The Outlier in All of Us: Why Implementing Pharmacogenomics Could Matter for Everyone

Abstract

The field of pharmacogenomics originally emerged in the 1950s from observations that a few rare individuals had unexpected, severe reactions to drugs(1). As recently as just six years ago, prominent views on the subject had largely remained unchanged, with authors from the FDA citing the purpose of pharmacogenetics as “tailoring treatment for the outliers”(2). It should not be surprising if this is the prevailing view – the best-studied pharmacogenomic drug examples are indeed just that, genetic explanations of extreme responses or susceptibilities among usually a very small fraction of the human population. Thiopurine methyltransferase (TPMT) deficiency as a cause of severe myelosuppression upon treatment with azathioprine or mercaptopurine is found as a heterozygous trait in only approximately 10% of patients, and homozygous (deficiency) carriers are even more rare – occurring in fewer than 1 in 300 patients(3, 4). Malignant hyperthermia resulting from inhaled anesthetics and succinylcholine is believed to have a genetic incidence of only about 1 in 2000 people(5).

The above statement about ‘outlier therapy’ was made in the context of discussing warfarin pharmacogenetics, and since then, the picture for warfarin has become even more complicated(6) with publication of the first two large prospective randomized trials examining the impact of pharmacogenomic results on therapeutic parameters for warfarin(7, 8). A common perception of these conflicting data around warfarin was one of a major blow to pharmacogenomics in general. Despite the fact that warfarin’s largest randomized evaluation fizzled – the poster-child was supposed to support the launch of mainstream pharmacogenomics into a widely-accepted orbit – the engineering work for bigger and better pharmacogenomic exploits has continued unabated, both on the discovery side and the implementation front. Pharmacogenomic associations have now been discovered (and replicated multiple times) for widely-used drugs like simvastatin(9), clopidogrel(10), and codeine(11). Positive pharmacogenomic studies have additionally been performed in tens of thousands of patients for other common drugs(12). Even drugs which could replace warfarin, like dabigatran, now have intriguing, compelling pharmacogenomic markers which might identify for prescribers those patients who would preferentially benefit from dabigatran over warfarin (the pharmacogenomic data on dabigatran suggest that a specific genetic subset of patients have significantly lower bleeding risk – without loss of efficacy - if treated with dabigatran rather than warfarin)(13).

Seeing the potential value of moving such findings to the clinic, a growing number of institutions (and commercial entities) are now offering clinical pharmacogenomic testing to individualize dosing of select drugs. In the U.S., Vanderbilt University has reported enrollment of >10,000 patients in their PREDICT (Pharmacogenomic Resource for Enhanced Decisions in Care and Treatment) program(14). St. Jude’s Children’s Research Hospital has enrolled >1500 patients in their PG4KDS effort(15). The University of Florida has implemented select pharmacogenomic clinical testing within their center (16). Indiana University has been funded to deploy a large testing program (with cost of care and outcomes evaluations compared to a control group of non-genotyped patients) across the state of Indiana (clinicaltrials.gov identifier NCT02297126). Multiple other examples exist(17–20). Pharmacogenetic screening (pre-testing for the HLA-B*57:01 allele) in patients being considered for abacavir therapy is recommended by multiple agencies and groups including the FDA (21) based on prospective, randomized data which demonstrate a reduction in hypersensitivity reactions with screening(22). At least one intriguing prospective study has already demonstrated a benefit to panel-based pharmacogenomic testing to guide prescribing among psychiatric patients compared to non-genomically guided prescribing(23). On a global scale, there is evidentiary support in Thailand for widespread pharmacogenomic ‘screening’ of patients before administration of carbamazepine(24), and efforts are underway in the Netherlands to attempt to integrate pharmacogenetic dose recommendations into computerized systems for drug prescribing and dispensing surveillance(25). At our institution, we have now enrolled >1000 patients into our clinical implementation program, with pharmacogenomic results for these patients having had the potential to impact prescription decision-making at approximately 3000 outpatient visits. We have found significant interest in pharmacogenomic testing among our patients, high applicability to the common drugs being prescribed, and very favorable responses by our physicians with respect to use and usefulness for decision-making(26).

It seems quite apparent that, despite a desire by some that multiple component utility and value analyses of each drug/gene pharmacogenetic pair be conducted before implementation(27), that implementation is going forward anyway(28, 29)– without such data. Indeed, it has been argued that a scientific standard of non-inferiority best applies to pharmacogenomic implementation(30)– that all that must be proven for initial implementation in clinical populations is that pharmacogenomic information must not add harm. Of concern, the pharmacogenomic algorithm which was applied in the aforementioned prospective warfarin trial appeared to in fact be detrimental to dose prediction in African Americans(8). Others have speculated that misinterpretation of pharmacogenomic results could be potentially harmful, for example, if a physician learns that a patient with significant cardiovascular disease carries the homozygous genotype for simvastatin myopathy risk (a risk which is associated with a <20% incidence of myopathy) and then avoids altogether treatment with a statin, that patient could be at an even higher risk for adverse clinical outcomes (like myocardial infarction) than if the statin had been used. These examples illustrate the fact that implementation must be performed thoughtfully, applied in populations supported by the discovery data and with appropriate decision-support so that genetic results are not considered in a vacuum but rather as part of the entire personalized medicine decision for a given patient.

As implementation efforts proceed, it will be important to measure clinical outcomes carefully, and implementation efforts must be iterative and adaptable in the sense that they must respond to new information and to feedback by those affected by implementation (prescribers and patients). As an example, at our institution, we stopped delivering warfarin pharmacogenomic results for self-identified African Americans within a week of the potential safety signal being reported in the prospective trial. We have also found that physicians are indeed ‘learning pharmacogenomics on the fly’ – meaning, physicians at our center are reporting significant increases in their understanding of pharmacogenomics merely by receiving and considering real results during the care of their patients. This should not be surprising, as physicians have a natural learning curve with any new medical innovation, be it a new drug or a new test. Many of the physicians within our institutional implementation program tell us they find it exciting to be considering such results for the first time, even without any formal training in pharmacogenomics, and that their approach is the same as with any other innovation – they evaluate the pharmacogenomic recommendation and its evidence and decide whether or not it applies to their patient. Clinical decision-supports (which prescribers will use as their immediate implementation guides) must therefore be carefully developed and written by experts in pharmacogenomic data evaluation. The work of groups like PharmGKB(31) and the Clinical Pharmacogenetics Implementation Consortium(32) has set important precedents in this regard.

Pharmacoeconomic considerations abound with respect to implementation. On the positive side, as genotyping costs continue to fall, it becomes challenging to imagine a near-future health care scenario in which pharmacogenomic testing would not be cost-effective—but of course rigorous studies must be performed. Some analyses are already suggesting cost-effectiveness and a possible financial benefit to the health care system for select pharmacogenomic testing (33–37), and this is despite the fact that such studies have mostly used single drug/gene examples as the test cases. Implementation of broad comprehensive preemptive pharmacogenomic testing for many markers at once would likely stand to have an even greater potential cost-savings upside, over the lifetime of prescribing for any given patient. Even now, a broad pharmacogenomic test panel can be run in a Clinical Laboratory Improvement Amendments (CLIA) setting for an order of magnitude less cost than a single CT scan, and, unlike the CT scan, the pharmacogenomic results can be used over and over again for the rest of the patient’s life. In an emerging environment of bundled care considerations, one wonders whether pharmacogenomics could potentially have a lot to offer.

But the latter issue begets the thorny question of insurance reimbursement for pharmacogenomic testing, which currently is unsupported by most organizations for the large majority of germline markers since most markers – even if well-established – do not meet criteria of the payors for medical necessity. Preemptive “pharmacogenomic screening” testing across multiple genes at once seems not even on the reimbursement radar –with the Centers for Medicare and Medicaid Services focused on evaluations of one gene/one drug at a time (www.cms.gov) - (even if the preemptive approach potentially makes the most economic sense given the ability to perform elective batch testing at a fraction of the cost of genotyping of a single gene prompted by the need to prescribe a single drug. The determination against coverage typically centers on a requirement for prospective randomized data demonstrating benefit. We argue that, not only is this standard pragmatically problematic for genomic implementation in general(38), but coverage considerations must appropriately consider the composite value of all elements in a possible pharmacogenomic (preemptive testing) screening model. Such pharmacogenomic screening could arguably be conducted in everyone at birth (to inform a lifetime of prescribing), or, in a cost-conscious system, could be reserved for those individuals identified as most likely to be prescribed medications in the near future (e.g., by age, or by type or number of comorbidities).

Before closing, we want to challenge the notion that we started with – that pharmacogenomics is a field of study about outliers. We recently analyzed the genotype information of >800 patients who have been tested across our broad pharmacogenomic (CLIA laboratory) test panel in our institutional pharmacogenomic implementation program. The panel contains germline markers clinically associated with response or toxicity risk for >40 different medications. For each patient genotype, we translate the result into a ‘risk signal’ for any drugs impacted by that genotype(26). Genotypes that increase toxicity (or non-response) probability are assigned a cautionary yellow-light signal – these are typically heterozygote risk genotypes. Highest risk genotypes (which are more rare, and are typically homozygotes having the risk allele) are assigned red lights. Genotypes that are favorable are assigned drug-specific green lights (low relative risk of toxicity or high probability of achieving intended effect). The results of this large cohort analysis show perhaps surprising results (Figure 1) which illustrate the potential power of preemptive, broad (composite) pharmacogenomic genotyping. First, 100% of patients carry a genetic ‘cautionary’ risk genotype which could impact toxicity or response outcomes for at least 1 of the >40 drugs analyzed. Furthermore, 85% of all patients have 7 or more cautionary drugs in their risk profile, and a staggering 44% of patients have more than 10 such cautionary drugs (Fig. 1A). If we restrict our analysis to only the highest risk red-light genotypes (typically these patients are homozygotes for the risk allele for whom the at-risk drug likely should be avoided altogether), 42% of all patients have at least 1 high-risk drug among the 9 high potential risk drugs that we analyzed, and 26% of all patients have at least 2 such high-risk drugs (Fig. 1B). These numbers hardly seem like outlier statistics. Two other groups recently reported analogous findings via slightly different analysis approaches considering an even fewer number of drugs across their respective institutional genotyped populations(14, 39), underscoring the potential applicability and relevance of pharmacogenomic testing to almost all people, not just a few ‘outliers’.

Figure 1.

Prevalence of pharmacogenomically ‘at-risk’ drugs in a large population of genotyped patients. (A) The percentage of patients having ‘potentially increased risk’ genotype-drug interactions that could increase toxicity (or non-response) probability if the drug was prescribed are shown, derived from actual genotype data for >800 patients that were genotyped across a broad preemptive pharmacogenomic panel. The number of ‘potentially increased risk’ drugs (out of 40 analyzed drugs all having clinically actionable known pharmacogenomic associations) that would be impacted by the genotype results for each individual are indicated across the x-axis. The faint yellow shading in each bar represents the exact percentage of patients with that number of affected “increased risk” drugs (e.g. 14.3% of patients have between 4-6 such at-risk medications), while the dark yellow shading in each bar represents the cumulative percentage of patients having at least that many affected drugs (e.g., 99.3% of patients have at least 4 affected increased risk medications). The analyzed drugs with potential increased risk associations included amlodipine, aspirin, atenolol, atorvastatin, azathioprine, benazepril, budesonide, candesartan, carvedilol, clopidogrel, colestipol, doxorubicin, duloxetine, felodipine, fenofibrate, fluticasone propionate, fluvastatin, hydralazine, hydrochlorothiazide, irbesartan, irinotecan, isosorbide dinitrate, lansoprazole, mercaptopurine, metformin, metoprolol, montelukast, nifedipine, omeprazole, pantoprazole, pegylated interferon, perindopril, pravastatin, ribavirin, rosuvastatin, sildenafil, simvastatin, sunitinib, triamcinolone, and warfarin. Note that the percentage of affected individuals would further increase as additional pharmacogenomically actionable drug/gene pairs (e.g., codeine/CYP2D6, and others) are added to the preemptively assessed information. (B) Prevalence results for high-risk genotype-drug interactions, from the same population. These highest risk genotypes (typically homozygote patients having two copies of the risk allele) often mean that the at-risk drug likely should be avoided altogether in such a patient. The faint red shading in each bar represents the exact percentage of patients with that number of affected “high risk” drugs (e.g., 16.4% of patients have exactly 1 high risk medication), while the dark red shading in each bar represents the cumulative percentage of patients having at least that many affected drugs (e.g., 42.0% of patients have at least 1 affected high risk medication). The analyzed drugs with potential high risk associations included azathioprine, mercaptopurine, clopidogrel, warfarin, 5-fluorouracil, capecitabine, omeprazole, lansoprazole, and verapamil.

When considered this way, pharmacogenomics is not just about finding the uncommon non-responder or the rare ‘extreme sensitivity’ patient in our practices – it potentially impacts all of our patients, and, as a human population, all of us. Osler seemed to understand this more than 100 years ago when he said, “one of the first duties of the physician is to educate the masses not to take medicine”(40), but of course, he also likely knew that that advice was not practical. Perhaps now, in an era of emerging and ever-widening pharmacogenomic implementation, we can do better – by finding the outlier within each of us. Perhaps, Osler might argue, it is in fact our imperative to do so.