Population-scale CYP star-allele frequencies, translated through CPIC Activity Scores into a geographic drug-dosing guidance map
Pharmacogenomic dosing guidance is only as reliable as the population data underneath it. Before 2015, the overwhelming majority of pharmacogenomic discovery cohorts were of European ancestry, producing dosing algorithms with unknown or poor transferability elsewhere. The last decade of large-scale, ancestrally diverse biobank sequencing has begun to close that gap — but unevenly.
A dosing recommendation derived entirely from a European-ancestry discovery cohort can fail silently in other populations — not because the biology differs, but because the causal allele's frequency, or even its identity, differs by ancestry:
• CYP2D6*10 (reduced-function): near-absent in Europe/Africa (<2%) but present in 41–70% of East Asian chromosomes — a variant invisible to European-only panels becomes the dominant driver of metabolizer status in East Asia. • CYP2D6 gene duplication (increased-function, ultrarapid): 1–2% in Northern Europe, but 16–29% in Ethiopia and parts of Saudi Arabia — first characterized only after Ethiopian cohorts were sequenced in the 1990s–2000s. • CYP2C19*17 (gain-of-function): common in Northern Europe and the Horn of Africa (20–40%), rare in East Asia (<2%) — inverse of the *2/*3 loss-of-function pattern that dominates East Asian CYP2C19 variation.
Without population-representative sampling, a "one-size-fits-all" dosing table simply encodes the ancestry composition of whoever happened to be sequenced first.
Seven resources anchor the current reference panel:
• 1000 Genomes Project (IGSR): 2,504 genomes at 30× WGS, the historical backbone for population allele frequency, still the reference for the 5 super-population / 26 sub-population framework (AFR, AMR, EAS, EUR, SAS). • gnomAD v4 (Broad Institute): 730,947 exomes and genomes pooled from >20 studies, 8 major genetic ancestry groups, now the default frequency reference for clinical variant interpretation (ACMG/AMP). • Biobank Japan: 200,000 participants, hospital-recruited, dense East Asian pharmacogenomic phenotype linkage. • UK Biobank: 500,000 participants, predominantly European but with linked longitudinal prescribing and outcomes data enabling real-world dosing validation. • All of Us (NIH, USA): 245,000+ sequenced participants, explicitly designed so >50% self-identify with a group historically underrepresented in biomedical research. • H3Africa: 75,000+ genomes across 18 African countries — Africa harbors the deepest human genetic diversity on Earth yet was, until H3Africa, the least-sequenced continent per capita.
Ancestry is assigned computationally (ADMIXTURE, principal component analysis against reference panels) since self-reported ethnicity and genetic ancestry frequently diverge, especially in admixed populations of the Americas.
CYP genes are genotyped and reported using the PharmVar star-allele nomenclature, a standardized system that collapses a gene's full haplotype — every SNP, indel, and structural variant in phase — into a single named allele such as *4 or *17. CYP2D6 is the hardest of the clinically actionable pharmacogenes to call correctly, because its locus sits beside two non-functional pseudogenes it readily recombines with.
Each PharmVar star allele represents a specific, phased haplotype and is assigned a functional category and a numeric Activity Score (AS) by CPIC:
• No function (AS = 0): e.g. CYP2D6*4, *5 (whole-gene deletion), *6 — null alleles that abolish enzyme activity • Decreased function (AS = 0.25 or 0.5): e.g. *10, *17, *41 — reduced but non-zero catalytic activity • Normal function (AS = 1.0): e.g. *1, *2 — wild-type-equivalent activity • Increased function (AS = 2.0 per copy): gene duplications such as *1×2, *2×3, *35×2
A patient's diplotype score is the sum of the two allele scores — e.g. *1/*4 = 1.0 + 0 = 1.0 — and that sum is later translated into a metabolizer phenotype (Stage 4). Because AS assignments are periodically revised as new functional evidence accumulates (CPIC published a major CYP2D6 AS revision in 2019), phenotype calls made on the same raw genotype can change over time as the underlying science improves.
CYP2D6 sits on chromosome 22q13.2 flanked by two nonfunctional pseudogenes, CYP2D7 and CYP2D8, with >95% sequence identity. This creates three genotyping hazards that short-read arrays cannot resolve on their own:
• Gene deletion (*5): the entire CYP2D6 gene is absent; array-based SNP calls at that locus read as homozygous reference for a gene that isn't there • Gene duplication (*1×N): tandem copies multiply the Activity Score of whatever allele is duplicated — an individual carrying a duplicated *2 (AS 1.0×2 = 2.0) is Ultrarapid, while the same *2/*2 without duplication is Normal • CYP2D6/2D7 hybrid alleles: unequal crossover between CYP2D6 and its pseudogene neighbor creates chimeric genes with mixed function, frequently misread as standard alleles by short-read variant callers
PacBio HiFi long-read sequencing generates reads averaging ~15 kb — long enough to span the entire CYP2D6 locus in a single read — enabling direct phasing of the two parental haplotypes and unambiguous resolution of copy number and hybrid structure without statistical inference. Droplet digital PCR or MLPA is used as an orthogonal confirmation assay for detected copy-number variants before they are entered into the frequency database.
Once diplotypes are called across tens of thousands of individuals per population, the raw counts are converted into stable per-population allele and phenotype frequencies and rendered as a geographic choropleth. This stage is standard population genetics — Hardy-Weinberg checks, sample-size thresholds, and differentiation statistics — applied specifically to pharmacogenes.
Raw diplotype counts must clear several quality gates before entering the map:
• Hardy-Weinberg equilibrium (HWE) testing: within each population, observed genotype counts are compared to Hardy-Weinberg expectation; systematic deviation flags genotyping artifacts (e.g. undetected CYP2D6 CNVs skewing homozygote counts) rather than true biology • Minimum sample size: populations with n<50 are excluded from the published frequency map, since binomial sampling error at that size produces a confidence interval too wide (routinely >±10%) for clinical use; n≥400 is preferred for rare (<2%) alleles • Relatedness and cryptic admixture correction: PLINK/KING kinship estimation removes first- and second-degree relatives, and PCA-based admixture proportions are used to either stratify or explicitly model within-population ancestry heterogeneity, particularly important for recently admixed populations of the Americas
Frequencies are computed simply as allele count / (2 × n) per population, then aggregated into the diplotype and phenotype tables used downstream.
Wright's FST — a measure of how much genetic variance is explained by population membership — reaches 0.09 for CYP2D6 across the five 1000 Genomes super-populations, comparable to well-known highly differentiated loci like skin-pigmentation genes. Most of the genome shows FST well under 0.05; CYP2D6's elevated value reflects strong, geographically distinct selective or demographic history acting on a gene that also happens to control the metabolism of >20% of all prescribed drugs.
The resulting frequency surface is not random noise — it traces recognizable geographic gradients:
• CYP2D6*10 (decreased function): dominant across East Asia (41–70% allele frequency), rare elsewhere — the single largest driver of the unusually high Intermediate Metabolizer fraction (~45–48%) seen in East Asian populations for CYP2D6-metabolized drugs • CYP2D6*17 and *29 (decreased function): common across sub-Saharan Africa (up to 20–34% combined), largely absent outside Africa • CYP2D6 gene duplication (increased function): concentrated in the Horn of Africa and Arabian Peninsula — up to 29% of Ethiopian and 21% of Saudi Arabian chromosomes carry a duplication, plausibly linked to historical qat and khat alkaloid exposure hypotheses, versus 1–2% in Northern Europe • CYP2C19*2/*3 (loss of function): carried by 55–70% of East Asians (at least one allele) versus ~15–20% of Europeans, making CYP2C19 Poor Metabolizer status roughly 10× more common in East Asia • CYP3A5*1 (functional, expresser): the ancestral, functional allele is common in African ancestry populations (55–85% carry at least one copy) but the loss-of-function *3 allele has risen to >85–90% frequency in Europeans — the reverse of the naive assumption that "wild type" is universally common
A diplotype is a laboratory fact; a metabolizer phenotype is a clinical decision category. CPIC and the Dutch Pharmacogenetics Working Group (DPWG) jointly standardized, in 2016 and refined in 2019, the translation from summed Activity Score to one of five phenotype classes — Poor, Intermediate, Normal, Rapid, and Ultrarapid Metabolizer — that clinicians actually act on.
For CYP2D6, the consensus translation from summed diplotype Activity Score (AS) to phenotype is:
• AS = 0 → Poor Metabolizer (PM): no functional enzyme from either allele • 0 < AS ≤ 1.0 → Intermediate Metabolizer (IM): one reduced- or no-function allele paired with a normal- or reduced-function partner • AS = 1.0–1.5, both alleles ≥ normal → borderline Normal/Intermediate, resolved by specific diplotype lookup rather than score alone • 1.25 ≤ AS ≤ 2.25 → Normal Metabolizer (NM): the reference range containing most *1/*1, *1/*2, and similar diplotypes • AS > 2.25, achieved only via gene duplication of a functional allele → Ultrarapid Metabolizer (UM)
This is a lookup-table system, not a free calculation — CPIC publishes and periodically revises the full diplotype-to-phenotype table (currently version 3, 2019) precisely because edge cases (e.g. specific hybrid alleles) do not translate cleanly from AS arithmetic alone and require expert curation.
Phenotype misclassification, or failure to test at all, has directly measurable clinical consequences for prodrug and high-therapeutic-index-margin medications:
• Codeine (CYP2D6 prodrug, activated to morphine): PM patients (~7% of Europeans, up to ~1% of East Asians, but reported up to 20%+ in parts of North Africa) get little or no analgesia at standard dose — clinically indistinguishable from "the drug not working." UM patients rapidly over-convert codeine to morphine; the FDA issued a boxed warning in 2013 and contraindicated codeine in breastfeeding UM mothers after a 2006 case of neonatal opioid toxicity death. • Clopidogrel (CYP2C19 prodrug, antiplatelet): PM patients under-activate the drug, leaving them under-protected against stent thrombosis after percutaneous coronary intervention — clinically important because CYP2C19 PM/IM combined affects up to 55–65% of East Asian patients, versus roughly 20–25% of Europeans, driving region-specific antiplatelet drug selection guidelines in China, Japan, and South Korea. • Tacrolimus (CYP3A5 substrate, immunosuppressant): CYP3A5 expressers (common in African-ancestry patients) clear the drug faster and are systematically under-dosed by fixed mg/kg protocols originally calibrated on predominantly non-expresser European cohorts, risking transplant rejection from sub-therapeutic exposure.
The final step converts population science into an operational clinical decision support (CDS) system: a population-weighted dosing recommendation delivered inside the electronic health record at the moment a clinician places an order, informed by whichever phenotype frequencies apply to that patient's inferred or tested ancestry group.
Modern preemptive pharmacogenomic programs test broad panels once, before a prescribing need arises, and store the result for the patient's lifetime:
• CDS Hooks and SMART on FHIR standards let the EHR query a decision-support service in real time when a clinician orders a covered drug, returning a structured alert with the patient's specific phenotype and a CPIC-sourced dose recommendation, rather than a generic pop-up • St. Jude PG4KDS: pediatric preemptive panel covering >20,000 patients, testing for CYP2D6, CYP2C19, TPMT, DPYD and others at first hospital contact, banking results before any relevant prescription is ever written • Vanderbilt PREDICT and Mayo RIGHT protocols: among the earliest large-scale U.S. preemptive genotyping programs, both demonstrating that upfront genotyping is cheaper per actionable result than reactive single-gene testing after an adverse event • Population phenotype frequency tables (Stage 4) feed the pre-test probability used to prioritize which genes go on a given population's standard panel, and to set the expected yield when justifying program cost to a hospital formulary committee
The Ubiquitous Pharmacogenomics (U-PGx) PREPARE study (2023, Lancet), a prospective cluster-randomized trial across 7 European countries with 6,944 patients, found that patients receiving genotype-guided dosing had 30% fewer clinically relevant adverse drug reactions than those receiving standard-of-care dosing across 39 drug-gene pairs — one of the strongest prospective outcome datasets pharmacogenomics has produced to date.
Yet the evidence base underneath these systems remains unevenly built: a 2024 PharmGKB ancestry audit found approximately 82% of variant-outcome association data underlying current dosing guidelines derives from European-ancestry cohorts, despite Europeans comprising well under 20% of the global population. For genes like CYP2D6 and CYP3A5, where allele composition differs sharply by ancestry (Stage 3), a guideline's confidence interval is mechanically widest exactly where its clinical stakes are often highest — sub-Saharan African, Oceanian, and Indigenous American populations remain the least represented in the reference panels the CDS alerts are built on.
The clinical promise of population allele frequency mapping is real and measured — a 30% reduction in adverse drug reactions in a 6,944-patient European trial is not a modest effect. But that same map is a mirror of where sequencing effort has been spent: expanding H3Africa-scale, Oceanian, and Indigenous American reference cohorts is now the rate-limiting step for making population-informed dosing guidance equally trustworthy everywhere it is deployed, not just where the original discovery cohorts happened to be recruited.