and toxicity, researchers may focus on that variant to study its relevance in new cohorts (e.g. by resequencing other samples). For example, some LCLs were used in an earlier study to determine the sequence diversity of CYP3A (Cytochrome P450, family 3, subfamily A) enzymes [71], which have been known to act in drug metabolism, influence circulating steroid levels and response to half of all oxidatively metabolized drugs. Another good example for candidate genes in pharmacology is the VIP list assembled by the PGRN (Pharmacogenetics Research Network) [72] members. Of the current list of 39 PGRN-VIP genes (Jan, 2009), most were annotated with evidence of either in vitro or in vivo functional effect on drug response (http://www.pharmgkb.org) [53]. Some successful examples of this approach in chemotherapy include: 1) The identification of genetic polymorphisms in TPMT (thiopurine S-methyltransferase), which lead to decreased TPMT enzyme activity and subsequently increase 6-mercaptopurine toxicity [73]; 2) Decreased activity of UGT1A1*28 polymorphism is associated with increased risk of neutropenia associated irinotecan treatment [5] and; 3) There is an effect of the 2677G>T/A polymorphism of ABCB1 (ATP-binding cassette sub-family B member 1) on progression-free survival inovarian cancer patients who are treated with a taxane/carboplatin, which is dependent on the extent of residual disease, with abetter prognosis for patients with the 2677T/A allele and minimalresidual disease [74]. Furthermore, using TPMT 719>G as an example, Jones et al. showed that HapMap LCL resources were useful tools for pharmacogenetic discovery when the candidate gene is known, by testing whether the TPMT 719A>G SNP could be identified as predicting TPMT phenotype [75]. The HapMap CEU and YRI samples were also successfully used to determine the functional consequences of genetic variants in DCK (deoxycytidine kinase), which is a rate-limiting enzyme in the activation of nucleoside analogs such as Ara-C (cytarabine arabinoside), gemcitabine, clofarabine, and others [76].

Because the candidate gene approach requires the information from well-known physiological, biochemical or functional knowledge of a gene, which is generally finite or sometimes not available at all, it is not suitable for the discovery of new candidate genes or variants. This limitation is especially important for pharmacogenomic studies. Because drug response and toxicity is most likely a complex trait, focusing on a single gene (or a few genes) may not allow one to see the complete picture of the genes involved in the drug response process. Therefore, a broader and unbiased approach may be needed for identifying novel pharmcogenetic loci.

2. Whole-genome approaches

Whole-genome approaches, including the GWA and linkage analysis, do not require a priori knowledge of a gene, which is necessary for the candidate gene approach. As mentioned in the previous section, the currently available LCL resources include genome-wide genetic variation data as well as whole-genome gene expression data. During the past several years, genetic variants and genomic regions associated with toxicity to some anticancer drugs have begun to be identified using these whole-genome approaches.

Because the CEPH LCLs include family members from three generations, early studies were able to estimate the heritability of drug response and apply linkage analysis to identify genomic regions linked to drug response. For example, the LCLs derived from 10 CEPH pedigrees were characterized for the degree of cisplatin sensitivity [27]. The authors estimated that the heritability of susceptibility to cisplatin-induced cytotoxicity to be 0.47, indicating an appreciable genetic influence. Watters et al also used the CEPH LCLs to discover genetic determinants of 5-fluorouracil and docetaxel-induced cytotoxicity [77]. Cytotoxicity as a result of exposure to these chemotherapeutic agents was shown to be heritable. Genome-wide linkage analysis was also used to map a QTL (quantitative trait locus) influencing the cellular effects of 5-fluorouracil to chromosome 9q13–q22, and two QTLs influencing the cellular effects of docetaxel to chromosomes 5q11–21 and 9q13–q22. More recently, more than 300 CEPH LCLs (24 pedigrees) were evaluated for toxicity to daunorubicin [78] and etoposide [79]. Similarly, heritability analysis showed a significant genetic component contributing to the cytotoxic phenotypes and some linked genomic regions (e.g. 4q28.2 to 4q32.3 for daunorubicin [78]) were identified using a whole-genome linkage scan.

Recently, with the availability of a dense map of human genetic variation, the HapMap LCLs have been begun to be used in pharmacogenomic studies to identify genetic loci responsible for drug response and toxicity (Table 2). The exon array gene expression data (~9,000 expressed genes) [51] generated from the 176 CEU and YRI HapMap LCLs were used to identify genetic variants, acting through their effect on regulating expression, associated with the cytotoxicities to some anticancer drugs, including etoposide [80], daunorubicin [81], carboplatin [82], cisplatin [83] and Ara-C [84]. A genetics-driven “triangle” approach [80] was used to integrate both gene expression and genotypic data in predicting chemotherapeutic-induced cytotoxicity. The first step of the “triangle” approach aims to identify significant SNPs (based on the HapMap SNP data) whose genotypes are associated with drug response (e.g. IC₅₀, the concentration at which 50% cell growth inhibition occurs). The second step then aims to find significant gene expression phenotypes associated with the identified SNPs from the first step (i.e. eQTLs). Finally, the third step determines whether the gene expression phenotypes from the second step are correlated with drug response. In addition to the HapMap LCLs, non-HapMap LCLs were used as a replication step to increase confidence in the genetic variants identified in the discovery phase [81, 84]. These results demonstrated the usefulness of the HapMap LCLs as a model for pharmacogenomic discovery. The identified genetic variants contribute to a substantial proportion of the overall variation in cell sensitivity (e.g. cisplatin [83]). Furthermore, this general whole-genome approach using the HapMap samples can be used to elucidate the eQTLs contributing to a wide range of cellular phenotypes (e.g. apoptosis).

Table 2.

Whole-genome pharmacogenomic studies using the HapMap LCLs.

Druga	Mechanism of Action	Samplesb	Validation Set	Reference
Etoposide	Topoisomerase II inhibitor	87 CEU and 89 YRI lines		(Huang et al.) 2007 [80]
Daunorubicin	Topoisomerase II inhibitor	87 CEU; and 89 YRI lines	50 non-HapMap CEPH	(Huang et al. 2008) [81]
Carboplatin	Alkylating agent	89 YRI lines		(Huang et al. 2008) [82]
Cisplatin	Alkylating agent	87 CEU and 89 YRI lines		(Huang et al. 2007) [83]
Ara-C	RNA/DNA antimetabolite	87 CEU and 89 YRI lines	49 non-HapMap CEPH	(Hartford et al. 2008) [84]
Bortezomib	Proteasome inhibitor	Experiment failed		(Choy et al. 2008) [91]
Rapamycin	mTOR inhibitor	Experiment failed		(Choy et al. 2008) [91]
5-Fluorouracil	RNA/DNA antimetabolite	90 ASN, 84 CEU and 85 YRI lines		(Choy et al. 2008) [91]
Methotrexate	RNA/DNA antimetabolite	90 ASN, 81 CEU and 85 YRI lines		(Choy et al. 2008) [91]
6-Mercaptopurine	DNA antimetabolite	90 ASN, 79 CEU and 85 YRI lines		(Choy et al. 2008) [91]
SAHA	Histone deacetylase inhbitor	90 ASN, 85 CEU and 85 YRI lines		(Choy et al. 2008) [91]
Simvastatin	Lipid-lowering agent	90 ASN, 87 CEU and 85 YRI lines		(Choy et al. 2008) [91]

^aArac-C, Cytarabine arabinoside; SAHA, Suberoylanilide hydroxamic acid.

^bCEU, CEPH samples from Utah, USA; YRI, Yoruba people form Ibadan, Nigeria; CHB, Han Chinese from Beijing, China; JPT, Japanese from Tokyo, Japan; ASN, Asian HapMap cell lines including the CHB and JPT samples.

Furthermore, because the HapMap LCLs include samples derived from individuals of different ancestries, these cell lines can be used to study differences in drug response between human populations. These samples may also be used to compare the identified pharmacogenetic loci between populations to evaluate the contribution of genetic variants to drug response in a particular population. For example, using the HapMap LCLs, significant differences in carboplatin and daunorubicin IC₅₀s were shown when comparing the YRI and CEU samples [64]. However, this population difference in treatment-induced cytotoxicity was not seen for either cisplatin or etoposide. By comparing pharmacogenetic loci between populations, some interesting patterns were observed. For example, in a study of the pharmacogenomics of Ara-C using the HapMap YRI and CEU LCLs, population-specific signatures were secondary to either: 1) Polymorphic SNPs in one population but monomorphic in the other; 2) Significant associations of SNPs with cytotoxicity or gene expression in one population but not the other [84].

3. The use of cell lines to further study clinical findings

LCLs are also a convenient model to further study clinical findings. For example, using the HapMap LCLs, French et al. demonstrated that acquired genetic variation in leukemia cells has a stronger influence on MTXPG (methotrexate polyglutamates) accumulation than inherited genetic variation by comparing the expression and inherited SNPs between leukemia cells and the HapMap LCLs [85]. In another study, using 90 CEPH cell lines, Pottier et al. identified a SNP in the promoter of SMARCB1(SWI/SNF related, matrix associated, actin dependent regulator of chromatin, subfamily b, member 1), a candidate gene from a previous study in ALL (acute lymphoblastic leukemia), which alters SMARCB1 expression [86]. Hartford et al. validated that adhesion biological pathway is implicated in secondary leukemia using a genome scan for etoposide-induced leukemogenic MLL [myeloid/lymphoid or mixed-lineage leukemia (trithorax homolog, Drosophila)] chimeric fusions in 15 HapMap cell lines for a previous study using patient samples [87]. More recently, in order to further address the hypothesis that cytotoxic effects of etoposide are not necessarily linked to the drug-induced leukemogenic rMLL (MLL rearrangement), Yang et al. screened 144 CEPH cell lines to identify lines with differing intrinsic sensitivity to etoposide [88]. The HapMap LCLs were also used as a validation set when exploring the genetic contribution to gene expression concordant across primary leukemia cells and normal leukocytes [89]. In an attempt to investigate the relationship between CYP3A4 (cytochrome P450, family 3, subfamily A, polypeptide 4) copy number and expression, Lamba et al. used LCLs representing eight ethnically diverse populations to compare with other samples (e.g. normal livers, primary and secondary liver tumors) [90].

LIMITATIONS AND POTENTIAL CONFOUNDING FACTORS

LCLs, particularly the HapMap samples, provide a useful model for pharmacogenomic discovery of cellular phenotypes induced by drugs; however, there are some important limitations and challenges that need to be considered when analyzing and interpreting the results using these samples. While some are “intrinsic,” some of these limitations could be addressed by the advancement of technologies and development of ongoing research efforts. A recent study using the HapMap samples for the genetic analysis of drug response suggested that biological noise and other confounding factors may reduce power and have the potential to create spurious association [91]. Because these cell lines have just recently been used in pharmacogenomic studies, debates exist for which drugs are suitable, how to measure drug response as well as how to account for confounding factors and interpret results.

1. Tissue type limitation

One assumption of the current cell line-based model is that LCLs can be used to represent tissues of toxicity. Genetically, LCLs have the same genetic make-up of autosomes as other tissue types in a human body. Therefore, LCLs are suitable for associations using genetic variation data. However, because of the tissue specificity of human gene expression [92], a large proportion of known genes are not reliably expressed in LCLs, including some well-known drug metabolizing enzymes and drug targets such as EGFR (epidermal growth factor receptor) [51]. Previous studies have estimated that LCLs reliably express ~50% of all known genes [40, 51]. In addition, the current gene expression data are at mRNA-level; proteomics data will be necessary to evaluate the relationship between mRNAs and protein products in LCLs. Hence, care should be taken when making conclusions on associations involving gene expression data.

2. Small sample size and rarer genetic variants

The current pharmacogenomic discoveries made using LCLs are generally hypothesis-generating, partly because of the limited samples available for these studies. For example, there are only 60 unrelated HapMap CEU samples, which are used to represent Caucasians of northern and western European ancestry, while there are only 45 unrelated Han Chinese and 45 unrelated Japanese to represent eastern Asians. Because there are more than 10 million SNPs in the human genome [93], the relative small sample size is especially challenging when the differences in complex traits are being compared between human populations. With small sample sizes, significant difference are more likely to be spurious than with large samples. Though the Phase 2 HapMap SNP data is believed to cover approximately 25–35% of common SNP variation in the populations surveyed [41], rare genetic variants with potential biological functions may not be captured and analyzed in these samples.

3. Technical limitations

The high throughput profiling of gene expression and genotyping relies on the accuracy and efficiency of microarray platforms. Different platforms may have specific technical designs and limitations. For example, the oligonucleotide probes of the Affymetrix exon array differ from those standard 3′ expression arrays (e.g. the Affymetrix focus array, U133 and U95 arrays as well as the Illumina platform) in their design, density, and coverage [46, 51, 94, 95]. In contrast to the 3′ arrays, the exon array interrogates the entire mRNA transcript and uses DNA targets. Another potential concern with the use of oligonucleotide arrays is the possibility that SNPs located within probes could affect hybridization efficiency [96] and lead to false eQTLs [97]. Earlier studies using these LCLs did not consider this potential problem, while a couple of recent studies proposed slightly different approaches based on the dbSNP database [98] to account for this factor [99, 100]. Though some results showed that the overall reproducibility of the Affymetrix oligonucleotide arrays and the Illumina BeadChip is good [101, 102], more comparisons, both statistically and experimentally, will be necessary for a comprehensive evaluation of the reliability of these platforms.

4. Other potential confounding factors

The CEPH LCLs [38] (including the HapMap CEU samples) were collected in the 1980’s, decades earlier than other HapMap collections [3, 4]. Cell line transformation, culture techniques and protocols have evolved during the past twenty years. In addition, the issue of being in culture and/or frozen may affect some biological parameters. Therefore, some caution is warranted when analyzing and interpreting the data from these Caucasian samples. For example, Akey et al. suggested that batch effects could be a confounding factor when interpreting results from a study of population differences in gene expression between the CEU and ASN (Asian HapMap cell lines including the CHB and JPT) samples [46, 103].

The LCLs were derived from B-lymphocytes which were transformed with EBV, a member of the herpesvirus persisting in most adults [104]. When EBV infects B-lymphocytes in vitro, LCLs eventually emerge that can proliferate indefinitely. The growth transformation of these cell lines is the consequence of viral protein expression [105]. EBV viral proteins have been found to induce differential gene expression in host cells. For example, LMP1 (latent membrane protein 1) was found to induce differential expression of genes that are characteristically expressed in inflammatory and hyperproliferative epidermis, including chemokines, cytokines and their receptors in cultured keratinocytes [106]. EBNA2 (Epstein-Barr virus nuclear antigen 2), a transactivator, was found to control expression of several viral and cellular genes [107]. Therefore, it is likely that EBV transformation may also affect the gene expression patterns of LCLs. However, the effects and scope of EBV transformation and its affect on expression have not been comprehensively evaluated in these cell lines. A systematic investigation on the effects of baseline EBV copy number and EBV lytic replication following treatment with drugs under study would help ascertain whether EBV influences cellular sensitivity to drug.

Another potential confounding factor is variation in growth rate of LCLs over time as well as differences between individuals. Most chemotherapeutic agents work by impairing mitosis, and therefore their efficacy depends on the rate of cell division. For example, increased proliferation was found to be associated with increased in vitro sensitivity to several anticancer agents in ALL [108]. Therefore, cell lines growing faster are likely to be more sensitive to chemotherapy. Though results from one recent study showed limited evidence for the heritability of growth rate [91], more comprehensive evaluation (e.g.—a GWA study between growth rate and SNPs or an evaluation of growth rate in large pedigrees) may be necessary to illustrate the genetic contribution to this factor. If growth rate is genetic and associated with specific polymorphisms, then these polymorphisms may also be associated with drug sensitivity through this intermediate phenotype. Whether growth rate is genetic or non-genetic, considering this factor would be especially useful when interpreting the pharmacogenomic discoveries of anticancer drugs that target cell division.

CURRENT DEVELOPMENTS AND OUTLOOK

Current cell-based studies for pharmacogenomic discovery have largely been utilizing the HapMap Phase 1/2 LCLs and focusing on a particular type of genetic variation, i.e. SNPs as well as gene expression (mRNA-level whole-transcript expression). As discussed in the previous section, these HapMap LCLs are a model, and therefore have some important limitations. In this section, we will highlight the current developments of some ongoing research efforts and high-throughput technologies that will have an impact on the next wave of pharmacogenomic studies using cell lines and facilitate the realization of personalized medicine.

1. Next-generation sequencing platforms

Recently, several new sequencing instruments, referred to as “next-generation” or “massively parallel” sequencing platforms, are becoming available to researchers. In contrast to the conventional capillary-based sequencing, the biggest advantage of these next-generation sequencers is their ability to process millions of sequencing reads in parallel rather than 96 at a time. Currently, there are three commercially available and relatively mature technologies, including the Roche (454) GS-FLX sequencer (454 Life Sciences, Branford, CT), the Illumina Genome Analyzer (Illumina, San Diego, CA) and the Applied Biosystems SOLiD (Sequencing by Oligo Ligation and Detection) (Applied Biosystems, Foster City, CA). These three platforms differ significantly in terms of cost, sequencing chemistry, amplification approach and performance (e.g. length of reads). Though the accuracy of their sequencing reads and associated quality are not comprehensively evaluated, these technologies are already transforming the field of genomics research [109, 110], as evidenced in the achievements of several ongoing deep resequencing projects.

2. Deep resequencing projects

Though extensive, the International HapMap Project data aimed to cover only common genetic variants (MAF>0.05) [3, 4]. The recently released Phase 2 data of the HapMap Project is comprised of more than 3.1 million SNPs including approximately 25–35% of common SNP variation in the populations surveyed [41]. On the other hand, potential pharmacogenetic loci associated with drug response and toxicity with smaller allele frequencies could be missed from the current dataset. The several large-scale deep resequencing projects covering the HapMap samples such as the ENCODE (Encyclopedia of DNA Elements) Project [111], the SeattleSNPs Project (http://pga.gs.washington.edu/) and the NIEHS (National Institute of Environmental Health Sciences) Environmental Genome Project (http://egp.gs.washington.edu/) [112] could be used to identify those relatively rare genetic variants in certain target genes. For example, the ENCODE Project launched by the NHGRI (National Human Genome Research Institute) aims to provide a more biologically informative representation of the human genome by using high-throughput methods to identify and catalogue the functional elements it encodes [111]. Currently, the HapMap-ENCODE Coordination (http://www.hapmap.org/downloads/encode1.html.en) aims to produce a dense set of genotypes across large genomic regions in 48 unrelated HapMap LCLs (16 YRI, 8 JPT, 8 CHB, and 16 CEU samples). Approximately 20,000 SNPs were identified in the 10 500kb HapMap-ENCODE regions. Some of these were already represented in the dbSNP [98, 113] (a public SNP database maintained by the NCBI) database, while others were discovered during the resequencing. Another interesting project is the 1000 Genome Project (http://www.1000genomes.org), which is an international research effort aimed at establishing a detailed catalogue of human genetic variation using the next-generation sequencing technologies [114]. The specified goal of this project is to identify > 95% of the variants with a MAF > 1% in parts of the human genome that can be sequenced, with > 95% certainty as well as to identify > 95% of the variants with a MAF > 0.1–0.5% in exons, with > 95% certainty [115]. The sequence data generated will provide a comprehensive reference for human genetic variation. Under the first pilot project, researchers are sequencing 60 HapMap samples from three different populations at low coverage. The second pilot involves high-coverage sequencing of two trios of the CEU and YRI samples. The third pilot project aims to sequence 1,000 genes in 1,000 individuals at high coverage [115]. Prospectively, once integrated with other HapMap resources (e.g. gene expression), these efforts will greatly benefit the next wave of association studies and data mining using these cell lines [116].

3. Integrating more “-omics” data

In addition to SNPs, other genetic or non-genetic elements have roles in regulating gene expression, which in turn may influence more complex cellular or whole-body phenotypes (e.g. drug response and toxicity). For example, CNVs contribute to a substantial fraction of gene expression variation [117] and have been implicated in drug response [118, 119]. Epigenetic changes such as DNA methylation at the CpG sites of the promoter regions can also affect gene expression [120]. Recently, the role of DNA methylation in regulating gene expression in drug response have been investigated [121–123]. For instance, the efficiency of the DNA repair enzyme O⁶-methylguanine DNA methyltransferase (MGMT) has been associated with an increased susceptibility to alkylating agent toxicity [124–126]. MGMT is transcriptionally silenced by promoter hypermethylation in several human cancers such as diffuse large B-cell lymphoma [127] as well as head and neck squamous cell carcinoma [128].

Besides gene-level or transcript-level expression, the differences in alternative splicing, i.e. transcript isoform variation, reflect another layer of complexity of gene regulation. The disruption of specific alternative splicing events has been implicated in several human genetic diseases including cancer [129–131]. Along with new technologies providing more and more “-omics” data, improved bioinformatics tools, statistical methods and systems biology approaches need to be developed so that researchers can comprehensively incorporate the complex networks of various relationships [132] (e.g. gene-environment, gene-drug, gene-gender, gene-age, cis- or trans- regulators) that affect complex traits such as drug response and toxicity. By integrating all of the relevant “-omics” data such as SNPs, CNVs, gene-level expression, exon-level gene expression (Table 1), DNA methylation changes [133] and other putative elements for gene regulation like STRPs [45], a comprehensive model for predicting drug response and toxicity could be built using these cell lines in the future.

4. More than the HapMap Phase 1/2 LCLs

More cell lines beyond the 270 HapMap Phase 1/2 LCLs will be necessary to evaluate the current findings of the pharmacogenomic discovery in replicate samples and other populations. Therefore, the recently released HapMap Phase 3 sample collection may present more opportunities for applying the same cell-based approach to identify pharmacogenetic loci. The HapMap Phase 3 cell line collection is comprised of 1,301 samples from 11 diverse populations which include the original 270 Phase 1/2 samples and 7 additional populations (ASW: African ancestry in Southwest USA; GIH: Gujarati Indians in Houston, Texas; LWK: Luhya in Webuye, Kenya; MEX: Mexican ancestry in Los Angeles, California; CHD: Chinese in Metropolitan Denver, Colorado; MKK: Maasai in Kinyawa, Kenya; TSI: Toscans in Italy) (http://www.hapmap.org). Extensive SNP genotypic and frequency data (~1 million SNPs) on these samples (draft 2 released on January 7^th, 2009) can be publicly accessed at the International HapMap Project website (http://www.hapmap.org). Except for the CEU samples, the HapMap Phase 3 cell lines are available through the NHGRI Sample Repository for Human Genetic Research at the Coriell Institute.

Many drug metabolism-related genes have distinct tissue-specific expression patterns. For example, tissue-specific expression patterns have been observed for UGT1A isoforms (UDP glucuronosyltransferase 1 family, polypeptide A cluster), which belong to a superfamily of microsomal enzymes responsible for glucuronidation of numerous endogenous and exogenous compounds [134]. Therefore, a more comprehensive understanding of variation of complex traits including drug response and gene expression might require cell lines derived from other tissues. For example, because the human liver is a metabolically active tissue that is important in drug metabolism as well as a number of common diseases (e.g. obesity, and diabetes), the recently published whole genome expression data on >400 normal livers [135] will be of tremendous value for pharmacogenomic discovery if these same liver tissues could be utilized for drug response profiling.

Furthermore, the NCI-60 data on a panel of 60 different human cancer cell lines (9 tumor types) used for the NCI (National Cancer Institute) Tumor Cell Line Drug Screen (http://dtp.nci.nih.gov/) represent another useful resource for cancer pharmacogenomics [136]. The current data release (March, 2007) is comprised of toxicity data on more than 40,000 compounds and whole-genome molecular profile data such as gene expression [137, 138], microRNA abundance [139], SNP genotypes [140] and CNVs [140] (http://dtp.nci.nih.gov/mtargets/download.html). Though with limitations such as the small sample sizes for the available tumor types (ranging from 2 to 9 cell lines), the collection of NCI-60 pharmacologic and molecular profile data has allowed both GWA and candidate gene approaches to identify genetic variants responsible for toxicities to anticancer drugs [141, 142] and could be used to compare with the findings from the cell lines derived from normal individuals.

5. Beyond discovery: validation and translational research

The current findings of pharmacogenetic loci through both the candidate gene and GWA approaches are largely at the discovery stage, i.e. the identification of associated genetic variants that could be either causal or linked to causal genetic elements determining drug response and toxicity. Because these discoveries are usually made using a limited number of samples (e.g. the 270 HapMap LCLs), validation of the identified pharmacogenetic loci on independent replicates or patient samples will be necessary to establish their real biological function. For example, the findings using the CEU samples could be evaluated for association in other independent Caucasian samples. Replication in other independent samples could also help eliminate any false discoveries because of the usual large number of tests being made during GWA studies. To validate their findings for gene expression relationships with toxicity to Ara-C using the HapMap CEU samples, Hartford et al replicated 2 of 3 significant relationships between gene expression (as identified using the exon array on the initial samples and real-time PCR on the second set of samples) and cytototoxicity on an additional set of 49 unrelated CEPH cell lines [84].

Finally, smooth communication and close collaboration between basic science researchers and clinical scientists need to be established to bring the pharmacogenomic discovery using cell lines into personalized medicine that attempts to benefit patients through individualized treatment. The cell lines do not consider all the confounding variables that are present in humans such as concomitant medications, diet, disease, etc. Therefore, the findings in cell lines need to be validated in a clinical setting for one to appreciate the advantages and limitations of the model.

In summary, cell line-based models have begun to allow researchers to identify genes and genetic variants that can be used to predict drug response and toxicity. The advances in sequencing and other “-omics” profiling technologies will facilitate the next wave of pharmacogenomic discovery, which in turn may help the realization of personalized medicine.

ABBREVIATIONS

AA

African American

ABCB1

ATP-binding cassette sub-family B member 1

ADR

Adverse drug reaction

ALL

Acute lymphoblastic leukaemia

Ara-C

Cytarabine arabinoside

ASN

Asian HapMap cell lines including the CHB and JPT samples

ASW

African ancestry in Southwest USA

CEPH

Centre d’Etude du Polymorphisme Humain

CEU

CEPH samples from Utah, USA

CHB

Han Chinese from Beijing, China

CHD

Chinese in Metropolitan Denver, Colorado

CNV

Copy number variant

CYP3A

Cytochrome P450, family 3, subfamily A

CYP3A4

Cytochrome P450, family 3, subfamily A, polypeptide 4

DCK

Deoxycytidine kinase

DGV

Database of Genomic Variants

EBNA2

Epstein-Barr virus nuclear antigen 2

EBV

Epstein-Barr Virus

EGFR

Epidermal growth factor receptor

ENCODE

Encyclopedia of DNA Elements

eQTL

Expression quantitative trait locus

FDA

Food and Drug Administration

GIH

Gujarati Indians in Houston, Texas

GWA

Genome-wide association

IC50

Half maximal inhibitory concentration

JPT

Japanese from Tokyo, Japan

LCL

Lymphoblastoid cell line

LMP1

Latent membrane protein 1

LWK

Luhya in Webuye, Kenya

MAF

Minor allele frequency

MEX

Mexican ancestry in Los Angeles, California

MGMT

O⁶-methylguanine DNA methyltransferase

MKK

Maasai in Kinyawa, Kenya

MLL

Myeloid/lymphoid or mixed-lineage leukemia (trithorax homolog, Drosophila)

MTXPG

Methotrexate polyglutamates)

NCI

National Cancer Institute

NHGRI

National Human Genome Research Institute

NIEHS

National Institute of Environmental Health Sciences

PGRN

Pharmacogenetics Research Network

QTL

Quantitative trait locus

rMLL

MLL rearrangement

SAHA

Suberoylanilide hydroxamic acid

SMARCB1

SWI/SNF related, matrix associated, actin dependent regulator of chromatin, subfamily b, member 1

SNP

Single nucleotide polymorphism

SOLiD

Sequencing by Oligo Ligation and Detection

STRP

Short-tandem-repeat polymorphism

TPMP

Thiopurine S-methyltransferase

TSI

Toscans in Italy

UGT1A

UDP glucuronosyltransferase 1 family, polypeptide A cluster

VIP

Very Important Pharmacogenes

WGTP

Whole Genome TilePath

WRN

Werner Syndrome

YRI

Yoruba people from Ibadan, Nigeria