Association, prediction and regulatory evidence

B13: Reusable teaching module

Learning objectives

Define marker dosage and phenotype contrasts; identify confounding and linkage disequilibrium; distinguish GWAS, eQTL, TWAS, fine-mapping and prediction.

Visualize candidate regulatory relationships

Concept diagram; detailed explanation and suggested reasoning are in the visual guide.

Use the visual guide to explain the assumptions in this diagram.

Biological question and measurements

GWAS evaluates association between variants and a phenotype. An eQTL analysis evaluates association between variants and expression. TWAS uses expression-related genetic information in a trait-association framework. These approaches connect measurements, but none automatically identifies a causal variant, causal gene or mediation mechanism.

Relatedness and population structure can induce association when genetic groups differ in environment or phenotype. Linkage disequilibrium allows a marker to tag nearby variation; multiple correlated associations need not imply multiple mechanisms. Fine-mapping compares candidate explanations under an explicit model and requires appropriate LD, coverage, allele alignment and prior assumptions. A posterior probability is conditional on that model and supplied candidates.

Genomic prediction evaluates the ability to predict new animals, rather than selecting the smallest association p-value. The deployment population and training relationships determine the relevant validation design.

Model and assumptions

A basic unrelated-sample model is \(y_i=\beta_0+\beta_g g_i+\gamma^T c_i+e_i\), where \(g_i\) counts a stated allele and c contains prespecified covariates. Residual independence and the design matter. With correlated relatives, a mixed-model or other suitable framework is needed.

For two marker vectors, their correlation describes association of the supplied dosages. It does not establish ancestry, phase, a causal allele or a calibrated fine-mapping probability.

Worked example

The inputs below are hypothetical. Run the example, predict one change, then compare your prediction with the result. All executable code uses base R.

set.seed(1301)
ancestry <- rep(c(0,1),each=20)
g <- rbinom(40,2,ifelse(ancestry==0,.2,.8))
y <- 3*ancestry+rnorm(40)
unadjusted <- lm(y~g)
adjusted <- lm(y~g+ancestry)
c(unadjusted_marker=coef(unadjusted)["g"],adjusted_marker=coef(adjusted)["g"])

unadjusted_marker.g   adjusted_marker.g 
         1.44495380         -0.02545743 

stopifnot(qr(model.matrix(adjusted))$rank==3)
plot(g,y,col=ifelse(ancestry==0,"#1b9e77","#d95f02"),pch=19,
 xlab="Counted-allele dosage",ylab="Hypothetical phenotype")

The generating model contains no direct marker effect. Group differences can nevertheless create a marginal marker association. One realization does not make the adjusted estimate exactly zero.

Define the association at the measurement level

A GWAS relates genotype to a stated phenotype; an eQTL analysis relates genotype to expression in a tissue and context. A cis label usually denotes a defined genomic neighborhood, while trans refers to more distant relationships under the study’s convention. Proximity is a location rule, not proof of direct regulation.

A TWAS signal relates genetically informed expression to a trait under a specified procedure. Shared genetic associations can arise through LD or pleiotropic effects rather than expression mediation. Colocalization and fine-mapping ask additional questions under their own assumptions about variants, signals and LD. They should not be collapsed into a single claim that a gene causes the trait.

Analysis	Direct target	Interpretation still conditional on
GWAS	Variant–trait association	Structure, relatedness, phenotype and allele alignment
eQTL	Variant–expression association	Tissue, assay, design and normalization
TWAS	Expression-related genetic association with trait	Prediction weights, LD and association model
Fine-mapping	Candidate causal configurations	Candidate coverage, LD, model and priors
Prediction	Performance for new observations	Information timing and deployment population

LD links signals without isolating a mechanism

Correlated markers can carry similar association evidence when one tags another. A near-perfect correlation also makes separate regression effects uncertain, even when joint prediction is stable.

set.seed(1302)
g1 <- rbinom(100,2,.4)
g2 <- g1;g2[1:12] <- rbinom(12,2,.4)
y <- .8*g1+rnorm(100)
c(dosage_correlation=cor(g1,g2),
 marginal_g1=coef(lm(y~g1))[2],marginal_g2=coef(lm(y~g2))[2])

dosage_correlation     marginal_g1.g1     marginal_g2.g2 
         0.9347726          0.7386323          0.7341350 

coef(summary(lm(y~g1+g2)))

              Estimate Std. Error    t value  Pr(>|t|)
(Intercept) -0.1954570  0.1402732 -1.3934023 0.1666828
g1           0.2715971  0.4062394  0.6685641 0.5053615
g2           0.4867802  0.3957958  1.2298771 0.2217178

The supplied generator gives g1 a direct role; in an observational dataset that truth is unknown. A strong g2 association does not independently identify another mechanism. In real analyses, LD depends on the population and can differ from a reference panel.

Relatedness and phenotype definition matter

Livestock samples often share family relationships and management. A residual covariance structure can represent polygenic relatedness, while fixed covariates describe selected environmental differences. Neither should be added mechanically: the phenotype and target comparison determine what belongs in the model.

For example, adjusting for body weight changes a feed-intake comparison relative to analysing intake alone. If a covariate is affected by treatment or shares unmeasured causes with the outcome, conditioning can change the causal interpretation. State the estimand rather than assuming every available covariate improves the model.

Allele alignment is a scientific consistency check

Effect estimates require the counted allele. Flipping allele coding reverses the coefficient and changes the intercept but can preserve fitted values. Reference/alternate status is not the same as favorable/unfavorable status. For cross-study integration, check assembly, strand, allele pairs and frequencies before combining signs.

An expression-associated variant can have different effects across tissues or environments. Failure to observe an effect in another tissue can reflect context, limited power or assay differences. Report what was measured and how precisely before asserting biological specificity.

Move from a signal to a follow-up

Choose follow-up to discriminate explanations: replication in a new population addresses generalization, improved mapping addresses candidate ambiguity, and a perturbation addresses a proposed mechanism. A validated predictor can still be a proxy, while a mechanistic result may not improve prediction at the decision time.

Candidate-gene challenge

The nearest gene to a GWAS peak also changes expression with diet. Does the agreement establish mediation? No. Position, expression response and trait association are distinct evidence. Examine LD, tissue/context, temporal order and competing explanations before a causal claim.

Interpretation and scope

Adjustment can reduce specified confounding but cannot guarantee that all confounding has been measured. Do not infer causality from conditional association, a TWAS signal or agreement between assays. Fit and validate appropriate models before applying any teaching formula to related livestock data.

Exercises

1. Why can a variant associate with a trait without changing it?
2. What is the eQTL response variable?
3. Why can randomly splitting relatives exaggerate deployment performance?

Suggested answers

1. LD with causal variation or association with ancestry/environment can produce the signal.
2. A specified expression measurement in a stated tissue/context.
3. Close relatives in training and test can make the task easier than prediction in a new family or population.

Further reading

Methods of integrating data to uncover genotype–phenotype interactions, Nature Reviews Genetics, offers a broader conceptual reading. Use it to distinguish data integration from identification of a causal path.

Practicals and course pathways

Association and regulatory variation, trait association and prediction.

Use the unified map to locate B13 in both harmonized outlines. Return to the building-block index.