From markers to genomic prediction

A young selection candidate may have no phenotype for a late, sex-limited, or destructive trait. It can still be genotyped. The breeding problem is to use marker information, together with records from other individuals, to predict the candidate’s breeding value before its own phenotype becomes available.

The preceding page defined breeding value as a population-specific additive genetic value. Markers do not measure that value directly. They become useful when a statistical model learns how marker patterns in a training population relate to phenotypes and then applies the fitted relationship to selection candidates.

Markers carry information about genetic differences

A quantitative trait locus (QTL) is a genomic region whose variation contributes to variation in a quantitative trait. The causal variant within a QTL is not automatically observed on a marker panel. A nearby marker can still carry information because alleles at the marker and causal locus may be associated in the population.

This population association is called linkage disequilibrium. Recombination tends to break associations over generations, while physical linkage, population history, selection, and drift shape the LD that remains. A marker can therefore predict a causal chromosome segment without causing the trait. The marker-QTL relationship is population-dependent and may weaken in a different breed, generation, or genetic background.

Genome-wide prediction does not require every fitted marker to be a known QTL. Dense markers can collectively track chromosome segments carrying causal variants. Some models express this information through estimated marker effects. Others use genomic similarity among individuals. Both routes learn a relationship between the observed genotypes and the prediction target.

Training and candidate populations have different roles

The training population contains individuals with both phenotypes and genotypes. Their phenotypes provide information about the trait, while their genotypes connect that information to inherited chromosome segments. Fixed effects such as sex, management group, or contemporary group must be handled by the model when they affect the phenotype but are not part of breeding value.

The candidate population contains the individuals to be ranked for selection. Candidate genotypes are supplied to the fitted model. Their trait phenotypes may be absent, unavailable at selection time, or deliberately withheld during validation. A close genetic and environmental relationship between the training and candidate populations usually gives the model more relevant information than a poorly matched training set.

The project data provide 160 training records and 40 held-out records, with both groups represented within each of ten full-sib families. Because the dataset is simulated, phenotypes and true breeding values remain stored for the held-out individuals so prediction can later be evaluated. A real deployment candidate would not need a known phenotype or true breeding value at the time its prediction is made.

Follow the information through the prediction process

The complete flow is

phenotypes + genotypes
          |
          v
      fit model
          |
          v
 estimated parameters
          |
          v
 candidate genotypes
          |
          v
         GEBV
          |
          v
        ranking

Each transition has a distinct meaning.

  1. Phenotypes and genotypes enter the fitted model. The training phenotypes contain genetic and non-genetic variation. The model relates their adjusted genetic signal to genome-wide marker patterns while representing relevant fixed effects and residual variation.

  2. Model fitting produces estimated parameters. Depending on the method, these may include marker effects, variance components, genomic relationships, shrinkage parameters, or posterior distributions. An estimate is uncertain because it is learned from a finite training sample.

  3. Candidate genotypes receive the training-derived representation. The same allele coding, marker order, centring values, and other preprocessing must be used. Re-estimating preprocessing from candidates can change the meaning of the fitted parameters and can leak information during model evaluation.

  4. The model produces a genomic estimated breeding value. A genomic estimated breeding value (GEBV) predicts the candidate’s additive genetic merit relative to the model’s reference population. It is neither an observed phenotype nor a directly measured genetic quantity.

  5. Breeders use GEBVs in a selection decision. Ranking candidates by GEBV can support selection for one trait. Practical breeding objectives may combine several traits, uncertainty, inbreeding, diversity, economic weights, and mating constraints. A high GEBV alone does not specify a complete breeding programme.

Marker effect does not mean causal effect

An estimated marker effect describes a statistical contribution under a particular coding, model, and training population. It may reflect LD with one or more causal variants, correlations with other markers, and shrinkage from the fitted model. Its sign can reverse if the counted allele is reversed. Its magnitude can change when markers are centred or scaled.

These properties separate prediction from biological mechanism. A marker can improve prediction without being causal, while a causal variant can receive a poorly estimated effect when it is unobserved, rare, or weakly tagged by the available markers. Establishing causality requires evidence beyond a fitted prediction coefficient or association statistic.

Three genomic analyses answer different questions

Marker-assisted selection, genomic selection, and genome-wide association studies all use markers, but their objectives and outputs differ.

Approach Main question Typical marker scope Primary output How it informs breeding
Marker-assisted selection Does a candidate carry one or more targeted alleles or regions with established relevance? A limited set of markers linked to known genes or QTLs Marker status or targeted genetic score Selects or introgresses specific regions, often for comparatively simple or major-effect targets
Genomic selection What is the candidate’s breeding value when genome-wide marker information is considered jointly? Dense markers distributed across the genome GEBV Ranks candidates using predicted additive genetic merit, including for polygenic traits
Genome-wide association study (GWAS) Which loci show statistical evidence of association with the trait? Genome-wide markers tested for association Locus-level effect estimates, uncertainty, and significance evidence Identifies regions for investigation; association alone does not establish causality or provide a validated selection index

The boundaries can overlap. A robust association discovered by GWAS may later supply a marker for targeted selection. A genomic prediction model may also use external association evidence to weight markers. The distinction rests on the immediate goal: tracking specified loci, predicting total breeding value, or finding associations.

Prediction must match its deployment population

A fitted model learns from the allele frequencies, LD patterns, relationships, environments, and phenotype definitions represented in its training records. Applying it to distant candidates changes the information available to connect markers with causal variants. Prediction accuracy can therefore decline across breeds, generations, or management systems.

Model evaluation must imitate the intended use. Holding out close relatives within the same generation asks a different question from predicting a later generation or another population. The project split is useful for teaching within-family prediction, but its result should not be described as evidence of performance in unrelated populations.

Exercises

  1. A breeder wants to rank 500 unphenotyped juveniles for a polygenic growth trait using dense genome-wide markers. Is the immediate task MAS, genomic selection, or GWAS?

The immediate task is genomic selection because the desired output is a genome-wide prediction of breeding value for each candidate. A GWAS would instead seek locus-level associations, while MAS would target a limited set of previously established loci.

  1. A model has been fitted with phenotypes only, and candidate genotypes are available. Which part of the prediction flow is missing?

Training genotypes are missing. Without them, the fitted model cannot learn how training phenotypes relate to marker patterns that can also be observed in candidates.

  1. A candidate has the highest GEBV for growth. What does this establish, and what does it leave unresolved?

It establishes that the fitted model predicts the highest additive genetic merit for growth among the candidates being compared. It does not reveal the candidate’s true breeding value, guarantee its offspring mean, establish that any marker is causal, or account automatically for other traits and inbreeding constraints.

The flow depends on representing candidate and training genotypes in exactly the same numerical form. The next page develops marker dosage, allele frequency, centring, scaling, and multiallelic encoding before those representations enter genomic relationship and prediction models.