Fig. 2. Reference-based analysis (RBA) is unsuitable for inferring global genetic architecture.

a The apparent complexity of genetic architecture varies depending on the choice of wild-type genotype. A first-order RBA model is directly fitted to the genetic architecture in Fig. 1a, with (A, A), (B, A), or (B, B) as wild-type. λ₀, wild-type phenotype; λ₁, λ₂, first-order effects of mutations at sites 1 and 2; empty circle, predicted phenotype of the double mutant; r², fraction of phenotypic variance explained by the model. b (Left) Absolute error of RBA terms computed from the simulated measurements with a small amount of noise in Fig. 1c. Dashed lines, mean absolute error of individual phenotypes. (Right) Fraction of phenotypic variance explained by the true and computed RBA terms. c Regression yields incorrect estimates of RBA terms and overestimates the fraction of phenotypic variance they explain. The first panel fits the first-order RBA model defined with respect to (A, A) by regression; the estimated terms, predicted phenotype for each genotype, and r² are shown. The next two panels repeat the analysis by choosing (B, A) or (B, B) as wild-type. d (Left) Fraction of phenotypic variance explained by the true and regression-estimated RBA terms for the simulated measurements in Fig. 1c. (Right) RBA terms estimated by fitting the full model, compared with their true values.