Integration of Multi-Modal Data to Guide Classification in Studies of Complex Diseases

Having entered the big data era, the unprecedentedly fast-growing volume and variety of biological data have swiftly transformed the landscape of biomedical research. Meanwhile, classification methods as a powerful bioinformatics tool have greatly empowered researchers to uncover new aspects of the complex biological systems. This thesis addresses the statistical and methodological challenges that often exist in different stages of biomedical multi-modal data integration with a focus into the application of classification methods in studies of complex diseases. Data generated from mass spectrometry (MS) platforms are inherently susceptible to systematic biases. Widespread missing values, where certain compounds cannot be identified or quantified, pose a prominent challenge to MS data normalisation. We propose a novel normalisation approach for high-dimensional MS data, called ruvms. This novel method is a one-step procedure that is able to handle missing values in input data and does not require imputation. We also explore a challenging situation in multi-modal data integration where not all types of data of interest are available within the same cohort. In brain studies, brain tissue samples are generally inaccessible from the same brain for which fMRI data can be obtained. We propose a gene-expression-guided fMRI network classification method that distinguishes patients of neurological diseases from the healthy control, called brainClass. brainClass links functional connectivity features to potentially involved biological pathways, to bridge the gap between functional biomarkers of neurological disorders and their underpinning molecular mechanisms. We also introduce a post-hoc interpretation framework to provide gene-expression-guided biological interpretations for predictive functional connectivity features identified by existing generic network classifiers applied to fMRI data