Deconstructing and Approaching Heterogeneities in the Biomedical Field via Computational Modeling

Abstract

Natural variation between human characteristics as well as differences across collected datasets in disparate medical or research centers on various levels (e.g., semantical and technical) lead to high heterogeneity in terms of patients and data in the biomedical field. These heterogeneities not only impede understanding of disease pathology and clinical diagnosis but also their implications in the treatment of disease are substantial. Moreover, these heterogeneities limit the impact of computational solutions on clinical practice in spite of their high potential in bringing significant advances in the biomedical domain. <br /> In this thesis, we address the aforementioned issues in the context of complex diseases, namely Alzheimer’s disease (AD) and multiple types of cancers. First, in an in-depth study, we shed light on hurdles derived from heterogeneities and out-line how they can restrict the impact of computational models in clinical practice with special focus in AD. Then, to demonstrate the findings of the preceding work, we present a comparative study on characterizing the order of pathological markers by applying a computational model, more specifically a data-driven one, to multiple independent datasets collected in different research centers. In this work, we investigate how heterogeneity across datasets can result in disparities among the ordering of changes in AD biomarkers and influence the models’ impact on clinical practices. Further, to provide a more meaningful biological context into AD pathology, we use a pure knowledge-driven approach to showcase different mechanisms of disease development and progression that genetic variants may cause. Finally, we conclude this thesis by proposing a novel methodology to address heterogeneity among cancer patients in the context of disease treatment. In this publication, with the help of highly predictive machine learning models and an innovative scoring algorithm, we evaluate whether a given sample that was formerly classified as diseased could be predicted as normal after treatment with a given drug taking into account the corresponding molecular signatures of that particular sample. <br /> In summary, this thesis presents the challenges and their implications brought on by heterogeneities in the biomedical domain in order to understand disease pathology and possible treatments, and attempt to uncover avenues to tackle the hindrances. Such advances have numerous applications in the biomedical field, ranging from patient stratification to drug discovery and achieving the ideal of precision medicine.