Study Of Cancer Prognosis And Drug Response Prediction Based On Multi-data Fusion

To cancer patients, cancer prognosis and drug response predictions are the preconditions for individual therapy. As cancer is a kind of complex diseases with multi-gene abnormalities, it is a tendency to study its prognosis and drug responses from the systematical perspective by reconstructing the biological network (In this thesis, we take breast cancer as example). However, there are a few problems in the present works. The first one is the lack of using the heterogeneous data sets which are useful to the study of cancer. The second one is ignoring the biological phenomena related to cancer, resulting in the losing sight of the biological mechanism behind the alteration of the phenotypes. Therefore, for cancer prognosis, it’s essential to reconstruct the biological network which can reveal the metastasis mechanism by using reasonable mathematical model based on cancer related biologically phenomena. As to drug response prediction, it can be set as the controllability problem of the drug to the disease system. Therefore, it’s important to solve the controlling problem by establishing a reasonable mathematical model.Based on the gene dependency phenomenon which the influence of a gene to the phenotype is dependent on another gene,we merged the gene expression data,clinic information as well as the PPI, using the conditional mutual information as calculation method, to reconstruct the gene dependency network specific to the alteration of the phenotype (direct network).By the functional analysis of the key nodes (hubs) and the gene dependency relations of the network, we found the network can indeed uncover the biological mechanism in cancer prognosis. In addition, the selected discriminative hubs (43-gene signature) can distinguish the prognosis of cancer patients across several data sets.RNAs can influence each other by completing for the limited miRNAs,and the phenomenon is called ceRNA phenomenon. In addition, the ceRNA phenomenon can play important roles in cancer and other disease. Base on the ceRNA phenomenon, we merged the miRNA expression data, gene expression data and miRNA targets relations, using hyper geometric cumulative distribution function test, to reconstruct the breast-cancer-specific ceRNA network (undirected network). In addition, the hub genes and the communities can be annotated to cancer hallmarks, which show that our ceRNA network can reveal some biological mechanism related to cancer. At last, we selected15discriminative hubs as15-gene signature to construct the prognosis model and the results show that the discriminative hubs in the ceRNA network can distinguish metastasis risks of cancer patients.MicroRNA is a class of small RNA which can play important roles in cancer metastasis. However, there are almost no methods studying cancer prognosis by using miRNA regulation network.Based on the hypothesis that miRNA can influence the prognosis of cancer patients by regulating specific biological process, we integrated the gene expression data,miRNA targets relations and Go Term gene sets, using t-test to estimate the miRNA activity on specific biological process (CoMi activity). The subsequent validation experiments show the correctness of our CoMi calculation method.After that, we can combine all the CoMi activities into a miRNA activity network which contains two kinds of nodes, the miRNAs and the biological processes, showing the miRNA regulation activity on specific biological process (abstract network). At last, by the hypothesis that several miRNAs can regulate the same biological process to influence the outcome of cancer patients and several biological processes can influence the prognosis of cancer patients by working together, we set the sub network that several miRNA action the same Go Term (biological process) as a module, using all the miRNA activities in the module as features to construct a sub classifier to predict the outcome of cancer patients and the sub classifiers with classification capabilities were combined as an ensemble classifier by majority voting strategy. The validation on several independent data sets shows our ensemble classifier outperformed the published methods. In addition, the discriminative modules can reflect some miRNA regulation mechanisms in cancer prognosis. All this results show our method can be used to predict the outcome of cancer patients as well as reveal the biological mechanism in cancer prognosis. In the above works, based on three biological phenomena or hypothesis related to cancer, we integrated different high-throughput data to reconstruct three forms of network (direct network, undirected network and abstract network) and used the networks to study cancer prognosis from the three angles. In the first two works, we selected two gene signatures which can be used as candidates for cancer therapy. In the last work, we construct an ensemble classifier with good and robust performance in cancer prognosis.As to the drug response prediction, we solved this problem using controllability theorem of complex system. We used the disease network from a database as the network structure of the system, and used the gene expression levels of each cancer patient as the disease state. In addition, the control sample’s gene expression levels were set as the healthy state. And the drug targets were set as the driver nodes in the complex system. And then, we proposed a two-state controllability theorem (Two-state ε-controllability for practical application) based on the controllability theorem to judge whether the disease state can be driven to the healthy state by the drug. The simulation and real data validate our theorem. At last, we applied our new theorem in the drug response prediction and found that our method works well. All this results prove that our two-state controllability theorem can be used to predict the drug response of cancer patients.