Keeping pace with CYP2D6 haplotype discovery: innovative methods to assign function

Abstract

The discovery of haplotypes with unknown or uncertain function in the CYP2D6 pharmacogene is outpacing the capabilities of traditional in vitro and in vivo approaches to characterize their function. This challenge will undoubtedly grow as pharmacogenomic research becomes more inclusive of globally diverse populations. As accurate phenotypic assignment is paramount to the utility of pharmacogenomics, high-throughput technologies are needed for this complex pharmacogene. We describe the evolving landscape of innovative approaches to assign function to CYP2D6 haplotypes and possibilities for adopting these technologies into cohesive processes. Promising approaches include ADME-optimized prediction frameworks, machine learning algorithms, deep mutational scanning and phenoconversion predictions. Implementing these approaches will lead to improved personalization of treatment for patients.

Tweetable abstract

Wondering how #pgx can keep pace with the rapid discovery of CYP2D6 haplotypes? See this review of innovative methods to assign function to #CYP2D6 haplotypes. #pharmacogenomics #pharmacogenetics #precisionmedicine #pgxread @ericawoodahl @karenbrown_mt @umontana

The CYP2D6 gene is a complex pharmacogene that is highly polymorphic with a large number of single nucleotide variants (SNVs), insertions and deletions (indels) and structural variants identified. The clinical importance of predicting CYP2D6 phenotype cannot be overstated. CYP2D6 is responsible for metabolizing approximately 20% of drugs and interrogation of CYP2D6 diplotypes and accurate prediction of functional status is crucial for pharmacogenetic-guided drug therapy. Worldwide resources have attempted to define a valid approach to utilizing pharmacogenetic information in clinical practice. The Pharmacogenomics Knowledge Base [1,2] annotates dosing guidelines published by several groups, including the Clinical Pharmacogenetics Implementation Consortium (CPIC) [3,4] and the Dutch Pharmacogenetics Working Group (DPWG) [5]. CPIC alone has published guidelines on 18 CYP2D6 drug–gene pairs spanning several therapeutic areas, including attention-deficit and hyperactivity disorder, cancer, depression, nausea/vomiting and pain [6–14]. Clinical implementation of pharmacogenomics continues to increase worldwide, and it is imperative that pharmacogenetic tests accurately assign CYP2D6 diplotypes and activity predictions in order to convey dosing recommendations to healthcare providers and their patients.

There are nearly 150 CYP2D6 star alleles defined by the Pharmacogene Variation Consortium (PharmVar) and more haplotypes are curated every year [15,16]. Adding to the complexity, accurate assignment of these haplotypes is complicated by structural variation in CYP2D6 such as gene deletions, multiplications and gene rearrangements between CYP2D6 and the highly homogenous pseudogenes in the CYP2D locus (CYP2D7 and CYP2D8). Haplotypes of CYP2D6 are assigned function based on their alteration in CYP2D6 enzymatic activity or protein expression and given a value: no function (value of 0), decreased function (value of 0.25 or 0.5), normal function (value of 1) and increased function (multiplication of the value associated with a single haplotype by the number of gene copies) [17]. The activity score (AS) system translates genotype to phenotype based on the sum of the values assigned to each haplotype, assigning individuals to four phenotype classifications: poor metabolizer, intermediate metabolizer, normal metabolizer and ultrarapid metabolizer. Yet the phenotype prediction of CYP2D6 haplotypes is a dynamic landscape. The current AS system is the product of a collaboration between CPIC and DPWG to standardize genotype to phenotype translation, reduce inconsistency in phenotype predictions, and address the sometimes discordant guidance from CPIC and DPWG [17,18]. While these changes to the AS system were supported by expert consensus, logistical challenges remain a hurdle for clinical settings incorporating the updates into their practice. Additionally, the method of binning an individual's CYP2D6 activity into one of four metabolizer status categories – which has the benefit of ease of use and familiarity – may be problematic, as binning may result in a loss of phenotypic variability explained by CYP2D6 genotype when compared with using a continuous phenotype score that does not require binning. In the CPIC and DPWG process of standardizing genotype to phenotype, several experts advocated for a continuous scale presented as a percent of CYP2D6 activity for more precise activity assignments [17]. While this ‘percentage activity’ system was not adopted, it is plausible that it – or another system for assigning CYP2D6 function – could be adopted in the future, presenting a challenge for clinical laboratories and health systems to keep up with the pace of changing recommendations.

A number of genotyping platforms are used clinically to interrogate CYP2D6 haplotypes, yet they may fall short of accurately assigning haplotypes for a number of reasons. These include targeted genotype panels that cover only a limited set of known variants, nonspecific structural variation assignment leading to ambiguous diplotype calls (e.g., detection of a gene duplication, but an inability to discern which allele is duplicated), and false-genotype calls in diverse populations for which there is insufficient characterization of the CYP2D locus [19–21]. Another challenge in implementing clinical pharmacogenetics is the at times lack of transparency among commercial labs offering pharmacogenetic test results curated through proprietary star allele calling algorithms [22]. Expert analysis of the evidence utilized to support phenotype predictions is often impossible, presenting a challenge for clinical decision-making.

In recent years, there has been a rise in the use of next-generation sequencing and bioinformatic tools to interrogate CYP2D6 haplotypes. Next-generation sequencing captures the full range of CYP2D6 variation and has the benefit of identifying rare variants and variants whose function have yet to be described [23]. Coupled with bioinformatics tools, next-generation sequencing can also characterize CYP2D6 structural variation important in the prediction of CYP2D6 activity [24]. A limitation of these approaches; however, can be sample quality and DNA integrity that can complicate robust haplotyping. Several bioinformatics tools have been developed to assign CYP2D6 haplotypes from next-generation sequencing data including Aldy, Astrolabe (formerly Constellation), Cyrius, Stargazer, and PGx-POP [25–29] with each algorithm having unique strengths and limitations [30]. With these technological innovations comes the identification of a large number of CYP2D6 haplotypes with unknown or uncertain function. Novel combinations of CYP2D6 variation continue to be discovered at a rapid rate, as shown by a recent study of UK Biobank participants, which identified 159 unique CYP2D6 combinations in 6.1% of study subjects [29]. Considering the growing number of SNVs and structural variants discovered among relatively well-studied European-descendent populations, the challenge is only exacerbated as additional variation is discovered among diverse populations of non-European descent [31]. Further, observed differences in linkage disequilibrium among diverse populations may be problematic for star allele assignments that rely on linkage disequilibrium [32].

These challenges highlight the need for innovative approaches to assign function to CYP2D6 haplotypes if the benefit of pharmacogenetic testing is to be fully realized for all patients. Traditional methods of characterizing CYP2D6 enzymatic activity or protein expression through in vitro or in vivo studies are onerous. Relying on this approach to assign function to novel CYP2D6 haplotypes is not feasible considering the rapid identification of CYP2D6 variation. In this paper, we review innovative technologies to assign CYP2D6 phenotype, including an absorption, distribution, metabolism, elimination (ADME) optimized prediction framework, machine learning algorithms and deep mutational scanning (DMS). We also discuss the use of physiologically based pharmacokinetic (PBPK) models for evaluating phenoconversion with concomitant drug administration. These innovative methods will not only lead to more efficient pharmacogenetic research, but also serve to drive clinical implementation.

ADME-optimized prediction framework to predict deleterious CYP2D6 variants

Numerous computational tools have been developed to predict whether genetic variation influences protein function utilizing in silico functionality prediction methods, evolutionary conservation scores based on sequence alignment or ensemble scores [33]. Researchers evaluated the ability of 18 commonly used prediction algorithms in 43 pharmacogenes, including CYP2D6 and found that these algorithms performed poorly in predicting the function of deleterious variations, as defined by a twofold reduction in in vitro intrinsic clearance compared with wild-type [34]. This was an anticipated finding, as these algorithms generally provide better predictions of the functional consequences of variants in more highly conserved genes associated with disease rather than in poorly conserved pharmacogenes. Therefore, the authors empirically derived an ADME-optimized prediction framework based on a combination of scores from individual algorithms that produced a continuous prediction score. This framework resulted in improvement over current methods with overall >90% sensitivity and specificity for predicting no function and normal variants for 337 SNVs distributed across all pharmacogenes. Notably, the ADME-optimized prediction framework was better at distinguishing no function and normal variants for variants with minor allele frequency <1% compared with variants with minor allele frequency ≥1%. Specifically for CYP2D6, the framework correctly predicted function for 15/18 (∼83%) CYP2D6 variants, confirmed by in vitro functional data with N-desmethyltamoxifen, bufuralol or dextromethorphan as CYP2D6 substrates. While this study focused on missense SNVs, the ADME-optimized prediction framework may be able to assess an array of CYP2D6 variation discovered by next-generation sequencing, including indels and synonymous SNVs.

Machine learning algorithms to assign function to star alleles with limited experimental data

Supervised machine learning algorithms including linear regression, logistic regression, neural networks and support vector machines have become increasingly popular as tools for making phenotype predictions from genotype data [35]. In the field of pharmacogenomics, these algorithms are informed by known genotype–phenotype training data to iteratively learn an optimized function for mapping high-dimensional input genotype data to a predicted phenotype. Thus, machine learning algorithms offer an in silico method to predict CYP2D6 metabolic activity from the ever-increasing number of CYP2D6 haplotypes with unknown or uncertain function due to limited experimental data. Two groups have demonstrated the utility of supervised machine learning for predicting CYP2D6 metabolic activity using neural network models [36,37].

The first group developed Hubble.2D6 as a bioinformatics tool to assign function to CYP2D6 haplotypes as either normal, decreased or no function [36]. The neural network model underlying Hubble.2D6 was trained using both simulated CYP2D6 diplotype data and experimental CYP2D6 activity data in human liver microsomes with probe substrate, dextromethorphan [24]. Hubble.2D6 first analyzes a genotype matrix representing the full CYP2D6 sequence with annotations and then the model assigns two scores for each haplotype: probability that the given haplotype is a no function allele and probability that the haplotype is a normal function allele. The two probability scores are then combined to determine the final CYP2D6 metabolic activity – normal, decreased or no function. Hubble.2D6 does not include CYP2D6 structural variation and thus cannot predict increased function haplotypes. Hubble.2D6 demonstrated high predictive accuracy for star alleles with known function by correctly predicting function for 88% of 25 star alleles that were used for model validation. For 16 star alleles with unknown function, Hubble.2D6 explained 47.5% of the variability observed in in vitro activity using N-desmethyltamoxifen as the CYP2D6 substrate [38]. Hubble.2D6 is a promising method for predicting CYP2D6 metabolic activity that will likely improve as additional training data from in vitro experiments are utilized.

Another group utilized a neural network model to assign CYP2D6 function on a continuous scale rather than using the conventional AS system that relies on categorical phenotypes [37]. The model was trained using data from a prospective clinical study evaluating CYP2D6 genotype in female patients receiving tamoxifen for the treatment of breast cancer (CYPTAM) [39]. The model evaluates full CYP2D6 sequence and assigns a predicted ratio of tamoxifen metabolites, endoxifen to N-desmethyltamoxifen, as a measure of CYP2D6 activity. The model makes a prediction first by assigning scores to individual CYP2D6 variants and then aggregating those scores into maternal and paternal CYP2D6 haplotype scores to arrive at a predicted metabolic ratio phenotype. This model explained approximately 79% of variability in CYP2D6 activity in patients enrolled in the CYPTAM study, an improvement from the conventional AS system that explained approximately 66% of variability using AS assignments and approximately 54% of variability using categorical metabolizer phenotypes. These results were validated in independent tamoxifen and venlafaxine cohorts demonstrating improved CYP2D6 activity predictions across CYP2D6 substrates. Evaluating additional gene–drug pairs will further strengthen the predictions provided by neural network models in assigning CYP2D6 activity.

Both neural network approaches described here have similar limitations. Neither model included sequences in distal CYP2D6 regulatory elements that may contribute to CYP2D6 activity. Of note, however, PharmVar also excludes these regions in star allele definitions at this time. Additionally, these methods may not perform well for rare variants due to the limited availability of training data. Non-genetic contributors to interindividual variability such as lifestyle, comedication, disease states and epigenetics are also not incorporated into CYP2D6 activity predictions. These models included individuals largely of European ancestry and may not be generalizable to non-European populations. Despite these limitations, machine learning algorithms are a powerful tool for assigning function to CYP2D6 haplotypes with unknown or uncertain function and will undoubtedly provide improved predictions with additional studies.

DMS to measure phenotypic consequences of variation in pharmacogenes

DMS is another innovative tool that can be used to predict the function of CYP2D6 SNVs through high-throughput characterization [40]. While DMS has not yet been employed for CYP2D6 phenotype prediction, this approach has been used for prediction of variation in CYP2C9 and CYP2C19 [41,42]. Similar to CYP2D6, new haplotypes of these pharmacogenes are rapidly being discovered and at faster rates than can be studied using traditional in vitro functional assays. One group used a multiplexed DMS assay approach to predict the function of 8091 missense CYP2C9 variants, which represent 87% of the possible missense variants in the gene [41]. In parallel, they measured the activity of these CYP2C9 variants in yeast and the protein abundance in human cell lines to characterize the phenotypic consequences. The combination of the two assays estimated whether a decrease in enzyme function was due to a decrease in protein abundance versus a change in substrate binding. The authors found that almost two-thirds of the studied CYP2C9 variants led to decreased activity. The predictions from this DMS method largely agreed with the CPIC-assigned functional statuses for CYP2C9 variants [43,44]. Another group characterized the effect of 109 missense CYP2C9 variants on protein content relative to wild-type by calculating a protein abundance score and classifying variants as either ‘severely damaging’, ‘damaging’ or ‘tolerated’ [42]. The researchers identified 19 variants with expression levels <25% of wild-type, a level that may be clinically important. In the same study, the researchers also characterized the effect of 121 missense CYP2C19 variants and identified 36 variants with expression levels <25% of wild-type. Abundance scores were generally in agreement with independent western blot assays for both CYP2C9 and CYP2C19 protein content, validating the DMS results. Additional efforts have focused on using DMS to predict the function of variation in other pharmacogenes, including NUDT15, SLCO1B1, TPMT and VKORC1 [45–49]. While DMS can predict variants that alter expression and activity, it may not provide the resolution that traditional in vitro metabolic activity assays provide, such as the ability to determine intrinsic clearance. Additionally, the method can only be applied to missense variants in the open reading frame, excluding variants in the promoter and intronic regions of the gene. Nevertheless, there is an exciting opportunity to expand DMS approaches to predict function of CYP2D6 variants using a validated and high-throughput method.

Prediction of CYP2D6 function is complicated by phenoconversion

Along with the high degree of CYP2D6 genetic variation, another challenge in accurately predicting CYP2D6 function in an individual is potential phenoconversion – a metabolizer phenotype that is different from that predicted by an individual’s diplotype – due to a gene–drug–drug interaction. While the impacts of drug–interactions are important to consider, predicting the prevalence and degree of phenoconversion is difficult. The US FDA provides resources to classify the potency of CYP2D6 inhibitors (e.g., weak, moderate or strong) [50]; however, the authors of a scoping review identified several CYP2D6 inhibitors – celecoxib, cimetidine, desvenlafaxine, fluvoxamine and mirabegron – with data contradicting the FDA classifications and provided alternative recommendations [51]. Further, the authors argued that independently evaluating every potential drug–drug interaction is not feasible – and appropriate evaluation of every gene–drug–drug interactions is even more impractical – leading to the need for innovative methods of assessment. PBPK modeling offers one promising approach. A group of researchers used PBPK modeling to predict gene–drug–drug interactions when CYP2D6 substrates (dextromethorphan, tolterodine and risperidone) are coadministered with CYP2D6 inhibitors (duloxetine, paroxetine and fluoxetine) incorporating CYP2D6 diplotypes [52]. The results were then compared with observations from in vivo studies. While the PBPK model resulted in an accurate prediction of dextromethorphan–duloxetine interaction, the model underestimated time-dependent inhibitors, fluoxetine and paroxetine. With continued improvement in these models, PBPK modeling may become a valuable approach for evaluating the contribution of CYP2D6 phenoconversion and genetic variation.

Future perspective

Several emerging methods are currently available or in the pipeline to drive higher accuracy in predicting the function of CYP2D6 haplotypes and subsequent genotype–phenotype associations. Improved predictions will allow pharmacogenetics research and clinical implementation efforts to move forward as the rapid pace in discovery of CYP2D6 SNVs and structural variation continues. While we welcome the necessary shift toward improving inclusion of diverse populations in pharmacogenomic research, the resulting discovery of novel variants and haplotypes with unknown function will further complicate widespread implementation of pharmacogenetic testing into standard of practice. The number of commercial labs offering pharmacogenetic tests and reporting phenotype predictions with proprietary algorithms will continue to present a challenge for clinical decision-making, increasing the demand for accurate and transparent methods to assign function to complex pharmacogenes. The dynamic landscape of expert consensus recommendations may lead to a lack of congruency and the traditional AS system may require continued modifications.

The innovative methods described here may ideally be integrated into research and clinical practice in the future, providing a refined prediction of CYP2D6 function. Once accomplished, CYP2D6 DMS can allow for rare, novel and known genetic variation to be rapidly assessed for changes in protein function. Results from DMS can be synthesized with predictions from the ADME-optimized prediction framework and machine learning algorithms to improve agreement across in silico models, simultaneously predict star allele function and experimentally validate predictions at a rapid pace. Predictions from these methods can also be incorporated into PBPK models to evaluate gene–drug–drug interactions, improving comprehensive medication management through genetically informed, personalized treatment regimens. While several barriers to clinical implementation of pharmacogenetics remain, technologies that offer reliable identification of clinically actionable haplotypes are imperative for accurate interpretation of results. The promise of precision medicine can only be realized when high-throughput methods to accurately predict CYP2D6 haplotype function are validated and deployed.

Executive summary.

Challenges in assigning function to CYP2D6 haplotypes

• CYP2D6 is a complex and highly polymorphic pharmacogene and it is imperative to accurately predict CYP2D6 phenotypes to support adoption of pharmacogenetic testing.

• With increased use of next-generation sequencing, the discovery rate of CYP2D6 haplotypes with unknown or uncertain function is rapidly increasing.

Innovative methods to assign function

• Researchers utilized the ADME-optimized prediction framework to effectively predict deleterious effects of CYP2D6 variants.

• Machine learning models provide an effective tool for measuring the phenotype consequences of single nucleotide variants and researchers have successfully developed and trained algorithms to improve prediction of CYP2D6 activity.

• Researchers have successfully applied deep mutational scanning to predict the function of haplotypes in similarly complex pharmacogenes.

Improved models for evaluating phenoconversion

• Phenoconversion is an important consideration when accurately predicting CYP2D6 function.

• Physiologically-based pharmacokinetic modeling provides an efficient method to evaluate the clinical impact of gene–drug–drug interactions.

Summary & future directions

• This review outlines innovative methods to provide high-throughput characterization of CYP2D6 haplotypes with unknown or uncertain function.

• The use of these methods is a necessity in the field of pharmacogenomics as identification of novel variants continues to increase.

• Optimized processes to identify clinically actionable haplotypes will allow the promise of personalized medications to be fully realized.