Re-fit variance components from your data, run the BayesC variable-selection sampler to identify which SNPs are actually driving each trait, and validate predictive accuracy on held-out records. The defaults work without ticking any box — toggle features on for the deeper analysis.
Variance components drive every BV — the defaults are industry averages. Calibrating against your own data tightens accuracy.
BayesC PIP tells you which SNPs drive each trait — GBLUP cannot.
Cross-validation reports how well the BVs predict held-out records. The number you actually trust.
Misztal 2014. An approximation for large genotyped populations; not yet demonstrated at scale.
Restricted maximum likelihood by expectation-maximisation. Monotone in the REML likelihood by construction, and the estimate it returns does not depend on the prior it starts from. It returns no asymptotic standard error — EM does not produce one, and none is invented.
Its estimates are currently REJECTED and never reach your breeding values. EM-REML has not passed the known-heritability recovery gate, so its runs keep the supplied prior heritability. Selecting this changes what is reported, not what is delivered.
Lamb survival, fertility, dystocia are binary — Gaussian BLUP gives wrong BVs. A probit threshold model is the appropriate estimator; Genemap's has not passed a recovery gate and is experimental.
Per-animal polynomial coefficients describing growth/milking trajectories. Replaces separate BWT/WWT/PWT/YWT models with one continuous model of weight-at-age.
Erbe 2012. Each SNP belongs to one of (zero, small, medium, large) effect classes. Better than BayesC for traits with major-effect loci on a polygenic background. The method behind Australian dairy MGBSI.
Holds out 20% of phenotypes, predicts them from the remaining 80%, reports correlation. The accuracy metric used in Daetwyler 2010.