Figure 3

Synthetic response data, average ROC curves. Number of true positives plotted against number of false positives for Simulations 1, 2 and 3. Proteomic data from Sachs et al.[10] were used to create response data with true underlying model known to favour a particular prior: Simulaton 1 - distance prior with positive λ; Simulations 2 and 3 - either distance prior or number of pathways prior with negative λ. Legend - ‘BVS’: Bayesian variable selection; ‘+int’: linear model with interaction terms; ‘-int’: linear model without interaction terms; ‘EB’: empirical Bayes used to select and weight pathway-based priors automatically; ‘flat’: flat prior; ‘incorrect’: wrong prior with respect to true, underlying protein set (see main text for details); ‘MRF prior’: Markov random field prior[19]; ‘Lasso’: Lasso linear regression (curve produced by thresholding absolute regression coefficients, whilst marker ‘X’ is single model obtained by taking only predictors with non-zero coefficients); ‘Li&Li’: penalised-likelihood approach proposed by Li and Li[21] that also incorporates network information (see main text for details); ‘corr’: absolute Pearson correlations between each protein and response. Area under the (average) ROC curve ("AUC") appears in parentheses.