← All methods
Genomic Selection

Genomic Selection (GBLUP, BayesB)

Predict the genetic merit of unphenotyped plants from their genotype, with cross-validated prediction accuracy.

How it works

Genomic Selection (GS) trains a whole-genome prediction model on individuals with both genotype and phenotype, then predicts genomic estimated breeding values (GEBVs) for unphenotyped candidates. We support GBLUP (genomic best linear unbiased prediction) using the realized G-matrix, and BayesB for sparse-effect traits. We report cross-validated predictive accuracy (Pearson r between predicted and observed phenotypes) so you know exactly how trustworthy the predictions are before deploying them in your breeding program.

Formula

GBLUP: y = μ + g + e, where g ~ N(0, Gσ²g). G is the genomic relationship matrix from centered, scaled marker scores.

What you get

  • ▸GEBVs for every individual in the dataset
  • ▸Variance components (σ²g, σ²e, narrow-sense h²)
  • ▸k-fold cross-validated prediction accuracy (r)
  • ▸Top markers by absolute effect (for BayesB)

When to use it

  • ▸You want to rank candidates without phenotyping every plant
  • ▸You're running a breeding program and want to accelerate generations
  • ▸You need to evaluate prediction accuracy before trusting the model

Inputs

Training genotype + phenotype
Aligned genotype matrix + phenotype CSV for lines with both
Target genotypes
Same-schema genotype matrix for candidates lacking phenotypes

Parameters

NameDefaultDescription
ModelGBLUPGBLUP for polygenic traits; BayesB when a few large-effect markers dominate.
KinshipVanRadenVanRaden 2008, Astle-Balding, or Endelman IBD — see the kinship tutorial.
CV scheme5-fold randomRandom k-fold, leave-one-family-out, or forward prediction.
Folds / reps5 × 10Bootstrap standard error is reported alongside mean r.

Workflow

  1. 1. Build kinship
    G = ZZᵀ / (2Σpᵢ(1−pᵢ)) with centered marker matrix Z.
  2. 2. REML variance components
    σ²g and σ²e estimated once on training data.
  3. 3. Solve MME
    Henderson's mixed-model equations yield BLUPs for every ID in G.
  4. 4. Predict candidates
    GEBVs read off directly for target IDs included in G.
  5. 5. Cross-validate
    Chosen CV scheme rerun for accuracy estimate with bootstrap SE.

Interpreting results

  • ▸r ≥ 0.5 is generally strong for a complex trait; 0.3–0.5 is usable for early-stage selection.
  • ▸Compare all three CV schemes — a big gap between random k-fold and forward prediction means the model won't hold up between seasons.
  • ▸Narrow-sense h² caps the achievable r; if h² is 0.4, expect r ≤ ~0.63.

Common pitfalls

  • ✕Reporting only the best fold instead of mean ± SE overstates accuracy.
  • ✕Training and test lines from different families (uncontrolled) can inflate random k-fold accuracy.
  • ✕Using BayesB when effects are truly polygenic wastes signal; GBLUP wins for most yield traits.

Worked example

Maize-282 GEBV run
282 lines × 44k SNPs. GBLUP with 5×10 CV returns r = 0.48 (± 0.04). Ranking the top 20% by GEBV recovers 72% of the actual top-20% by phenotype.

Try it — interactive example

Predict expected genomic-prediction accuracy for GBLUP using the Daetwyler formula: r ≈ √(N·h² / (N·h² + Me)).

Expected outputs
Expected prediction accuracy (r)
0.408
Reliability (r²)
0.167
Training N to reach r = 0.70
2402
Solve Daetwyler for N given target accuracy.

Numbers are quick analytical estimates for planning — actual runs incorporate the full data, covariates, and QC pipeline.

FAQ

Do I need imputation?
Not always — GBLUP tolerates sparse markers well; BayesB benefits more from dense SNPs.
Can I combine multiple traits?
Yes; the multi-trait GBLUP mode borrows strength across genetically correlated traits.

References

Run Genomic Selection on your data

Open the module and upload a CSV.

Open module