Fig. 2. CONTENT is powerful and well-calibrated in simulated data.

Accuracy of each method to predict the genetically regulated gene expression of each gene-context pair for different correlations of intra-individual noise across contexts. Mean adjusted R² across contexts between the true (A) full (context-specific + context-shared), B shared, and (C) specific genetic components of expression and the predicted component for each method and for different levels of intra individual correlation. The context-by-context approach and UTMOST output only a single predictor, and we show the variability captured by this predictor for each component of expression. CONTENT, however, generates predictors for all three components of expression, and notably, CONTENT(Specific) and CONTENT(Shared) capture their intended component of expression without capturing the opposite (i.e., the predictor for CONTENT(Specific) is uncorrelated with the true shared component of expression and vice versa). We show here the accuracy for each component and method on gene-contexts with both context-shared and context-specific effects, but show in Supplementary Fig. 4 the accuracy for all gene-contexts pairs.