The Invisible Epistasis

Polygenic traits — height, blood pressure, disease risk — are shaped by hundreds or thousands of genetic variants, each contributing a small effect. The standard simplification treats these loci as additive: the phenotypic effect of each allele is independent of which alleles sit at other loci. Epistasis — the interaction between alleles at different loci — is acknowledged in principle but ignored in most quantitative genetics because its contribution to phenotypic variance appears small under stabilizing selection.

This paper (arXiv:2603.27255, March 2026) shows that the simplification is correct at the phenotypic level and wrong at the genetic level, simultaneously.

Using diffusion theory for a diploid population under stabilizing selection with symmetric mutations in linkage equilibrium, the authors identify parameter regimes where epistatic interactions substantially reshape allele frequency distributions at individual loci. Below a threshold effect size, allele frequencies are unimodal — a smooth distribution centered on intermediate values. Above that threshold, they become bimodal — alleles tend toward fixation or loss, with the intermediate region depleted. The transition mirrors the deterministic case where stable equilibria bifurcate into bistability.

The counterintuitive finding: these changes in the microscopic allele frequency landscape leave macroscopic phenotypic statistics essentially unchanged. The mean deviation from the phenotypic optimum and the genic variance are well captured even when epistatic interactions are entirely neglected. The alleles redistribute, but the redistribution cancels at the phenotypic level. What happens at one locus is compensated by what happens at another, and the aggregate phenotype — the quantity that selection actually sees — barely notices.

The mechanism is compensation across loci. Under stabilizing selection, the phenotype is constrained near an optimum. If epistasis shifts an allele at one locus toward higher frequency, it simultaneously constrains alleles at other loci to compensate, maintaining the phenotypic sum near the optimum. The degrees of freedom are genetic; the constraint is phenotypic. The system has many ways to hit the same target, and epistasis reshapes which of those ways the population uses without changing the target it hits.

This creates a specific kind of invisibility. Any measurement at the phenotypic level — quantitative trait loci mapping, genome-wide association studies, heritability estimates — will correctly conclude that epistasis is negligible. The standard models will fit. The predictions will work. But the underlying allele frequency architecture — which alleles are common, which are rare, which loci are polymorphic — will be wrong. The map is correct for navigation and wrong for geology. The surface matches; the subsurface does not.

The structural lesson is that phenotypic equivalence does not imply genetic equivalence. Two populations can have identical phenotypic distributions — same mean, same variance, same response to selection — while differing completely in their allele frequency architectures. This matters when the question shifts from "what does the population look like?" to "what can the population become?" — because the genetic architecture, not the phenotypic summary, determines the accessible evolutionary trajectories. A bimodal allele frequency distribution and a unimodal one respond differently to novel selection pressures, even if they produce the same current phenotype. The invisible epistasis is invisible only to the questions we're currently asking.