Computational prediction of cell fates by systems modelling of cellular signalling pathways

In the last few decades, extensive computational studies have been conducted on signaling pathways with both logic modeling (knowledge-based) and data-driven modeling. The increasing availability of high-throughput molecular data and multiplex techniques for uncovering cellular systems has made predictive in silico modeling more accurate which is important for understanding and engineering cellular functions. Generally, signal transduction inside a cell involves various modifications, including protein phosphorylation. Through a cascade of biochemical reactions, the signals are transmitted downward to the nucleus or other cellular organelles to regulate physiological functions and finally control the cellular phenotypes (e.g., apoptosis, proliferation and cell cycle). It is widely believed that the dysregulation of signal transduction is one of the most important pathogeneses of many human diseases including cancer, which makes the study of signaling pathways crucial for discovery of new anticancer therapies. However, there remain several challenges for the systems modeling of signaling pathways, such as how to integrate prior knowledge with real data into a context-specific model. This thesis presents the studies I have carried out with my collaborators to address some of the challenges. We first propose a generalized logic model that is able to simulate the graded responses to degradations and the effects of scheduled perturbations to the cells. A network is constructed for computational simulation based on the knowledge extracted from databases and literatures. In the network, the nodes and the edges represent the signaling proteins and the phosphorylation interactions, respectively. Given the inputs, the activity level of each node is updated synchronously based on their own previous states and the incoming signals from the parent nodes. Different combinations of perturbations were applied to the network. The evaluations of the simulation results with real data demonstrated that our simulator has the ability to grasp the main dynamical trends of signal transduction. Moreover, compared with existing simulators, our model achieved better performance in predicting the state transitions of signaling networks. Secondly, to study how the cellular signals control the cell fates (e.g. cancer cell death), a nonlinear power-law model is proposed to relate the activities of all the measured signaling proteins to the probabilities of cell fates. In our nonlinear function, the independent and dependent variables denote the activities of signaling proteins and the probability of cell death, respectively. The parameters, which indicate the contributions of the signaling proteins to the cell death, are identified from the training dataset and subsequently used to predict the probability of cell death on the testing dataset. Compared with the linear model on three cancer datasets with phosphoproteomics and cell fate measurements, the present nonlinear model has superior performance in the cell fates prediction. Moreover, our model is able to capture cell line specific information to distinguish one cell line from another in cell fates prediction. By in silico experiments of virtual protein knock-down, the proposed model is able to reveal the drug effects which can complement traditional approaches such as binding affinity analysis. After linking the cellular signals with the cell fates, we further investigate how to increase the selectivity of the drugs to the cancer. It is believed that if two genes have synthetic lethality (SL) relationship with one of them being a cancer-specific mutated gene, the drug that targets the partner gene would be able to give rise to SL and kill the tumor cells selectively. Therefore, a hybrid model, which combines the data-driven methods with prior knowledge of the underlying mechanisms (e.g., signal transduction), is proposed to study the potential drug targets by predicting the change of the probability of cell death caused by the gene knock-down. The single gene knock-down and the double genes knock-down are simulated to identify the human essential genes and the potential SL pairs of genes, respectively. A pair of genes is considered as an SL candidate if the double genes knock-down highly increases the predicted probability of cell death while the knock-down of either single gene does not. The promising performance of the hybrid model opens new directions for balancing the experimental data and the prior knowledge in future.DOCTOR OF PHILOSOPHY (SCE