Introduction

One of the goals of precision medicine is to use pharmacogenomics to optimize treatment efficacy and minimize adverse drug reactions. Barriers to the implementation of pharmacogenomics-guided therapy include the turnaround time for obtaining a pharmacogenetic (PGx) result1 and the clinical utility of returning PGx variants.2 One recommendation for avoiding treatment delays is to implement preemptive PGx testing.1 Current PGx testing uses array-based genotyping platform—e.g., Affymetrix DMET Plus (Drug Metabolizing Enzymes and Transporters array)—screens for a predefined set of PGx variants.3,4 Genomic testing platforms such as exome sequencing (ES) or genome sequencing (GS),5 also called massively parallel sequencing (MPS), have potentially wider utility than the aforementioned genotyping platforms, and this begs the question of whether MPS sequence data could be used for preemptive PGx testing. Part of the larger challenge for the field of medical genomics is to identify all potential uses of sequencing so that the cost of these assays can be amortized across multiple applications, thereby decreasing the effective cost of the test. Prior studies with small sample sizes showed high ES genotype concordance rate with other platforms (99.6% with MiSeq and 98.9% with iPLEX ADME PGx panel)6 and variable (60–80%) ES coverage of DMET Plus PGx variant positions depending on the capture kit used.7 An extensive analysis with a larger data set was needed to assess the capability of ES in detecting clinically relevant PGx variants. We set out to assess MPS concordance and coverage of annotated PGx variants compared with a current genotyping platform to determine whether MPS could serve as a genotyping source for preemptive PGx testing.

Materials and Methods

Participants

This study was performed at the NIH Clinical Center as part of the ClinSeq project and included 973 participants enrolled between 45 and 65 years of age who were consented for baseline clinical tests, ES and/or GS, return of genetic results, and iterative phenotyping based on an individual’s genetic variants.8,9 The National Human Genome Research Institute Institutional Review Board reviewed and approved this study. See Supplementary Methods online.

Selection of clinically relevant pharmacogenetic variants for comparison

We identified 50 Pharmacogenomics Knowledgebase (PharmGKB) level 1A and 1B PGx variants (https://www.pharmgkb.org/) and 154 Clinical Pharmacogenetics Implementation Consortium (CPIC) (http://www.pharmgkb.org/page/cpic) variants from 40 gene–drug pairs with level A evidence (two promoter variants were located at the same genomic position) for a total of 203 PGx variant positions. We evaluated coverage of these 203 PGx variant positions from 973 exomes, 5 genomes, and 5 chip data. Three HLA-B variants (HLA-B*52:01:01, HLA-B*57:01:01, HLA-B*58:01:01) were excluded because they were not amenable to genotyping by the chip, ES, or GS. MPS genotype concordance was determined by comparing five individuals with ES, GS, and DMET Plus genotypes (hereafter referred to as the chip). The chip has been previously shown to have high genotype concordance (91 to 99%) compared with six orthogonal genotyping platforms.10 We selected CYP2D6 for copy-number variant (CNV) analysis because 1–2% of individuals carry more than two functional copies that may have an ultrarapid metabolizer phenotype that can lead to codeine toxicity.11

Laboratory methods

See Supplementary Methods online.

Results

Detection of 203 CPIC/PharmGKB variant positions by exome versus genome versus chip

Five individuals were examined for 203 curated variants (132 coding, 71 noncoding positions) by ES, GS, and chip-based testing. One would ideally like to detect a total of 1,015 genotypes (203 × 5). A total genotype count regardless of genotype quality from five individuals is shown in Figure 1 . GS detected 998/1,015 genotypes (657/660 coding and 341/351 noncoding). In the coding positions, 129/132 positions were covered in five individuals and 3/132 were covered in four individuals. In the noncoding positions, 63/71 positions were covered in five, 5/71 in four, and 3/71 in two individuals. For ES, 117/203 positions were targeted by two capture kits for five individuals (Agilent38Mb n = 2, TruSeqV2 n = 3), 12/203 were targeted only by Agilent38Mb, and 14/203 were targeted only by TruseqV2. The expected total genotype count is 651 ((117 × 5) + (12 × 2)) + (14 × 3)). The targeted genotype detection rate was 647/651. Of the positions targeted by both capture kits, 114/117 variant positions were covered in five individuals, 2/117 in four individuals, and 1/117 in three individuals. All 26 positions targeted by only one of the capture kits had complete coverage. The total ES genotype count was 849 (647 targeted and 202 off-target). The chip targeted 46/132 coding and 14/71 noncoding positions, and the targeted detection rate was 225/230 (coding) and 70/70 (noncoding) ( Figure 1 , Supplementary Tables S1–S3 online). The in-house cost per genotypable site was $43.77 ($8,710/199 positions) for genomes, $4.79 ($810/169 positions) for exomes, and $9.31 for the chip ($549/59 positions). These figures may not reflect clinical costs.

Figure 1
figure 1

Genotype count of five genomes, five exomes, and five chip data at 203 pharmacogenetic variant positions in five individuals. Total genotype count for five individuals at 203 pharmacogenetic variant positions regardless of genotype quality from genome, exome, and chip. The total genotype count is 1,015 for genomes: 651 targeted genotypes for exomes and 300 targeted genotypes for the chip (represented by the horizontal dashed line). Genomes detected 998/1,015, exomes detected 849/1,015 (647 targeted represented by the diagonally striped area and 202 off-target represented by the light gray area), and the chip detected 295/1,015. Chip, Affymetrix DMET Plus (Drug Metabolizing and Transporters array).

Full size image

We next examined the detection rate of high-quality genotypes (GQ ≥50) per individual at 203 positions. We included 973 exomes captured with four kits (Agilent38Mb, Agilent50Mb, TruSeqV1, TruSeqV2). ES, GS, and chip data were grouped as coding and noncoding variants (intergenic, intronic, promoter, or 3ʹ untranslated region). GS and ES detected an average of 101 and 120 genotypes per individual at coding positions, respectively. At noncoding positions, GS and ES detected an average of 55 and 27 genotypes per individual, respectively ( Figure 2a , b ; ES average based on TruSeqV1 and V2 data). ES coverage was the highest in coding regions and the TruSeqV2 kit had the highest average (122); the chip captured 45 genotypes per individual ( Figure 2a ). ES coverage in noncoding regions was low. Among the 71 noncoding positions, TruSeqV1/V2 had the highest average (27) and the Agilent38Mb kit and the chip had the lowest average (14) of genotypes per individual ( Figure 2b ). GS coverage was outperformed by the Agilent50Mb and TruSeqV1/V2 kits in coding regions ( Figure 2a ; Supplementary Table S4 online).

Figure 2
figure 2

Comparison of variant coverage between exomes, genomes and chip. (a) Exome capture kits versus genome versus chip coverage of 132 coding pharmacogenetic variant positions. (b) Exome capture kits versus genome versus chip coverage of 71 noncoding pharmacogenetic variant positions. Total number of variant positions represented by the horizontal dashed line. Bar graphs shows the average number of high-quality variants per individual shown by four exome capture kits (Agilent 38Mb (n = 393), Agilent 50Mb (n = 318), Illumina TruSeqV1 (n = 147), Illumina TruSeqV2 (n = 115)) versus genome sequence (n = 5) versus chip data (n = 5). The tops of the bars indicate the average number of high-quality (GQ score ≥50) variant(s) detected per individual for exomes, genomes, and chip. The whiskers above the bars represent the standard error of the mean. See Supplementary Table S4 online for mean, SEM, and N. 3ʹUTR, 3 prime untranslated region; Chip, Affymetrix DMET Plus (Drug Metabolizing and Transporters array); CPIC, Clinical Pharmacogenetics Implementation Consortium; ES, exome sequence; GQ, genotype quality; GS, genome sequence; Mb, megabase; N, number of individuals tested per platform; PGx, pharmacogenetic; PharmGKB, Pharmacogenomics Knowledgebase; SEM, standard error of the mean.

Full size image

Detection of CPIC and PharmGKB pharmacogenetic variants and rare loss-of-function variants in known pharmacogenes

ES identified 36 star (*) allele variants with CPIC recommendations for change in therapy, including individuals homozygous for CYP2C19 *2 (n = 18), TPMT *3B, TPMT *3C (n = 5),12 SLCO1B1 *5 (n = 21),13 and individuals heterozygous for DPYD*13 (n = 2) and rs67376798 (n = 6) (Supplementary Table S5 online).14 Twenty individuals with rare loss-of-function and eight with splice variants were identified in eight known pharmacogenes (Supplementary Table S6 online).

Genotype concordance between exomes, genomes, and genotyping chip

The chip had 1,929 unique variant positions and identified 9,598 genotypes for the five samples tested.

Of 8,040 genotype calls made by chip–ES, 7,258 homozygous/hemizygous and 639 heterozygous calls were concordant and 143 (1.8%) calls were discordant. Of the chip–ES discordant calls, the chip called 89/143 heterozygous and 54/143 homozygous, 83/143 of the discordant calls had ES GQ <50, and 77/83 are noncoding. For discordant calls with ES GQ ≥50, 57/60 were concordant in ES–GS (12/57 heterozygous and 45/57 homozygous) (Supplementary Tables S7 and S8 online).

Of 9,543 genotype calls made by chip–GS, 8,411 homozygous/hemizygous and 1,029 heterozygous calls were concordant and 103 (1.1%) were discordant. Of the discordant chip–GS calls, the chip called 19/103 heterozygous and 84/103 homozygous/hemizygous, 29/103 had GS GQ <50, and 74/103 had GS GQ ≥50. More than two-thirds (20/29) were discordant coding calls. Among the discordant calls with GS GQ ≥50, 52/74 were concordant between ES and GS (12/52 heterozygous, 40/52 homozygous) (Supplementary Tables S7 and S8 online).

Of 8,013 genotype calls made by ES–GS, 7,267 homozygous/hemizygous and 649 heterozygous were concordant and 97 (1.2%) were discordant. Of the discordant ES–GS calls, the chip called 78/97 heterozygous and 19/97 homozygous, 80/97 had ES GQ <50, and 73/80 were noncoding. The majority (76/97) of the discordant ES–GS calls were concordant between chip and GS (75/76 heterozygous, GS GQ ≥50; 1/76 homozygous, GS GQ <50). A few (17/97) of the discordant ES–GS calls had ES GQ ≥50 and 14/17 were concordant between chip and ES (2/14 heterozygous, 12/14 homozygous) (Supplementary Tables S7 and S8 online).

Detection and validation of CYP2D6 CNVs using eXome hidden Markov model

CYP2D6 CNVs were detected in 57/973 exomes (duplication n = 39, deletion n = 18) (Supplementary Table S9 online). XHMM quality scores (QS) ranged from 38 to 99. Seven individuals with the highest XHMM QS of 99 (duplication n = 6, deletion n = 1) were selected for validation with real-time quantitative polymerase chain reaction (qPCR) and all samples were confirmed (Supplementary Table S10 online). An additional 19 samples with XHMM QS ranging from 38 to 99 (duplication n = 17, deletion n = 2) were selected for a second round of validation with qPCR and all were confirmed (Supplementary Table S10 online). Of the 26 samples tested, 11 samples showed agreement across all CNV regions, nine samples were inconclusive (XHMM does not make CNV calls in noncoding regions), and six samples (168397, 136439, 181872, 181608, 185076, 196659) showed breakpoint discrepancies between the XHMM predictions and the qPCR results. This was not a surprising finding because their XHMM Q_exact scores (confidence measure of the predicted CNV breakpoint) were low, ranging from 4 to 18 (data not shown). Nine samples (142307, 175100, 187383, 140601, 190031, 190871, 194883, 131340, 167715) with predicted whole-gene duplication showed a normal copy number with the 5ʹ probe (Supplementary Table S10 online). This may be due to the CYP2D6 duplication not extending into the 5ʹ region. High sequence identity (96.9%) of CYP2D6 and the CYP2D7P pseudogene (NM_000106.5 and NR_002570.2, respectively) can result in CYP2D6-2D7P hybrid genes.15 For these nine individuals, ES data analysis did not find paired-end reads mapping to CYP2D6-CYP2D7P. Our paired-end reads (89 bp) and inserts (180 bp) are short; therefore, the absence of detecting paired-end reads mapping to CYP2D6-CYP2D7P does not rule out the presence of fusion/hybrid genes.

Discussion

Adoption of PGx-guided therapy has been limited by insufficient data to support clinical utility and cost-effectiveness, knowledge gaps in pharmacogenomics, and the inherent delay engendered by PGx testing. We propose leveraging existing MPS data by extracting PGx variants preemptively based on two premises. The first is that thousands of patients are currently undergoing clinical ES and GS, and these data comprise a valuable resource for pharmacogenomics. The second is that the extraction of PGx variants from ES/GS data is part of a larger effort to maximize the utility of ES/GS testing results. Studies have demonstrated how ES data can be used to extract variants for the secondary screening of susceptibility to cancer, malignant hyperthermia, cardiomyopathy, cardiac dysrhythmias, and aortic dissection.16,17,18,19 We assessed the capability of MPS for preemptive PGx testing by comparing the coverage of 203 important PGx variants in 973 exomes with that of a widely used PGx chip.

ES and GS had several advantages over chip-based testing. The genome-wide coverage of ES and GS allowed coverage of more PharmGKB class 1A, 1B, and CPIC gene–drug level A variants than the genotyping chip and identified both known and yet to be discovered PGx variants in one test.

CYP2D6 is a good example for exploring the ability of ES to interrogate CNVs. XHMM detected complete and partial CYP2D6 deletions and duplications, but the chip only detects deletions.

Limitations of this study include the fact that 399/973 of the ES sequences were generated with the Agilent38Mb capture kit, which accounted for the majority of the NC in the ES data, thus decreasing the coverage of some variant positions. The use of four capture kits provided us an opportunity to assess variance in capture-kit coverage.

Our results showed that high exome genotype concordance rates and higher coverage with the TruSeq capture kits (using GQ ≥50) are consistent with findings from recent studies evaluating exome capability for pharmacogenomics screening.6,7 An updated array targeting PharmGKB level 1A and 1B and CPIC level A variants may be a more cost-efficient initial screen than exomes; however, panel testing and enhanced exome capture with additional targets in noncoding regions20 will require periodic updating of the test platform and repeat testing of subjects for future discoveries. Although our results showed that exomes can be used to extract PGx variants, we are not advocating ordering an exome primarily for pharmacogenomics screening because our analyses did not answer the question of whether there is clinical utility and validity for using MPS for preemptive PGx screening for these variants.

We have demonstrated the utility of MPS data for the detection of single PGx variants and CYP2D6 CNVs. Currently, no tools are available to extract and annotate PGx variants from MPS data. We conclude that tools should be developed to extract PGx variants from existing ES and GS data for research and potential future use.

Disclosure

D.N., C.S.H., L.N.S., J.J.J., and J.C.M. declare no conflict of interests. L.G.B. is an uncompensated adviser to the Illumina Corp, receives royalties from Genentech, Inc., and receives honoraria for editing from Wiley-Blackwell, Inc.