Evaluating the frequency and the impact of pharmacogenetic alleles in an ancestrally diverse Biobank population

The vision of precision medicine involves using genomic information to guide preventive measures and tailor treatment for patients. A major component of this vision is pharmacogenomics (PGx), which aims to provide pharmacotherapy guidance for patients carrying genetic alleles impacting response to medications, with the goal of genetically tailored prescribing at the point of care. Many medications, such as opioid analgesics or the anticoagulant warfarin, require careful dose titration to achieve a therapeutic dose, and consideration of PGx information can help select the most appropriate drug at the right dose, avoiding harmful adverse drug reactions. The Clinical Pharmacogenetics Implementation Consortium (CPIC) has published guidelines for 71 drugs based on a large body of evidence showing the impact of PGx alleles which provide specific drug recommendations to alter prescribing practices. [1‐3] CPIC also assigns preliminary prioritization levels (called “provisional”) based on cursory review of evidence including, but not limited to, actionability in other professional society guidelines, recommendations from FDA-approved drug labels and assessment of evidence from PharmGKB clinical annotations. [4, 5] Level A refers to gene-drug pairs where genetic information should be used for prescribing decisions and alternative therapies, or dosing are highly likely to be effective and safe. [1] Level B refers to pairs where genetic information could be used to change prescribing and alternative therapies or dosing are extremely likely to be as effective and as safe as non-genetically based dosing. [1] As of June 2022, there are 96 gene-drug pairs with a CPIC level A or B guideline. Many of the alleles in these genes are common in the population and associated with a large effect size and a clearly identifiable drug response phenotype such as metabolizer status. However, one barrier to PGx implementation is the uncertainty about the true clinical burden of PGx alleles, particularly when accounting for ancestry-specific allele frequencies. For example the CYP2C19*2 no function allele (with rs4244285 as a key SNP) is twice as common in Asian ancestry individuals as those with European ancestry, which will have ethical and legal implications for testing prioritization [6] Understanding the prevalence of PGx alleles and the frequency of prescribing drugs impacted by these alleles in multiple ancestral populations will aid in targeting clinical implementation efforts for drug-gene pairs with the greatest impact.

Electronic Health Records (EHRs) are a digital record of a patient’s diagnosed conditions as well as information about the care they have received. EHRs have also become a robust source of real-world data that when linked to large DNA repositories can validate the association of known PGx alleles with drug response phenotypes but also further enable discovery of novel PGx associations. There have been several successful examples using the EHR to identify novel genetic association with drug side effects. [7‐9] The University of Pennsylvania Health System (also known as Penn Medicine) is one of the primary health care providers in Philadelphia and the surrounding region. Penn Medicine consists of over 3.3 million patients whose data are stored in an EHR since 2008. In our study period of 2012–2020, we counted 3,384,922 unique patients. Approximately 150,000 of these patients have enrolled in the Penn Medicine BioBank (PMBB) research program as of July 2022 and thus far 43,359 have been genotyped for research purposes on a genome-wide SNP array and sequenced using whole exome sequencing (WES); all participant genetic data is then linked to their EHR data. Prior studies have shown that  > 95% of individuals carry one or more alleles that interact with drugs from CPIC Level A guidelines [10‐12], but the number of patients that will actually be prescribed the drugs impacted by the alleles they carry is challenging to predict. In consideration of the implementation of preemptive PGx into clinical care, we sought to understand the population of Penn Medicine patients that are impacted by PGx alleles and are prescribed PGx affected medications. Considering this, we evaluated the prescribing patterns in the Penn Medicine EHR for a set of medications that are included in CPIC level A and B guidelines. Actionable PGx alleles are ones for which a medication prescribing change is recommended. Therefore, we also evaluated the frequency of actionable PGx phenotypes, the set of alleles that contribute to a clinically meaningful and observable drug response trait (e.g., drug metabolizer status), and are associated with strong or moderate therapeutic recommendations in the CPIC guidelines, in the genotyped PMBB participants. The large number of ancestrally diverse participants in the PMBB also allowed for ancestry stratified analyses to better understand how drug prescription practices based on an individual’s ancestry may influence the burden of unwanted drug response outcomes.

The identification of many clinically actionable alleles and their interactions with medications has led investigators to critically think about implementing PGx testing of these alleles into clinical practice. Our study examined approximately 3.3 million individuals from Penn Medicine and identified that 21.3% of patients have been prescribed at least one CPIC guideline medication and 9.3% were prescribed two or more. This evidence suggests that the clinical testing of pharmacogenetic alleles could help in informing patient care at Penn Medicine and other large healthcare systems. Our analyses on ~ 43,000 genotyped individuals from PMBB demonstrate that the frequency of participants with one or more clinically actionable alleles is very high. 100% of participants in our biobank population are carriers of at least one non-referent allele in a gene with a CPIC Level A or B guideline, out of which 98.9% of participants have a PGx phenotype that would lead to change in the prescription or dosage of at least one drug. More than 14% of patients were prescribed a drug for which they carry an actionable allele in our 8-year study period (2012–2020). These results suggest strong evidence for the benefits of integrating pharmacogenomic testing with current clinical practice for precision health care. Pharmacogenomic testing can be beneficial in reducing the number of drug adverse effects and nonresponse outcomes as well as minimize trial and error for dose specification. As more pharmacogenetic associations are discovered for more drugs, and as existing PGx-affected medications continue to grow in usage, the utility and clinical impact of pharmacogenomic infrastructure will only grow. At scale, these interventions can be a powerful and cost-effective method for improving patient care.

In this study, we aimed to observe the landscape of prescription trends of drugs that have a pharmacological impact in the Penn Medicine EHR to make necessary decisions about where and how to implement pharmacogenomics. Similar findings have also been reported by several other previous studies [11, 12, 16, 20] but to our knowledge, our study is unique in analyzing pharmacogenomic alleles among already genotyped individuals across multiple ancestry groups for identifying individuals that could benefit from pharmacotherapy. We evaluated the potential impact of PGx testing by retrospective analyses in our PMBB population and compared our findings to similar studies in one or multiple health systems. We identified different challenges in collecting and mining data from the EHR to observe these trends. Our analyses aimed at identifying individuals who could benefit from pharmacogenomic testing while highlighting essential considerations for clinicians and researchers for implementing pharmacogenomics in clinical practice. Genotyping individuals with a one-time panel of PGx genes would impact medication prescribing over the course of care. Among the most prescribed medication classes affected by PGx alleles include pain medications, anti-emetics, and proton pump inhibitors. These medications are often prescribed in the post-operative or acute care setting and represent a high-risk patient population that will benefit from implementation of pre-emptive PGx panel testing [21]. In one retrospective analysis in patients who underwent panel based PGx testing, the presence of a gene-drug interaction for CPIC guideline medications was associated with an increased risk of 90-day hospital readmission by more than 40% [22]. An ongoing prospective study will determine whether PGx panel testing is feasible in the acute care setting [23].

Our study examined the evidence based CPIC guidelines and applied them in a clinical setting to quantify their burden on actual patients. Based on the widespread prevalence of PGx alleles and PGx-affected medication, there would be substantial benefit to preemptive, universal PGx testing in both patient outcomes and healthcare costs. Warfarin is a prime example for a drug in widespread use where genetic information may inform dosing in a large proportion of patients. Prior knowledge of warfarin responsiveness can reduce trial and error in dose titration towards a safe and effective dose. Three randomized controlled trials of PGx guided warfarin dosing have been performed and one found PGx testing to increase the time in therapeutic range [24] and one of the studies showed an improvement in the composite clinical outcome of major bleeding, supratherapeutic International Normalized Ratio (INR), venous thromboembolism, or death [25]. While the use of reactive PGx testing may not be clinical feasible, the use of preemptive testing for PGx genes to guide cardiovascular medications including warfarin was found to be cost-effective [26]. In addition to dose-modifying effects, pharmacogenomic analyses have identified many gene-drug interactions such as clopidogrel where patient carrying an alternate allele is entirely indicated against taking a drug due to high risk of nonresponse leading to adverse outcomes. Nonresponse to clopidogrel treatment resulting in stroke or myocardial infarction is a severe outcome for patients that leads to dramatic economic costs but is preventable. Information on an individual patient’s drug metabolizer status can aid clinicians in making informed decisions regarding which drug regimen should be selected for patients based on their genetic information. Owing to the ancestry diversity of the PMBB, we were able to demonstrate that the burden of PGx alleles disproportionately affects certain ancestry groups. We report notable differences in phenotype frequency for many of the studied genes in each ancestry group. In aggregate, European ancestry individuals have significantly lower average counts of actionable alleles than individuals of African ancestry (p < 0.0001), as well as compared to all non-European ancestry combined. Furthermore, patients of Asian ancestry taking clopidogrel therapy have greatly increased rates of treatment modifying CYP2C19 reduced metabolizer phenotypes than individuals of European ancestry, quantifying a burden of PGx in actual patients. IM and PM phenotypes may result in increased risk of major adverse cardiovascular events in clopidogrel users, and therefore this demonstrates an unequal burden of PGx-related harms in clopidogrel patients of Asian ancestry. We found similar trends of differential PGx burden across ancestry groups for other widely used drugs including tacrolimus and voriconazole. Despite the observation of ancestry-specific patterns in PGx, we strongly discourage the use of ancestry as a proxy for PGx testing. In the case of clopidogrel, although a majority of clopidogrel-prescribed patients of Asian ancestry have actionable PGx alleles in CYP2C19, 42.3% still do not. Furthermore, in clopidogrel-prescribed patients of European ancestry, a still-notable 27.0% of patients have actionable CYP2C19 alleles. These findings demonstrate how inaction towards the implementation of pharmacogenetic testing harms all patients, but disproportionately harms patients of non-European ancestry.

Although our study highlighted a large proportion of pharmacogenomic alleles with CPIC guidance in the PMBB population, several limitations should be noted. We identified variations based on an integrated call set of genotype chip and whole exome data in VCF format, which does not capture PGx-relevant copy number or structural variations (in CYP2D6) and is also not well suited to calling HLA polymorphisms. As a result, we used presumptive CYP2D6 calls from PharmCAT which must be interpreted carefully. Another limitation is that although we were able to collect data on the prescribed medications during encounters at Penn Medicine, we are unable to confirm if the prescription is new or was prescribed at another clinical care setting. Thus, identification of patients with adverse reactions to drugs with CPIC Level A or B guidelines from the EHR has also been challenging. To address this, future work could involve developing algorithms for identifying previously administered drugs by mining through patient notes and EHR fields, and mining through longitudinal EHR data, allergy fields, and notes for adverse reactions and nonresponse to drugs. This would allow us to further quantify the burden of PGx as a measure of actual health outcomes. It should also be noted that although the integrated call set allowed us to genotype and average of 66% of PGx sites included in CPIC Level A or B guidelines, the missing sites in some cases limit our ability to call phenotypes resulted in some patients having missing or ambiguous calls for certain genes. Targeted PGx panels or whole-genome sequencing could largely avoid this issue, but targeted panels would eventually become obsolete as new alleles are discovered. Our study is unique in that it specifically explores how ancestry-specific patterns in PGx allele frequencies are reflected in patients receiving drugs for which they have a gene-drug interaction. Although we have more than 500 patients in every studied ancestry group which is enough to confidently demonstrate differences in phenotype frequencies, the relatively small proportion of patients who have been prescribed each drug limits our power to draw conclusions about ancestry-specific drug-gene interactions. As more patients are enrolled and genotyped by PMBB, and other data becomes available in resources like All of Us, we will gain more power to disentangle these ancestry-specific associations, which carry major implications for the effect of PGx across different ancestry groups [27]. It is important to consider relevant PGx alleles in all ancestry groups as more health systems consider clinical implementation of PGx; broad, diverse inclusion of multiple ancestry groups in clinical implementation has the potential to reduce health disparities. Lastly, the CPIC guidelines are easily accessible, but implementation of these guidelines poses several challenges in a clinical setting. One important challenge is the ability to return timely genotyping results prior to prescription of the drug. Based on our findings, it seems appropriate to perform genetic testing on patients preemptively and universally before they are prescribed a relevant PGx-informed drug. However, this poses many challenges which include but are not limited to cost effectiveness, insurance coverage, future discovery of new actionable genes and alleles, integration of real time clinical decision support, availability of resources, patient-prescriber education, and review of genetic reports among others.

With the increasing availability of clinical decision support tools like PharmCAT, it will soon be possible to add pharmacogenetic information to EHR systems such as Epic. With proper education and resources, along with increased adoption of PGx testing, healthcare professionals will be able to utilize this information to make genetics-informed drug dosing and prescription. Addressing these barriers will be imperative before widespread adoption of clinical PGx can be achieved. Implementation of PGx into clinical workflows has the potential to affect clinical care for a very large proportion of the health care population and is a significant contributor to fully realizing the practice of precision medicine.