Interpreting Genetic Principal Components

What a Genetic Principal-Component Plot Can Tell You

A genetic principal-component analysis (PCA) reduces many genotype measurements to a smaller set of summaries that capture prominent patterns of variation in the analyzed data. A plot of the first two components can make those patterns visible: samples close together have more similar scores on those axes, while samples farther apart have less similar scores on them. But a plot does not, by itself, say what produced the pattern. A separation may be consistent with ancestry differences, geographic structure, technical artifacts, or a combination of these.

Interpretation therefore requires more than naming the apparent groups. Relate scores to the study’s ancestry labels and geographic metadata, examine how much variation the plotted components summarize, and check whether technical variables align with the same pattern. The interpretation should describe the observed structure without turning an axis into a literal ancestry measure for each person.

Scores, Eigenvalues, and Variance Explained

Suppose the genotype matrix has been centered, scaled, and LD-pruned as appropriate for the analysis. Let \(X\) denote the resulting matrix, with individuals in rows and variants in columns. PCA finds directions of variation in this matrix. For component \(k\), the eigenvector \(v_k\) gives the direction across variants, and the score for individual \(i\) is the projection of that individual’s genotype vector onto that direction. If the eigenvalue for component \(k\) is \(\lambda_k\), it describes the amount of variance along that component under the chosen preprocessing and PCA convention.

A score is not an ancestry proportion. Its numerical value depends on the data and preprocessing, and its units are the score units of that PCA. The same person can have different scores if the sample, variants, or preprocessing change. Scores should therefore be interpreted in the context of the analysis that produced them, not as universal quantities.

A scree plot is a descriptive summary, not a test that identifies a unique number of biologically meaningful components. An apparent bend or “elbow” can help describe where successive eigenvalues become smaller, but its location is not a proof that earlier components are ancestry-related or that later ones are noise. Small components may still reflect real structure, while a prominent component can reflect technical variation.

Relating Scores to Labels and Geographic Metadata

An ancestry label is metadata attached to a sample, not an intrinsic interpretation of a PC score. Labels may summarize self-identification, recruitment site, or a reference classification; these are not interchangeable. First identify precisely what the label represents. Then ask whether samples with a shared label tend to occupy similar regions of the score plot, whether labels overlap, and whether geographic metadata show a corresponding pattern.

A pattern in which scores vary gradually across geographic locations can be consistent with broad population structure rather than a set of sharply separated groups. Conversely, substantial overlap among labels does not establish that the labels are incorrect: the PCA may capture only some variation, labels may be broad, or the sampled groups may share genetic variation. A useful interpretation reports what is visible and avoids implying that a two-dimensional plot fully represents ancestry.

Geographic metadata can add context to a score plot. If locations show an orderly pattern across a region, the scores may reflect broad geographic population variation. That is an interpretation to evaluate, not a fact encoded by the axis. If there are multiple sampling sites, compare the pattern with the actual locations and consider whether recruitment or genotyping procedures also differed by site. Geographic agreement is informative, but it does not rule out technical explanations that happen to align with geography.

Why the Sign and Order of an Axis Are Not Biological Labels

The sign of a principal-component direction is arbitrary. If \(v_k\) is an eigenvector for an eigenvalue \(\lambda_k\), then \(-v_k\) is also an eigenvector for the same eigenvalue. The corresponding scores change sign, but the underlying separation is unchanged.

This gives a direct reason not to assign a fixed meaning to “positive PC1” or “negative PC1.” If a plotting program reverses the sign, the same samples appear on opposite sides of zero, but their relative positions and the pattern of separation are preserved. Likewise, the component number describes the order of variance captured under the PCA, not a named ancestry dimension. A component may correlate with an ancestry label in one dataset and have a different pattern in another.

Checking for Technical Structure

Before interpreting a visible split as population structure, compare the scores with available technical metadata. Examples include genotyping batch, plate, array or sequencing platform, missingness, and processing site. A strong alignment between a PC and a technical variable raises concern that the component reflects technical differences. It does not prove an artifact on its own, because technical groups can also differ in ancestry; the design and data quality must be considered.

Useful checks include plotting or tabulating PC scores by technical category, checking whether ancestry labels are balanced across batches, and examining whether the pattern persists after appropriate quality-control decisions. If the technical variable is confounded with geography or ancestry labels, the data alone may not cleanly distinguish those explanations. State that limitation rather than choosing the preferred interpretation without evidence.

A Practical Interpretation Sequence

A disciplined interpretation keeps the description of the plot separate from the explanation for it. The following sequence helps make that distinction explicit.

Common Mistakes and Why They Matter

• Calling PC1 “the ancestry axis.” A component is a direction of variation in the analyzed data, not a universal ancestry dimension. Describe its observed relationship with labels or metadata instead.
• Reading a score as an ancestry percentage. Scores are coordinates in the PCA; they are not proportions that sum to one and should not be reported as ancestry percentages.
• Treating the sign as meaningful by itself. Reversing an eigenvector reverses all its scores without changing the pattern. Interpret relative positions, not positive versus negative as fixed categories.
• Taking an elbow as proof of biological structure. A scree plot summarizes eigenvalues. It does not determine whether a component reflects ancestry, technical variation, or another source.
• Assuming label separation establishes causation or discrete groups. Labels and scores may be associated, but a plot does not establish why. Report the association and evaluate alternative explanations.
• Ignoring technical alignment. A component that tracks a plate or platform may be misleading if interpreted biologically without checking data quality and study design.