Wright was right.
In over 1,600 line-cross datasets spanning plants and animals, epistasis was detected in the majority. The Fisher vs. Wright question from the 1930s now has an empirical answer. Gene interactions are the rule, not the exception, and the assumption of strict additivity is rarely defensible.
Why epistasis matters
Foundations Epistasis is the interaction between genes The phenotypic effect of an allele at one locus depends on the genotype at another.
When allelic effects combine independently, populations tend to evolve smoothly toward a single optimum, the picture behind Fisher's infinitesimal model. When they do not, populations can get trapped on local fitness peaks, the response to selection depends on genetic background, and the adaptive landscape becomes rugged and multi-peaked. Bateson (1909) coined the term "epistasis" to describe the masking of one gene's effect by another. A century later, it is one of the central variables in quantitative genetics, speciation biology, and evolutionary medicine.
Types Magnitude, sign, and reciprocal sign epistasis Same direction, flipped direction, and the architecture that guarantees multiple fitness peaks.
Magnitude epistasis changes the size of an allele's effect. Sign epistasis reverses its direction. Reciprocal sign epistasis, where both alleles flip depending on the other, is the necessary and sufficient condition for multiple fitness peaks in two-locus systems (Poelwijk et al. 2011). In higher dimensions, rugged landscapes can also arise from accumulated magnitude or sign epistasis on their own. Either way, when the right kind of epistasis exists, there is no single-step mutational path from one peak to another that is always uphill. The population must cross a valley.
Fisher vs. Wright, a 1930s debate
Fisher's view Selection on additive variance in large, panmictic populations Epistasis as statistical noise, broken up by recombination.
Fisher argued that evolution is primarily driven by natural selection acting on additive genetic variance in large, freely interbreeding populations. His Fundamental Theorem of Natural Selection states that the rate of increase in fitness equals the additive genetic variance in fitness. In this framework, epistasis contributes little to the response to selection because recombination breaks up favorable gene combinations every generation. Evolution is a smooth, deterministic, hill-climbing process. Drift is negligible in populations of realistic size, and population structure is an unnecessary complication.
The key works: Fisher, "The Correlation between Relatives on the Supposition of Mendelian Inheritance" (Trans. R. Soc. Edinb. 1918, 10.1017/S0080456800012163), and The Genetical Theory of Natural Selection (Clarendon Press, 1930).
Wright's view Structured populations, drift, and the shifting balance Three phases: drift explores, selection refines, migration spreads.
Wright saw a different picture. Real populations are structured into small, partially isolated demes. In small demes, genetic drift can push a population off its current fitness peak. Gene interactions create multiple adaptive peaks, and drift, migration, and selection work together to let populations explore and ultimately find the highest ones. This is the shifting balance theory.
Phase 1, Random drift. Within small demes, allele-frequency fluctuations move a population off its current local peak and into the basin of a different one.
Phase 2, Mass selection within demes. Ordinary natural selection pushes the population uphill toward the new peak.
Phase 3, Interdeme selection. Demes on higher peaks produce more emigrants, and their migrants spread favorable gene combinations across the metapopulation toward the global optimum.
The foundational papers: Wright, "Evolution in Mendelian Populations" (Genetics 1931, 10.1093/genetics/16.2.97), and "The Roles of Mutation, Inbreeding, Crossbreeding and Selection in Evolution" (Proc. Sixth Int. Congr. Genet., 1932), where the adaptive landscape diagram first appeared.
The modern exchange Coyne, Barton & Turelli vs. Wade & Goodnight, 1997 to 2000 A masterclass in how scientific arguments are constructed, attacked, and defended.
Coyne, Barton, and Turelli (Evolution 1997, 10.2307/2411226) argued that drift is too weak to move populations off fitness peaks in realistic populations, and that interdeme selection has never been convincingly demonstrated. Wade and Goodnight replied (Evolution 1998, 10.2307/2411328) with Tribolium experiments and the theoretical insight of variance conversion: bottlenecks expose previously masked epistatic effects as additive variance, changing the substrate on which selection acts. The exchange continued in 2000 (Coyne, Barton & Turelli, 10.1111/j.0014-3820.2000.tb00033.x; Goodnight & Wade, 10.1111/j.0014-3820.2000.tb00034.x).
The honest resolution: the two sides were arguing about different scales. Fisher's fundamental theorem is exact for the additive component regardless of how much epistatic variance also exists (Turelli & Barton 1994, Hansen 2013). Modern line-cross data show that epistatic composite effects are pervasive in generation means, while epistatic variance in segregating populations can still be modest because interaction effects partially absorb into the additive component as allele frequencies shift. Both frameworks are right in their domain, Fisher's for short-term within-population response to selection, Wright's for the architecture that channels long-term evolution across rugged landscapes.
Fisher (1918) recognized that interactions between non-allelic genes (epistasis) represent the most important deviation from a simple additive scheme. Paraphrased from R.A. Fisher, Trans. R. Soc. Edinb. 52: 399 to 433.
Measuring epistasis, line cross analysis
The design Parents, F1, F2, and backcrosses, with the trait measured every generation The pattern of generation means reveals the genetic architecture.
Cross two divergent lines (populations, species, or artificially selected lines), create a series of composite generations (F1, F2, BC1, BC2, and potentially further generations), and measure the trait in each. Under pure additivity, the F1 falls at the midparent value, the F2 matches the F1 in expectation, and backcrosses fall midway between the F1 and the corresponding parent. Dominance shifts the F1 off the midparent value but leaves the F2 and backcrosses predictable relative to it. Epistasis breaks these expectations: F2 means deviate from F1 means, and backcross means become asymmetric. The specific pattern tells you which types of epistasis are at play, additive × additive, additive × dominance, or dominance × dominance.
The classical statistical framework is the joint-scaling test (Cavalli-Sforza 1952; Hayman 1958; Mather & Jinks, Biometrical Genetics, Chapman & Hall, 1982). It fits nested models and tests whether each additional parameter significantly improves fit, with the usual problems of sequential hypothesis testing: multiple-testing inflation, model-selection bias, and the arbitrariness of p-value thresholds.
Design choices Classical F2, RILs, and MAGIC populations No single design is universally best.
The classical P/F1/F2/BC design is powerful for detecting net additive, dominance, and digenic epistasis, but loses power for higher-order interactions and small-effect components. Recombinant inbred lines (RILs) trade repeated selfing for fully homozygous, replicable genotypes that enable QTL-by-environment work at higher per-locus detection power. MAGIC populations (multiparent advanced generation intercross) introduce alleles from many founders and shuffle them through several generations of intermating, giving fine-mapping resolution two-parent crosses cannot reach. Sample size, generation time, expected effect size, and the question being asked all matter. The field has moved well beyond the F2-only era, and any modern survey should account for that heterogeneity.
SAGA, an information-theoretic upgrade
Method AICc, all models at once, averaged estimates Statistical Analysis of Genetic Architecture, built in the lab.
SAGA replaces sequential chi-squared testing with an information-theoretic approach based on corrected Akaike Information Criterion (AICc). Instead of asking "is this model significantly better than the simpler one?" it asks "what is the relative support for each model given the data?" All candidate architectures, every combination of additive, dominance, and epistatic effects, are evaluated simultaneously, with a natural complexity penalty that scales with sample size. The output is a confidence set of models within a few AICc units of the best, plus model-averaged parameter estimates that incorporate model uncertainty.
The methods papers: Blackmon & Demuth, "An information-theoretic approach to estimating the composite genetic effects contributing to variation among generation means" (Evolution 2016, 70(2): 420 to 432, 10.1111/evo.12844); Armstrong, Anderson & Blackmon, "Inferring the potentially complex genetic architectures of adaptation, sexual dimorphism, and genotype by environment interactions by partitioning of mean phenotypes" (J. Evol. Biol. 2019, 32(4): 369 to 379, 10.1111/jeb.13421).
The Burch 2024 findings
1,600+ datasets The largest survey of epistasis to date Epistasis is the rule across plants and animals.
In Burch et al. 2024 we applied SAGA2 to more than 1,600 line-cross datasets spanning plants and animals and covering morphological, physiological, life-history, and behavioral traits. Epistasis was detected in the majority of crosses. Purely additive models were rarely the best-supported architecture for any trait category. The assumption of strict additivity that still sits at the base of most breeding, genomic-prediction, and evolutionary models is rarely defensible.
What "epistasis detected" means in this analysis: a model with an epistatic component fits the generation means better than a purely additive-plus-dominance model (information-theoretic comparison via SAGA2). It does not mean epistatic variance dominates the response to selection. In many crosses the additive component still explains the largest single share of trait differentiation. The right reading: epistatic composite effects in line crosses are pervasive, and how much they matter for any specific evolutionary or breeding question depends on their magnitude and on allele frequencies in the target population.
Burch, Chin, Fontenot, Mandal, McKnight, Demuth & Blackmon, "Wright was right: Leveraging old data and new methods to illustrate the critical role of epistasis in genetics and evolution" (Evolution 2024, 78(4): 624 to 634, DOI: 10.1093/evolut/qpae003).
Recognition SSE 2025 President's Award for Outstanding Dissertation Paper Society for the Study of Evolution.
Burch et al. 2024 received the 2025 Society for the Study of Evolution President's Award for Outstanding Dissertation Paper, which recognizes the most impactful dissertation-based publications in the field.
The question is no longer whether epistasis exists (it clearly does), but how to quantify its contribution to adaptation, speciation, and disease across the diversity of life.
Implications for speciation, disease, adaptation
Speciation Dobzhansky-Muller incompatibilities are epistasis by another name No epistasis, no genetic reproductive isolation.
Dobzhansky-Muller incompatibilities, alleles that work in their home genetic background but cause reduced fitness, sterility, or inviability when placed into a hybrid background, are epistasis at their core. Without epistasis, speciation through accumulated genetic incompatibilities would be impossible. See Orr, "Testing Natural Selection vs. Genetic Drift in Phenotypic Evolution Using Quantitative Trait Locus Data" (Genetics 1998, 149: 2099 to 2104, 10.1093/genetics/149.4.2099).
Disease and missing heritability Additive GWAS models underestimate risk alleles that live in epistatic networks Accounting for interactions may close part of the gap.
The missing-heritability problem, where identified GWAS variants explain only a fraction of the heritability estimated from family studies, may partly reflect undetected epistatic interactions. If a risk allele's effect depends on variants at other loci, additive GWAS underestimates its contribution. Modern pairwise GWAS interaction scans and polygenic-adaptation frameworks (Pritchard, Coop, and others) are now routine, though they carry an n² computational burden and demand very large samples. Reviews: Mackay, Nat. Rev. Genet. 2014, 15: 22 to 33 (10.1038/nrg3627); Phillips, Nat. Rev. Genet. 2008, 9: 855 to 867 (10.1038/nrg2452); de Visser & Krug, Nat. Rev. Genet. 2014, 15: 480 to 490 (10.1038/nrg3744).
Adaptation and drug resistance Context-dependent fitness effects, heterosis, and combination therapies Epistasis everywhere you ask the question carefully.
The response to selection is context-dependent: an allele that is beneficial in one population may be neutral or harmful in another, depending on epistatic context. Heterosis in hybrids involves non-additive genetic effects, and the relative contributions of dominance and epistasis remain debated but both matter. Combinations of resistance mutations can interact in unexpected ways, sometimes more than the sum of their parts (positive epistasis, accelerating resistance) and sometimes less (negative epistasis, potentially constraining it). Understanding the epistatic landscape of resistance matters for designing combination therapies and predicting the evolutionary trajectory of pathogens and cancers.
A century of accumulating data increasingly suggests that both Fisher and Wright captured important truths, but Wright's emphasis on epistasis was more prescient than the field long acknowledged. Our analysis of more than 1,600 line-cross datasets adds to the evidence that epistasis is a pervasive feature of genetic architecture, not a statistical curiosity to be swept into the error term.