Written Before the Test: Why American Prescriptions Still Ignore Pharmacogenomic Evidence
The scenario is familiar to any clinical geneticist. A patient carries a variant that makes a standard dose of a common drug ineffective or dangerous. The evidence is unambiguous, the guideline is published, and the recommendation is a single sentence long. The prescription is written exactly as it would have been a decade ago, because nobody ordered the test.
Pharmacogenomics has spent twenty years being described as imminent. It has produced real clinical victories — abacavir and HLA-B57:01, carbamazepine and HLA-B15:02, fluoropyrimidines and DPYD, thiopurines and TPMT or NUDT15, clopidogrel and CYP2C19 — and a supporting infrastructure of peer-reviewed guidelines and regulatory labeling. What it has not produced is a routine place in American prescribing.
Evidence Is Not the Bottleneck
The science has moved well past the proof-of-concept stage. The Clinical Pharmacogenetics Implementation Consortium maintains dosing guidance covering dozens of gene-drug pairs, and the Food and Drug Administration publishes a table of pharmacogenetic associations that grows most years. Major health systems operate successful programs in oncology, cardiology, and psychiatry.
The gap is between those programs and everyone else. Most prescriptions in the United States are still written in settings where no pharmacogenomic result exists, and where ordering one would require a specialist referral, an out-of-pocket explanation, and a wait measured in days. For a drug started in an urgent care clinic or an outpatient visit lasting twelve minutes, that sequence is not realistic.
The Reimbursement Knot
Payers have been the most consequential gatekeepers, and their posture has been inconsistent. Some cover testing when a guideline recommendation is specific and the clinical scenario is narrow — a pretreatment HLA-B*57:01 test before abacavir is the archetype. Coverage thins considerably when the test is preemptive, meaning it is ordered before a drug is chosen, or when it panels multiple genes at once.
That asymmetry creates a predictable distortion. Testing flourishes where a single gene maps to a single drug and a bad outcome is expensive, and stalls where the value lies in avoiding a long sequence of trial-and-error prescribing. Health systems that run their own laboratories sometimes absorb the cost internally; independent practices cannot.
The economics may shift as panel prices continue to fall, but low prices do not automatically produce coverage decisions. Payers want to know who benefits and over what time horizon, and the evidence base for many gene-drug pairs is strongest for surrogate endpoints rather than for hard clinical outcomes. That is a research gap as much as a policy one.
Alert Fatigue and the Limits of the Electronic Record
Even where results exist, they often fail to reach the moment of prescribing. Results live in scanned PDFs, in specialty clinic notes, or in laboratory portals that the ordering clinician may never reopen. Efforts to embed pharmacogenomic data into electronic health records as structured, computable fields have advanced, but implementation varies by vendor and institution.
Then comes the paradox of success. When pharmacogenomic alerts are layered onto systems already saturated with drug-interaction and allergy warnings, clinicians learn to dismiss them. A well-designed alert fires rarely, explains itself, and offers an alternative. A poorly designed one fires constantly and becomes noise. The difference between the two is rarely a scientific question, and it is almost never funded as one.
Education, and the Confidence Gap
Clinicians report low confidence in interpreting pharmacogenomic results, and the reason is not mysterious. Most medical school curricula devote a few hours to the topic, residency training rarely includes it, and the results themselves are often reported in language built for specialists — activity scores, diplotype translations, and phenotype categories that require a legend to read.
Continuing education helps, but it cannot substitute for the most effective intervention available: a result that arrives already interpreted in the context of the drug being prescribed, with a recommended action.
Who Actually Gets Tested
Uptake is unevenly distributed. Testing concentrates in academic medical centers, in oncology and cardiology, in the Veterans Health Administration's integrated system, and among patients who can pay out of pocket or who have generous coverage. Patients in rural areas, in safety-net clinics, and in primary care settings where most prescriptions originate are least likely to be tested.
That pattern matters beyond fairness. If pharmacogenomic evidence is generated and validated predominantly in well-resourced populations, the dosing recommendations that emerge may not generalize. Allele frequencies differ across ancestry groups, and a variant that is rare in one population may be common in another. Equity in testing is therefore a scientific requirement, not only a distributive one.
What Would Actually Move the Prescription
Three changes would matter more than any new discovery.
First, coverage rules tied to guideline concordance rather than to narrow prior-authorization criteria. If a consortium recommends testing before a drug is started, that recommendation should be sufficient.
Second, results that live in the record as structured data and surface as decision support at the point of ordering, with a small number of high-specificity alerts.
Third, testing that happens before it is needed. Preemptive panels ordered once and consulted repeatedly convert pharmacogenomics from an emergency consult into a reference document, which is closer to how the rest of laboratory medicine works.
None of these is a breakthrough. They are implementation decisions, and implementation is where the field has been stuck. The test exists. The prescription has simply not caught up.