Fig. 1: Simulation study.
From: Bayesian reassessment of the epigenetic architecture of complex traits
Boxplots of distribution of scores, the line in the middle of the box represents the median, upper, and lower bounds of the box represent first and third quartiles respectively, whiskers represent datum up to 1.5 interquartile distance from box bounds. a Estimation of phenotype-epigenetic associations using five recent approaches, BayesRR in Brown, to OSCA-moa in magenta, ReFACTor in gray, LFMM-Lasso in Cyan and LFMM-Ridge in Blue; where probes are associated with cell-type proportion variation and the norm of the correlation vector between the phenotype and the cell-type proportions have two different values either 0.08 or 0.25. Row panels provide results for different metrics of performance: the correlation between true effects and estimates (\(\rho (\beta ,\hat{\beta })\)), the slope of a regression of the estimates on the true effects (\({\beta }_{\hat{\beta } \sim \beta }\)), the number of genome-wide significant probes identified (loci), the mean square error (MSE), the MSE of the genome-wide significant probes (MSEsig), the false discovery rate (FDR), the norm of the correlation vector between a individual-level predictor made from the probe effects and the cell-type proportions (\(| | \rho ({\bf{R}},\hat{{\bf{g}}})| |\)), the correlation between the first principal component of the probe data and the difference between the estimated and true effect (∣∣ρ(P, R)∣∣) and the phenotypic variance attributable to the probes (\({\sigma }_{cg}^{2}\)). Black lines give the true value across panels. b Comparison of BayesRR with just the methods, which fit probes jointly (multi-probe methods) either accounting for latent factors (LFMM-Lasso in Cyan and LFMM-Ridge in blue) or not (GLMNET-Lasso in dark-blue and GLMNET-Ridge in dark-yellow). c Simulation results of methylation marker effects for a phenotype influenced by both 100 methylation probes and 1000 SNP markers, showing the difference between the true and the estimated phenotypic variance explained by genetic markers (\({\sigma }_{G}^{2}-\hat{{\sigma }_{G}^{2}}\)) and epigenetic probes (\({\sigma }_{cg}^{2}-\hat{{\sigma }_{cg}^{2}}\)). d Comparisons of approaches that do not fit latent factors within the model when the underlying epigenetic architecture is less sparse (phenotype is influenced by 1000 probes, rather than only 100).
