Network-based Research Of Complex Diseases

Complex diseases are rarely caused by the mutation or abnormality of a single gene, but by the complicated interactions among many genes or even between genetics and the environment, the sophisticated mechanisms bring certain difficulty to prevent, diagnose and treat complex diseases effectively. With the development of the high-throughput techniques, a large amount of omics data has been produced, which provides an opportunity to further study the pathogenic mechanisms, the occurrence and devel-opment, and the prevention and control of complex diseases. Based on the multi-omics high-throughput data, this dissertation studies the identification of dynamic nelwork biomarkers (DNBs),disease genes and modules respectively from two different network-levels, i.e., the levels of dynamic network and multi-layer network. The research work in this dissertation will provide new approaches and new ideas for revealing the pathogenic mechanisms of complex diseases and detecting early warning signals for sudden deteri-oration of complex diseases from the system level. The main work for this dissertation are presented as follows:1. Using a nonlinear ordinary differential equation model to describe the relation-ships among biomolecules, we develop a model-based framework for the construction of dynamic networks by integrating high-throughput gene expression data into protein and protein interaction (PPI) data. First, the problem of constructing dynamic net-work is converted into an optimal problem of identifying time varying parameters, that is, a dynamic optimization problem that the parameters are changed with the time. Then, piecewise linear functions are used to approximate the nonlinear relationships and the spline interpolation is used to obtain new data containing the desired samples. The choice of regularization parameters is based on the Bayesian Information Criterion. All parameters of the piecewise linear models that are determined by the optimization method are referred to as the edges in the dynamic networks at each time point. Finally, the accuracy and the robustness of the constructed dynamic networks are evaluated by average errors and leave-one-out cross-validation (LOOCV) method, respectively.2. Through the defined two quantitative indexes, i.e., the similarity score of two modules and the influence index of a module, we present a novel framework for detecting DNBs from the constructed dynamic networks based on the idea of modularization. The performance of the proposed method is testified by real data from four types of complex diseases, including influenza caused by either H3N2 or H1N1, acute lung injury and type 2 diabetes mellitus. Through the comparisons with other methods in publishes literatures, the effectiveness of the proposed method is demonstrated. Function and pathway analyses reveal that the identified DNBs are significantly enriched during key events in early disease development. Furthermore, we present two concepts, i.e., local information flow and global information flow. Correlation and information flow analyses reveal that DNBs effectively discriminate between different disease processes and that disproportional information flow may contribute to the dysfunctional regulation and increased disease severity.These studies offer new insights into early diagnoses and the design of drug target of complex diseases, and also provide a general paradigm for revealing the deterioration mechanisms of complex diseases.3. A systematic method is developed to construct a multi-layer network by inte-grating multi-level prior knowledge with multi-dimensional biological data and is ap-plied to the construction of a multi-layer network of influenza A virus infection. Five high-throughput gene expression datasets of the two different strains(H1N1 and H3N2) for influenza A virus are used and the data of viral proteins interactions, virus-host interactions, protein-protein interactions and transcriptional regulatory relationships are integrated. Each layer of the multi-layer network is obtained by the method of traditional single-layer network construction. The specific interactions between lay-ers are determined by combining database mining, literature retrieving and optimiza-tion method, and the significant interactions are selected by using "Z-score" statistical method. Therefore, a three-layer virus-host interaction network,i.e., the viral protein layer, the host protein layer and the host gene layer, is obtained.4. Based on the topology of the constructed multi-layer network and the bioinfor-matics databases, the essential conserved modules across multiple datasets are detected. Combining functional enrichment analysis and module similarity, we find that a module, which consists of 44 genes or proteins,plays a key role in three steps(fusion and uncoat-ing, vRNP transport from the nucleus to the cytoplasm, assembly and budding) of the two different strains'replication cycle and launches its effect mainly in the cytoplasm. However, the specific modules for two different influenza viruses H3N2 and H1N1 are mainly functioned in the cell nucleus. To more comprehensively characterize the in-fluenza virus replication cycle, singular value decomposition is used to detect essential components in multi-layer networks. The overlap and its significance between the es-sential components and the above module are calculated based on inverse cumulative hyper geometric distribution and the results show that identified essential components are concentrated in the essential common module, which verifies the reliabilities of the identified disease module. Our study provides a theoretical prediction for revealing the molecular mechanisms of influenza A virus infection and replication.The methods for constructing dynamic networks and multi-layer networks proposed in this dissertation and the obtained quantitative indexes based on network analysis can be further extended to other biological problems and the integration of biological big data.