Statistical Reverse-Engineering of Signal Transduction Pathways

Inferring cell signaling networks from high-throughput data is a challenging problem in systems biology. Signaling networks are stochastic in nature. They are not entirely predictable based on the limited information available on the system. Recent advances in cytometric technology enable us to measure the abundance level of a large number of proteins at the single-cell level across time. Traditional network reconstruction approaches usually consider each time point separately resulting in inferred networks that strongly vary across time. In order to account for the possibly time-invariant physical coupling within the signaling network we extend traditional graphical Lasso with an additional regularizer that penalizes network variations. We provide the corresponding expression for the Bayesian information criterion (BIC) that allows to determine the regularization parameters from data. Receiver operating characteristic (ROC) evaluation of the method on in silico data shows higher reconstruction accuracy compared to standard graphical Lasso. We then perform single cell cytometry experiments of IFNγ pathway stimulation in the THP1 cell and apply the novel method on the panel of 26 phospho-proteins. Besides well-documented interactions, such as the wiring of the JAK/STAT pathway, the method also provides novel candidate interactions. The reconstruction results are ultimately validated by performing dedicated perturbation experiments with PI3K, MEK1/2 and AMPK inhibitors. Causal inference is an effective tool for signaling network reconstruction to identify cause-effect relationships among biomolecules. Most traditional directed network reconstruction approaches deal with linear relationships among the variables via a Gaussian or a nonparanormal distributional assumption. These assumptions are not tenable to the stochastic biological systems. Moreover, traditional directed network reconstruction approaches overlook possibly nonlinear relationships among variables. To account for nonlinear relationships, we consider a mutual information based approach. In particular, we apply a mutual information based directed structure learning method to an additive predictor with multiplicative noise model under in silico study. We validate the performance of our method using area under ROC (AUROC) curve scores. We consider nearest neighborhood based mutual information estimation technique for large networks to avoid computational complexity caused by kernel density based methods