Network perturbation analysis of gene transcriptional profiles reveals protein targets and mechanism of action of drugs and influenza A viral infection

Noh, Heeju; Shoemaker, Jason E; Gunawan, Rudiyanto

The performance of ProTINA, like any other networkbased analytical methods, depends on the fidelity of the network used in the analysis. Uncertainty in the PGRN model, both in the structure and the edge weights, is expected to negatively affect the accuracy of the target prediction. Here, structural uncertainty is associated with the reliability of the information used to construct the PGRN, which in our study, comes from online databases of PIN and TF-gene networks. On the other hand, the uncertainty in the edge weights is associated with multiple factors, including the information content of the gene transcriptional profiles and the mathematical formulation used for the weight inference. The information content of the gene expression data is in turn related to measurement uncertainty and richness in the experimental perturbations. Keeping the same number of treatments, datasets with more replicates and less correlated gene expression profiles (i.e. the treatments induce more distinct perturbations to the network), would have a higher degree of information. Meanwhile, we have previously shown that the validity of the model assumption (e.g. pseudo steady-state condition) has an effect on the accuracy of the inferred weights and thus the target prediction accuracy (10). While we have circumvented the issue arising from the violation of the pseudo steady-state assumption in ProTINA (see Equation (7)), (in)validating all model assumptions may be difficult, if not impossible, in practice. A common strategy, as implemented in this study, is to test the performance of the method against benchmark datasets (13) . The results of applying ProTINA to drug treatment and influenza A viral infection datasets give confidence to the suitability of the mathematical formulation used in this work.