Why Genotype Imputation Accuracy Fails in Breeding Populations

Home
Resource
Article Hub
Why Genotype Imputation Accuracy Fails in Breeding Populations

Why Genotype Imputation Accuracy Fails in Breeding Populations

Genotype imputation troubleshooting workflow

Genotype imputation is widely deployed across plant and animal breeding programs to scale up low-density arrays, genotyping-by-sequencing, and low-coverage whole-genome sequencing to comprehensive whole-genome marker sets. However, bioinformaticians and quantitative geneticists frequently encounter unexpected imputation accuracy failures—manifested as poor genomic prediction reliability, phantom QTL peaks in association mapping, or low genotype concordance across breeding families. When imputation underperforms, teams often struggle to determine whether the bottleneck stems from insufficient target marker density, genetic divergence from the reference panel, software parameter misconfiguration, or baseline phasing errors. This troubleshooting guide breaks down the biological, mathematical, and algorithmic causes of imputation failure in agricultural populations, provides a structured diagnostic matrix, and answers the critical project question: should you add more target markers or expand your reference panel?

Key takeaways

Imputation accuracy is strongly influenced by genetic relatedness between target samples and the reference panel; simply increasing reference size may not compensate for missing ancestral lineages or mismatched linkage disequilibrium (LD) phases.
Raw genotype concordance can mask imputation errors at low-frequency and heterozygous alleles; it should be interpreted alongside Non-Reference Discordance (NRD), Dosage R-Squared (DR²), and MAF-stratified performance.
Target marker density should be evaluated against the population's LD decay; genetically diverse and outcrossing populations often need denser scaffold information than narrow inbred populations.
Software parameterization can affect imputation performance. Effective population size (N_e), reference composition, and marker density should be tested empirically rather than transferred unchanged from unrelated populations.
A structured diagnostic workflow (Problem → Cause → Metric → Action) allows teams to resolve imputation bottlenecks without wasteful re-sequencing.

The hierarchy of imputation accuracy drivers

What determines whether genotype imputation succeeds or fails in a breeding program? Imputation accuracy is governed by four primary structural drivers and four secondary operational factors. Understanding this hierarchy allows teams to identify the true root cause of failure rather than making trial-and-error adjustments.

For foundational guidelines on constructing a high-fidelity reference panel before deployment, consult the companion guide on how to build a genotype imputation reference panel for crop breeding.

Key drivers of genotype imputation accuracy

Primary structural drivers

Target-to-Reference Genetic Distance: Sharing of recent identity-by-descent (IBD) chromosome tracts between target lines and reference accessions is one of the strongest determinants of accuracy. If a target line contains foreign or recombinant haplotypes that are poorly represented in the panel, imputation may remain unreliable even after changing software or tuning parameters.
Reference Panel Phasing Quality: Imputation models rely on phased haplotype sequences. Switch errors in the reference panel introduce corrupted haplotype templates that systematically misinform target allele estimation.
Target Marker Density and Distribution: The physical spacing of genotyped backbone markers must be tight enough to unambiguously identify which reference haplotype spans each chromosomal window.
Minor Allele Frequency (MAF) Spectrum: Common alleles are generally easier to impute than rare or low-frequency alleles. Performance at the rare end of the spectrum depends heavily on whether relevant carrier haplotypes are represented in the reference set.

Secondary operational factors

Effective Population Size (N_e) Parameterization: Some imputation algorithms use N_e to model haplotype transitions. A default value developed for a very different population may be suboptimal, so plausible settings should be tested against held-out or independently genotyped samples.
Local Recombination Rate Variation: Centromeric and pericentromeric regions with suppressed recombination behave differently from telomeric recombination hotspots, creating localized accuracy fluctuations.
Missingness and Genotyping Error in Target Data: High sporadic missingness or miscalled heterozygous loci in low-density arrays and genotyping-by-sequencing (GBS) data induce phase ambiguity.
Structural Variation and Copy Number Ambiguity: High-copy paralogs, presence-absence variants (PAVs), and chromosomal inversions violate collinearity assumptions and cause regional imputation failure.

Root causes of imputation failure and diagnostic metrics

When an imputation run finishes, assessing output quality using overall genotype concordance alone can be misleading. In datasets dominated by common homozygous genotypes, headline concordance may remain high even when performance is weaker at heterozygous, alternate, or low-frequency sites. Validation should therefore separate overall agreement from non-reference and MAF-stratified performance.

For a complete breakdown of evaluating raw array quality reports before imputation, see the operational guide on how to read a genotyping array QC report.

The thresholds and actions below are diagnostic starting points rather than universal pass/fail criteria. Appropriate cutoffs should be calibrated to species, ploidy, population structure, reference-panel composition, sequencing design, and downstream application.

Failure symptom	Underlying root cause	Primary diagnostic metric	Recommended corrective action
Systematic accuracy drop across entire cohort	Reference-target genetic divergence or unrepresented founder line	Kinship matrix (G), PCA distance, MAF-stratified DR²	Consider adding missing parental or closely related lines with sufficiently confident genotypes
High concordance but poor genomic prediction (GEBV)	Failure on rare/heterozygous alleles masked by high reference homozygosity	NRD, MAF-stratified r², and locus-level DR²	Apply validated marker-quality filters and consider adding relevant carrier accessions
Localized drops in accuracy on specific chromosomes	Structural variation, chromosomal inversion, or assembly misorientation	Regional LD decay, localized switch error spikes	Check coordinate and strand alignment; evaluate SV-aware or pangenome-aware representation where appropriate
Widespread errors across low-pass sequencing targets	Coverage insufficient for the specific reference-target design, high duplication, or uneven read sampling	Downsampling accuracy curve, read depth distribution, and NRD	Evaluate a higher coverage tier and/or a likelihood-aware workflow such as GLIMPSE2
Excessive switch errors in inbred crop lines	Inappropriate N_e or phasing parameterization for the study population	Mendelian inconsistency in pedigree trios, high switch rate	Test plausible N_e settings and select parameters using empirical validation
High missingness after post-imputation filtering	Target marker density too sparse for rapid LD decay in outcrossing crop	Inter-marker distance vs. LD r² half-decay distance	Evaluate a denser array or low-coverage WGS for the target cohort

Genotype concordance versus Non-Reference Discordance (NRD)

To accurately evaluate imputation reliability, bioinformaticians must separate concordant reference homozygotes from non-reference genotypes. Non-Reference Discordance specifically isolates discrepancies at heterozygous and homozygous alternate positions:

NRD = (e_RR + e_RA + e_AA) / (e_RR + e_RA + e_AA + m_RA + m_AA) × 100

Where e denotes mismatch counts between imputed and true genotypes, and m denotes true positive matches. NRD should be interpreted together with coverage, allele frequency, and the intended downstream use rather than treated as a universal pass/fail threshold. In rainbow trout, Liu et al. (2024) reported 98.7% genotype concordance at 0.5× coverage, declining to 97.8% at 0.2× and 96.7% at 0.1× as coverage decreased. The study also showed that imputation performance differed between breeding populations, reinforcing the importance of reference-target genetic representation. These results were reported in G3: Genes, Genomes, Genetics.

Dosage R-Squared (DR²) as a quality filter

Dosage R-Squared (DR²) is a marker-level measure used to summarize imputation certainty or agreement in allele dosage. The appropriate cutoff depends on allele frequency, validation design, and downstream analysis. In a 2026 soybean study, imputed variants were filtered using DR² ≥ 0.80 and MAF > 0.03 before downstream GWAS and local haplotype analysis, providing a crop-specific example of how dosage-quality filtering can be incorporated into an analysis workflow, as reported by Mohamedikbal et al. in Theoretical and Applied Genetics.

Decision framework: more markers or more reference samples?

When an imputation pilot yields substandard accuracy, breeding program managers face a major budgetary question: Should we invest in higher marker density on the target screening panel, or should we sequence more reference individuals?

The answer depends on the primary bottleneck identified during diagnostic QC. For strategic comparisons between fixed arrays and sequencing-based density tiers, explore the analysis in choosing marker density for breeding cohorts.

Marker density versus reference panel decision matrix

When increasing target marker density is worth testing

Rapid Linkage Disequilibrium Decay: In genetically diverse or outcrossing populations, local LD may decay rapidly. If backbone-marker spacing is wider than the LD scale relevant to the target population, evaluating a denser marker panel or low-coverage whole-genome sequencing (lc-WGS) can improve haplotype assignment.
Advanced Recombinant Generations: In multi-generation breeding schemes, accumulated recombination can fragment founder haplotypes into shorter segments. Denser target information may therefore be needed to resolve breakpoint boundaries reliably.
Plateaued Accuracy Across Reference Panels: If a substantial increase in well-matched reference samples produces little measurable improvement, this suggests that target marker density, data quality, or another bottleneck may be limiting performance.

When expanding or re-curating the reference panel is worth testing

Target-to-Reference Genetic Isolation: If PCA or genomic relationship matrices show that target breeding families are poorly represented by the reference panel, increasing target marker density alone may not resolve the problem. Consider adding missing parental lines or closely related reference material.
Poor Imputation of Rare and Donor Alleles: If common variants impute well but low-frequency introgressed alleles perform poorly, adding donor carriers or closely related haplotypes may improve representation of those genomic segments.
High Baseline Switch Error Rates: If validation on high-density target samples reveals persistent switch errors despite adequate marker density, first verify reference genotype quality and sample identity. If reference uncertainty contributes to the problem, higher-confidence genotype calling followed by re-phasing may be warranted. Browning et al. (2021) describe an efficient two-stage phasing framework in The American Journal of Human Genetics.

Software parameter optimization and algorithmic pitfalls

Even with high-quality reference and target data, algorithmic parameterization can influence imputation fidelity. In cattle, pigs, and chickens, Jiang et al. (2022) showed that Beagle performance varied with reference size, target marker density, relatedness, and the effective population size (N_e) setting, supporting empirical parameter testing rather than automatic use of defaults. The study is reported in Frontiers in Genetics.

Effective population size (N_e) tuning

Imputation tools may use N_e when modeling transitions between reference haplotypes. Because breeding populations can differ greatly in relatedness, historical recombination, and selection history, N_e should be treated as a dataset-specific tuning parameter rather than a universal species constant.

Inbred and Pedigree-Structured Crops: Narrow breeding populations may differ substantially from the populations for which software defaults were designed. Test plausible N_e settings using held-out high-confidence genotypes or pedigree-consistent validation rather than assigning one fixed crop-wide value.
Diverse and Outcrossing Germplasm: Broader germplasm collections may require different transition assumptions because of greater haplotype diversity and recombination history. Where population-specific estimates are unavailable, sensitivity testing across plausible settings is preferable to relying on a single default.

Algorithm selection for low-coverage target data

For low-coverage sequencing, converting sparse read evidence into hard genotype calls before imputation can discard useful uncertainty information. Likelihood-aware workflows such as GLIMPSE2 retain genotype uncertainty while using phased reference haplotypes for imputation. Rubinacci et al. (2023) introduced GLIMPSE2 for scalable imputation of very-low-coverage sequencing data using large reference panels in Nature Genetics. Transfer to agricultural datasets should still be validated for the species, ploidy, reference panel, and coverage design.

Building robust quality gates before genomic selection

Imputed marker data should never enter downstream genomic prediction or selection index calculations without strict quality gating. Misimputed markers introduce non-random noise into genomic relationship matrices (G), biasing estimated breeding values (GEBV) and reducing genetic gain. Standardized protocols for formatting and filtering imputed outputs for downstream quantitative genetics are outlined in building GS-ready datasets from array and sequencing outputs. Broader principles governing training population reliability are covered in training population design for genomic selection.

Post-imputation QC and filtering workflow

How CD Genomics can help

Troubleshooting and optimizing genotype imputation pipelines requires experience across diverse crop mating systems, polyploid genomes, and high-throughput bioinformatics architectures. CD Genomics provides agricultural genomics and quantitative genetics support to help research teams diagnose imputation accuracy failures, refine reference panels, and evaluate low-coverage whole-genome sequencing (lc-WGS) and array-based workflows. Our bioinformaticians assess population structure, test phasing and imputation parameters, and perform held-out validation to determine whether resulting datasets are suitable for the intended downstream analysis. Explore our comprehensive suite of sequencing, array, and computational solutions via the agricultural genomics services overview, or review our specialized agricultural genomic data analysis capabilities. All services described in this guide are provided strictly for Research Use Only (RUO) in agricultural, plant, and animal genetics research, and are not intended for clinical or human diagnostic use.

Frequently asked questions (FAQ)

Q1: Why is my overall imputation concordance above 98% while genomic prediction accuracy is still low?
A: High concordance can be driven by common homozygous genotypes that dominate the evaluated sites. If performance is weaker at heterozygous, alternate, or low-frequency genotypes, important variation for association or genomic prediction may still be affected despite a high headline value. Evaluate NRD, dosage-quality metrics, and MAF-stratified accuracy rather than relying on overall concordance alone.

Q2: How does effective population size (N_e) affect Beagle imputation in crops?
A: The N_e parameter influences the model's expected haplotype transitions. There is no single crop-wide value that is appropriate for every breeding population. When N_e is configurable, test plausible settings and compare them using held-out or independently genotyped samples rather than relying automatically on a default value.

Q3: When should I choose low-coverage WGS (lc-WGS) over low-density SNP arrays for imputation targets?
A: lc-WGS may be preferable when marker coverage from an available array is too sparse for the target population, when germplasm diversity is broad, or when no mature species-specific array exists. Low-density arrays can remain practical for standardized screening in well-characterized breeding populations. Coverage and platform choice should be piloted against the intended reference panel before large-scale deployment.

Q4: Can a reference panel built from wild accessions accurately impute elite commercial cultivars?
A: Not reliably when elite haplotypes are poorly represented. Wild accessions can broaden diversity, but a panel intended for elite breeding populations generally benefits from including contemporary parents, recurrent parents, or other genetically related elite material. The balance should be tested empirically for the target cohort.

Q5: What is the minimum recommended Dosage R-Squared (DR²) threshold for downstream GWAS and GS?
A: There is no universal DR² cutoff for every breeding dataset. A threshold of DR² ≥ 0.80 has been used as a stringent filter in some crop studies, including the 2026 soybean example cited above, but the appropriate cutoff should be validated against allele frequency, sample size, the intended downstream analysis, and independent or held-out genotype data.

References

Jiang, Yifan, Hailiang Song, Hongding Gao, Qin Zhang, and Xiangdong Ding. "Exploring the Optimal Strategy of Imputation from SNP Array to Whole-Genome Sequencing Data in Farm Animals." Frontiers in Genetics, vol. 13, 2022, Article 963654.
Rubinacci, Simone, Robin J. Hofmeister, Bárbara Sousa da Mota, et al. "Imputation of Low-Coverage Sequencing Data from 150,119 UK Biobank Genomes." Nature Genetics, vol. 55, 2023, pp. 1088–1090.
Liu, Sixin, Kyle E. Martin, Warren M. Snelling, Roseanna Long, Timothy D. Leeds, Roger L. Vallejo, Gregory D. Wiens, and Yniv Palti. "Accurate Genotype Imputation from Low-Coverage Whole-Genome Sequencing Data of Rainbow Trout." G3: Genes, Genomes, Genetics, vol. 14, no. 9, 2024, jkae168.
Browning, Brian L., Xiaowen Tian, Ying Zhou, and Sharon R. Browning. "Fast Two-Stage Phasing of Large-Scale Sequence Data." The American Journal of Human Genetics, vol. 108, no. 10, 2021, pp. 1880–1890.
Mohamedikbal, Shameela, Hawlader A. Al-Mamun, et al. "Dissection of Local Haplotype Diversity at Soybean Rust Loci Reveals Resistance-Associated and Context-Dependent Variation Patterns in Diverse Germplasm." Theoretical and Applied Genetics, vol. 139, no. 4, 2026, Article 107.