Control Principles Of Complex Networks For Personalized Cancer Multi-omics Data

With the rapid development of high-throughput sequencing and biotechnology,large number of genomic data and bio-molecular expression polymorphic data containing complex disease information has emerged.These data bring new opportunities and challenges to develop the efficient computational methods for exploring the pathogenic mechanism and treating the cancer patients.Due to small samples,high dimensional and large noise characteristics of individual patient data,it is difficult for traditional statistical methods to excavate effective information.In this thesis,the control principles of complex networks for discovering the personalized key genes related with phenotype transitions in cancers are studied.The main contributions are as follows:1.Considering that existing network structure-based control methods do not take into account the prior information of driver nodes which can be imposed the control signal,we introduce a concept of the constrained target controllability(CTC)of complex networks to investigate the ability for driving any state of target nodes to their desirable state with the control signals to the driver nodes from the set of the constrained control nodes.To efficiently explore CTC of complex networks,we design a novel graph-theoretic algorithm(namely CTCA)to estimate the ability of controlling the targets by choosing driver nodes from the set of the constrained control nodes.We evaluated the constrained target controllability on numerous real complex networks.The results show that the biological networks with higher average degree are easier to be controlled than the biological networks with lower average degree,while the electronic networks with lower average degree are easier to be controlled than the web networks with higher average degree.We also confirm that our CTCA can more efficiently produce driver nodes for target controlling the networks than the existing state-of-the-art methods.Moreover,CTCA was also used to analyze two expert-curated bio-molecular networks,and the results show that CTCA can more efficiently identify the proven drug targets and new potentials than other state-of-the-art methods.2.Existing researches only focus on controlling the system through minimum driver-node set,ignoring the existence of multiple candidate driver-node sets.In fact,when expecting to control the system with objectives optimization,the different driver-node sets may not equally participate in controlling the networks.Therefore,we frame the target control problem with objectives-guided optimization(TCO),that is,for a system,how could we control the interested variables(or targets)with the optional driver nodes by minimizing the total quantity of drivers and meantime maximizing the quantity of constrained nodes among those drivers.Then we develop an efficient algorithm(TCOA)to find the optional driver nodes for controlling targets in complex networks.The main consideration of TCOA is to optimize an outlier measurement,which is introduced to balance the minimum number of driver nodes and the maximum number of driver nodes within the set of pre-selected nodes.Results on several real-world networks and the experiment show that the performance of TCOA is superior to other existed network control approaches.Especially,the resulst on the signaling transduction networks in pancreatic cancer and inflammatory bowel disease show that TCOA can effectively identify the drug targets for leading the phenotype transitions of underlying biological networks.3.Most of existing methods ignore the patient-specific network information(i.e.,topology or edges)to determine the model parameters,which potentially lead to many false positives.Therefore,from network controllability perspective,we presented a novel single-sample controller strategy(SCS)to identify personalized driver mutation profiles.SCS mainly contains three parts: i)For each patient,obtaining the personalized differentially expressed genes(DEGs)by comparing the expression profiles of the tumor sample and normal samples,then extracting the individual gene regulations from the expert-curated databases.ii)Identifying the minimal number of individual mutations which cover maximal DEGs using network controllability.iii)Based on the dynamic network control theory,obtaining the consensus modules consisting of confidence control paths from the driver genes to the DEGs and assessing the impact of driver genes according to the scores of consensus modules.We widely validate the driver mutation profiles mined with SCS from following three aspects: i)The improved precision for predicting known cancer driver genes compared with other methods.ii)The effectiveness for discovering the personalized driver genes.iii)The risk assessment through integration of the driver mutation signature and expression data on the five benchmark datasets.The results show that SCS not only can help to discover the personalized causal mutations from those mutations which are obscured by tumor heterogeneity,but also is effective to integrate cancer omics data in tumor pathology,clinical stratification and personalized therapy.4.For lacking an efficient framework to control the undirected networks with nonlinear dynamics,we frame the nonlinear control problem of undirected networks(NCU),that is,how we choose the minimum driver nodes to control the network from the initial attractor to the targeted attractor with nonlinear dynamics.Then we proposed a novel graphic-theoretic algorithm(named Weight-NCUA)to identify the minimum driver nodes based on the feedback vertex sets by adopting the following process: i)Constructing a bipartite graph with node weighted in which the nodes of the top side are the nodes of the original undirected network and the nodes of the bottom side are the edges of the original undirected weighted network.ii)Determining the dominating set among the top side nodes to cover the bottom side nodes in the bipartite graph by using integer linear programming..The results of Weight-NCUA on synthetic scale-free networks,real-world networks and TCGA cancer data sets show that the nonlinear control characteristic is related with both network structure and the choice of network controllers.Furthermore,we also used Weight-NCUA to identify the personalized drug targets and quantify the control role of personalized combinational drugs which drive the individual molecular system from a disease state to a normal(or sub-normal)state.We validated the effectiveness of Weight-NCUA on two benchmark cancer datasets of breast invasive carcinoma(BRCA)and lung squamous cell carcinoma(LUSC).The results show that the performance of Weight-NCUA is superior to other exisitng methods in terms of the predictive accuracy of drug combinations.