Figure 5.

Stability selection on real DNA methylation and gene expression data. The stability selection logistic-LASSO with the future lung cancer status as outcome and epigenetic markers of smoking as predictors is calibrated by maximising the stability score (a). The selection proportions in the calibrated model and explanatory performances of unpenalised logistic models where the predictors are sequentially added by decreasing selection proportion are showed (b). The three blocks of a multi-OMICs graphical model integrating DNA methylation and gene expression markers of tobacco smoking are calibrated separately using models where the other blocks are weakly penalised (${\lambda }_0=0.1$), while ensuring that ${\text{PFER}}_{\mathrm{MB}}<150$ overall (c). The stability selection model includes edges that are stably selected in each block (d).