Breeding programs with Genomic selection

Traditional selection depends on phenotypes. Accuracy is limited by a trait's heritability, and it struggles with traits that are lowly heritable, sex-limited (e.g. egg production), expressed late in life, or hard or expensive to measure (disease resistance, meat quality). It also cannot see Mendelian sampling: we don't know which alleles a parent actually passed on, so relatives only take us so far.

Gene-detection (QTL) studies found a few large-effect genes but left most variation unexplained, the “missing heritability”: most traits are driven by very many genes of tiny effect. The breakthrough came with cheap SNP chips (genotyping 50,000 to 700,000 markers per animal). With markers blanketing the genome, at least one is in linkage disequilibrium with each causal variant. Meuwissen, Hayes and Goddard (2001) proposed using all markers at once to predict a genomic estimated breeding value (GEBV), even before any phenotype is recorded on the candidate.

The case for genomic selection is clearest where the old approach is slowest. Traditional progeny testing of a dairy bull takes about five years, because he must wait for daughters to be born, grow up and complete a lactation before his breeding value is known; and sex-limited traits (milk, egg number) give the sire no record of his own. Genomic prediction breaks this bottleneck by estimating merit from DNA at birth, so a young bull or heifer can be selected before any phenotype exists, roughly halving the generation interval. This is exactly why dairy cattle breeding was transformed once dense SNP chips became affordable around 2008, and why the economics turn on the cost of genotyping versus the value of selecting earlier and more accurately.

A tempting but wrong approach is to test each SNP, keep the significant ones, and add up their effects. This suffers the winner's curse: the SNPs that reach significance have over-estimated effects, and together they explain only a small slice of the variance. Meuwissen et al. showed the fix is to fit all markers simultaneously.

But with tens of thousands of markers and far fewer animals, this is the “big p, small n” problem, ordinary regression cannot invert the equations. Two equivalent families of solutions are used:

• SNP-BLUP / ridge regression, every marker is assumed to explain the same small variance; effects are shrunk toward zero.
• Bayesian methods (BayesA, BayesB), allow markers to have different variances; BayesB lets many markers have exactly zero effect, matching the biology of few large and many tiny effects.

The third route gives the same answer as ridge regression but is easier to implement at scale: build a genomic relationship matrix (G) from the markers and use it in the standard mixed model in place of the pedigree relationship matrix A. This is GBLUP. G captures the realised relationships between animals, two full sibs are not always related by exactly ½, so “some animals are more equal than others”, which is extra information traditional pedigrees miss.

It helps to see that the marker model and the GBLUP model are two views of the same thing. One can either estimate an effect for every SNP at once (SNP-BLUP / ridge regression, with all markers sharing a small shrunk variance) or summarise the markers into a genomic relationship matrix G and predict animals directly (GBLUP); under the infinitesimal assumption the two give identical predictions. The Bayesian alphabet (BayesA, BayesB and relatives) relaxes the equal-variance assumption, letting a few markers carry large effects while most carry none, which helps for traits influenced by a major gene but matters less for highly polygenic traits. In every case the effects are shrunk toward zero, and that deliberate shrinkage, not selecting a handful of “significant” markers, is what makes genomic prediction robust.

Genomic prediction is best understood as the successor to an earlier idea. For two decades breeders tried marker-assisted selection: map a few quantitative trait loci (QTL) by linkage, then select on those markers. It helped for the rare trait controlled by a gene of large effect, but most production traits are highly polygenic, so a handful of significant QTL captured only a small slice of the genetic variance and progress disappointed. Whole-genome prediction succeeds precisely by abandoning the search for individually significant markers and instead fitting all markers simultaneously, letting thousands of small effects sum to an accurate prediction. Recognising that genomic selection is the polygenic generalisation of marker-assisted selection explains both why it works and why selecting only the top SNPs (the old approach) reintroduces the very bias it was designed to avoid.

The accuracy of a GEBV is its correlation with the true breeding value. We estimate it by validation: train the prediction on a reference population, predict a separate validation set, and compare predictions with later phenotypes, checking both correlation (accuracy) and the regression slope (bias; a slope of 1 means unbiased).

The Daetwyler equation captures what drives accuracy:

r = √( Nh² / (Nh² + M_e) )

where N is the number of reference animals with phenotypes, h² the heritability, and M_e the effective number of independent chromosome segments (larger in populations with a big effective size N_e and long genome). Accuracy rises with a larger reference population and higher heritability, and falls when M_e is large. The practical messages: low-heritability traits need much larger reference populations; predicting across breeds is hard because linkage disequilibrium differs; and cheap low-density genotypes can be imputed up to high density to add information at low cost.

The accuracy of a GEBV is captured by a simple relationship (the Daetwyler equation): it rises with the number of phenotyped reference animals (N) and the heritability, and falls with the effective number of independent chromosome segments (M_e), which is larger when the effective population size and genome are larger. Two consequences follow. Low-heritability traits need much larger reference populations to reach the same accuracy, and breeds with high N_e (more independent segments) are harder to predict than tightly related ones. Accuracy also decays over generations as recombination breaks the marker–gene associations the model learned, so the reference population must be continually refreshed with new genotyped-and-phenotyped animals rather than trained once.

Accuracy depends not just on reference size but on relatedness: predictions are most accurate for candidates closely related to the reference animals. Comparing genomic and pedigree relationships also serves as quality control, large mismatches (Mendelian inconsistencies) flag pedigree or genotype errors.

The phenotypes that go into the reference matter as much as the genotypes: they should be accurate and unbiased (for dairy bulls, for example, de-regressed proofs are used). Where some animals are genotyped and others are not, single-step GBLUP combines pedigree and genomic relationships into one matrix (H), so all animals are evaluated together, no information is wasted. Different BLUP variants (pedigree BLUP, GBLUP, single-step) can be compared on the same data to see what genomic information adds.

This is why relatedness between the reference population and the selection candidates is the single biggest practical driver of accuracy: a reference made of close relatives predicts well but narrowly, while pooling animals across herds and countries builds the large, diverse references that numerically small breeds could never assemble alone. A common misconception is that genomic selection removes the need to record phenotypes; the opposite is true, because the reference must be continuously phenotyped to keep predictions current. That makes investment in recording novel, hard-to-measure traits, feed efficiency, methane emission, disease resistance, the real frontier, since a trait can only be selected genomically once a reference population has been phenotyped for it.

Response to selection follows from the breeding goal (H = g′v), the genetic parameters, the information used for breeding-value estimation, and the selection and mating decisions (including the generation interval). Selection-index theory predicts response, and the SelAction software implements it. A GEBV enters the index as an extra information source with its own accuracy.

Genomic selection helps breeding programs in two big ways: it gives accurate predictions on young animals before they have records, which shortens the generation interval, and it improves accuracy for difficult traits. Both raise the rate of genetic gain.

Two cautions are essential. The Bulmer effect: selection reduces the additive genetic variance (often by around 25%) because selected parents are less variable, so realised response is lower than a naive calculation suggests, and after accounting for it, extra pedigree information adds little. And when judging the benefit of GS (for example, the value of genotyping cows), do not compare against a parent-average prediction as if it were accurate, once the parents themselves are selected, the parent average is almost useless, so honest comparisons must account for selection.

Adopting genomic selection is not a bolt-on; it redesigns the breeding programme. Young bulls can enter service on their genomic test without a progeny-test wait, so schemes use more sires for shorter periods and turn generations over faster. The female pathway changes too: genotyping heifers, and even screening IVF embryos, lets breeders select intensely on the dam side that progeny testing never could. Because faster, family-based selection also accelerates inbreeding, modern programmes pair these gains with optimum-contribution selection to manage diversity, the subject of the next module. The net effect is the same levers as always, intensity, accuracy and generation interval, but reset to deliver substantially more gain per year.

Selection changes allele frequencies directionally and leaves signatures of selection in the genome, detectable through statistics such as F_ST (differentiation between populations), reduced heterozygosity, and haplotype-based measures (iHS, EHH) around selective sweeps. But not all change is selection: genetic drift, random change that is larger in small populations (inversely proportional to effective size N_e), also shifts frequencies and erodes diversity, and must be separated from true selection signals.

This matters for sustainability. Genomic selection captures close family relationships, so without care it can increase inbreeding per generation relative to progeny testing. The remedy is genomic control of inbreeding: managing matings to restrict the increase in marker-based kinship while still selecting for merit. Maintaining genetic diversity is not a side issue, it protects long-term response, adaptability and resilience, which is exactly where breeding meets biodiversity.

The practical tool that reconciles gain with diversity is optimum-contribution selection (OCS): instead of simply mating the highest-merit animals, OCS chooses how many offspring each candidate contributes so as to maximise genetic gain for a fixed, pre-set rate of inbreeding. With genomic data the kinship being constrained can be measured directly from markers rather than pedigree, and realised inbreeding can be audited from runs of homozygosity (ROH), long stretches of homozygous genotype that reveal recent common ancestry. Setting an explicit ceiling on the rate of inbreeding, typically well under one percent per generation, and letting OCS optimise within it, is how a programme keeps the fast genomic gains of the previous modules sustainable.