advanced 8 min read
Life sciences · Topic
Epistasis
linear algebra · hypothesis testing · probability theory · information theory
Epistasis refers to the interaction between alleles at different loci such that the effect of one locus depends on the genotype at another. Distinguishing statistical epistasis (departure from additive models in a population) from biological epistasis (biochemical pathway interaction) is critical for correct interpretation of GWAS results and variance partitioning.

Variance Decomposition

Fisher’s partition of genetic variance decomposes the genotypic value $G_{ij}$ at two loci into orthogonal components:

\[G_{ij} = \mu + \alpha_i + \alpha_j + \delta_{ij} + (\alpha\alpha)_{ij} + (\alpha\delta)_{ij} + (\delta\delta)_{ij}\]

where $\alpha$ are additive effects, $\delta$ dominance deviations, and the interaction terms constitute epistatic variance $V_I$:

\[V_G = V_A + V_D + V_{AA} + V_{AD} + V_{DD} + \cdots\]

In large outbred populations, $V_{AA}$ (additive-by-additive epistasis) dominates because allele frequencies weight interaction terms differently.

Statistical Epistasis in Linear Models

A two-SNP interaction test adds a product term:

\[y_i = \mu + \beta_1 g_{i1} + \beta_2 g_{i2} + \beta_{12}\,(g_{i1} \cdot g_{i2}) + \varepsilon_i\]

Significance of $\hat{\beta}_{12}$ indicates statistical epistasis. The multiple testing burden is severe: with $M = 10^6$ SNPs, all pairwise tests number $\binom{M}{2} \approx 5 \times 10^{11}$, requiring thresholds of order $p < 10^{-13}$ for FWER control.

Biological vs. Statistical Epistasis

Aspect Biological epistasis Statistical epistasis
Definition Physical interaction in a pathway Non-additive term in a statistical model
Allele-frequency dependence No Yes
Detectable by GWAS interaction test Not necessarily By definition
Contributes to $V_I$ Partly Depends on model

A biological interaction can appear additive at the population level if interacting alleles are rare or at extreme frequencies.

FWER Inflation and Power

Exhaustive epistasis searches inflate type-I error dramatically. Common strategies include:

  • Restricting search to cis-pairs within LD blocks
  • Testing known pathway gene pairs (biologically informed)
  • Using LASSO-penalised regression to select interaction candidates
  • Bayesian approaches (BSLMM) that shrink interaction effects genome-wide