Next-generation sequencing in pharmacogenomics – fit for clinical decision support?

ABSTRACT

Introduction

The technological advances of sequencing methods during the past 20 years have fuelled the generation of large amounts of sequencing data that comprise common variations, as well as millions of rare and personal variants that would not be identified by conventional genotyping. While comprehensive sequencing is technically feasible, its clinical utility for guiding personalized treatment decisions remains controversial.

Areas covered

We discuss the opportunities and challenges of comprehensive sequencing compared to targeted genotyping for pharmacogenomic applications. Current pharmacogenomic sequencing panels are heterogeneous and clinical actionability of the included genes is not a major focus. We provide a current overview and critical discussion of how current studies utilize sequencing data either retrospectively from biobanks, databases or repurposed diagnostic sequencing, or prospectively using pharmacogenomic sequencing.

Expert opinion

While sequencing-based pharmacogenomics has provided important insights into genetic variations underlying the safety and efficacy of a multitude pharmacological treatments, important hurdles for the clinical implementation of pharmacogenomic sequencing remain. We identify gaps in the interpretation of pharmacogenetic variants, technical challenges pertaining to complex loci and variant phasing, as well as unclear cost-effectiveness and incomplete reimbursement. It is critical to address these challenges in order to realize the promising prospects of pharmacogenomic sequencing.

KEYWORDS:

• Variant effect prediction
• gene-drug associations
• targeted sequencing
• genomic medicine
• precision medicine

1. Introduction

Inter-individual variability in drug response and toxicity constitutes a common phenomenon in clinical care that poses a significant burden on the healthcare system. A multitude of factors contribute to unintended treatment outcomes including drug–drug interactions, physiological and pathophysiological factors, environmental cues, adherence, and genetic differences, the latter of which are estimated to account for an estimated 20–30% of this variability [1]. The majority of genetic variations that affect drug response or the risk of adverse drug reactions (ADRs) localize in loci involved in drug absorption, distribution, metabolism, and excretion (ADME) or in genes that encode drug targets. Genetic variants in these genes that robustly associate with drug efficacy or safety can provide biomarkers for the individualized guidance of pharmacotherapy and further identification of such pharmacogenomic associations thus constitutes a key aim of precision medicine [2–4]. To date, around 180 drug labels by the US Food and Drug Administration (FDA) and European Medicines Agency (EMA) have actionable pharmacogenomic labels [5]. Importantly, >90% of all individuals carry at least one actionable genotype [6–8] and 58% of all adult patients are prescribed at least one medicine with available pharmacogenomic guidance, which increases to 89% for patients over 70 years of age [9,10]. Nevertheless, the use of pharmacogenomic biomarkers in clinical practice is limited and only few associations are routinely used in clinical care outside of targeted oncology.

From twin studies, it is clear that candidate variants that are typically evaluated in association studies only explain a part of the overall hereditary variability in pharmacogenetic traits [11]. While consideration of common variations can improve the outcomes of pharmacotherapy, considerable variability up to 100-fold can remain even within genotype-stratified groups [12,13]. These findings strongly support the hypothesis that a considerable fraction of heritable drug response variations remains unexplained. This missing heritability can be due to rare or personal variants not typically considered in genotyping-based analyses, epigenetic mechanisms or caused by epistatic phenomena in cis or trans [14].

Currently, pharmacogenomic testing is mainly conducted via the genotyping of pre-selected variants in a panel of pharmacogenes. The advancement of sequencing technologies during the past 20 years has, however, revealed that there are tens of thousands of different variations across the human pharmacogenome. Specifically, population-scale sequencing projects revealed that > 90% of all identified pharmacogenomic variants are rare with minor allele frequencies (MAFs) below 1% and each individual carries multiple rare functional variants across ADME and drug target genes [15–17]. Overall, rare coding and non-coding variations are estimated to account for approximately 30% and 20% of the genetically encoded functional variability in pharmacogenes, respectively [18,19]. These estimates suggest that consideration of comprehensive sequencing data that includes such rare variations could considerably improve individualized predictions of pharmacophenotypes.

In this review, we discuss the advantages and limitations of Next-Generation Sequencing (NGS) and genotyping in clinical pharmacogenomics and compare the existing panels designed for pharmacogenomic sequencing in either research or clinical settings. Subsequently, we provide a current overview of sequencing studies that either repurpose clinical sequencing data or conducted targeted pharmacogenomic sequencing to predict drug response or explain extreme phenotypes. We conclude that NGS provides a powerful and versatile tool for precision medicine whose implementation into routine clinical care, however, requires further developments at multiple fronts.

2. Genotyping versus sequencing

Pharmacogenomic testing via genotyping has been widely adopted in many clinical institutions and pharmacogenomic programs [20–22]. Genotyped variants usually have strong evidence for their association with drug efficacy or toxicity. Therefore, the test results are easier to interpret and, when respective guidelines are available, can be translated into dose adjustments or alternative drug selection.

In contrast to genotyping, which can only cover a limited number of candidate variants, sequencing provides comprehensive genetic information about the interrogated loci. Three different sequencing strategies can be distinguished: whole-genome sequencing (WGS), whole-exome sequencing (WES) and targeted sequencing of selected candidate loci. From a technical standpoint, pharmacogenomic variants can by now be accurately called from sequencing data and results show high concordance with targeted genotyping results [23–25]. While genotyping of candidate polymorphisms detects most clinically significant pharmacogenomic variants, up to 13% of variants per pharmacogene are not covered in conventional genotyping panels and could only be detected by sequencing-based approaches [26]. It is furthermore important to consider that conventional short-read sequencing also has technical limitations, particularly with regard to variant detection in complex pharmacogenetic regions that contain repetitive sequences, copy number variations (CNVs) or nearby homologous pseudogenes, such as CYP2D6, CYP2B6 and the human HLA locus [27]. These issues can be addressed by using NGS with longer reads. For instance, while five out of nine CYP2B6 exons contained inaccessible regions when using 100bp single-end reads, all exons could be fully interrogated when 200bp reads were used instead [28]. Alternatively, long-read sequencing is increasingly utilized, which produces continuous reads of 20-100kb, thus allowing to produce phased information even in highly complex loci. While not yet in routine clinical use, current developments in software and workflows have lowered the barriers for implementation [29,30].

Importantly, while some clinical testing facilities already use sequencing as a modality for pharmacogenomic testing, they only extract known variations with available guidelines and do not interpret results pertaining to pharmcogenomic variants of unknown significance. The main reason is that ACMG/AMP guidelines are designed for the interpretation of variants encountered through clinical sequencing for hereditary disorders [31]. Specifically, they recommend the functional classification of variants based on population data, computational variant effect predictions, as well as functional and segregation data. While these categories are useful for the classification of variant pathogenicity, they are not applicable to pharmacogenomic variants that are often not disease associated. No similar guidance documents for the interpretation of pharmacogenetic sequence variants are presently available. Furthermore, if variants of unknown significant coincide with variants or haplotypes for which pharmacogenomic guidance documents are available, it is currently unclear how such information should be reported. Thus, while the added technical capabilities of pharmacogenomic sequencing are evident, its main added value pertaining to the detection of rare or novel variations is currently not leveraged. Investigations into whether and how this additional information can be clinically utilized are thus an important frontier of pharmacogenomic research.

3. Pharmacogene selection

Pharmacogenomic variability can be examined i) by analyzing preexisting genetic data that was generated for other purposes, such as diagnostic sequencing or ii) by actual sequencing of pharmacogenes. Analysis of preexisting data, often from WGS or WES, is commonly conducted in a clinical setting. This approach requires only a reinterpretation of already available data and, as such, virtually any number of genes that are of interest in the analytical context can be assessed. In contrast, sequencing for pharmacogenomic applications is predominantly conducted in a research setting. In this context, targeted panel-based approaches are typically preferred over WGS or WES due their favorable balance between cost, throughput, and coverage.

Multiple panels for such sequencing-based interrogations of pharmacogenetic variation in large cohorts have been presented, including PGRN-Seq [32], ClinPharmSeq [33], PGxSeq [34] and PKSeq [35]. These panels differ considerably in scope, ranging from 59 to 340 included genes, as well as in gene composition (Figure 1a, Table 1, Supplementary Table S1). Only 18 genes are included in all of the pharmacogenomic sequencing panels, whereas 261 and 91 were only included in one and two panels, respectively (Figure 1b). Genes encoding drug metabolizing enzymes were more extensively covered, whereas drug target genes were underrepresented. Of the 26 actionable pharmacogenes, i.e. genes encoded in the nuclear genome with guidelines from the Clinical Pharmacogenetics Implementation Consortium (CPIC) and the Dutch Pharmacogenetics Working Group (DPWG), 12 were covered by all sequencing panels (Figure 1c). These mostly included polymorphic CYPs and other phase I or phase II enzymes, as well as the drug transporter gene SLCO1B1 and VKORC1, encoding vitamin K epoxide reductase, the target enzyme of warfarin. Furthermore, NAT2, POR, SLC15A2, SLC22A2, SLCO2B1, and UGT1A4 were also evaluated by all panels, despite a lack of pharmacogenetic guidelines. In contrast, multiple loci associated with adverse events, such as HLA-A, HLA-B and CACNA1S, or treatment efficacy (IFNL3) were only covered in one or two of the sequencing panels. Combined, these results suggest that clinical actionability per se is not a major focus of most current pharmacogenomic sequencing panels, despite the ambition that panel-based approaches might eventually transition from research tools to clinical use.

Figure 1. Analysis of pharmacogenomic targeted sequencing panels. (a) The concordance of genes included in pharmacogenomic sequencing panels. Highest concordance is indicated in red and lowest in blue. (b) Column plot showing the overlap of pharmacogenes across the six analyzed panels. Note that pharmacogenes included in 4–6 panels are almost exclusively encoding factors involved in the absorption, distribution, metabolism and excretion (ADME) of drugs, whereas most drug target genes are only included in 1–3 panels. (c) Column plot showing the panel coverage of pharmacogenes with established actionable guidelines by CPIC or DPWG. While 12 out of 26 actionable genes are included in all panels, a total of nine genes are included in less than half of all panels.

Table 1. Number of pharmacogenes included in NGS-based pharmacogenomic studies.

Panel name	Total number of pharmacogenes	ADME genes	Target genes	Reference
PGRNseq	84	43	41	[32]
ClinPharmSeq	59	38	21	[33]
PGxSeq	100	89	11	[34]
PKseq	100	99	1	[35]
Han et al.	114	72	42	[36]
Klein et al.	340	337	3	[37]

4. Repurposing of diagnostic and population-scale sequencing data

Due to the plummeting cost and increasing accessibility of NGS, large amounts of human sequencing data have accumulated over the past decades. Where accessible, these datasets can be tapped into to conduct retrospective pharmacogenomic analyses. To this end, sequencing data is mostly interpreted as targeted genotyping tests, i.e. only a few common variants with known clinical significance are extracted and functionally interpreted to infer metabolic phenotypes, whereas rare and novel variants without actionable guidance are not considered. In this section, we first discuss the potential and limitations of pharmacogenomic repurposing of such clinical sequencing data generated for diagnostic applications. Subsequently, we highlight the importance of biobanks and large population-scale sequencing projects for pharmacogenomic analysis.

WES and WGS are powerful tools for the molecular diagnostics of congenital and rare diseases [38,39]. Repurposing of such clinical sequencing data for pharmacogenomic analyses can provide population frequencies as well as useful information for personal drug response predictions (Table 2). Analysis of WES and targeted exome sequencing data from 5,001 Spanish or Latin American individuals identified 280 alleles in 11 pharmacogenes that can be used to inform pharmacogenomic phenotypes impacting over 95% of the studied cohort [40]. In addition, 197 novel variants were discovered, of which 26 were considered loss-of-function. Similarly, pharmacogenomic profiling of WES data from 1,116 Hong Kong Chinese revealed that almost all individuals carried at least one actionable pharmacogenomic variant across 19 pharmacogenes and, when taking prescription data into account, 13.4% of the population would benefit from pharmacogenetically guided drug prescriptions [42]. Furthermore, above 93% of the individuals carry at least one putatively deleterious rare variant across 108 pharmacogenes. While repurposing of WES data resulted in meaningful pharmacogenomic profile, it suffers from difficulties in inferring functionality of UGT1A1, CYP3A5, CYP2C19 and CYP2D6, due to the lack of coverage of important non-exonic variations and problems pertaining to copy number variant calling. Consequently, functional interpretation of these loci requires additional validation before these data can provide guidance for clinical use [41]. Such issues can be overcome by using WGS and multiple studies have successfully translated WGS data into pharmacogenomic phenotypes, mostly in pediatric cohorts [46,47].

Table 2. Studies repurposing clinical sequencing data for pharmacogenomic analysis. WES = whole-exome sequencing; WGS = whole-genome sequencing; TS = targeted sequencing.

Data type	Cohort size	Variants/alleles analyzed	Result based on well-characterized common variants	Result based on novel and rare variants	Ref
WES or TS	5001	788 alleles in 19 pharmacogenes, actionable; novel variants in 11 pharmacogenes	Each individual carried in average 2.2 alleles; 95% of the cohort could be informed of at least one actionable pharmacogenetic phenotype.	197 novel variants in 11 selected pharmacogenes	[40]
	1583	42 variants in 11 pharmacogenes	At least one actionable phenotype was present in 86% of the individuals.	-	[41]
	1116	133 variants in 19 pharmacogenes; rare variants in 108 pharmacogenes	Nearly all (99.6%) subjects carried at least one actionable pharmacogenetic variant	93.5% of subjects harbored at least one rare deleterious pharmacogenetic variant	[42]
	94	12 variants in 3 pharmacogenes	91% had at least one variant allele reported; 20% had potential immediate implications on current medication use	-	[43]
	94	48 variants in 11 pharmacogenes	49 out of 94 participants enrolled had one or more pharmacogenomic variants identified for return	-	[44]
WGS	547	45 variants in 11 pharmacogenes	37 out of 45 variants were detected, on average, 6 of which occurred per genome	14 additional variants (including 10 novel variants)	[45]
	98	67 variants in 19 pharmacogenes	95 out of 98 patients had at least one clinically actionable pharmacogenetic variant	-	[46]
	50	Star alleles in 16 pharmacogenes	24 out of 50 carried variants in pharmacogenes with drug-gene guidelines for one or more of their current medications	-	[47]

NGS is furthermore recommended for a variety of solid tumors and hematological malignancies and has become standard in clinical oncology. While recommendations regarding the sequencing approach and scope can differ between tumor type and stage [48], even comprehensive WGS has been shown to be feasible in clinical practice in a comprehensive cancer center setting [49]. The majority of cancer patients receive NGS only on tumor samples to identify potential therapeutic vulnerabilities. However, whether somatic sequences can be reliably used to infer germline genotypes of drug response variants remains controversial [50–52]. In recent years, there is an increase in paired somatic-germline sequencing. This approach allow offers multiple advantages, such as better guidance for treatment selection and the opportunity to reduce risk in family members [53,54]. Furthermore, due to the resulting availability of genomic germline data, such information can be repurposed for pharmacogenomic applications for instance, to make evidence-based treatment recommendations for common chemotherapeutics [55].

While the aforementioned evidence indicates the feasibility and utility of repurposing diagnostic sequencing data for pharmacogenomic purposes, the seamless integration of this information into electronic health records (EHRs) remains challenging. A 2015 proof-of-concept study identified two key challenges – the accurate interpretation of incidental findings in agreement with pharmacogenomic guidelines and identification of those findings that are clinically meaningful [44,56]. For well-characterized variations, the first challenge has been largely addressed by pharmacogenomic variant calling and interpretation tools, such as Stargazer [57], Aldy [58] and PharmCAT [59], whereas the interpretation of rare variants with unknown functional significance remains an open issue. Furthermore, a dedicated pharmacogenomic reporting system, named lmPGX, was recently presented that fully automates the translation of clinical short-read sequencing data into succinct recommendations for 11 clinically actionable pharmacogenes that can be readily interpreted by clinicians [60]. The second challenge, however, remains largely unsolved since the functional interpretation of an individual’s pharmacogenomic profile remains, with few exceptions, imprecise.

In recent years, biobank evolved from collections of biological samples to complex multi-dimensional and often decentralized repositories in which diverse samples are stored and annotated with clinical information and, often, complex omics data [61]. Many national biobanks were established that contain in-depth genetic and health information for population-specific research purpose. For example, the UK Biobank currently contains genotype data from 500,000 individuals, WES data from 470,000 individuals and WGS data from 200,000 individuals that are all linked to EHRs [62]. Similarly, the Estonian Biobank includes genetic data from 200,000 individuals, WES data from 2,500 individuals and WGS data from 3,000 individuals accompanied by lifestyle and genealogical information [63]. These biobanks serve as valuable resources for large-scale pharmacogenomic studies and have been used to generate comprehensive pharmacogenomic variation profiles. For example, analysis of the genetic variability across 14 pharmacogenes with CPIC guidelines based on the UK Biobank indicated that 99.5% of population are at risk of an atypical response to at least one drug [8]. Similarly, preexisting genotype data of over 44,000 participants from the Estonian biobank, in comparison with WES and WGS data, showed that 99.8% of studied individuals carry at least one actionable variant and pharmacogenomic dosing adjustments will affect at least 50 daily drug doses per 1000 inhabitants [7].

In addition, consolidated and harmonized international sequencing databases, such as the 1000 Genomes Project and The Genome Aggregation Database (gnomAD) have emerged as powerful tools for pharmacogenomic research. These resources not only allow analysis of pharmacogenomic germline variations across different ethnogeographic groups [64–66] but also provide opportunities to estimate functional consequence of naturally occurring variants at the population-scale by integrating frequency information with functional data from other sources [67]. Moreover, germline variation data from sequencing data stored in cancer databases, such as The Cancer Genome Atlas (TCGA), can be utilized to predict therapeutic effect of different cancer drugs [68]. Even in the absence of EHR annotations, these resources constitute important repositories to assess global pharmacogenomic variability and, combined with established gene–drug relationships allow to estimate drug response phenotypes at the population scale.

5. Pharmacogenomic sequencing

In addition to retrospective pharmacogenomic analysis of preexisting sequencing data, NGS is increasingly applied prospectively (Table 3). More than 10 years ago, the RIGHT Study (The Right Drug, Right Dose, Right Time) was initiated that aimed to develop individualized treatment protocols using genomic data from cohort samples stored in the Mayo Clinic Biobank [82]. A RIGHT pilot study was conducted in which 1,013 subjects were sequenced using the 84-gene PGRN-seq panel, followed by targeted analysis of SLCO1B1, CYP2C9, CYP2C19 and VKORC1. In addition, CYP2D6 was separately genotyped using a customized genotyping method. The result revealed that 99% of the subjects carried an actionable pharmacogenomic variant in as least one of the five genes [69]. Functional consequences of novel variants identified in CYP2C9 and CYP2C19 were interrogated by in vitro and in silico methods in a follow-up study and the authors conclude that if novel variant interpretation can be generalized using high-throughput functional assays and highly accurate predictive algorithms, DNA sequencing will ultimately be preferable to genotyping for the clinical implementation of pharmacogenomic variants [70]. Recently, the complete Mayo-Baylor RIGHT 10k Study was published, presenting the largest preemptive DNA sequence-based pharmacogenomic testing to date [71]. In the frame of this study, targeted sequencing of 77 pharmacogenes was performed for 10,077 individuals. The results revealed that, in addition to clinically actionable variations, each individual carried on average 127 additional SNVs and indels, which would have been missed by genotyping. Importantly, 3.3 variants per individual were interpreted as putatively deleterious and more than 30 individuals carried rare functionally deleterious variants in CYP2C9 and CYP2C19 that were confirmed by deep mutational scanning. To facilitate clinical implementation, electronic consultations were sent to primary care providers of risk variant carriers to suggest dose adjustment or alternative therapy. These adjustments resulted in full remission or alleviation of ADRs in some cases. While clearly promising, stringent statistical evaluations of the results are required to demonstrate clinical utility.

Table 3. Pharmacogenomic sequencing studies. AUC = area under the curve; NGS = next-generation sequencing; WES = whole-exome sequencing; WGS = whole-genome sequencing; TS = targeted sequencing.

Study objective	N	Therapeutic area	Sequencing method (analyzed genes)	Findings beyond well-characterized common variants	Conclusion	Ref
Evaluation of pharmacogenomic workflow in the clinical setting	1,013	General	TS (84)	14 novel variants in CYP2C9 and CYP2C19 were identified. 11 caused reduced catalytic activity.	High-throughput functional assays and highly accurate predictive algorithms are needed to improve personalized predictions.	[69,70]
	10,077	General	TS (77)	Each individual carried on average 3.3 deleterious variants in addition to star alleles.	Preemptive pharmacogenomic panel sequencing can be clinically useful if delivered at the point-of-care.	[71]
	100	General	WES (1800 for rare variant analysis; 9 for star allele analysis)	4539 rare variants (including 115 damaging non-synonymous) were identified	While a comprehensive clinical PGx profile could be successfully extracted from WES data, available tools to interpret these data demonstrated inconsistencies that complicate clinical application.	[72]
Resequencing and variant interpretation	137	General	TS and WGS (3)	Six novel alleles and 38 novel sub-alleles were identified. Several ambiguous genotype calls were resolved using NGS	Many novel alleles are yet to be discovered, even in highly characterized genes.	[73]
	304	Psychiatry	TS (100)	Novel pharmacogenetic missense variants that are predicted to be deleterious, significantly alter CYP2D6 and CYP2C19 activity	NGS-derived data coupled with computational predictions as well as with experimental validations can improve phenotype predictions.	[74]
Explaining extreme phenotypes	Case study	Oncology	WGS (1)	The patient carries a heterozygous rare missense variant in DPYD that likely explains the observed toxicity.	Genomic sequencing along with in silico prediction and functional testing can identify the molecular basis of rare ADRs.	[75]
	224	Oncology	TS (1)	Burden of rare DPYD variants was significantly higher in patients with toxicity compared to controls.	Consideration of rare variants by DPYD-sequencing could improve prevention of severe toxicity of fluoropyrimidines.	[76]
	1,441	Respiratory	WGS (genome-wide)	Common and rare variants explained 4–8% of bronchodilator drug response variation.	WGS identified genetic variants important for bronchodilator drug response in asthmatic children	[77]
Drug response prediction	561	Oncology	TS (1)	Model trained on CYP2D6 long-read sequences achieved 25% increase in explained phenotype variability compared to the conventional star alleles.	A continuous scale and a completely phased genotype could improve prediction of CYP2D6 enzyme activity and enable more accurate prediction of personalized drug response.	[78]
	60	Cardiovascular	TS (20)	10% of the 60 variants that were used to construct the prediction model would not have been found by genotyping.	Comprehensive analysis of several genes is required to predict atorvastatin AUC.	[79]
	699	Oncology	TS (1)	Rare variants accounted for 17.8% of the genetic impact of SLCO1B1 on methotrexate clearance and had larger effect sizes than common variants.	Rare variants can have important effects on pharmacophenotypes.	[80]
	420	General	WGS (231)	More than 1500 variants with frequency > 20% were not present on the DMET-array. Novel ADME variants occur more frequently in certain ethnic groups.	WGS can improve the analysis of personal pharmacogenomic profiles compared to conventional genotyping methods	[81]

In a different study, a similar workflow from sequencing to pharmacogenomic result reporting was explored [72]. Specifically, WES data from 100 individuals were analyzed for both star alleles and rare variants. The former was translated into individualized medication safety cards while the latter underwent complex computational assessments. Notably, the authors observed drastic discordance between different haplotype calling algorithms and proposed that, in the absence of more accurate and precise prediction methods, results from multiple tools should be integrated to obtain more confident results.

To harness the potential advantage of sequencing over genotyping, functional interpretation of rare and novel variants is of central importance. Resequencing of 137 DNA samples from the Genetic Testing Reference Material (GeT-RM) identified novel variants in CYP2C8, CYP2C9 and CYP2C19 that would have been missed in targeted genotyping assays, highlighting that many novel variants are yet to be discovered even in extensively studied pharmacogenes [73]. To further address the issue of variant interpretation, an analytical workflow was developed in which 13 variants in CYP2D6 and CYP2C19 identified by targeted sequencing of 304 individuals were first assessed using in silico predictors and subsequently expressed in 293 FT cells for experimental analyses [74]. The results demonstrated a high concordance between computational predictions and in vitro data, and emphasized the functional impact of novel variants for genotype–phenotype translations.

Besides prediction of drug response, pharmacogenomic data can also be utilized in forward genetics to identify genetic variations that explain ADRs or other pharmacophenotypes. For instance, WGS identified a novel rare missense variant in DPYD, p.R235Q, in a patient who developed life-threatening toxicity after capecitabine treatment after commonly tested risk variant could not explain the observed phenotype [75]. This variant was predicted to be deleterious by the DPYD-specific prediction algorithm DPYD-Varifier [83] and the prediction was validated in vitro. Combined, these mechanistic insights provide strong evidence for the conclusion that the identified variant is responsible for the observed phenotype despite the fact that the low frequency of the variant did not allow for epidemiological association studies. In another study, sequencing of DPYD coding and untranslated regions was conducted for 120 patients with grade 3–5 fluoropyrimidine toxicity and 104 matched controls [76]. The authors found that rare variants were significantly enriched in the ADR group and that carrying at least one very rare DPYD variant increased the risk of developing acute fluoropyrimidine toxicity by fourfold (p = 0.017). In a separate example, a machine-learning algorithm trained on CYP2D6 long-read sequences from 561 breast cancer patients quantitatively predicted alterations of CYP2D6 enzyme activity [78]. The model showed a 25% increase in the fraction of explained CYP2D6 variability compared to conventional approaches using star alleles, exemplifying the added value of using comprehensive sequencing for phenotype and drug response predictions.

Importantly, WGS also allows untargeted searches for gene-phenotype associations. In a study with 1,441 children with asthma who exhibited extreme albuterol bronchodilator drug response (BDR), rare variants in loci associated with lung capacity, immunity, and β-adrenergic signaling were shown to significantly contribute to genetically encoded variations in BDR [77]. In addition to explaining outlier phenotypes, novel variants identified by targeted gene sequencing were also demonstrated to have important roles in predicting pharmacokinetic parameters, such as drug plasma concentration and clearance, and, as a result, can inform optimized treatment strategies [79–81].

6. Expert opinion

Over the last decades, large-scale genetic association studies have clearly shown that common variants in ‘classical’ pharmacogenes can only explain a fraction of the genetically encoded inter-individual variability in drug response observed in the clinics. Sequencing technologies thus offer exciting opportunities to increase analytical breadth and detect associations with novel variants as well as additional loci that would not be included in classical genotyping arrays. Particularly pharmacogenomic sequencing panels allow to capture the full spectrum of pharmacogenomic variations and are suitable for large-scale profiling. Current panels differ in scope and composition and there is limited consensus as to which pharmacokinetic and pharmacodynamic genes are to be included in such panels. For instance, important pharmacogenes, such as HLA-A and HLA-B are not consistently covered in these panels, at least in part due to technical difficulties in analyzing these loci. The incorporation of sequencing data into consolidated databases as well as the increasing availability of biobanks can further fuel profiling of the human pharmacovariome at population-scale.

Despite the impressive developments in the research domain and promising prospects for clinical application of sequencing-based pharmacogenomics, significant hurdles remain to be overcome [84–87]:

1. Sequencing-based solutions require appropriate processes and workflows to handle variants of unknown significance if the added value of these technologies is to be leveraged. These could include computational pipelines for the functional interpretation of identified variants, which can, as one piece of evidence, assist clinical decision-making, for instance, by flagging the respective variant carriers for increased surveillance.

2. Specific attention needs to be paid for genes that are challenging to evaluate, such as CYP2B6, CYP2D6 or HLA genes. This includes both the use of appropriate sequencing methods with sufficiently long read-lengths and the use of specific analytical software packages that allow for the accurate calling of genotypes in these genes [88,89].

3. When one or more heterozygous variants are detected, the correct phasing of variants can be important for the inference of functional phenotypes [90]. While imputation can increase the likelihood of correct phasing, only long-read sequencing can fully overcome these issues.

4. The inclusion of sequencing needs to be feasible in clinical routine, i.e. the necessary infrastructure and expertise must be available, and workflows must be established to generate results with an acceptable turnaround time.

5. The cost-effectiveness of sequencing must be demonstrated. While the majority of studies report that genotype-guided dosing might be cost-effective or even dominant in different disease areas [91–93], most studies utilized genotyping and the cost-effectiveness of sequencing remains unclear. With declining sequencing costs, other factors pertaining to data analytics and storage are becoming increasingly more relevant [94]. In parallel to declining sequencing costs, emerging cloud storage solutions have resulted in a rapid reduction of storage costs from hundreds of USD to ten-year costs of 14 USD per genome and 0.71 USD per exome [95]. However, analytical costs, including bioinformatics, interpretation and results reporting, remain relatively constant and can be as high as one-third of the sequencing costs [96].

6. The reimbursement of pharmacogenomic tests remains incomplete and variable. A recent study from a tertiary academic medical center in the US indicated that panel tests are being reimbursed at a significantly higher rate than single-gene tests [97]. However, there are considerable differences over time as well as between indications, genes tested, and the specific payer, likely limiting the generalizability of these results. Considering the most important genomic tests as medical devices has been suggested to stimulate test reimbursement [98]. While not specific to NGS-based pharmacogenomics, the higher out-of-pocket payments for sequencing magnifies the importance of this aspect compared to genotype-based tests. However, such decisions are complex and need to consider how the costs of the approval process propagate to the diagnostic market. Thus, whether genomic tests should be considered as medical devices requires robust health economic evaluations and pricing systems in order to avoid an eventual increase in test costs and decrease in test availability.

While each of the aforementioned points is complex and their in-depth discussion would go beyond the scope of this article, they can nevertheless illustrate the multitude of challenges that remain to be overcome to implement clinical sequencing-based pharmacogenomics. As a consequence, the use of NGS for prospective pharmacogenomic applications remains at present mostly limited to research settings.

To unlock its full potential, parallel progress has to be made in several directions. We specifically highlight the need for efficient and accessible incorporation of sequencing data into electronic health records so that the responsible physicians can make maximal use of the available data when needed. Furthermore, more reliable computational tools are needed that support the interpretation of pharmacogenetic variants with unknown significance to extend clinical actionability to the entire personal genomic information. Finally, appropriate ethical frameworks need to be in place at the respective study site, which ensure that utilization of sequencing data balances patient privacy with anticipated healthcare benefit. Only such integrated developments promise to help a highly efficient translation of sequencing data into personalized health recommendations in pharmacogenomics and beyond. Judging from recent developments, we anticipate that the aforementioned technical challenges can be largely overcome within the next five to 10 years. We are thus optimistic that, with time, pharmacogenomic utilization of sequencing data will become increasingly common and be well positioned to facilitate more accurate translation of an individual’s genetic variability into pharmacophenotypes and personalized treatments.

Article highlights

• Sequencing shows benefits over genotyping regarding genomic information provided and utility over time. Lack of guidance with regard to the interpretation of rare and novel variants as well as unclear cost-effectiveness remain the main challenges for its broader clinical use.

• Current pharmacogenomic targeted sequencing panels are used primarily for research applications and differ drastically in scope and composition.

• Actionability is not major design criterion for the inclusion of genes in most pharmacogenomic sequencing panels.

• NGS data from clinical projects, biobanks and sequencing databases provide ample resources for large-scale pharmacogenomic variation profiling and rare variant analysis.

• Sequencing instead of genotyping is becoming increasingly prevalent for the investigation of extreme phenotypes and drug response predictions.

Declaration of interest

VM Lauschke is co-founder, CEO and shareholder of HepaPredict AB. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

The laboratory received funding from the Swedish Research Council [grant numbers 2019-01837 and 2021-02801], the Knut and Alice Wallenberg Foundation [Grant VC-2021-0026], Cancerfonden [Grant 23-0763PT] and the Robert Bosch Foundation, Stuttgart, Germany.