When to Build a Population-Specific Imputation Reference Panel: Design, Sample Selection, and Validation

In modern population genomics, evolutionary biology, and complex disease genetics, genotype imputation is a widely used computational approach for increasing variant density in SNP microarray or low-pass whole-genome sequencing (lpWGS) datasets. Large public reference resources such as TOPMed and the Haplotype Reference Consortium (HRC) provide broad haplotype coverage, but global reference resources remain uneven in their representation of many regional and underrepresented populations. When the target cohort is genetically distant from the populations represented in a public panel, imputation accuracy can decline, particularly for low-frequency and rare variants (MAF < 5%).

For population-genetics consortia, regional biobanks, and project managers, this raises a fundamental strategic question: When does a study justify building a dedicated, population-specific imputation reference panel, and how should it be engineered?

Recent empirical studies show that genetic proximity can outweigh sheer reference sample size for some underrepresented populations, especially when the study focuses on low-frequency and rare variants. In appropriate cohorts, a smaller population-matched reference panel can outperform a substantially larger but less closely matched public panel for selected frequency ranges and downstream analyses. This guide details the design, sample selection, sequencing depth optimization, phasing, and validation framework used to evaluate and construct a population-specific imputation reference panel.

TL;DR

• Population Match Can Outweigh Sheer Size: A population-matched reference panel can outperform a substantially larger public panel for selected low-frequency and rare variants when the target cohort is underrepresented in the public reference.
• Recognize When Public Panels Underperform: Consider custom panel construction when public-panel imputation shows persistently weak quality metrics, poor empirical concordance, or marked population-specific allele-frequency and haplotype mismatches. Thresholds should be defined for the study rather than applied universally.
• Optimize Sample Selection via PCA: Sample across the breadth of population structure, representing relevant geographic and genetic subgroups while managing cryptic relatedness according to a pre-defined kinship threshold.
• Balance High and Low Depth WGS: Mixed-depth designs can combine a high-depth core with additional lower-depth genomes to expand haplotype representation within a fixed sequencing budget; the optimal allocation is cohort- and project-dependent.
• Enforce Modern Phasing Standards: Assemble panels using advanced phasing algorithms like SHAPEIT5 or Beagle 5.4, validating panel accuracy with empirical leave-out WGS benchmarking (r², IQS, and NRD).

The Reference Panel Dilemma: Mega-Scale Public Panels vs Custom Population-Specific Panels

Genotype imputation operates on the principle of haplotype matching: unobserved genotypes in a target study sample are statistically inferred by identifying shared or closely matching haplotypes represented within a densely characterized reference panel.

1.1 Why Public Panels Can Underperform in Underrepresented Populations

When target cohorts originate from East Asian, Southeast Asian, African, Indigenous American, Middle Eastern, or Oceanian populations, public panels present three major vulnerabilities:

• Haplotype Sparsity in Distant Populations: Even very large public panels may contain limited representation of specific regional populations. When closely related haplotypes are sparse, imputation of population-specific low-frequency and rare alleles can be less accurate (Cengnata et al., 2024).
• Missing Population-Specific Variants: Population-specific alleles that are absent from a reference panel cannot be recovered from that panel through standard genotype imputation. This limitation becomes increasingly important as analyses move toward lower-frequency variants.
• Quality-Metric Calibration in Multi-Ancestry Panels: Estimated imputation-quality metrics can be upwardly biased relative to empirical dosage accuracy for ancestry groups that are less well represented in a multi-ancestry panel. This makes empirical validation especially important when relying on marginal-quality variants (Shi et al., 2024).

1.2 The "Build vs. Borrow" Decision Matrix

Project managers and consortia should evaluate the following objective criteria when deciding whether to construct a dedicated reference panel:

When evaluating sequencing scaffolds or assessing baseline array technologies, review our guide on SNP Arrays vs Low-Pass and Deep WGS in Population Genomics.

Figure 2. Strategic PCA-guided donor selection capturing maximal ancestral diversity across regional populations.

Cohort Sampling and Genetic Diversity: Selecting the Ideal Reference Anchor Individuals

The ultimate accuracy of a custom reference panel is governed by the genetic diversity and representativeness of its constituent samples. Selecting biased, geographically clustered, or cryptically related individuals severely degrades the panel's generalizability across the broader target population.

2.1 PCA-Guided Stratified Sampling

Rather than randomly selecting available biobank samples, project leaders must apply stratified sampling across genetic principal component (PC) space:

• Capture the Breadth of PC Space: Select individuals that represent the major clusters, gradients, and less common regions of the population's PC space. Sampling should reflect the genetic and geographic structure relevant to the study rather than assuming that social, linguistic, or administrative categories map directly to genetic groups.
• Avoid Geographic or Sampling Imbalance: Biobanks can over-represent readily accessible recruitment sites. Where scientifically justified and ethically appropriate, donor selection should cover regional groups and genetic substructure that are relevant to the target cohort (Yang et al., 2025).
• Leverage Multi-Dimensional Clustering: Integrating Population Structure Analysis Service and PCA QC for GWAS: Outlier & Stratification Detection Guide can help identify unintended ancestry imbalance, outliers, and population-structure mismatch relative to the intended panel design.

2.2 Relatedness Filtering and Family Trios

• Cryptic Relatedness Pruning: For standard population reference panels, identify related sample pairs with tools such as KING or PLINK 2.0 and exclude or down-weight relatives according to the project design. In KING-style interpretation, kinship values below approximately 0.044 are commonly treated as more distant than third-degree relationships, while higher values indicate progressively closer relationships.
• Strategic Inclusion of Parent-Offspring Trios (When Feasible): While unrelated individuals maximize population representation, a subset of parent-offspring trios can provide transmission-informed phasing and an empirical benchmark for switch-error assessment. The number of trios should be selected according to panel size, sample availability, and validation goals.

2.3 Reference Donor Selection Checklist

Before committing biospecimens to deep whole-genome sequencing, every candidate donor must pass the following verification criteria:

Sequencing Strategy and Depth Allocation: High-Depth vs. Low-Depth Blends

A common misconception in reference panel construction is that every donor must be sequenced at the same very high depth. Empirical work in Japanese population reference panels shows that mixed-depth designs can expand haplotype representation efficiently within a fixed sequencing budget, although the optimal balance depends on cohort composition and analysis goals (Flanagan et al., 2024).

3.1 The Power of Mixed-Depth Reference Architectures

In the Biobank Japan benchmark (Flanagan et al., 2024), researchers evaluated four population-specific reference panel configurations against TOPMed:

• 1KG augmented with 1,037 Japanese samples at 30×: Improved population representation relative to 1KGP alone, although performance gains varied across allele-frequency bins.
• 1KG augmented with 3,256 Japanese samples at 15–30×: Improved imputation performance for low-frequency variants in the Japanese target population compared with less closely matched public resources.
• 1KG augmented with 4,216 Japanese samples at approximately 3×: Showed that a larger low-pass contribution can capture useful population-specific haplotypes when supported by appropriate variant calling and phasing.
• 1KG augmented with 7,472 Japanese samples across mixed depths: Produced the strongest overall performance among the evaluated population-specific configurations, with reported gains over TOPMed for selected imputation metrics and downstream association analyses.

Key Architectural Takeaway: When project funds are constrained, a mixed-depth design that combines a high-depth core with a larger set of medium- or low-depth genomes can be evaluated as an alternative to sequencing every donor deeply. The optimal allocation should be tested for the target population rather than copied directly from another cohort. For sequencing support, see Whole Genome Re-sequencing for Population Genetics.

3.2 Library Chemistry: PCR-Free Illumina vs Long-Read Supplements

• PCR-Free Short-Read WGS: PCR-free library preparation reduces PCR-associated amplification bias and can improve coverage uniformity, although GC content, DNA quality, library construction, and sequencing platform can still influence difficult regions.
• Long-Read Integration: Long-read sequencing can improve discovery and haplotype representation of structural variants, segmental duplications, and difficult repeat regions. If long-read genomes are incorporated into a reference-panel strategy, any downstream gain in structural-variant imputation should be evaluated empirically for the target cohort.

For high-level institutional infrastructure and long-term cohort scaling, see our overview on Biobank Sequencing Strategy: From Array to WGS.

Figure 3. Conceptual resource-allocation comparison between uniform high-depth sequencing and mixed-depth population reference-panel strategies.

High-Accuracy Phasing and Panel Assembly Workflows

Once high-quality variant calls (SNPs and short InDels) are generated and filtered, the unphased diploid genotypes must be computationally phased into discrete, chromosome-length maternal and paternal haplotypes. Phasing errors directly introduce chimeric switch errors that degrade downstream imputation accuracy.

4.1 Modern Phasing Algorithms: SHAPEIT5 vs Beagle 5.4

• SHAPEIT5: A phasing framework designed for large sequencing datasets (Hofmeister et al., 2023). It uses separate workflows for common and rare variants and was shown to phase very rare variants with low switch-error rates in large UK Biobank sequencing datasets; observed error rates depend strongly on allele frequency, sample size, and data quality.
• Beagle 5.4: Highly efficient hidden Markov model phasing that natively handles unphased reference VCFs, providing rapid multi-threaded processing.
• Genetic Recombination Maps: Use an appropriate recombination or genetic map when required by the selected phasing workflow, ideally one that is suitable for the target population and genome build. Linkage-disequilibrium patterns can provide complementary population context through our Linkage Disequilibrium Analysis Service, but an LD analysis is not itself a substitute for a validated meiotic recombination map.

4.2 Handling Complex Genomic Loci

Specialized chromosomal regions require dedicated pre-processing:

• The MHC / HLA Region (Chr 6p21): The human leukocyte antigen locus harbors extreme polymorphism and long-range linkage disequilibrium. Phase the MHC locus as an isolated high-density chunk with customized genetic distance parameters.
• Sex Chromosomes (Chr X and Y): Males must be phased as haploid in the non-pseudoautosomal regions (non-PAR) and females as diploid. Pseudoautosomal regions (PAR1 and PAR2) must be split and processed as autosomes.

Rigorous Panel Validation & Empirical Benchmarking

A newly assembled reference panel must be empirically validated before release to ensure it delivers genuine accuracy gains over existing public resources.

5.1 Validation Methodologies

• Masked WGS Leave-Out Validation: Set aside a representative high-depth WGS test subset that is not included in the reference panel. Mask or down-sample markers to emulate the intended target dataset, impute with the custom and comparator panels, and compare imputed dosages against observed WGS genotypes across frequency bins and genomic regions. The number of held-out samples should be selected according to cohort size and validation objectives (Mauleekoonphairoj et al., 2023).
• Key Quantitative Validation Metrics:
  • Concordance r²: The squared Pearson correlation between imputed dosage and true diploid genotype (0, 1, 2).
  • Imputation Quality Score (IQS): An accuracy metric based on true/false positive/negative rates that corrects for chance agreement on rare alleles.
  • Non-Reference Discordance Rate (NRD): The proportion of discordant non-reference genotypes or alleles, interpreted by allele-frequency bin and the validation design rather than against a single universal threshold.
  • Rare Variant Performance: Evaluate empirical dosage r², concordance, and recovery across low-frequency and rare-variant bins using thresholds defined for the intended downstream analysis.

To downstream post-imputation data filtering protocols and dosage extraction standards, refer to our technical companion guide on Post-Imputation QC for Large Cohorts: INFO/R², MAF, Concordance, and Dosage Filtering.

Figure 4. Conceptual leave-out validation framework comparing empirical imputation performance across allele-frequency tiers.

End-to-End Roadmap and Production Release Checklist

Constructing an institutional or population-scale imputation reference panel requires coordinated sampling, sequencing, variant processing, phasing, validation, and data governance. A documented operational checklist improves traceability and downstream reuse.

Where custom genotyping arrays are utilized in conjunction with imputation panels, cross-platform harmonization can be evaluated via SNP Genotyping Service workflows. For panel construction, deployment, phasing, and imputation support, see our Genotype Imputation & Haplotype Phasing Service. In downstream association studies, Genome-wide Association Analysis Service (GWAS) can support population-specific association testing. For complex livestock, crop, or polyploid species, exploring Combining Reduced Representation Genome Sequencing with Imputation: When It Works and How to Validate provides additional validation frameworks.

Planning note: Sample numbers, sequencing depths, kinship cutoffs, imputation-quality metrics, and validation thresholds in this article are examples for study planning rather than universal acceptance criteria. Appropriate values depend on species, population structure, reference-panel composition, sequencing technology, genome build, allele-frequency range, and downstream analysis objectives. Project-specific criteria should be defined and validated before full-scale panel deployment.

FAQs

Next steps: If you're planning a custom reference panel project or evaluating panel feasibility for an underrepresented cohort, explore our Genotype Imputation & Haplotype Phasing Service to review target data, reference-panel options, phasing strategy, and validation requirements.

  Related Services

Genotype Imputation & Haplotype Phasing

Sequencing

Whole Exome Sequencing for Population Genetics

Bioinformatics

Genetic Diversity Analysis

Genome-wide Association Analysis Service

QTL Location Analysis Service

Whole Genome Re-sequencing for Population Genetics

  Download

Power Your Regulatory Discoveries with DAP-seq: A MYB-GLK Case Study

EM-seq Services for High-Resolution DNA Methylation Analysis

Multi-Omics Strategies to Decode Molecular Mechanisms at Every Level

Explore Customized Bioinformatics Services: Case Studies & Insights