Imputation is what makes low-pass and low-density genotyping economical. Genotype or sequence many individuals sparsely, then use a reference panel of known haplotypes to infer the genotypes you did not directly observe.
Most of the effort in an imputation project goes into the target samples: how deep to sequence them, which array to use, which software to run. The reference panel often gets less attention, even though it sets a hard ceiling on everything downstream. A variant that is absent from the panel cannot be imputed. A haplotype phased incorrectly in the panel carries that error into every sample imputed against it.
If your imputation accuracy is lower than expected, the panel is the first place to look.
What Imputation Actually Does
Individuals in a population share long stretches of chromosome inherited from common ancestors. Within those stretches, alleles travel together, a pattern called linkage disequilibrium (LD). Imputation software such as Beagle, Minimac, IMPUTE5, and GLIMPSE, which was designed for low-coverage sequencing data, compares the sparse observations in each target sample with the phased haplotypes in the reference panel, identifies the best-matching haplotype segments, and fills in the missing genotypes from them.
The quality of that inference depends on three properties of the panel: which variants it contains, how accurately its haplotypes are phased, and how well those haplotypes represent the population being imputed.
Why Reference Panels Underperform
They Inherit the Blind Spots of the Data Used to Build Them
Most panels are built from genotyping arrays or short-read sequencing. Whatever those methods miss, the panel misses. Variants in repetitive regions, duplicated gene families, and the homeologous segments of polyploid genomes are underrepresented, and structural variants are often absent altogether. Imputation cannot recover what the panel never contained.
Statistical Phasing Introduces Switch Errors
A short read rarely contains more than one heterozygous site, so short-read panels are phased statistically, by inferring which alleles travel together across the population. Statistical phasing produces switch errors: points where the inferred haplotype jumps from one parental chromosome to the other. Each switch error is a place where the panel gives the imputation software incorrect information.
They Underrepresent Diverse Populations
Imputation accuracy drops when target samples carry haplotypes that are rare or missing in the panel. This is why accuracy is often lower in landraces, wild relatives, and genetically diverse breeding germplasm, as well as in human populations underrepresented in large reference resources. Diverse populations also tend to have shorter LD blocks, so each observed site carries less information about its neighbors. A panel built from a narrow set of elite lines or a single ancestry group will be least accurate exactly where new diversity is being introduced.
What Long Reads Change About Panel Construction
Physical Phasing Instead of Statistical Inference
A HiFi read spanning many kilobases often carries several heterozygous sites at once. Those sites are phased by direct observation, because they were read from the same DNA molecule. Overlapping reads extend those phased segments into longer blocks. Read-backed phasing reduces reliance on statistical inference, and with it, the switch errors that degrade downstream imputation.
Structural Variants Become Part of the Panel
Because long reads span insertions, deletions, inversions, and copy number changes, those variants can be genotyped directly in panel samples and phased onto the same haplotypes as the surrounding SNPs. Once a structural variant sits on a phased haplotype, it can be imputed into target samples from the SNPs around it, even when the target data could never detect that structural variant directly.
This is the basis for imputing structural variants at population scale: sequence a representative set of individuals with long reads, then impute their structural variation into the full cohort.
Better Coverage of Hard-to-Map Regions
Long reads align confidently in repeats and duplicated regions where short reads are filtered out. For instance, in a peanut breeding population, long-read low-pass sequencing covered 55% of the genome at about 1.6x depth, compared with 17.3% for short-read low-pass at similar depth (Lee et al., G3). Regions covered in the panel are regions that can be imputed.
A Practical Framework for Building the Panel
Choose Panel Samples for Haplotype Diversity
The panel should capture the haplotypes present in the population you intend to impute. In structured breeding populations, that usually means founders and parents: the founders of a MAGIC or NAM population, the parents of a biparental family, or a core collection that represents a germplasm bank. In diverse populations, prioritize breadth over depth. Adding distinct haplotypes improves accuracy more than resequencing similar ones.
Include the Variant Classes You Want to Study
Decide at the design stage whether the panel must carry structural variants, small insertions and deletions, or presence/absence variation. If it must, the panel data has to be able to detect them, which in practice means long reads.
Match the Panel to the Target Data
Target samples can be low-pass long reads, low-pass short reads, array genotypes, or legacy data from earlier studies. Align everything to the same reference assembly, use a consistent variant representation, and confirm that the sites observed in the target data overlap well with the panel.
Measure Accuracy by Variant Class
Validate by masking known genotypes in a subset of samples randomly across the whole genome, imputing them, and measuring concordance. Report accuracy separately by allele frequency and by region type. A high overall concordance can hide poor performance on rare variants and complex regions, which are often the variants a study was designed to find.
Limits Worth Knowing
Rare variants are imputed poorly by any panel, because they are carried on few haplotypes. Detecting them reliably requires direct observation.
Structural variant imputation works best for common variants in strong LD with nearby SNPs. Accuracy should be validated variant class by variant class.
Imputation tools for polyploid species are less mature than those for diploids. Polyploid projects benefit from direct long-read genotyping of more samples and a more conservative use of imputation.
Where Reference Panels Are Heading
The field is moving from single linear references toward pangenome graphs that encode the haplotypes and structural variants of an entire population, and reference panels are following the same path. A panel built from phased long reads is a natural input for a pangenome, and each new population sequenced adds haplotypes that improve imputation for the next study. Built this way, a panel becomes a cumulative research asset rather than a one-time project deliverable.
The Bottom Line
A reference panel sets the ceiling on imputation accuracy. Variants missing from the panel cannot be imputed.
Panels built from arrays or short reads inherit their blind spots, depend on statistical phasing, and often lack structural variants.
Imputation accuracy is lowest in diverse populations because their haplotypes are underrepresented in most panels.
Long-read low-pass (LRLP) sequencing phases variants physically, adds structural variants to the panel, and covers regions short reads cannot map.
Validating by masking known genotypes at random across the genome measures accuracy by variant class and allele frequency, not just overall.
Veil Genomics offers the first commercially available long-read low-pass sequencing service on PacBio HiFi. If imputation accuracy is limiting your study, we can help you think through panel design for your population.
Talk to a scientist or request a quote.
