A functional ancestry-linked regulatory haplotype influences CYP2D6 expression

Abstract

Cytochrome P450 (CYP) enzymes metabolize 70–80% of therapeutic agents, yet regulatory variants affecting drug metabolism remain poorly characterized across populations. Although numerous studies have examined coding variants in pharmacogenes, comprehensive analysis of regulatory variation across tissues and ancestries is lacking. Expression quantitative trait loci (eQTLs) for 54 CYP genes were analyzed using GTEx data, identifying CYP2D6 as harboring the most extensive regulatory variation, with 249 variants including two high-effect variants in strong linkage disequilibrium. In silico functional analysis predicts that one variant disrupts thousands of transcription factor binding sites in a region of high chromatin accessibility. Both variants exhibit population stratification, occurring at 45% frequency in East Asian populations versus 2.5% in Europeans, and show strong linkage with the reduced-function CYP2D6*10 allele. These results indicate that ancestry-specific regulatory variants should be incorporated into pharmacogenetic testing strategies.

Introduction

Cytochrome P450 (CYP) enzymes catalyze the metabolism of approximately 70–80% of therapeutic agents, rendering them critical determinants of drug efficacy and toxicity [1]. CYPs facilitate the biotransformation of diverse xenobiotics and endogenous compounds through oxidative reactions, directly influencing drug clearance, bioactivation, and the generation of reactive metabolites. Genetic polymorphisms in CYP genes, particularly CYP2D6, CYP2C9, and CYP3A4, have been incorporated into clinical pharmacogenomic guidelines and directly influence dosing recommendations for numerous medications, including warfarin, clopidogrel, and selective serotonin reuptake inhibitors [2, 3]. However, while protein-coding variants have been extensively characterized, the contribution of regulatory variation to CYP expression remains poorly understood, particularly across tissue types and populations.

Expression quantitative trait loci (eQTLs) are a major source of regulatory variation affecting gene expression and account for a substantial proportion of the heritability of complex traits [4]. These variants can modulate transcription through alterations in transcription factor binding, chromatin accessibility, or enhancer-promoter interactions, often in a tissue-specific manner [5]. Unlike coding variants, which typically affect protein function uniformly across tissues, regulatory variants can exhibit remarkable tissue specificity, activating or silencing gene expression depending on the cellular context and available transcriptional machinery [6, 7]. For drug-metabolizing enzymes expressed in multiple organs, including the liver, intestine, and kidney, understanding tissue-specific regulatory programs is essential for predicting drug disposition and individual treatment responses [8, 9].

Recent population genomic studies have revealed substantial differences in the frequencies of pharmacogenomic variants across ancestries; however, most regulatory variant discoveries have focused on European populations [10]. These biases limit understanding of how regulatory variation contributes to population-specific differences in drug metabolism capacity and may partially explain the observed disparities in drug response across ethnic groups. The well-documented differences in warfarin dosing requirements between Asian and European populations, for instance, extend beyond known coding variants, suggesting that uncharacterized regulatory mechanisms contribute to these clinical observations [11, 12]. The Genotype-Tissue Expression (GTEx) project, now in its tenth release with over 19,000 samples from 943 donors across 50 tissues, provides an unprecedented opportunity to systematically map tissue-specific eQTLs and begin addressing these critical gaps in pharmacogenomic knowledge [13].

To address these knowledge gaps, a comprehensive analysis of regulatory variations in cytochrome P450 genes using GTEx v10 data was conducted, identifying rare high-effect eQTLs and integrating functional genomics approaches, including ATAC-seq and transcription factor binding analysis, to distinguish causal variants from linked polymorphisms. Global distribution of these regulatory variants using gnomAD population data further revealed striking ancestry-specific frequency patterns that may contribute to known pharmacogenomic disparities in drug responses and adverse reactions across populations.

Methods and materials

GTEx data acquisition and processing

Expression quantitative trait loci (eQTL) data were obtained from the Genotype-Tissue Expression (GTEx) Project version 10 release (https://gtexportal.org/) [13]. Cis-eQTL summary statistics were downloaded with focus on liver tissue. All cis-eQTLs for cytochrome P450 (CYP) family genes were extracted by identifying all CYP gene identifiers using the HUGO Gene Nomenclature Committee (HGNC) symbols matching the pattern “CYP*.” Variants within 1 Mb of the TSS were extracted (nominal p < 0.05). Effect sizes were converted to absolute values representing fold changes (2^|β|). The final dataset comprised 4812 variant-gene pairs across 54 CYP family genes. For each eQTL, the following data were retained: variant ID, gene symbol, effect size (slope), standard error, nominal p-value, and allele frequency. Genetic variants were stratified by minor allele frequency (MAF) using population-specific thresholds: rare (MAF < 0.01), low-frequency (0.01 ≤ MAF < 0.05), and common (MAF ≥ 0.05). High-effect variants were defined as those with absolute effect sizes |β| > 0.5.

GWAS catalog phenotype association analysis

Phenotype associations for variants within 1 Mb of the CYP2D6 locus (chr22:41,125,000–43,125,000, GRCh38) were obtained from the NHGRI-EBI GWAS Catalog (accessed November 2024). Effect sizes were standardized: odds ratios were log-transformed to log(OR), beta coefficients were retained as reported. Standard errors were calculated from confidence intervals or p-values.

Linkage disequilibrium analysis with CYP2D6 star alleles

Linkage disequilibrium (LD) between rs5758628 and eight CYP2D6 star alleles (*2, *3, *4, *6, *9, *10, *17, *41) was calculated using 1000 Genomes Project Phase 3 data across East Asian (N = 504), European (N = 503), and African/African American (N = 661) populations [14]. Star allele definitions were obtained from PharmVar. Both R² (correlation-based LD) and D’ (normalized LD coefficient) metrics were calculated using LDlink version 5.x. The deletion variant rs56984362 could not be directly assessed due to technical limitations with multi-base indels in VCF reference panels; rs5758628 was used as a proxy based on allele frequency concordance. Haplotype frequencies were obtained from 2 × 2 contingency tables (n11, n12, n21, n22) provided by LDlink, representing the four haplotype combinations.

Transcription factor binding and chromatin accessibility analysis

To investigate potential regulatory mechanisms, transcription factor binding site (TFBS) analysis was performed using position weight matrices (PWMs) from the JASPAR 2022 CORE vertebrate non-redundant database, which contains 841 high-quality transcription factor binding profiles from ChIP-seq experiments [15]. Flanking sequences (±8 bp for SNVs, ±12 bp for deletions) were extracted. Position-specific scoring matrices (PSSMs) were calculated using pseudocounts to avoid log(0) errors. Binding affinity changes (Δ scores) were calculated as the difference between the alternative and reference allele scores. Changes with |Δ| > 2 were considered significant. For each transcription factor with multiple binding sites, the site with the maximum absolute Δ score was selected as the representative. To assess chromatin accessibility at variant sites, ATAC-seq data from human liver tissue were analyzed using Gene Expression Omnibus (GEO accession GSE164870, sample GSM5021286). For each variant, ATAC-seq signal data were extracted within ±15 kb for overview visualization and ±2 kb for detailed analysis.

Pharmacogenomic annotation

Variants were annotated using dbSNP build 155, gnomAD v3.1 [16, 17], and PharmGKB databases. The CYP2D6 gene region (chr22:42,052,000–42,220,000, GRCh38) was examined for regulatory features, including promoter (<2 kb from TSS), proximal (2–10 kb), and distal (>10 kb) regulatory elements.

Population frequency analysis

Population-specific allele frequencies were compared across major ancestry groups using the gnomAD database v3.1.

Large language model usage

Claude Sonnet 4 (Anthropic) was used to assist with code documentation, data visualization formatting, and data assembly. The LLM was not used for data analysis, statistical interpretation, or generation of scientific conclusions. All analytical methods, results interpretation, and scientific content were developed independently by the authors.

Statistical analysis

All analyses were performed using Python 3.8 with pandas (v1.3.3), NumPy (v1.21.2), SciPy (v1.7.1), and Matplotlib (v3.4.3). Effect sizes represent log₂ fold changes in gene expression per alternative allele, derived from GTEx linear regression models adjusted for population structure, sex, platform, and the top five principal components of the genotype. For GWAS associations, odds ratios were log-transformed prior to analysis. Spearman’s correlation was used to assess the relationship between allele frequency and effect size.

Results

Multi-tissue eQTL analysis of cytochrome P450 genes

Analysis of the GTEx v10 data revealed extensive regulatory variations in cytochrome P450 (CYP) genes across tissues. CYP1A2, CYP2D6, CYP2E1, CYP2C, CYP3A, and CYP4 subfamilies exhibited distinct expression patterns across the analyzed tissues (Fig. 1A). CYP21A2 showed 418 eQTLs in the liver and 570 in the intestine, with no eQTLs in the kidney. CYP2D6 exhibited 234 eQTLs in the liver, 385 in the kidney, and 317 in the intestine. CYP4F12 showed the highest liver eQTL count of 634, with no eQTLs in the kidney or intestine. Several genes, including CYP2C8, CYP2C18, and CYP27C1, showed tissue-specific eQTL patterns, with expression limited to one or two tissues (Fig. 1B). Liver tissue showed enrichment for variants with higher effect sizes in the 0.6–0.8 range, while kidney tissue was dominated by high effect eQTLs peaking around 0.8–1.0. Intestinal tissue showed a broader distribution, with effect sizes ranging from 0.2 to 1.4 across tissues (Fig. 1C). Variants with frequencies between 0.05–0.15 exhibited effect magnitudes exceeding 1.0, representing a 2-fold change in gene expression (Fig. 1D). The analysis identified effect sizes from near 0 to 1.75, with CYP2D6 variants showing enrichment at moderate (0.5–0.75) and high (>1.0) effect thresholds (Fig. 1E).

Fig. 1. Multi-tissue eQTL analysis reveals tissue-specific regulatory patterns in cytochrome P450 genes.

A Heatmap of z-score normalized expression for 54 CYP genes across 50 GTEx tissues. Hierarchical clustering using complete linkage with Euclidean distance. Drug-metabolizing CYP families indicated by color-coded bars. B eQTL count matrix showing significant cis-eQTLs (p < 0.05, ±1 Mb from TSS) for CYP genes across liver (n = 261), kidney (n = 103), and intestinal (n = 206) tissues. Numbers indicate eQTL counts per gene-tissue combination. C Violin plots showing distribution of absolute eQTL effect sizes ( | β | , log₂ fold change) across three tissues. Horizontal lines indicate median values. D Scatter plot of variant allele frequency versus eQTL effect magnitude (2^|β | ) for 4812 variant-gene pairs. E Histogram comparing effect size distributions for CYP2D6 variants (red, n = 249) versus other CYP genes (blue, n = 4563). Vertical dashed lines indicate high-effect ( | β | =0.5) and ultra-high-effect ( | β | =1.0) thresholds.

CYP2D6 regulatory variant landscape

The distribution of effect sizes across CYP2D6 and other major drug-metabolizing CYP genes showed that CYP2D6 (highlighted in orange) had the highest median effect size near 1.0, with maximum effects of 1.664. CYP2D6 contained 249 variants, substantially more than other analyzed genes, including CYP21A2, CYP4A, CYP2C, CYP3A, CYP4F, CYP2A, CYP1, CYP2B, CYP17, CYP11, CYP2E, and CYP27 (Fig. 2A). Regional association mapping on chromosome 22 revealed multiple variants reaching genome-wide significance (p < 5 × 10⁻⁸), with stars indicating novel variants absent from PharmGKB and ClinVar databases (Fig. 2B). High-effect variants spanned a frequency range from rare to common, with novel candidates highlighted (Fig. 2C). Effect sizes ranged from 0.2 to 1.6 across genomic regions categorized as proximal (near TSS), distal, or novel GWAS-significant (Fig. 2D). The genomic locus diagram shows CYP2D6 gene structure, neighboring genes, and the spatial distribution of regulatory eQTLs (Fig. 2E).

Fig. 2. CYP2D6 exhibits extensive regulatory variation.

A Violin plots comparing absolute effect size distributions ( | β | ) for major drug-metabolizing CYP genes in liver tissue. Horizontal lines indicate median and quartiles. CYP2D6 shown in orange. B Regional association plot of CYP2D6 locus (chr22:41,125,000–43,125,000) showing -log₁₀(p-value) for variants. Stars indicate genome-wide significant variants (p < 5×10⁻⁸) absent from PharmGKB and ClinVar databases. C Scatter plot of allele frequency versus effect size for CYP2D6 variants. Points colored by gradient scale. D Effect sizes ( | β | ) plotted against distance from transcription start site (TSS). Variants categorized as proximal, distal, or novel (stars). E Genomic locus diagram showing CYP2D6 gene structure and regulatory eQTL distribution by chromosomal position. Bar heights indicate effect magnitude.

Transcription factor binding disruption analysis

Analysis of transcription factor (TF) binding disruption revealed distinct patterns for the two CYP2D6 variants. The variant at chr22:42,177,950 (G > T) demonstrated alterations in the binding affinity for 4540 TF-binding sites, with 2353 sites showing significant changes (|Δ| > 2) (Fig. 3A). The deletion variant at chr22:42,196,754 (TGA > T) exhibited a more extensive disruption of TF-binding, affecting 9859 binding sites with 4779 significant changes. This variant exhibited a more pronounced asymmetric distribution. Specific changes in TF binding were quantified for each variant. For the chr22:42,177,950 variant, top gained binding included HOXA5 (+6.66), PRRX2 (+6.66), and HOXD3 (+6.66), whereas top lost binding included Sox6 (−6.66) and KLF15 (−6.66). The chr22:42,196,754 deletion showed more dramatic changes, with increased binding for PAX9 (+25.42), ZNF410 (+23.85), and REST (+23.11), and decreased binding for HOXD12::ELK1 (−28.06), ELK1::HOXB13 (−23.93), and ZNF528 (−23.46) (Fig. 3B). Expression of several transcription factors was observed in the liver (Supplemental Fig. 1A).

Fig. 3. High-effect CYP2D6 regulatory variants show predicted disruption of transcription factor binding and population stratification.

A Histograms showing predicted transcription factor (TF) binding site disruption for rs5758628 (chr22:42,177,950 G > T, top) and rs56984362 (chr22:42,196,754 TGA > T, bottom). Orange bars: gained binding sites; blue bars: lost binding sites. Analysis using 841 TF profiles from JASPAR 2022 CORE database. Significant changes defined as |Δ| > 2. B Waterfall plots ranking TFs by binding affinity changes (Δ scores). Top gains (positive) and losses (negative) shown for each variant. C Binding affinity curves for top three gained and lost TFs per variant, comparing reference (cyan) versus alternate (orange) allele sequences. D ATAC-seq signal from human liver tissue (GEO GSE164870) showing chromatin accessibility across ±15 kb windows centered on each variant position. E Detailed ATAC-seq signal profiles in ±2 kb windows. Signal values quantified from bedGraph data. F Population-specific allele frequencies from gnomAD for East Asian (EAS), European non-Finnish (EUR), African/African American (AFR), South Asian (SAS), and overall gnomAD.

TF binding affinity curves and chromatin accessibility

High-contrast TF-binding affinity curves demonstrated position-specific binding changes across variant regions. The G > T variant showed moderate changes while the deletion showed dramatic binding shifts (Fig. 3C). ATAC-seq analysis using publicly available liver tissue data (GSE164870) revealed differential chromatin accessibility at the two variant positions. The SNP at chr22:42,177,950 showed minimal accessibility, whereas the variant at chr22:42,196,754 demonstrated substantially higher signals and was marked with distinct accessibility peaks, with the functional TGA > T deletion occurring in a region of particularly high chromatin accessibility (Fig. 3D). The chr22:42,177,950 G > T variant showed a relatively low but consistent accessibility signal across surrounding regions. In contrast, the chr22:42,196,754 TGA > T functional variant displayed a prominent accessibility peak precisely at the variant position, with signal values of 231.5 at the deletion site (Fig. 3E).

Population frequency comparison

Allele frequencies revealed striking population patterns. The chr22:42,177,950 G > T variant (rs5758628) showed a dramatically elevated frequency in East Asian populations (45.0%) compared to European (non-Finnish) populations (2.5%). African/African American populations showed intermediate frequencies of 8.1%, South Asian populations 8.1%, and an overall gnomAD frequency of 5.9%. The chr22:42,196,754 TGA > T deletion (rs56984362) demonstrated identical population frequencies: East Asian 44.8%, European 2.5%, African/African American 11.4%, South Asian 8.2%, and overall 6.9%. Near-perfect concordance suggests linkage disequilibrium (LD). Both variants showed a consistent frequency of 2.87% in the GTEx study cohort, reflecting the predominantly European ancestry of GTEx donors and resulting in frequencies closely matching the European-specific gnomAD data rather than the global average (Fig. 3F).

Population-specific star allele linkage architecture

To evaluate the relationship between regulatory eQTLs and functional CYP2D6 coding variants, LD analysis was performed using 1000 Genomes Project data across East Asian (EAS), European (EUR), and African (AFR) populations. The deletion variant rs56984362 could not be directly assessed likely due to technical limitations in processing multi-base indels from reference panels; rs5758628 was used as a proxy for LD GWAS calculations. Regional association analysis of the CYP2D6 locus identified 576 variant-trait associations within 1 Mb, with the top genome-wide significant signals including rs199525628, rs143525991, and rs79681549 (Fig. 4A). Within 500 kb, Drug Metabolism (n = 24), Psychiatric (n = 23), and Cognitive (n = 14) phenotypes accounted for the most high-confidence associations, while the full 1 Mb window showed enrichment for Hematologic (n = 45), Psychiatric (n = 37), and Drug Metabolism (n = 28) traits (Fig. 4B). LD decay patterns revealed Drug Metabolism traits maintained mean across distances up to 805 kb, while Psychiatric and Pain/Opioid traits showed variable patterns (Fig. 4C).

Fig. 4. Linkage disequilibrium analysis reveals population-specific regulatory-coding haplotype architecture.

A Regional GWAS association plot for CYP2D6 locus ( ± 1 Mb) showing -log₁₀(p-value) for 576 variant-trait associations from NHGRI-EBI GWAS Catalog. Top genome-wide significant variants labeled. Significance threshold: p < 5 × 10⁻⁸. B Trait category counts within 500 kb (left) and 1 Mb (right) windows. Categories include Drug Metabolism, Psychiatric, Hematologic, Cognitive, and others. C LD decay patterns showing mean R² across distances for Drug Metabolism, Psychiatric, and Pain/Opioid trait categories. D Linkage disequilibrium heatmap showing R² values between rs5758628 and eight CYP2D6 star alleles (*2, *3, *4, *6, *9, *10, *17, *41) across East Asian, European, and African populations from 1000 Genomes (N = 504 EAS, N = 503 EUR, N = 661 AFR). E Heatmap showing D’ (normalized LD coefficient) values for same variant-allele pairs across three populations. F Bar plot of haplotype frequencies showing percentage of individuals carrying both regulatory variant and *10 (dark bars) or *10 alone (light bars) in each population.

The regulatory variant rs5758628 demonstrated population-specific LD with CYP2D6 star alleles (Fig. 4D). Strong LD with *10 was observed in EAS (R² = 0.672, D’ = 0.981) and AFR (R² = 0.637, D’ = 0.861) but not EUR (R² = 0.083, D’ = 0.946). *41 showed minimal LD (R2 < 0.04), while *2 showed moderate LD only in EAS (R2 = 0.147). No-function alleles (*3, *4, *6, *9) and increased-function *17 were absent or showed no measurable LD. The deletion variant rs56984362 displayed similar patterns, with *10 LD values of R² = 0.661 (EAS), R² = 0.079 (EUR), and R² = 0.443 (AFR). High |D’| values (0.86–0.98) across all populations indicated no historical recombination between rs5758628 and *10 (Fig. 4E). Haplotype frequency analysis demonstrated that 83.7% of East Asian *10 carriers also possessed the regulatory variant, compared to 10.8% in Europeans and 76.5% in Africans (Fig. 4F). In EAS (N = 1008), 47.8% carried both variants versus 9.3% with *10 alone. EUR (N = 1006) showed inverted frequencies: 2.2% both, 18.0% *10 alone. AFR (N = 1322) displayed intermediate patterns: 8.6% both, 2.6% *10 alone.

Discussion

This study advances the current understanding of pharmacogenomic regulation by providing a comprehensive analysis of regulatory variations across cytochrome P450 (CYP) genes using a tissue- and population-resolved framework. While CYP2D6 has long been recognized for its pharmacogenetic importance, the integration of eQTL architecture, transcription factor disruption, and population frequencies reveals an underappreciated regulatory complexity and highlights clinically relevant population stratification [18]. These findings provide mechanistic support for long-standing clinical observations and inform future directions for implementing precision medicine [19, 20].

Tissue-specific regulatory landscapes

The tissue specificity of CYP gene regulation has been well documented, particularly in the liver, which remains the central organ for drug metabolism [21]. However, recent studies leveraging multi-tissue expression datasets, such as GTEx, have begun to uncover how regulation varies across different tissues [22, 23]. This analysis corroborates and expands upon this literature, demonstrating that kidney tissue harbors a distinct regulatory landscape characterized by high magnitude eQTLs across a broad allele frequency spectrum. This is consistent with emerging findings that renal CYP isoforms, including CYP4A11 and CYP4F2, contribute to metabolite clearance and endogenous signaling (e.g., 20-HETE synthesis) rather than to canonical hepatic drug metabolism [24, 25]. The observed disparity in eQTL counts between the liver and non-hepatic tissues for genes such as CYP4F12 underscores that tissue-restricted transcriptional control is not merely a function of gene expression levels but reflects distinct enhancer-promoter configurations and transcription factor accessibility [26, 27]. This aligns with the growing evidence that the regulatory logic of drug-metabolizing enzymes is shaped by their functional roles in different organ systems, including xenobiotic defense, metabolism, and homeostasis [28].

Regulatory architecture beyond star alleles

While pharmacogenetic studies have long focused on star alleles and structural variants of CYP2D6, this analysis highlights the depth of regulatory variation not yet captured by clinical panels [29]. Prior functional studies have identified specific enhancer variants, such as rs5758550, located over 100 kb downstream, which significantly reduce CYP2D6 expression in hepatocyte models [30, 31]. The broader set of eQTLs uncovered here substantially extends this knowledge, suggesting that unrecognized regulatory elements may explain a portion of the “missing heritability” in CYP2D6 metabolism [32].

The bimodal effect size distribution observed mirrors that of previous studies that have demonstrated layered transcriptional control [33]. Recent ChIP-seq analyses have identified HNF4α as a key hepatic regulator of CYP2D6 expression that operates in conjunction with distal enhancer elements [34]. These insights support a broader reconceptualization of CYP2D6 as a gene whose pharmacogenetic profile is defined not only by coding variations but also by a complex modular regulatory architecture [30]. Such a framework is more consistent with the diverse metabolizer phenotypes observed clinically, which cannot be fully explained by the star alleles alone [35].

Molecular mechanisms of disruption and linkage architecture

Efforts to elucidate the predictive capacity of genetic regulatory mechanisms have been limited by challenges in connecting noncoding variations to functional consequences [36]. Transcription factor binding analysis predicts regulatory disruption by high-frequency non-coding variants in CYP2D6. The deletion at chr22:42,196,754 is predicted to affect thousands of TF binding sites, exceeding disruption by the G > T variant in a region of strong chromatin accessibility. These findings indicate that structural variants may exert larger regulatory effects than single nucleotide variants. Furthermore, this study aligns with ENCODE and Roadmap Epigenomics data, showing that pharmacogenetically relevant regulatory elements are enriched in accessible chromatin and coincide with long-range enhancer loops [37]. Disruption of the binding of repressive complexes, such as HOXD12::ELK1, suggests that deletion variants may relieve transcriptional repression, affecting CYP2D6 expression. This is supported by 3D chromatin conformation studies showing that CYP genes are regulated by distal enhancers, forming dynamic regulatory topologies [38].

LD analysis across populations revealed strong, population-specific linkage between these regulatory variants and the reduced-function CYP2D6*10 allele [39]. East Asians and Africans demonstrated both high correlation and absence of historical recombination between regulatory and coding variants, while Europeans showed weak correlation despite high normalized LD coefficients. This suggests a compensatory haplotype where the majority of East Asian *10 carriers possess the expression-enhancing regulatory variants, compared to a minority in Europeans. Integration with GWAS Catalog data revealed enrichment for drug metabolism, psychiatric, and hematologic phenotypes within the CYP2D6 locus. The co-localization of regulatory eQTLs with GWAS signals for clinically relevant traits provides evidence that these variants contribute to phenotypic variation beyond expression levels alone, supporting the hypothesis that regulatory variation contributes to unexplained heritability in drug response phenotypes [32]. These insights support haplotype-level interpretation of regulatory variation rather than single variant annotation.

Ancestry-specific regulatory variation

These findings contribute to the growing body of literature revealing population differences in the distribution of pharmacogenetically relevant regulatory variants [3]. The CYP2D6 regulatory haplotype evaluated here occurs at a frequency of 45% in East Asian populations but only 2.5% in Europeans, a nearly 20-fold difference with immediate clinical implications. Similar patterns were observed for other pharmacogenes. For example, CYP2C19*17, a gain-of-function variant, is prevalent in Europeans (21–25%) but rare in East Asians (<1%), whereas CYP3A53 displays the opposite pattern and is far more common in European than in African populations [40, 41].

These findings reinforce concerns regarding the Eurocentric bias of reference datasets, such as gnomAD, and underscore the importance of ancestry-aware variant interpretation [16, 17, 42].

A recent large-scale genomic analysis of 1136 pharmacogenomic variants using machine learning algorithms on 3714 individuals revealed distinct geographical patterns in drug toxicity risk, with Admixed Americans and Europeans showing higher risk proximity than East Asian populations [43]. Clinically, population-specific regulatory variations are associated with meaningful outcomes [44]. CYP2C19 loss-of-function carriers, for example, have more than twice the risk of cardiovascular events when treated with clopidogrel [45]. The absence of regulatory variants from pharmacogenetic algorithms may exacerbate these risks, particularly in non-European populations, where these variants are more prevalent [46]. These population differences in regulatory variant frequencies contribute to differential drug response patterns that current clinical testing strategies may not adequately capture. Current pharmacogenomic polygenic risk scores (PRS) rely predominantly on coding variants, potentially missing regulatory contributions. The 18-fold frequency difference for rs5758628/rs56984362 across ancestries suggests these variants could substantially impact PRS portability with CYP2D6*10 cross-ancestry prediction. Future PRS development should evaluate whether incorporating high-effect regulatory variants improves phenotype prediction in underrepresented populations.

Clinical implementation considerations

The translational importance of regulatory variations in pharmacogenetics is increasingly recognized; however, their integration into clinical practice remains limited [47, 48]. The *10 allele (rs1065852, P34S) reduces per-enzyme catalytic activity by approximately 50% and is classified as conferring intermediate metabolizer status [35]. However, in East Asian populations, 83.7% of *10 carriers also possess the expression-enhancing regulatory haplotype that increases mRNA levels 2.2-fold, potentially offsetting reduced enzymatic activity. In European populations, only 10.8% of *10 carriers have this compensatory regulation, meaning the reduced-function phenotype may operate largely unopposed.

Current clinical algorithms classify individuals based solely on star allele genotype, assigning activity scores that assume uniform expression across populations. East Asian *10/*10 individuals with the regulatory haplotype experience 4.8-fold increased expression, potentially shifting toward normal metabolizer status. This regulatory compensation is systematically absent from clinical testing panels, potentially explaining the well-documented variability in *10 phenotype penetrance across populations and the frequent discordance between predicted and measured metabolizer status [32, 35]. For drugs with narrow therapeutic indices and strong CYP2D6 dependence, including tamoxifen, codeine, and tricyclic antidepressants, misclassification of metabolizer status leads to therapeutic failure or toxicity [49]. Preemptive pharmacogenetic testing has demonstrated cost-effectiveness in 71% of evaluated scenarios but these analyses assume accurate phenotype prediction [50]. The impact of unmeasured regulatory variation on test sensitivity and specificity suggests that population-tailored panels incorporating both coding and regulatory variants would maximize clinical utility.

This study expands the known pharmacogenomic landscape by demonstrating that regulatory-coding haplotype architecture fundamentally alters the interpretation of CYP2D6 star alleles across populations. The strong linkage between expression-enhancing regulatory variants and the reduced-function *10 allele in East Asian and African ancestries creates a compensatory mechanism invisible to current genotyping strategies. The discordance between evolutionary history and predictive utility demonstrates that ancestry-aware testing approaches are essential rather than optional. The clinical implications highlight a need to revise pharmacogenetic implementation strategies to incorporate regulatory variation in a population-informed manner.

Author contributions

MM was responsible for designing the study protocol, conducting the data analysis, interpreting results, and writing the manuscript.

Data availability

All data used in this analysis were publicly available through the GTEx Portal, JASPAR database (https://jaspar.genereg.net/), and NCBI GEO (GSE164870) [50, 51]. The analysis code and processed datasets are available at 10.5281/zenodo.17468427.

Competing interests

MM is an employee of Takeda Development Center Americas, Inc. The author declares no other competing financial interests related to this work.

Ethics approval and consent to participation

No new animal or human data was generated in this study.

Supplementary information

The online version contains supplementary material available at 10.1038/s41397-026-00398-1.