The researchers benchmarked star-allele calls, the standardized haplotype labels used in pharmacogenetics, from three data sources against an analytical reference derived from whole-genome sequencing using the tool Aldy. The comparison covered 33 pharmacogenes in 1,450 UK Biobank participants drawn from four ancestry groups: 1,000 of European ancestry, and 150 each of East Asian, African, and South Asian ancestry, sampled in a stratified way to balance representation. Whole-exome sequencing alone or imputed array data alone reached at least 95 percent diplotype concordance with the whole-genome reference for 25 of the 33 genes on at least one platform, and the integrated call set matched that same 25-gene benchmark.

Integration proved more than the sum of its parts for a subset of genes. The combined call set improved or maintained concordance relative to the better-performing single platform for 29 of the 33 genes. The largest incremental gains among already high-performing loci appeared for CYP1A1, CYP3A4, and DPYD, where integration lifted concordance by 1.3 to 3.3 percentage points over the better single platform, a pattern the authors attribute to complementary recovery of non-coding and rare haplotype-defining variants. CYP3A4 climbed from 92.3 percent with exome data and 98.6 percent with the array to 99.9 percent integrated, and DPYD rose from 95.7 and 94.8 percent to 99.0 percent.

The logic behind these gains lies in the complementary blind spots of each platform. Whole-exome sequencing delivers high-resolution coverage of coding regions but routinely misses the non-coding regulatory and intronic variants that define many star alleles. Imputed genotyping arrays, which in UK Biobank span roughly 820,000 markers imputed against the Haplotype Reference Consortium, capture genome-wide common variation at low cost but carry no read-level information such as sequencing depth or structural-variant signals. The study illustrated the divide starkly: for VKORC1, whose key variants are non-coding, exome data managed only 12.6 percent concordance while the array reached 98.8 percent; for coding-driven loci such as CYP2C9, CYP3A7, TPMT, and NUDT15, exome data scored 98 to 100 percent. To make the approach reproducible and scalable, the team built PGxIntegrator, an open-source workflow that harmonizes exome and array data, performs coordinate liftover from GRCh37 to GRCh38, merges quality-controlled variant records, and runs parallelized Aldy-based star-allele calling across CPU cores.

Notably, the pipeline does not statistically impute missing genotypes or swap individual exome genotypes for array calls; when the same normalized variant record exists in both sources, the exome record is retained, and array records are added only where the exome call set lacks them. The authors note that generative artificial intelligence tools were used in a limited way to polish and refactor author-written scripts, with all design decisions, code changes, and outputs reviewed and validated by the authors. Not every gene cooperated. Five genes fell below 60 percent concordance even with integration: CYP2B6 at 47.4 percent, CYP2D6 at 51.8 percent, CYP2F1 at 40.1 percent, GSTM1 at 18.2 percent, and UGT1A1 at 44.2 percent. For CYP2B6, the culprit was a missing variant: rs2279343, which distinguishes the functionally important *6 allele from *9 and *4 from *1, was present in the source exome file but excluded because 31.86 percent of genotypes were missing, exceeding the study’s 5 percent threshold, and it was absent from the array and integrated call sets. As a result, whole-genome-defined *6 diplotypes were frequently represented as *9 and *4 as *1, showing that integration cannot compensate for a haplotype-defining variant that never enters the analysis.

CYP2D6, a gene marked by complex structural variation, resisted all non-WGS platforms evaluated. Among 1,446 participants with a callable whole-genome reference diplotype, 344, or 23.8 percent, carried at least one explicit structural indicator in the Aldy output, such as hybrid or conversion-associated alleles, duplications, or the *5 full-gene deletion notation. For none of these 344 participants did any non-WGS platform reproduce the structural configuration reported by the whole-genome-based analysis. Among the 1,102 participants without explicit structural notation, concordance was higher but still modest, at 68.1 percent for exome, 71.7 percent for array, and 68.0 percent for integrated calls. The authors stress a crucial caveat: the whole-genome reference came from a single computational caller, Aldy, and was not orthogonally validated, so these figures measure agreement with that reference rather than clinical typing accuracy. Under a stricter rule counting all multi-solution calls as discordant, exome concordance for CYP2D6 fell from 51.9 to 43.2 percent, while array and integrated estimates were essentially unchanged.

Ancestry shaped performance in revealing ways. CYP2C19 showed uniformly high concordance, at 97.9 percent or above, for array and integrated calls across all ancestry groups, but exome performance lagged in European, African, and South Asian participants, at 60.1 to 67.3 percent, because exome sequencing cannot detect the common non-coding promoter variant behind the *17 allele. East Asian participants, in whom *17 is rare, showed much higher exome concordance at 96.0 percent. Similarly, SLCO1B1 exome concordance dropped to 14.1 percent in African and 21.0 percent in South Asian participants while array performance stayed at 98.0 percent or higher everywhere, and CYP2D6 concordance ranged widely across platforms and populations, lowest in East Asian and highest in South Asian participants.

Beyond headline percentages, the team dissected discordance with a “union-of-FAIL” framework, isolating individuals whose call from at least one non-WGS platform mismatched the reference and tracing how diplotypes shifted across platforms using Sankey diagrams. Integration yielded its largest net gains over exome data for VKORC1, rescuing 1,253 previously discordant calls, followed by UGT2B7, SLCO1B1, IFNL3, CYP2E1, and NAT1, and over array data for CYP2B6, DPYD, CYP2A6, and CYP2C9. But integration occasionally hurt: CYP2D6 lost 41 calls relative to the array alone, GSTP1 lost 22, and SLCO1B1 slipped slightly, from 98.1 percent with the array to 97.2 percent integrated, suggesting that at loci where one platform is already near-perfect, merging mainly compensates for the weaker source rather than surpassing it. For UGT1A1, integration left concordance unchanged at 44.2 percent, with whole-genome-defined configurations involving *28, *60, *80, and *93 represented by simpler diplotypes in the non-WGS calls, indicating limited resolution of compound heterozygous configurations.

The practical message is gene-dependent rather than universal. For loci such as VKORC1, SLCO1B1, CYP3A4, and DPYD, array-derived or integrated calls reached concordance high enough to support large-scale retrospective pharmacogenetic research where whole-genome sequencing is unavailable. But the authors caution that their any-match concordance rule, which counts a call as concordant if any of several equally likely candidate diplotypes matches the reference, is appropriate for benchmarking analytical performance, not clinical implementation, since ambiguous diplotypes can correspond to different predicted drug-metabolism phenotypes. They add that loci with persistent diplotype ambiguity or complex variant architectures may require orthogonal confirmation or specialized pharmacogenetic assays before clinical implementation, and that the evaluated VCF-based workflows should not substitute for validated copy-number- and structure-aware CYP2D6 testing when clinical phenotype assignment is required. The study also carries cohort-specific limits: non-European sample sizes were small, results reflect UK Biobank’s particular sequencing and imputation pipelines, and evolving star-allele definitions may shift interpretations over time. For now, the work offers researchers a practical gene-level map of when the data they already have is good enough, and when deeper or structure-aware sequencing approaches remain necessary.

Subject of Research: Benchmarking exome, array, and integrated call sets against whole-genome sequencing for pharmacogenetic star-allele typing in UK Biobank

Article Title: Integrating whole-exome and genotyping array data for pharmacogenetic typing: a comparative evaluation with whole-genome sequencing in UK Biobank

Article References: Lee, C., Jeong, S., Lee, J.-Y., Lee, J., Park, J. H., Sung, J., Lee, S., & Kho, M. (2026). Integrating whole-exome and genotyping array data for pharmacogenetic typing: a comparative evaluation with whole-genome sequencing in UK Biobank. Frontiers in Pharmacology, 17, Article 1892708. https://doi.org/10.3389/fphar.2026.1892708

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1892708

Keywords: pharmacogenetics, pharmacogenomics, whole-exome sequencing, genotyping arrays, whole-genome sequencing, UK Biobank, star alleles, CYP2D6, Aldy, PGxIntegrator, precision medicine, diplotype concordance