Continuous Geographic Structure and Genetic Clines

Why Genetic Variation Can Form a Gradient

Genetic differences among sampled people do not always divide neatly into a small number of distinct groups. If people have moved, intermarried, and exchanged genes across neighboring regions over many generations, allele frequencies may change gradually with location. A map of such variation can show a gradient rather than a sharp boundary. Treating every gradient as evidence for discrete ancestry categories can give a misleading description of the data.

A genetic cline is a systematic change in an allele frequency, genetic summary, or other measured genetic characteristic across a geographic or environmental coordinate. A cline may be gradual across a large region or steep over a short distance. It describes a pattern in the sampled data; by itself, it does not identify the historical or biological process that produced it.

A one-dimensional coordinate is a convenient way to describe a route across a region, not a claim that geography is truly one-dimensional. In a wider area, location can be represented by two coordinates, such as longitude and latitude, and allele frequencies can vary smoothly over a map. The direction and steepness of a gradient can differ across loci: one allele may become more frequent toward the east, while another changes little or varies in a different direction.

Describing an Allele-Frequency Cline

For a simple transect, a linear model describes a constant change in allele frequency per unit of distance:

Here, \(p(s)\) is the frequency of the specified allele at position \(s\), \(a\) is the modelled frequency at \(s=0\), and \(b\) is the change in frequency per unit of distance. If distance is measured in kilometres, \(b\) has units of frequency per kilometre. Because a frequency must be between 0 and 1, a straight line is only a suitable description over a range where its predictions stay within those limits. It is not a sensible model to extrapolate indefinitely.

A logistic model keeps modelled frequencies between 0 and 1 and permits a changing rate of increase across the transect:

In this model, a one-unit increase in \(s\) changes the log odds of the specified allele by \(b\). The corresponding odds are multiplied by \(e^b\) per unit. The frequency itself does not change by a constant amount per unit, as it does under a linear model. Both models are descriptions of a pattern; choosing one does not, on its own, establish a cause for that pattern.

From Geographic Gradients to Principal-Component Clines

Principal component analysis (PCA) summarizes patterns of variation across genotype data. For a biallelic locus \(j\), let \(G_{ij}\) be the number of copies of a chosen allele for person \(i\), coded as 0, 1, or 2. If the allele frequency at that person’s location \(s_i\) is \(p_j(s_i)\), then the expected dosage is \(E(G_{ij}\mid s_i)=2p_j(s_i)\). This equality follows because the expected count is the sum of the probabilities of the two allele copies; Hardy–Weinberg equilibrium is not needed for the mean.

A PCA score is a weighted combination of genotype measurements across loci. Consequently, if many loci have allele frequencies that change smoothly across geography, their combined genetic scores can also vary smoothly across locations. The resulting pattern is often called a principal-component cline. A smooth trend in a PC score can summarize a broad direction of genetic variation; it does not mean that the score is a measurement of distance or a proportion of a particular ancestry.

Here, \(z_{ij}\) is the genotype measurement for person \(i\) at locus \(j\), after any centering or scaling used in the PCA, and \(w_j\) is the corresponding component weight. If the expected value of each contributing \(z_{ij}\) changes smoothly with location, then the expected weighted sum can change smoothly too. The PC sign is arbitrary: reversing every weight reverses the scores without changing the underlying direction of variation. A plotted gradient therefore needs to be interpreted using the locations, the genotype data, and the chosen sign convention.

A Short Justification for a Geographic Association

The simple linear cline also illustrates why a geographic trend in allele frequency can produce a geographic trend in expected genotype dosage. Suppose that, over the sampled range, \(p(s)=a+bs\), and let \(G\) be the allele dosage for a randomly selected person at location \(S\). The conditional mean is \(E(G\mid S)=2(a+bS)\). Using the covariance identity \(\operatorname{Cov}(S,G)=\operatorname{Cov}\{S,E(G\mid S)\}\), we obtain:

The constant \(2a\) contributes no covariance, and multiplying \(S\) by \(2b\) multiplies its variance with \(S\) by \(2b\). Thus, if \(b>0\) and locations vary, expected dosage and location have positive covariance; if \(b<0\), the covariance is negative. This is a statement about the modelled association, not a proof that geographic location caused the genetic pattern. It also depends on using the linear model over the range being considered.

What a Cline Does—and Does Not—Establish

A smooth pattern can arise through multiple histories and processes, including movement and mating among neighboring populations, past migration, admixture, or spatially varying selection. The shape alone does not distinguish among these explanations. Nor does a gradient necessarily imply that all loci follow the same pattern: a broad genome-wide trend and a steep cline at one particular locus are different observations that may call for different explanations.

Clines are useful because they describe genetic variation without forcing locations into a fixed number of groups. This matters when a clustering model or an ancestry label would imply sharper divisions than the data support. Conversely, a smooth map does not prove that a population is homogeneous, that geographic distance is the only relevant factor, or that there are no meaningful local differences. The description should match the resolution and distribution of the samples.

A principal-component cline is especially easy to overinterpret. PCA finds directions of variation in the analyzed genotype data; it does not take geographic coordinates as proof of a causal relationship. A PC score can be associated with geography because allele frequencies covary with location, but it is not inherently a geographic coordinate, an ancestry fraction, or a biological category. Its interpretation depends on which individuals and variants were analyzed and on how genotypes were prepared.

Common Mistake: Turning a Gradient into a Boundary

• Wrong reasoning: “PC1 separates people on the map, so the positive and negative scores represent two distinct ancestry groups.”
• Correction: A continuous change in allele frequencies can produce a continuous change in PC scores. Check whether scores vary gradually with location before describing them as evidence of a discrete division.
• Wrong reasoning: “The logistic slope is the allele-frequency increase per kilometre.”
• Correction: In a logistic model, the slope is the change in log odds per kilometre. The absolute frequency change depends on the current frequency.
• Wrong reasoning: “The pattern proves that geography caused the genetic differences.”
• Correction: A cline describes an observed spatial pattern. Additional evidence is needed to distinguish possible historical or biological explanations.

Descriptions also depend on where samples were collected. If only widely separated locations are sampled, a gradual change between them can appear to be a jump. If only a narrow region is sampled, a broader curved pattern may look nearly linear. Reporting the sampled geographic range and avoiding unsupported extrapolation help keep the interpretation tied to the evidence.