Fig. 3

Semi-synthetic experiment. A Comparison of FDR control and power for conditional independence testing at the gene level, averaged over five random weights initializations for the models of VI-VS, and across the five surface proteins of the dataset. Here, VI-VS uses a neural network with 64 units to compute importance scores. Left: FDR control comparison for the CRT, ordinary least squares (OLS) under t-tests, and marginal independence tests. Because the marginal test did not control the FDR, it was removed from the rest of the experiments. Center: Zoom on the previous figure. Right: Associated TPR. B FDR-TPR curves for different importance scores averaged over five random weights initializations for the models of VI-VS and across the surface protein of the dataset. Circles, squares, and rectangles respectively represent the models’ decisions for target FDR levels of 0.05, 0.1, and 0.2. C Associated held-out mean squared error of the different regression models used as importance scores. D Use of VI-VS as a calibration tool for GENIE3. After fitting the regression tree ensemble of GENIE3, we used their prediction error as importance scores for VI-VS, allowing one to detect conditionally dependent genes with statistical significance. E FDR/TPR levels of VI-VS using GENIE3 reconstruction losses as importance scores. In this experiment only, for scalability reasons, we considered a total of 100 genes in the experiment. In B and E, dashed lines denote target FDR levels