What a Genotype PCA Plot Shows
A genotype PCA plot places each sample at coordinates formed from its principal-component scores. The earlier material on principal components established how those scores are computed from a genotype matrix. Here the task is interpretation: what patterns in the plotted points can reasonably suggest, and what they cannot establish.
Each point represents one sample. In a plot of PC1 against PC2, its horizontal and vertical coordinates are that sample’s scores on the first and second principal components. These components summarize directions of variation in the analyzed genotype data; PC1 captures the greatest variance among the components, and PC2 captures the greatest remaining variance subject to being orthogonal to PC1. The axes are not geographic coordinates, ancestry percentages, or direct measures of relatedness.
These are descriptions of the plotted geometry, not conclusions about why the pattern exists. A separated cluster may be consistent with broad ancestry differences when it also agrees with suitable sample information, but PCA does not assign an ancestry label on its own. A gradual arrangement of points may indicate continuous structure in the analyzed data; it does not by itself identify the source of that structure. Likewise, an outlying point is a reason to investigate, not proof of a technical error or an unusual biological history.
Conditions for a Meaningful Reading
Interpretation depends on how the plot was produced. Samples should have scores computed from the same genotype representation, variant set, and preprocessing choices. If samples were projected using a different variant set or scaling procedure, their displayed positions may not be directly comparable. The plot should also include enough relevant samples to provide context: an isolated point can look less unusual when additional samples are shown.
Population structure is one possible source of genotype variation, but technical differences can also shape a plot. Differences in genotyping quality, missingness, or processing can align with sample groups and contribute to apparent separation. Check sample labels and available technical and biological metadata before interpreting a visual pattern. If a pattern is driven mainly by technical structure, describing it as ancestry would be misleading.
There are two further limits to keep in mind. First, the sign of a principal-component axis is arbitrary: flipping all PC1 scores from positive to negative leaves the underlying configuration unchanged. So “left” and “right” have no inherent biological meaning. Second, a two-dimensional plot omits other components. Samples that look close on PC1 and PC2 may differ on components not shown, while apparent separation in the displayed view does not capture every aspect of genetic similarity.
Clusters and Continuous Patterns
A compact cluster indicates that its members have similar scores on the components shown. If several clusters are visibly separated, the plot suggests that the analyzed samples contain distinct patterns of genotype variation captured by those components. When well-characterized reference samples are available, their positions can help contextualize the pattern. Even then, the appropriate conclusion is limited: the data show separation along the plotted components that is consistent with the relevant reference information. PCA alone does not establish a person’s identity, full ancestry, or the cause of the separation.
A continuous arrangement can be described as a gradient in the plotted scores. The gradual change is a feature of the sample configuration, not a direct measurement of a biological quantity. In particular, a point halfway between two visible groups is not automatically “half” of either group or a fixed proportion of ancestry. More detailed explanations of intermediate positions require additional evidence and are not supplied by the plot alone.
Worked Example: Describing Two Separated Clusters
Suppose six hypothetical samples have the following scores, with all coordinates in principal-component score units.
| Sample | PC1 | PC2 |
|---|---|---|
| A1 | -2.1 | 0.2 |
| A2 | -1.8 | -0.1 |
| A3 | -2.0 | 0.1 |
| B1 | 1.9 | 0.1 |
| B2 | 2.2 | -0.2 |
| B3 | 2.0 | 0.2 |
The mean coordinates of the first set are \(\bigl((-2.1-1.8-2.0)/3,\ (0.2-0.1+0.1)/3\bigr)=(-1.9667,\ 0.0667)\). For the second set, they are \(\bigl((1.9+2.2+2.0)/3,\ (0.1-0.2+0.2)/3\bigr)=(2.0333,\ 0.0333)\). The distance between these mean coordinates is
The centers differ by about 4.000 score units on PC1 and very little on PC2. The points in each set are also close to one another compared with the distance between the centers. The defensible description is that the plotted samples form two compact, well-separated groups, primarily along PC1. If independent sample information supports the relevant interpretation, that pattern may be consistent with broad ancestry structure. The coordinates alone do not supply group labels or explain the cause of separation.
Worked Example: Describing a Gradual Arrangement
Consider five hypothetical samples with scores \((-2,-0.5), (-1,-0.2), (0,0.1), (1,0.4), (2,0.7)\), where each ordered pair is (PC1, PC2). Each step increases PC1 by 1 and PC2 by 0.3. The change from the first sample to the last is \((2-(-2),\ 0.7-(-0.5))=(4,\ 1.2)\), so the plotted points make a steady diagonal progression rather than forming two obvious compact clusters.
A concise interpretation is that these samples exhibit a continuous pattern in the PC1–PC2 view. It would be unjustified to turn the middle sample’s position into an ancestry percentage: its coordinates describe its location in this PCA, not a calibrated mixture scale. The pattern also does not tell us by itself whether the source is population structure, sampling design, or some other feature of the data. Metadata and appropriate checks are needed before proposing an explanation.
Outliers and Close Family Relationships
An outlier is relative to the samples and components displayed. It may represent a sample with genuinely different broad genetic background, a technical problem, or a feature not represented by the main groups. Start by verifying the sample identity and checking available quality information, missingness, processing records, and relevant metadata. Then assess whether the point remains unusual in other useful views or when the sample set is considered in context. An outlier should not be removed simply because it is inconvenient; investigate and document the reason for any exclusion.
Close relatives can appear near one another because they share more inherited genetic material than unrelated individuals. But that does not make PCA a family-relationship test. The first components are constructed to summarize broad variation across the analyzed samples and variants. A pair of close points may share broad population structure without being close relatives; conversely, relatives may not form a conspicuous pair in the first two components if broader structure dominates those axes. PCA coordinates alone do not establish a pedigree or quantify a pair’s degree of relationship.
Worked Example: A Close Pair and an Isolated Sample
Suppose samples F1 and F2 have scores \((0.20,0.10)\) and \((0.28,0.08)\), while sample R has scores \((3.00,2.50)\). The distance between F1 and F2 in the displayed plane is
If the main cloud is near the origin, R appears far from it: its distance from \((0,0)\) is \(\sqrt{3.00^2+2.50^2}=\sqrt{15.25}\approx 3.905\). These calculations describe distances in the two plotted scores. F1 and F2 are close in this view, but that is not enough to call them relatives. R is an outlier in this view, but that is not enough to call it an error or assign an ancestry label. Each observation motivates a separate check of sample information and data quality.
A Practical Interpretation Workflow
A disciplined reading separates what is visible from what is inferred. Describe the plotted pattern first, then consider explanations and checks. This keeps the interpretation tied to the evidence and reduces the risk of treating a convenient visual story as a result.
Check that each point is a sample, identify the components and any stated preprocessing, and note which samples are included.
State whether the visible pattern consists of compact clusters, a gradual arrangement, isolated points, or some combination.
Use relevant sample labels and technical or biological metadata to assess plausible explanations; do not assign labels from position alone.
Use PCA to describe broad patterns in the analyzed data, not to establish close family relationships from a pair’s proximity.
Specify the components shown and make clear that the plot is a low-dimensional view whose geometry depends on the analyzed samples and variants.
Why the Plot’s Geometry Is Not a Relationship Test
The distance between two plotted points is easy to calculate, but it has a restricted meaning: it is distance between two score vectors on the displayed components. For samples \(i\) and \(j\), the two-component distance is
This is a useful description of the plot, but it is not a kinship coefficient and is not a calibrated measure of how closely two people are related. It ignores components not shown and depends on the PCA analysis that produced the scores. A common sign flip leaves the distance unchanged: replacing each PC1 score by its negative changes \(PC1_i-PC1_j\) to its negative, whose square is unchanged. The picture may flip left to right, but this distance does not. This illustrates why axis direction has no inherent ancestry interpretation.
Other approaches answer different questions. PCA is useful for visualizing major patterns of variation in a sample set. Sample metadata can help interpret those patterns, but metadata do not turn a visual position into a formal relationship estimate. Determining close relationships requires methods designed for genetic relatedness, rather than reading them off a PCA plot. The key distinction is the question being asked: broad structure in the plotted scores is not the same quantity as recent family sharing.
Common Mistakes and Limitations
- Assigning ancestry from an axis. A positive PC1 score has no universal ancestry meaning; its sign and orientation are arbitrary, and interpretation requires appropriate external context.
- Calling every separated group an ancestry group. A cluster is an observed pattern. Technical structure or sample selection may also contribute, so check relevant metadata and quality information.
- Reading a midpoint as a percentage. PCA coordinates are scores, not ancestry proportions. Intermediate positions alone do not supply a calibrated biological interpretation.
- Calling nearby samples relatives. Proximity on two components is not evidence sufficient to establish a close family relationship.
- Treating an outlier as a data error. An unusual point deserves investigation, but visual unusualness alone does not justify deleting a sample.
- Overlooking the projection. Two components show only part of the variation summarized by PCA; close points in this view need not be close in every relevant dimension.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A plot shows two compact groups separated mainly along PC1. What can be stated directly, and what additional information is needed to interpret the separation?
- One sample lies between two groups. Does its position establish that it has a particular ancestry percentage? Explain.
- Two samples are very close on PC1 and PC2. What does that establish, and what does it not establish?
- Why should an isolated point be investigated rather than automatically removed?
Solutions. (1) Directly, the plotted scores form two compact groups separated mainly along PC1. Interpreting this as broad ancestry structure requires relevant sample context and checks for technical or sampling explanations. (2) No. The position is a pair of PCA scores; it is not a calibrated ancestry proportion. (3) It establishes that the samples are close in the displayed two-dimensional score plot, but it does not establish that they are close relatives or that they are similar on components not shown. (4) The point could reflect a technical issue, a real difference in the samples, or another feature of the analysis. Quality information, identity, metadata, and context help distinguish possibilities before any exclusion decision.