Epigenetics

Phantom additivity and the unknown heritability of epistatic interactions

---

Abstract

Additive genomic models, such as GBLUP, are routinely used in livestock genetic evaluation programs. Although these models strictly assume effects as being additive, a proportion of the epistatic variance may be indirectly captured through the correlations between estimated additive effects and the true unknown statistical epistatic effects. This results in phantom additive effects, components that behave additively but only partially reflect the underlying non-additive variance. And of course, this still leaves an untapped proportion of the genetic variance unaccounted for and relegated to the residuals - the unknown heritability that arises from higher-order epistatic interactions. Having a better understanding of the architecture of complex traits is relevant, as it would enable us to move beyond additive methods for estimating breeding values, allowing the adoption of more nuanced approaches better tailored to the specific architecture of a given trait. This study aimed to evaluate how effectively additive models capture epistatic components across varying orders and types of interactions. We compared whether heritability estimates (h²) align with prediction accuracies, which we calculated based on the Pearson correlations between simulated phenotypes and EBVs achieved by GBLUP, and assessed the extent to which correlations between additive and epistatic effects are reflected in both heritability and predictive performance. For the simulations, we used a livestock cattle population of 281,558 animals genotyped at 110,220 SNPs. Phenotypes were generated by varying the number of QTL (10, 100, 1,000, 10,000), the order of epistatic interactions (0, 2, 4, 6), and the relative contributions of broad (H² = 0, 0.5, 1) and narrow (h² = 0, 0.5, 1) sense heritabilities. Scenarios with H² = 0 and h² = 0 served as controls. Simulations were based on real SNP chip data, and heritability estimates were obtained using GREML. Each scenario included a control with no genetic effects, pure additive scenarios, and pure epistatic scenarios with varying orders and numbers of QTL, all replicated to assess model performance. Results showed that increasing epistatic complexity (k = 2, 4, 6) led to a significant decline in GBLUP prediction accuracy (Fâ‚‚,₂₁ = 18.59, P = 2.25 à— 10⁻⁵; R² = 0.639). Estimated marginal means (average prediction accuracy per scenario) decreased from 0.488 (±0.031) at k = 2 to 0.390 (±0.031) at k = 4 (P = 0.035) and 0.226 (±0.031) at k = 6 (P < 0.001), representing a 53.7% relative reduction. All pairwise contrasts were significant (Tukey-adjusted), and model diagnostics confirmed assumption validity. These findings demonstrate that, although additive GBLUP kernels capture part of the epistatic variance and can be used to achieve genetic progress, they do not adequately capture complex genetic architectures, emphasizing the need for explicit modeling of non-additive effects in genomic selection when higher-order interactions predominate.

Keywords: 2026

How to Cite:

Razi, K. & Gondro, C., (2026) “Phantom additivity and the unknown heritability of epistatic interactions”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286902. doi: https://doi.org/10.31274/wcgalp.24158

Rights: 1