Network-based Analyses For Biomarker Identification

This thesis adopts the network-based analyses to deal with a hot problem – biomarker dis-covery.This multidisciplinary research makes it of great research value and scientific signifi-cance both in theory and application.Organisms are complex regulatory systems and the occur-rence of diseases are often caused by abnormal regulatory mechanisms.Therefore,designing the effective analyses to discover disease-related molecular regulatory mechanisms plays an important role in many aspects,including understanding of diseases,innovative drug develop-ment,improved treatment effect and medical technology development.This thesis focuses on the problem of biomarker discovery,and designs three types of effective mathematical analyses:single,double and multiple network-based analysis.The contribution are as follows:First,we proposes a brain network common harmonic discovery approach on Stiefel man-ifold.Generally,disease-relevant brain alterations follow brain networks.Thus,a powerful network analysis is indispensable to understand the mechanism of neuro-pathological events.Indeed,the topology of each brain network is governed by its native harmonic waves derived from the Eigen-system of the Laplacian matrix.To that end,we propose a novel harmonic analy-sis framework to detect frequency-based alterations relevant to brain disorders.Our framework is a novel manifold algebra for harmonic waves analysis that overcomes the limitations of us-ing classic Euclidean operations on irregular data structures.The excellent performance of the proposed method are verified on the synthetic and Alzheimer's disease data set.Second,we propose a higher order graph matching with multiple network constraints model for gene-drug regulatory modules identification.The emergence of large amounts of genomic and pharmacological data provides new opportunities and challenges.Identifying gene-drug associations is crucial to understand the molecular mechanisms of drug action and develop ef-fective treatments.However,accurately determining the complex associations among phar-macogenomic data remains challenging.Therefore,we propose a high-order graph matching framework with prior constraints,and design an effective sampling strategy to finally accu-rately discover the gene-drug regulatory modules.The experiment results on synthetic and drug gene data demonstrate our approach can discover closely related gene-drug regulatory modules.Third,we propose a higher order graph matching model for common and specific mi RNA-gene regulatory modules identification.Identifying regulatory modules between mi RNAs and genes can deeply understand the molecular mechanisms of cancer and facilitate the development of precise treatments.However,the genomic data usually relate to different cancer statuses.Therefore,there is an urgent need to develop a novel method to jointly analyze mi RNA and gene data of different cancer statuses to identify the background of the tumorigenesis(common modules)and subtype-specific regulatory mechanisms(specific modules).To this end,we de-velop a higher order graph matching model,and incorporate variational difference constraint to extract common and specific modules.The experimental results on synthetic,stomach adeno-carcinoma and breast invasive carcinoma data demonstrate the superiority of our method.Last,we propose a multi-graph matching with network constraints model for multidimen-sional regulatory modules identification.The accumulation of large amounts of multidimen-sional genomic data provides new opportunities to study multilevel biological regulatory as-sociations.Identifying multidimensional regulatory modules(md-modules)is crucial to un-derstand the regulatory mechanisms.Thus,we develop a multi-graph matching framework to accurately capture highly relevant md-modules by considering the relationships intra- and inter-multidimensional omics data.The proposed technique adopts a novel graph-smoothing similar-ity measurement for the highly contaminated genetic data.Experiments on synthetic data and cervical cancer data show that our proposed method can accurately identify the md-modules that are significantly enriched in GO terms and KEGG pathways.In conclusion,this thesis is devoted to the problem of biomarker discovery.For the sin-gle network-based analysis problem,a brain network common harmonic discovery approach is proposed.For the double network-based analysis problem,a higher order graph matching model for gene-drug regulatory modules identification is proposed.In addition,this method is extended to mi RNA and gene data with a view to discover common and specific mi RNA-gene regulatory modules.For the multiple network-based analysis problem,a multi-graph matching model for multidimensional regulatory modules identification is proposed.Theoretical analysis and experimental results show our methods well capture the nature of the problem,improve the performance notably,and enrich the research contents in bioinformatics.