Mixtures of Genetic Heterogeneity and Phenocopies

Why Separate Heterogeneity From Phenocopies?

A disease can appear in several relatives for different underlying reasons. In one family, a rare inherited variant at a particular locus may explain the pattern. Another family with a similar clinical diagnosis may have a causal variant at a different locus. A third family may include an affected person whose disease is unrelated to the familial genetic condition. A useful model must distinguish these possibilities rather than treating every affected relative as evidence for one shared genetic cause.

Locus-level genetic heterogeneity means that similar disease presentations can be caused by variants at different loci, with different families potentially belonging to different genetic subtypes. A phenocopy is a phenotype resembling the disease of interest that occurs in someone without the specified familial genetic cause. It may reflect environmental exposure, another genetic cause, or another disease process. The phenotype alone does not identify which explanation applies.

The key modeling choice is where the latent mixture occurs. A family-level mixture assigns an etiologic class to each family; a person-level model can instead allow different relatives in one family to have different sources of disease. Those are distinct assumptions and generally produce different likelihoods.

Family-Level Mixtures of Genetic Causes

Suppose the candidate genetic explanations involve \(K\) disease loci. In a family-level model, \(Z_f=k\) means the family’s disease pattern is modeled under the inherited-disease model for locus \(k\). A separate component, \(Z_f=0\), can represent a background or phenocopy-only model. The weights \(\pi_0,\pi_1,\ldots,\pi_K\) describe the proportions of source families assigned to these components under the model.

Within a genetic component, the family likelihood must account for inheritance and phenotype. If relatives’ genotypes are unobserved, they are summed over rather than treated as known. For family data \(D_f\), including observed phenotypes and any available genotype information, a generic component likelihood is

where \(\mathbf{g}_f\) is the collection of family genotypes. The genotype distribution follows the inheritance assumptions for that component. The phenotype term uses the component’s penetrance model: it gives the probability of each observed phenotype conditional on the genotypes. A background component can have likelihood \(L_{f0}=P(D_f\mid Z_f=0)\), specified using the relevant nongenetic disease risks.

This formulation relies on the model being specified before calculating its likelihood. In particular, it requires stated component weights, inheritance assumptions, and phenotype probabilities. A common simplification is that relatives’ phenotypes are conditionally independent given their genotypes and the component. This does not imply that relatives’ phenotypes are independent without conditioning: their genotypes are related by inheritance. Shared environmental effects or other residual dependence require additional modeling.

Phenocopies Within a Pedigree Model

A phenocopy risk is often represented through the phenotype probabilities within a genetic model. For example, a noncarrier may have a small probability \(q\) of being affected even when the family’s disease model is genetic. A carrier may have a larger probability of being affected, but not necessarily a probability of 1. These probabilities describe observed phenotype conditional on genotype; they do not reveal the biological cause of any one affected person.

Care is needed not to add a phenocopy risk twice. If the stated penetrance for a genotype already includes background cases, it is already an overall disease probability for that genotype. Adding a separate phenocopy probability to it without defining a new causal mechanism can produce an invalid risk, or a risk with no clear interpretation. A model may instead define separate mutually exclusive etiologic components, but then it must state precisely how those components generate phenotype.

Here is a simple pedigree calculation with genotypes treated as known solely to illustrate the phenotype likelihood. Suppose two relatives are carriers and affected, while two noncarriers are unaffected. Under the genetic model, let carrier penetrance be 0.70 and noncarrier phenocopy risk be 0.03. Assume conditional independence of phenotypes given these genotypes and the model. Under the background-only component, suppose every relative has affected probability 0.03. The identities of the two affected relatives are specified, so no combinatorial factor is needed.

Different Mixture Levels Give Different Likelihoods

A family-level mixture chooses one component for the family. Conditional on that shared component, relatives contribute to the same component likelihood. By contrast, a person-level mixture allows each relative to have an individual latent source. It is generally incorrect to switch between these interpretations without changing the likelihood.

For two relatives with observations \(D_1,D_2\), suppose a family-level class \(Z\) has weights \(\pi_z\), and the observations are conditionally independent within a class. The family-level likelihood is

If instead each person independently receives a component label with the same weights, the likelihood is

These expressions differ because the first shares a latent family class, whereas the second assigns latent classes separately. For example, with two equally weighted components and \(P(D_i\mid 1)=0.8\), \(P(D_i\mid 0)=0.1\) for each person, the family-level probability is \(0.5(0.8^2)+0.5(0.1^2)=0.325\). The person-level probability is \([0.5(0.8)+0.5(0.1)]^2=0.2025\). The data and component-specific probabilities are the same; the latent structure is not.

Mixtures Across Loci and Across Families

A genetic-heterogeneity model can have several locus-specific components, each with its own likelihood. The mixture does not average the loci’ penetrance parameters or treat them as one composite locus. It averages their probabilities for the observed data using the stated component weights. The following calculation illustrates how the evidence can update support for particular components.

For several families, the location of the mixture relative to the product also matters. If each family independently belongs to a component with population weights \(\pi_z\), then the likelihood across independent families is \(\prod_f \sum_z \pi_zL_{fz}\). If instead one common component is assumed to explain every family, the likelihood is \(\sum_z \pi_z\prod_fL_{fz}\). The first allows family-to-family locus heterogeneity; the second assumes one shared component for the entire set. The choice must match the scientific model.

When family ascertainment depends on component, conditioning must also respect the mixture. If \(A\) denotes the event that a family was selected, then under source-population weights \(\pi_z\),

Thus the component proportions among selected families need not equal the source-population proportions: components with higher selection probabilities can be overrepresented. Ignoring this difference can distort both the fitted mixture weights and the apparent evidence for genetic heterogeneity.

Common Mistakes and Interpretation

• Calling every affected noncarrier a phenocopy with certainty. A model may assign that observation a phenocopy probability, but the individual’s actual cause is not observed merely from genotype and phenotype.
• Confusing heterogeneity with a phenocopy. Different causal loci across families describe genetic heterogeneity. A phenocopy is a disease-like phenotype not explained by the specified familial genetic cause. Both can occur in one population and require distinct components or risk terms.
• Mixing component likelihoods at the wrong level. Summing over one class per family is not equivalent to assigning a separate class to every relative. State whether the latent class is shared by a family, an individual, or the whole sample.
• Reading a component weight as a posterior probability. The weight \(\pi_z\) is a model-level prior proportion; the posterior probability for a particular family also depends on its component likelihoods.
• Adding background risk to a penetrance that already includes it. Define whether penetrance is an overall phenotype probability or a probability for one etiologic mechanism before combining risks.