Latent Population Structure and Ancestry Proportions

Why Population Structure Matters

A genetic sample can combine people whose genomes trace, in different proportions, to populations with different allele frequencies. If those source differences are not directly observed, the resulting structure is latent: it is a feature of the data-generating population that must be described through a model, rather than a recorded group label. Recognizing this structure helps describe who is represented in a sample and what a genome-wide ancestry summary means.

This tutorial focuses on two connected ideas. Population structure refers to differences in allele frequencies across latent source populations. An individual’s ancestry proportions summarize the fractions of that individual’s genome attributed to those sources. These are model-based population-genetic descriptions; they do not, by themselves, establish social identity, geographic origin, or a person’s membership in a uniquely defined natural group.

Definition: Population structure is variation in allele frequencies across subpopulations or source populations represented in a sample. It is latent when the relevant source membership or ancestry is not directly observed. A genome-wide ancestry proportion \(q_{ik}\) is the fraction of individual \(i\)’s genome attributed, under a specified model, to ancestral source \(k\).

Mixture Proportions as a Genome-Wide Summary

Suppose a model specifies \(K\) ancestral source populations. For individual \(i\), write the ancestry vector as \((q_{i1},q_{i2},\ldots,q_{iK})\). Each entry is a proportion, so it cannot be negative, and the proportions across all sources must account for the whole genome:

q_{ik} \geq 0 \quad \text{for each } k, \qquad \sum_{k=1}^{K} q_{ik}=1.

For example, with two sources, an ancestry vector of \((0.65,0.35)\) describes a genome attributed 65% to the first source and 35% to the second, under the chosen model. The entries are not probabilities that the person “is” one source or another. They describe a mixture, and more than one entry can be positive.

The word ancestral here names a source component in a statistical description. The source populations must be defined for the analysis, and the proportions are interpreted relative to that definition. They are not automatically exact fractions of a person’s family tree, nor do they imply that human populations have sharp, universal boundaries. Different choices of source populations can produce different proportions for the same individual.

A genome-wide proportion is an aggregate summary across the genome. It does not specify which source is represented at each particular genomic location. In particular, a global vector cannot tell us whether ancestry is distributed evenly across the genome or concentrated in different stretches. That distinction matters, but a genome-wide summary alone does not resolve it.

Worked Example: Reading a Two-Source Ancestry Vector

Suppose a hypothetical analysis uses two defined source populations and reports the ancestry vector \((0.72,0.28)\) for one individual. Interpret the entries and check that they form a valid vector.

The first entry attributes 72% of the individual’s genome-wide ancestry to source 1; the second attributes 28% to source 2. The entries are nonnegative, and

$$ 0.72+0.28=1.00. $$

Thus the vector satisfies both constraints for ancestry proportions. The result does not mean there is a 72% probability that the individual belongs to source 1. It is a two-component genome-wide mixture summary, conditional on the source definitions and model used to produce it.

How Allele-Frequency Differences Define Structure

To describe structure, consider a particular allele, such as allele \(A\), at a specified genetic locus. Let \(p_k\) be the frequency of \(A\) in source population \(k\). A sample is structured with respect to this allele when its source populations have different frequencies; for instance, \(p_1=0.20\) and \(p_2=0.80\). If the frequencies are the same in all sources at that locus, there is no frequency difference there to reveal structure, even if other loci differ.

A basic mixture calculation follows from the definition of a mixture. For a randomly selected allele copy, suppose its source is \(k\) with probability \(w_k\), where the weights are nonnegative and sum to one. Conditional on source \(k\), suppose the probability that the allele copy is \(A\) is \(p_k\). Then the overall probability of observing \(A\) is the sum of the source-specific probabilities, weighted by the source proportions.

Mixture identity: If a randomly selected allele copy comes from source \(k\) with probability \(w_k\), and the frequency of allele \(A\) in that source is \(p_k\), then its marginal probability of being \(A\) is \(\sum_{k=1}^{K} w_kp_k\). The weights describe the mixture being averaged; they must correspond to the population or genomic summary under discussion.

This identity is a consequence of conditioning on source. The possible sources partition the event that an allele copy is selected, so the probability of \(A\) is the sum of the probability of selecting each source and then observing \(A\) within it:

P(A) = \sum_{k=1}^{K} P(A \mid \text{source } k)P(\text{source } k) = \sum_{k=1}^{K}p_kw_k.

The calculation is valid when the source weights and source-specific frequencies refer to the same mixture and the same allele. In a sample of people, weights might describe the fractions sampled from each source. For an individual’s genome, ancestry proportions describe a genome-wide mixture. They should not be substituted for the exact source composition at a particular locus unless the model or calculation justifies doing so.

Worked Example: A Pooled Allele Frequency Can Hide Structure

Imagine a hypothetical sample containing equal numbers of individuals from two source populations. At one locus, allele \(A\) has frequency \(0.20\) in source 1 and \(0.80\) in source 2. Calculate the pooled frequency and compare it with the two source-specific frequencies.

The source weights are \(w_1=0.50\) and \(w_2=0.50\). Using the mixture identity,

$$ P(A) = w_1p_1+w_2p_2 = (0.50)(0.20)+(0.50)(0.80) = 0.10+0.40 = 0.50. $$

The pooled frequency is \(0.50\), even though neither source population has that frequency: one has \(0.20\), and the other has \(0.80\). The pooled value is a correct summary for this mixture, but by itself it conceals the difference between sources. This is one reason a sample-wide frequency can be insufficient to describe a structured sample.

Genome-Wide Proportions and Weighted Summaries

An individual’s ancestry vector can be used to express a genome-wide weighted summary of source-specific allele frequencies. To make the interpretation precise, suppose each source has a specified average frequency \(p_k\) for allele \(A\) across the set of loci being summarized, and \(q_{ik}\) is the fraction of individual \(i\)’s genome attributed to source \(k\). Under that summary model, the expected genome-wide fraction of allele copies that are \(A\) is the weighted average \(\sum_k q_{ik}p_k\). This is an average over the stated genomic summary, not a claim that every locus has the same ancestry composition or allele frequency.

The assumptions are important: the source labels and source-specific frequencies must be defined; the ancestry proportions must describe the individual’s genome under those source definitions; and the frequencies and proportions must refer to compatible genomic summaries. If those conditions are not met, the weighted average may not have the intended interpretation.

Worked Example: A Weighted Genome-Wide Allele Summary

For a hypothetical illustration, suppose two defined sources have average frequency \(0.10\) and \(0.60\) for allele \(A\) across a specified set of loci. An individual’s genome-wide ancestry vector is \((0.70,0.30)\). Under the weighted-summary assumptions just stated, calculate the expected genome-wide fraction of allele copies that are \(A\).

The weights sum to one, \(0.70+0.30=1.00\), and both are nonnegative. The weighted average is

$$ (0.70)(0.10)+(0.30)(0.60) = 0.07+0.18 = 0.25. $$

The model-based genome-wide summary is therefore \(0.25\), or 25% of allele copies. This is an expected aggregate under the stated assumptions, not a claim that allele \(A\) occurs at frequency \(0.25\) at every locus in this individual’s genome. Source frequencies vary by locus, and a global ancestry vector does not describe the ancestry composition at each location.

What the Representation Does—and Does Not—Say

The mixture representation is useful because it makes two features explicit: source populations can differ in allele frequencies, and a sampled individual can draw ancestry from more than one source. It also separates the description of an individual’s genome-wide mixture from the description of the population-level allele frequencies that define structure.

The representation is conditional on the source populations selected for the model. It does not establish that those sources are the only meaningful groups, that they are internally uniform, or that ancestry has a single interpretation outside the statistical model. In particular, ancestry proportions should not be treated as measurements of social identity or as evidence of an individual’s personal history without additional information.

A global proportion also compresses variation. Two people can have the same genome-wide ancestry vector while differing in how ancestry is distributed across genomic locations. Conversely, their vectors can differ even when the source frequencies at a particular locus are similar. The vector is informative at the scale it summarizes, but it is not a complete description of genetic variation.

Common Mistakes

Treating proportions as mutually exclusive labels. A vector such as \((0.72,0.28)\) does not assign the person to one source with a 72% probability. It describes a mixture across sources.
Ignoring the sum constraint. If proposed proportions are negative or add to something other than one, they are not a valid complete ancestry vector as defined here.
Equating a pooled frequency with the frequency in every source. A pooled frequency is a weighted average. As the calculation above shows, it can differ from every source-specific frequency.
Reading a global summary as locus-by-locus information. Genome-wide proportions do not state which ancestry is present at any particular location.
Giving source labels more meaning than the model supports. The interpretation is relative to the specified sources; proportions alone do not define social identity or natural population boundaries.

Key takeaway: Population structure is expressed through allele-frequency differences among latent sources. A genome-wide ancestry vector records nonnegative source proportions that sum to one, but its interpretation depends on the source definitions and the genomic scale of the summary.

Check Your Understanding

Use the definitions and mixture calculations in this tutorial to answer the following questions.

A hypothetical individual has a three-source ancestry vector \((0.20,0.35,0.45)\). Is it valid? State the two checks you used.
Two source populations have allele \(A\) frequencies \(0.30\) and \(0.70\), and make up 25% and 75% of a sample. Calculate the pooled frequency.
Why does a pooled allele frequency of \(0.50\) not establish that every source population has frequency \(0.50\)?
Explain why an ancestry vector of \((0.60,0.40)\) is not a 60% probability that an individual belongs to the first source.
What information about genomic locations is not provided by a genome-wide ancestry proportion?