Massive UK Biobank reference panel sharpens genetic imputation worldwide

Researchers have built a haplotype reference panel from the whole-genome sequences of 490,319 UK Biobank participants and shown that it outperforms the most widely used imputation panel in the world for most human populations. The new resource, described in Nature Genetics, delivers lower phasing switch error rates than the TOPMed r3 panel, which contains 133,597 individuals, for all global superpopulations except American ancestry groups and for 25 of 26 subpopulations from the 1000 Genomes Project.

Genotype imputation is the computational backbone of modern genome-wide association studies. The richer and more diverse the reference panel, the more accurate the imputed genotypes become, particularly for variants that are rare in the population being studied.

The new panel was constructed from the GraphTyper population-level whole-genome sequencing variants released for 500,000 UK Biobank participants, a dataset comprising 151,561 packed VCF files with an average size of 9.96 gigabytes and a total volume of 1,509.29 terabytes. To handle this enormous resource, the team wrote a custom C tool called vcffilter that extracted variant positions, alleles, filter flags and genotypes while running on the UK Biobank Research Analysis Platform. The compressed output shrank to an average of 27.86 megabytes per file, or 4,123.70 gigabytes in total, small enough to download to the researchers’ own computing resources for further processing.

Phasing, the process of assigning each variant to one of the two parental chromosomes, was carried out with SHAPEIT5 in two stages. First, the team phased common variants with a minor allele frequency of at least 0.001 in chunks of roughly 20 to 30 megabases per chromosome, with about two megabases of overlap, and then ligated the separately phased chunks into a continuous scaffold. Second, they used the phase_rare tool to phase rare variants with a minor allele count of at least 4, anchoring them to the common-variant scaffold. Chromosome X required special handling because male samples are encoded as diploid in the UKB GraphTyper data even though they must be haploid in the non-pseudoautosomal region, so the pseudoautosomal regions 1 and 2 and the non-pseudoautosomal region were phased separately and the ploidy was fixed afterward with standard BCFtools plugins.

The resulting panel was converted into a custom Qref format that loads far faster during imputation than conventional .bcf or .vcf.gz files. The researchers also produced a second, filtered version restricted to variants with a minor allele frequency of at least 0.0001. Their benchmarks showed that variants with a relative population minor allele frequency below 0.0001 had mean estimated R2 values far below the 0.6 threshold typically used to filter imputed variants before association testing, meaning their contribution to a typical common-variant GWAS is small. Dropping these ultra-rare variants therefore allows much faster imputation without considerable loss of quality, while the complete panel remains available for analyses that specifically target rare variation.

Benchmarking against the TOPMed r3 panel revealed consistent advantages across the globe. Phasing switch error rates were lower for all superpopulations except those with American ancestry and for 25 of the 26 1000 Genomes Project subpopulations tested. Imputation accuracy, measured as mean absolute error, improved for four of five superpopulations and for twelve subpopulations for variants with minor allele frequencies down to 0.01 percent. Squared correlation, another standard accuracy metric, improved across all five superpopulations and 22 subpopulations, with the largest gains observed in East and South Asian populations. The benchmark datasets were built from 1000 Genomes Project data reduced to the variants on the Illumina Global Screening Array chip, the platform commonly used in large-scale genetic studies.

To demonstrate the panel’s practical value, the team reimputed genome-wide association study datasets for severe COVID-19 from Italy, Germany, Spain and Norway that had originally been analyzed with older reference panels. For the Italian dataset, comprising 835 cases with severe respiratory failure and 1,255 controls, the researchers reran the association analysis using their rpMAF-filtered panel with the EagleImp software, the BIGwas pipeline and PLINK2, applying ten principal components as covariates and filtering summary statistics at an estimated R2 of 0.6 and a minor allele frequency of 0.01 for comparability with the original study. The reimputation showed that common-variant imputation remains unsaturated, meaning that even well-studied common variants can still yield new discoveries when a better reference panel is used.

That demonstration extended to fine-mapping, the statistical effort to narrow a list of associated variants down to the likely causal one. SuSiE’s susie_rss method calculated posterior inclusion probabilities for each single-nucleotide polymorphism at the locus, assuming one credible set of variants, and generated the minimum set of variants containing the causal variant with at least 95 percent certainty. Improved imputation quality sharpens these probabilities, and the researchers reported that the new panel enhanced the discovery of genome-wide significant loci and improved fine-mapping resolution in the reimputed datasets.

Making the panel usable at scale required more than statistical work. The team implemented EagleImp-RAP as a DNAnexus applet on the UK Biobank Research Analysis Platform, compiling EagleImp, htslib and BCFtools for the platform’s target environment and adding a precompiled version of Intel’s libtbb library. The applet handles argument parsing, retrieval and indexing of input files, downloads the selected reference panel from a predefined location within the platform, and returns all output files to the user’s project space after imputation. Chromosome X inputs are split into their three segments, processed with the appropriate reference panel for each, and merged into a unified output file. The standalone EagleImp software is freely available under the GPL version 3 license, as are the vcffilter tools and the applet source code.

The work was conducted under UK Biobank application 139525, with imputation analyses performed exclusively within secure compute environments approved for that project. The COVID-19 GWAS benchmark datasets remain available only on request to the COVID-19 GWAS Group, while the 1000 Genomes benchmark datasets are openly downloadable. Because the panel is anchored in UK Biobank genomes, its strengths are greatest where UK Biobank ancestry representation is strongest, and the authors’ own benchmarks show the American-ancestry superpopulation as the exception to the phasing advantage over TOPMed. Even so, the consistent gains across four of five superpopulations, the improved resolution for rare variants down to 0.01 percent minor allele frequency, and the demonstration that common-variant discovery is not yet exhausted together mark a substantial step forward for a field whose accuracy depends on how well a reference panel mirrors the world’s genetic diversity.

Subject of Research: Construction and benchmarking of a haplotype reference panel from 490,319 UK Biobank genomes for improved genotype imputation across global populations

Article Title: A haplotype reference panel constructed from 490,319 UK Biobank genomes improves genotype imputation for global populations

Article References: Wienbrandt, L., Prieß, C., Schmöhl, M., Neff, V., Gretsova, M., Torres, G., Wacker, E. M., Franke, A., & Ellinghaus, D. (2026). A haplotype reference panel constructed from 490,319 UK Biobank genomes improves genotype imputation for global populations. Nature Genetics. https://doi.org/10.1038/s41588-026-02819-0

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02819-0

Keywords: genotype imputation, haplotype reference panel, UK Biobank, genome-wide association studies, phasing, TOPMed, 1000 Genomes Project, rare variants, fine-mapping, COVID-19 GWAS, EagleImp, population genetics