Can You Use Low-Coverage Sequencing for Genomic Selection?

Genomic selection has become standard practice in many plant and animal breeding programs. Train a model on lines that have been both genotyped and phenotyped, then use it to predict the breeding value of new lines from genotype alone. Selection happens earlier, cycles get shorter, and fewer lines need to be carried through expensive field trials.

Most genomic selection programs were built on SNP arrays or genotyping-by-sequencing. As sequencing costs have fallen, breeders are asking whether low-coverage whole-genome sequencing can replace those platforms. The short answer is yes, and published work in both livestock and crops supports it. The more useful question is what the prediction model can see once you make the switch.

How Genomic Selection Works, and What Limits it

A genomic selection model estimates the combined effect of genome-wide markers on a trait. Methods such as GBLUP, rrBLUP, and Bayesian regression models differ in their assumptions, but all of them depend on the same inputs: the size and relatedness of the training population, the heritability of the trait, the quality of the phenotypes, and how well the markers capture the genetic variation that actually drives the trait.

That last input is the one a sequencing decision changes. Markers do not need to be causal. They need to be in strong linkage disequilibrium (LD) with the causal variants. When a causal variant sits in a region the markers cannot tag, its contribution goes partly or entirely unmodeled.

Low-Coverage Sequencing Can Match Arrays

Multiple studies in livestock and crops have shown that low-coverage sequencing combined with imputation can deliver prediction accuracy comparable to SNP arrays. The approach brings practical advantages: no fixed marker panel designed around a handful of reference lines, less ascertainment bias toward variants common in the germplasm used to design an array, and a genotype dataset that can be reanalyzed as reference genomes and panels improve.

Matching an array, though, is a low bar. Once marker density saturates the LD in a population, adding more SNPs of the same kind adds little prediction accuracy. The real opportunity in sequencing-based selection is capturing different variants, rather than more of the same.

The Variants Your Model May Be Missing

Arrays and short-read methods under-sample the same parts of the genome: repeats, duplicated gene families, subgenome-specific regions in polyploids, and structural variants. Those regions carry real breeding value. Disease resistance genes frequently occur in clustered, duplicated families, and presence/absence variation and copy number changes affect many agronomic traits.

In a peanut breeding population, long-read low-pass sequencing identified 16 lines carrying an insertion associated with resistance to tomato spotted wilt virus. Short-read low-pass data from the same population could not align within the inserted segment (Lee et al., G3). A variant like this can only enter a prediction model through nearby markers, and only if those markers happen to be in strong LD with it. A variant that is genotyped directly can be modeled directly, and a known large-effect locus can be included as a fixed effect alongside the genome-wide markers.

Studies that add structural variants to genomic prediction models are still relatively new, and results so far are trait-dependent. Some traits show measurable gains, while others show little change. Gains are most likely where large-effect structural variants are known to act, which describes many of the disease resistance and adaptation traits breeders care about most.

Why Polyploid Programs Should Pay Close Attention

Polyploid crops add two complications. In allopolyploids such as peanut, wheat, and cotton, closely related subgenomes must be kept separate, or a variant on one subgenome is recorded as a variant on the other. In autopolyploids such as potato and alfalfa, prediction models benefit from allele dosage, which requires accurate allele counts at each locus.

Both depend on reads being placed correctly. In the same peanut population, alignment confidence averaged 92.9% for long reads and 54.9% for short reads at similar depth. Long reads span the near-identical sequence that leaves short reads ambiguous, so the genotypes feeding the model reflect the right subgenome.

Genotype Quality is Leverage on Your Phenotyping Budget

In most genomic selection programs, phenotyping is the largest cost. Multi-location field trials, disease screens, and quality assays cost far more per line than genotyping, and every phenotyped line in the training population represents a significant investment.

That reframes the genotyping decision. The genotype determines how much of the information in each expensive phenotype the model can use. A genotyping method that misses the variants driving a trait leaves part of every phenotype unexplained, however carefully it was measured. Evaluating genotyping platforms on per-sample price alone ignores the investment those genotypes are meant to leverage. Cost per answer is the better measure.

A Practical Design for Sequencing-Based Selection

  • Sequence the founders, parents, or a diverse subset of the training population with long reads to build a reference panel that includes structural variants and physically phased haplotypes.

  • Genotype the training population and selection candidates at low coverage, and impute against the long-read panel.

  • Keep historical array or GBS data in the model by imputing it against the same panel, so prior training data is not lost.

  • Compare prediction accuracy with and without structural variants through cross-validation on your own traits before changing your pipeline.

Where Selection is Heading

Genomic selection is moving from genome-wide markers as anonymous predictors toward models that incorporate known functional variation: structural variants, pangenome-informed markers, and biological information about which variants matter. Every one of those steps depends on first detecting the variation. Programs that build long-read reference panels now will have the genotype data those models require as they mature.

The Bottom Line

  • Low-coverage sequencing with imputation can match SNP array accuracy for genomic selection, and published studies in livestock and crops support it.

  • Beyond marker saturation, more SNPs add little. The gains come from capturing variant classes and regions arrays and short reads miss.

  • Long-read low-pass (LRLP) sequencing detects structural variants and resolves polyploid subgenomes, including a disease resistance insertion in peanut that short reads could not align to.

  • Phenotyping is the most expensive input to genomic selection. Genotype quality determines how much value each phenotype returns.

  • Test structural variant inclusion through cross-validation on your own traits, and keep legacy data in the model through imputation.

Source: Lee K, et al. Long-read low-pass sequencing enhances variant detection in a peanut MAGIC population. G3: Genes|Genomes|Genetics, jkag196. https://doi.org/10.1093/g3journal/jkag196


Veil Genomics offers the first commercially available long-read low-pass sequencing service on PacBio HiFi. If you are evaluating whether to move genomic selection from arrays or GBS to sequencing, we can help you design a comparison on your own germplasm.

See example per-sample pricing by genome size in our 2026 Pricing Guide, talk to a scientist or request a quote.

Related Reading

Previous
Previous

Can You Combine Short-Read and Long-Read Data in the Same Study?

Next
Next

How Do You Build a Better Reference Panel for Genotype Imputation?