Comparison of Models for Genome-Wide Association Mapping in Plants.

Genome-wide association studies (GWAS) are an alternative approach to linkage mapping of bi-parental populations and can provide high mapping resolution for complex trait variation. However, a major problem in GWAS is population stratification that can induce false positives. Controlling false positives can induce false negatives due to over fitting of the model. In this study, a number of statistical models were compared for three traits differing in heritability including oxygen isotope ratio (δ¹⁸O) (H = 20%), carbon isotope ratio (δ¹³C) (H = 60%), and canopy wilting (CW) (H = 80%) and simulated data. These models were: (i) single marker regression (SR, (ii) GLM with Q matrix (population membership estimates), (iii) GLM with PCA (Principle Component Analysis), (iv) MLM with Q + K (Kinship matrix for family relatedness estimates), (v) MLM with PCA + K, (vi) compressed MLM (CMLM), (vii) enriched compressed MLM (ECMLM), (viii) Settlement of MLM Under Progressively Exclusive Relationship (SUPER), and (ix) Fixed and random model Circulating Probability Unification (FarmCPU). Comparing Q-Q plots for δ¹⁸O, only FarmCPU model showed a sharp deviation from the expected P-value distribution in the tail area, indicating that false positives and negatives were adequately controlled whereas Q-Q plots from all other models did not show a sharp deviation. For δ¹³C and CW, a large number of false discovery rates were observed for SR, GLM+PCA, and GLM+Q models; however, MLM, CMLM, and ECMLM controlled the false positives but increased the false negatives. A sharp deviation in the Q-Q plot of FarmCPU model for δ¹³C and CW indicated that this model reduced both false positives and negatives. Similar results were observed for simulated data. Overall, this study indicated that the FarmCPU is an appropriate model for association analysis of traits that have either low or high heritability.