Preview

PDF, English Download (11MB) | Terms of use

Abstract

Schizophrenia is a complex mental disorder that has been extensively studied from various perspectives, emphasizing the complexity of factors influencing psychotic disorders. The unraveling of molecular genetics unshadowed a new epoch in the understanding of schizophrenia, redefining it as not only a neuropsychiatric abnormality but an complex disorder with a significant genetic underpinning. In addition, the nature of schizophrenia comorbidity covered a wide range of different conditions, of which the complexity involved emergent syndromes such as type 2 diabetes, cardiovascular disease, and bipolar disorders. In order to better understand the comorbidity of illness, it is important to explore various factors, including genetic variants, molecular profiles, and systematic biological networks.

The rapid development of high-throughput screening techniques has given rise to large-scale biomedical data, which offers the potential to characterize complex biological structures. One of the main challenges is to discover the molecular pattern from high-dimensional data. Secondly, the complexity of biomedical data introduced was compounded by effects of interest as well as different confounders, coming from the heterogeneity of the data itself, which required us to disentangle factors from various sources from the complex variance. And thirdly, multiple layers across various modalities to be analyzed simultaneously became feasible, resulting in the increased application of multi-modal integrative learning that utilizes multidimensional information simultaneously to improve understanding of biological systems.

To address this, factor-based techniques as one of the fundamental approaches in machine learning and data analysis, offer a chance to handle complex datasets to unveil hidden biological data structures. The landscape of factor-based techniques has evolved significantly for diverse datasets and applications, while the versatility of factor-based methods extends far beyond their original usage to current molecular data analysis. Therefore, in the thesis, we would further expand the landscape of the factor-based model and describe the development of our three different factor-based approaches applied in schizophrenia and psychotic disorder research. In Chapter 2, the development of a hierarchical factorization computation framework for signature-based disease comorbidity modeling for schizophrenia and type 2 diabetes was described. In Chapter 3, the deployment of a federated factorization approach was demonstrated, which projected the standard integrative factorization to a federated version based on the DataShield platform, thereby facilitating cross-modal and cross-institutional learning. In Chapter 4, a novel application based on the extended biological-informed interpretable factorized variational autoencoder was introduced for predicting genotype-to-function association in schizophrenia individuals, enabling the separation of the variance of diagnosis from the complex effects by controlling confounding factors as well as the prediction of biological processes with genetic variants. In the end, the prospects and outlook will be discussed in Chapter 5 regarding the further extension of the factor-based techniques and the research gap in schizophrenia and psychotic disorders.