INAUGURAL ARTICLE by a Recently Elected Academy Member:Epistasis dominates the genetic architecture of Drosophila quantitative traits

Our understanding of genetic architecture of complex traits has been greatly advanced by genome-wide screens for DNA variants associated with phenotypic variation. To date, genome-wide association studies (GWASs) have identified over 6,000 common SNPs robustly associated with human complex traits and common diseases (1). Two general findings have emerged from these studies. First, there are typically a large number of loci associated with each trait, each of which explains a very small fraction of phenotypic variation (2). Second, loci associated with each trait collectively account for only a small proportion of genetic variation, giving rise to the mystery of missing heritability (3). Fitting all SNPs simultaneously and additively in a linear model can substantially increase the fraction of genetic variation explained by DNA variants (4), suggesting the existence of many weak associations. Another potential explanation under active investigation is that rare alleles with large effects, non-SNP variants (e.g., structural variants such as copy number variations and small indels), and nonsequence epigenetic modifications together account for the missing heritability (3). However, heritability in human studies is usually estimated as two times the difference between the observed phenotypic correlation between monozygotic and dizygotic twins, which estimates the fraction of phenotypic variance caused by additive genetic variation as well as overestimates the variance caused by dominance and epistasis (5). The potential inflation of estimates of narrow sense heritability (i.e., heritability caused by only the additive component of genetic variance) because of genetic interactions in human studies could lead to substantial underestimates of heritability explained by DNA variants (6). Controversy about the relative importance of epistasis in the genetic architecture of complex traits began with early formulations of quantitative genetic theory (7, 8) and continues today (9, 10). The crux of the controversy stems from the disparate goals of assessing the extent to which interactions affect mean genotypic values vs. estimating the fraction of total genetic variance caused by epistatic interactions in outbred populations. There is extensive evidence for epistatic interactions among quantitative trait loci (QTLs) affecting mean genotypic values in Drosophila, mice, Arabidopsis, yeast, and chickens (2, 9). Epistatic effects can be as large as main QTL effects, and they can occur in opposite directions between different pairs of interacting loci and between loci without significant main effects on the trait. Knowledge of interactions between loci can be used to infer genetic networks affecting complex traits (11), greatly informing the underlying biology. Epistasis is also the genetic mechanism underlying canalization (genetic homeostasis) (12, 13) and speciation (Dobzhansky–Muller incompatibilities) (14, 15); therefore, identifying interacting loci segregating in natural populations is relevant to understanding both evolutionary stasis and change. Finally, knowledge of interacting loci will improve predictions of long-term response to selection and inbreeding depression (and its converse, heterosis) in agricultural animal and crop species and individual disease risk in humans. However, nonadditive gene action does not translate to nonadditive genetic variance. Pure dominance results in mostly additive variance across the entire range of allele frequencies (5); pure epistasis gives largely additive genetic variance when allele frequencies are low, and most frequencies are low (10). In practice, all estimates of additive genetic variance (and hence, narrow sense heritability) from resemblance among relatives include fractions of the interaction variance (5). Thus, the only approach to discover epistatic interactions is a genome-wide screen in a mapping population, which incurs a severe penalty for multiple testing and hence, requires unreasonably large samples. Here, we use the 168 sequenced inbred lines of the D. melanogaster Genetic Reference Panel (DGRP) (16) to evaluate the contribution of common and rare variants as well as additive and epistatic gene action to the genetic architecture of three complex life history traits. We constructed an outbred, advanced intercross population from 40 DGRP lines, and we assessed allele frequency changes between extremes of the distribution for each trait by deep DNA sequencing of pools of individuals. We compared the results of GWASs in the DGRP lines to changes of allele frequency between extreme scoring individuals of the outbred population. We expected that common SNPs with additive effects shared between the two populations would be significant in both and that SNPs associated with the traits that were too rare in the DGRP to include in the GWAS might be significant in only the outbred population. Furthermore, we hypothesized that common SNPs shared between the populations with epistatic effects would not replicate across populations, because with epistasis, allelic effects are expected to vary between populations with different background allele frequencies. Surprisingly, we find that all three traits have distinct genetic architectures in the two populations caused by epistasis and that genetic networks inferred from the epistatic interactions are highly interconnected.