Genomic Prediction
What is genomic prediction?
Genomic prediction (GP) is the use of dense marker data to predict the breeding value of an individual — that is, the additive genetic merit that the individual will transmit to its offspring. Where classical pedigree-based BLUP relies on expected co-ancestry inferred from family records, GP measures realised relatedness directly from the marker genome, capturing Mendelian-sampling deviations that pedigree records cannot see.
The trained model returns one number per candidate animal or plant — the genomic estimated breeding value (GEBV) — which downstream breeding decisions sort, threshold, and select on. GP rests on the infinitesimal model: trait variation is assumed to be driven by many loci of small effect, jointly captured by the genomic relationship matrix or the marker-effects prior.
When to use GP
| Situation | Use GP? |
|---|---|
| Selecting candidates that have no own phenotype yet (juveniles, eggs, embryos) | Yes — GP is the standard solution |
| Replacing a slow / expensive phenotyping step with a fast genotype call | Yes |
| Identifying which causal variants matter biologically | No — use GWAS instead |
| Trait with very few large QTL of known location | Use marker-assisted selection or include those QTL as fixed effects |
Prerequisites
- Training set: individuals with both phenotype + genotype (~ several hundred for low-heritability traits is the practical minimum; thousands for routine accuracy gains).
- Test / candidate set: animals/plants with genotype only.
- Genotype:
- SNP dosage matrix (rows = individuals, columns = SNPs, values 0/1/2), or
- Phased microhaplotype matrix from the
maspipelinepreprocessing pipeline.
- Fixed-effect covariates (sex, batch, season). A design matrix
Xbuilt withstats::model.matrix()is the standard form. - Optional: pedigree records, for PBLUP or single-step GBLUP via
build_A_ped().
GP vs GWAS at a glance
The two tasks share inputs but ask different questions:
| Dimension | Genomic prediction (this task) | GWAS |
|---|---|---|
| Output | \(n \times 1\) GEBV vector | \(p \times 1\) per-marker statistic |
| Goal | Rank breeding candidates | Identify causal variants |
| Use of marker effects | Aggregated into a single GEBV | Per-marker hypothesis tests / PIPs |
| Cross-validation? | Central — accuracy is the headline metric | Less common; power analysis instead |
| Typical follow-up | Select the top-ranked candidates | Annotate peaks, replicate, GWABLUP |
A single project usually does both: GWAS first to understand which parts of the genome drive the trait, then GP to translate that into selection decisions.
Methods covered
Both packages here implement the same statistical core — a linear mixed model with additive genetic effects — but differ in how marker effects are estimated.
| Approach | Prior on marker effects | Output | Package | Tutorial |
|---|---|---|---|---|
| GBLUP / Microhaplotype-GBLUP | Implicit ridge (all markers shrunk equally) | Aggregated GEBV | masreml |
→ with masreml |
| Threshold model (binary trait) | Same, on liability scale | Liability-scale GEBV + probability | masreml |
→ with masreml |
| BayesA | Scaled-\(t\) per marker (heavy tails) | Per-marker \(\hat\beta\) + aggregated GEBV | masbayes |
→ with masbayes |
| BayesR | 4-class mixture (zero / small / medium / large) | Per-marker \(\hat\beta\) + PIPs + aggregated GEBV | masbayes |
→ with masbayes |
What’s next?
Pick a package. Both tutorial pages use the same bundled demo data — you can run them in either order.
- Genomic prediction with masreml — REML mixed model, full-data fit, cross-validation, binary trait.
- Genomic prediction with masbayes — BayesA and BayesR (MCMC + EM), manual cross-validation, binary trait.