Extended Data Fig. 4: PDGrapher has stable performance on the synthetic datasets with various intensities of confounders added on the gene expression.
From: Combinatorial prediction of therapeutic perturbations using causally inspired neural networks
Performance of simulation analyses evaluated by percentage of accurately predicted samples (a) and nDCG (b) for the synthetic datasets with varying levels (0 to 1) of confounding bias introduced into the gene expression data. Gaussian noise, with distinct means and variances, was added progressively to random subsets of genes, simulating latent confounder effects in the treated gene expression data. The intensity of the confounding bias increases as more gene groups (representing network communities) are affected. This approach creates global, controlled variability in the gene expression data, paired with an unperturbed PPI network, allowing for the evaluation of algorithmic performance across different degrees of confounder noise. See Online Methods for more details on data generation.
