Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)



First Advisor

Matt Mahoney

Second Advisor

Dimitry Krementsov


Alzheimer’s Disease (AD) is a devastating neurodegenerative disorder affecting all tissues and cell types of brain leading to emotional dysregulation and cognitive dysfunction. From genome-wide association studies (GWAS), to date we have identified forty-two genome-wide significant genes for AD that influence overall disease risk or endophenotypes, including neuroimaging and gene expression profiles. Nevertheless, the currently known AD genes do not account for a significant proportion of the heritability of disease risk, implying the existence of many weak-effect variants in potentially thousands of genes as drivers of AD outcomes. This genetic architecture, composed of many small effects, is partly due to the complexity of molecular interaction networks, where the influence of individual genetic variants is attenuated by overlapping molecular pathways that are tuned by evolution for robustness. Thus, overcoming the limitations of GWAS for dissecting the mechanisms of AD requires methods to identify disease pathways that are enriched for weak genetic effects. Network-based functional prediction (NBFP) methods use machine learning in gene interaction networks to robustly learn pathways containing risk genes, which augments the raw statistical signal from GWAS with biological prior knowledge encoded in a tissue-specific gene network. NBFP methods have several benefits, including their robustness to statistical noise over raw GWAS statistics and they enable nomination of functionally relevant candidate genes that do not themselves carry risk polymorphisms. However, most NBFP methods are currently limited to single tissues, which is not optimal for complex disorders like AD that involve many functionally distinct cell types and brain regions. Moreover, there are now multiple studies of AD endophenotypes, including brain-region specific gene expression and whole-brain neuroimaging, both paired with genotypes, and single-cell gene expression data. Thus, there is an additional need for integrative tools that combine disparate data sources to nominate candidate genes for distinct pathophysiological processes. In this work, we developed new methods to rank candidate genes based on multiple disease-relevant networks and to combine gene rankings arising from multiple sources, including NBFP, imaging GWAS, and gene expression. In our first study, we applied NBFP to systematically rank AD-risk genes in the hippocampus and amygdala and developed a novel combined scoring method to integrate these scores with GWAS associations for low hippocampal and amygdalar volume in patients with AD. Our method nominated a novel set of region-specific candidate genes primarily involved in maintaining the stability of the synapse and regulating excitotoxicity. In our second study, we developed a multi-network-based functional prediction (mNBFP) to allow multiple source networks. Using three brain-cell-specific networks, our mNBFP approach outperformed single-network approaches in training performance and achieved high concordance with recently published AD-GWAS associations. In our third and final study we integrated our multi-network NBFP and combined scoring approaches with single-cell gene expression and the Library of Integrated Network-based Cellular Signatures (LINCS) database to identify potential drug repositioning candidates for AD. We identified the protein AP1B1 as having strong potential to target an early-AD gene expression signature, which may yield a novel mechanism for early therapeutic intervention.



Number of Pages

203 p.