Reading PCA and Kinship Together
A principal-components plot and a pairwise kinship table describe different aspects of genomic similarity. Principal component analysis (PCA) summarizes major directions of variation across samples: points that are close together have similar scores on the displayed components. A pairwise kinship estimate instead summarizes genetic sharing for a particular pair under an allele-frequency and estimation convention. The central diagnostic task is to decide whether a pattern in the plot is consistent with ancestry structure, close relatedness, duplicated samples, or more than one of these.
Neither display supplies a complete interpretation by itself. A cluster in a PCA plot is not automatically a family, and a close pair in the plot is not automatically related. Conversely, a pair of close relatives need not be conspicuous on the first two PCs if ancestry variation dominates those axes. Interpret the plot and pairwise estimates jointly, and treat any labels such as “ancestry group” or “duplicate” as hypotheses to check against the data and study context.
The expected kinship coefficient for unrelated individuals in a specified reference population is approximately zero under the usual convention. Common pedigree expectations are \(\phi=0.5\) for identical individuals (including monozygotic twins), \(\phi=0.25\) for parent–offspring and full-sibling pairs, and \(\phi=0.125\) for second-degree pairs such as grandparent–grandchild or half-siblings. These are expectations, not exact boundaries for every realized genomic estimate. Estimates vary with the estimator, marker set, allele-frequency choices, and sampling variation.
Assumptions and Interpretive Limits
A useful interpretation assumes that sample identifiers are correct, genotype calls are sufficiently reliable, and the PCA and kinship estimates were computed from a suitable set of markers. It also assumes the plotted PCs are based on comparable genotype data across samples. If missingness or technical batch effects align with sample groups, the PCs can reflect those artifacts rather than ancestry. If allele frequencies used for kinship estimation do not represent the relevant population structure, ancestry differences can also distort pairwise estimates. These possibilities should be checked before assigning a biological explanation.
PCA axes have no intrinsic ancestry names. Their signs can be reversed without changing the analysis, and a rotation within a tied-variance subspace can change displayed coordinates while preserving the represented structure. Thus, “positive PC1” is not itself a biological category. Clusters and gradients are patterns in the selected components, not proof of discrete populations or of a particular migration history. The appearance also depends on which samples are included, which PCs are displayed, and how the markers were selected.
Pairwise kinship values are likewise estimates, not labels. A threshold can be useful for flagging pairs for review, but no universal threshold separates every pedigree class for every estimator. Compare a value with the expected ranges relevant to the estimator and study, inspect its stability under sensible marker or frequency choices, and consider the PCA evidence without treating it as a substitute for the estimate.
Worked Diagnostic Cases
Worked Example: Separate Ancestry Clusters Without Inferring Families
Suppose a hypothetical PCA of four samples gives the following PC1 and PC2 scores: A \((-2.0, 0.1)\), B \((-1.8,-0.1)\), C \((1.9,0.0)\), and D \((2.1,0.2)\). A and B lie near one another, as do C and D; the two pairs are far apart on PC1. Estimated kinship coefficients are \(\phi_{AB}=0.004\), \(\phi_{CD}=0.006\), and \(\phi_{AC}=0.001\).
For a simple visual comparison, calculate the Euclidean distance in the displayed two-PC plane. For A and B, \(d_{AB}=\sqrt{(-2.0-(-1.8))^2+(0.1-(-0.1))^2}=\sqrt{0.04+0.04}=0.283\). For A and C, \(d_{AC}=\sqrt{(-2.0-1.9)^2+(0.1-0.0)^2}=\sqrt{15.21+0.01}=3.901\). So A and B are much closer in this projection than A and C.
That visual closeness does not establish a close pedigree relationship. The estimated \(\phi_{AB}=0.004\) is far below the usual expectation for a second-degree pair (\(0.125\)) and is consistent with little detectable close relatedness under this estimator. The plot supports a shared position in the displayed ancestry structure; it does not establish why A and B are similar. The same reasoning applies to C and D. Their respective kinship estimates are small, so the two visible clusters should not be described as family clusters on this evidence.
The plot is only a two-dimensional projection, and the distance calculation uses only PC1 and PC2. It is a visual summary, not a kinship estimator or a test of relatedness. The appropriate conclusion is that these samples separate into two groups on the displayed PCs, while the reported pairwise estimates do not indicate close relatedness for the within-group pairs.
Worked Example: Recognize a Gradient and Check Pairwise Kinship
In another hypothetical data set, five samples have PC1 scores \(-2.0,-1.0,0.0,1.0,2.0\), and PC2 scores \(0.1,0.0,-0.1,0.0,0.1\). They form an orderly progression along PC1 rather than two clearly separated groups. The kinship estimate for the two endpoints is \(\phi_{15}=0.003\), and estimates for adjacent samples range from \(0.002\) to \(0.009\).
The endpoint distance in the displayed plane is \(d_{15}=\sqrt{(-2.0-2.0)^2+(0.1-0.1)^2}=\sqrt{16}=4.000\). The adjacent distance between the first and second samples is \(d_{12}=\sqrt{(-2.0-(-1.0))^2+(0.1-0.0)^2}=\sqrt{1.01}=1.005\). The displayed scores therefore show a gradual change, not two isolated clusters.
The small pairwise estimates, including \(\phi_{15}=0.003\), do not indicate close relatedness between the endpoints. A plausible diagnostic description is a continuous ancestry-related gradient, subject to checks for technical patterns and the choice of included samples. It would be incorrect to call the samples a chain of relatives merely because each adjacent pair is closer on PC1 than the endpoints. PCA proximity reflects the displayed variation; it does not state a pedigree relationship.
A gradient is not proof of admixture in a particular historical sense. Reference samples, sampling design, marker choice, or a technical variable can affect the pattern. The cautious interpretation describes what the plot shows and uses kinship estimates to evaluate close relatedness separately.
Worked Example: Distinguish Duplicate Samples From Close Relatives
Consider three hypothetical samples. X and Y have nearly identical scores on the first two PCs: X is \((0.42,-0.17)\), and Y is \((0.43,-0.16)\). Their estimated kinship coefficient is \(\phi_{XY}=0.492\). Sample Z has scores \((0.45,-0.14)\), close to both X and Y, but \(\phi_{XZ}=0.247\) and \(\phi_{YZ}=0.251\).
The kinship estimates can be placed on the usual genomic relationship scale by multiplying by two. For X and Y, \(A_{XY}=2(0.492)=0.984\), close to the expected relationship value of 1 for identical samples. For X and Z, \(A_{XZ}=2(0.247)=0.494\), and for Y and Z, \(A_{YZ}=2(0.251)=0.502\), both near the expected value of 0.5 for parent–offspring or full siblings.
The evidence suggests that X and Y may be duplicate samples or otherwise genetically identical, while Z is closely related to each. Their nearly coincident PCA positions are compatible with this conclusion, but do not distinguish duplicates from monozygotic twins. The kinship values also cannot by themselves determine whether Z is a parent or a full sibling. Sample records, quality checks, and additional pedigree information are needed to resolve those distinctions.
This example illustrates why both displays matter: PCA position highlights overall similarity on the plotted axes, while kinship supplies pair-specific evidence on a relationship scale. Even a high estimate should prompt checks of sample identity and genotype quality rather than an automatic biological claim.
A Scale Check for Pairwise Relationships
The conversion between kinship and the usual additive genomic relationship scale is important because the two scales have different reference values. At a locus, the kinship coefficient is the probability of identity by descent for one randomly selected allele from each individual. The additive relationship scale counts the contribution from both alleles in the covariance of additive genetic values. Under the standard convention, that covariance is \(2\phi_{ij}\sigma_a^2\), where \(\sigma_a^2\) is the additive genetic variance at the locus or under the corresponding model. Dividing by \(\sigma_a^2\) gives \(A_{ij}=2\phi_{ij}\). Thus, a kinship estimate near \(0.25\) corresponds to an off-diagonal relationship value near \(0.5\), not \(0.25\).
This scale identity does not make all software outputs interchangeable. Confirm whether a reported value is a kinship coefficient, an additive genomic relationship, or another estimator-specific quantity before comparing it with pedigree expectations. A factor-of-two mistake can change a pair's apparent relationship class.
How to Diagnose a Plot-and-Estimate Pattern
State whether the samples form separated groups, a continuous gradient, isolated points, or nearly coincident pairs. Name the PCs shown; do not assign ancestry labels from axis signs alone.
For pairs that appear close or unusual, compare the estimated \(\phi_{ij}\) with the estimator's convention and expected pedigree values. Keep kinship and genomic relationship scales distinct.
A PCA cluster with low pairwise kinship is compatible with shared ancestry structure without close relatedness. A kinship estimate near a close-relative expectation can indicate relatedness even when the pair is not uniquely separated on the displayed PCs.
Review sample identity, missingness, batch, marker selection, allele-frequency choices, and reference composition. If the pattern changes substantially under reasonable checks, report that instability rather than giving it a firm biological interpretation.
Useful diagnostics include coloring the PCA plot by known technical batches and checking whether missingness or call quality tracks PC scores. Inspecting additional PCs can reveal separation not visible on the first two, but adding axes does not automatically make the interpretation more reliable. Similarly, a kinship estimate should be considered in relation to its estimation method and input data, rather than treated as a direct pedigree record.
Common Pitfall and Alternatives
Wrong reasoning: “A and B are close together on the PCA plot, so they must be close relatives.” This confuses similarity along selected population-variation directions with pairwise identity by descent. An ancestry cluster can contain many unrelated individuals. Correction: describe the PCA pattern as shared position on the displayed PCs, then evaluate close relatedness using pairwise kinship estimates and their estimation assumptions.
PCA and pairwise kinship are complementary, not competing procedures. PCA is especially useful for visualizing broad structure across many samples; pairwise kinship is more directly suited to identifying unusually related pairs. A kinship matrix alone can be hard to interpret as a global pattern, while a two-PC plot may conceal relatedness or variation on later components. Combining them helps identify cases that deserve review, but neither alone resolves every ambiguity.
The limitations are substantive. A two-dimensional plot discards variation outside its axes. Continuous structure can look like separate groups after selective sampling, and technical artifacts can mimic either groups or gradients. Relatedness estimates can be biased or unstable when their allele-frequency assumptions are poor, when data quality varies, or when markers provide limited information. Therefore, a diagnostic conclusion should say what the evidence supports, identify unresolved alternatives, and avoid asserting ancestry or pedigree beyond that evidence.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A pair lies in the same PCA cluster and has \(\phi_{ij}=0.006\). What does the combination suggest, and what does it not establish?
- Convert a kinship estimate of \(0.125\) to the usual off-diagonal genomic relationship scale. Which common pedigree expectation does the kinship value match?
- Two samples are nearly coincident on PC1 and PC2, but their kinship estimate is \(0.24\). Is that evidence for duplicate samples? Explain.
- A PCA plot shows a smooth line of points, and adjacent pairs have small kinship estimates. Give a cautious interpretation.
- Name two checks that could help distinguish a biological pattern from a technical artifact.
Solutions. (1) The samples share a position in the displayed PCA structure, and the low estimate does not indicate close relatedness under the usual convention. The cluster does not establish a particular ancestry history. (2) \(A_{ij}=2(0.125)=0.250\); \(\phi=0.125\) is the usual second-degree expectation. (3) No. A kinship estimate near \(0.24\) is much closer to the usual parent–offspring or full-sibling expectation of \(0.25\) than to the duplicate expectation of \(0.5\); check the estimator convention and sample records before naming the relationship. (4) The plot is consistent with a gradient in the displayed variation, not necessarily a chain of relatives; small adjacent estimates do not support close relatedness. (5) For example, compare PC scores with batch and missingness, and check whether conclusions are stable under reasonable marker or allele-frequency choices.