Vulnerable Brain Networks Associated with Risk for Alzheimer’s Disease

Abstract

AbstractBrain connectomes provide untapped potential for identifying individuals at risk for Alzheimer’s disease (AD), and can help provide novel targets based on selective circuit vulnerability. Age, APOE4 genotype, and female sex are thought to contribute to the selective vulnerability of brain networks in Alzheimer’s disease, in a manner that differentiates pathological versus normal aging. These brain networks may predict pathology otherwise hard to detect, decades before overt disease manifestation and cognitive decline. Uncovering network based biomarkers at prodromal, asymptomatic stages may offer new windows of opportunity for interventions, either therapeutic or preventive. We used a sample of 72 people across the age span to model the relationship between Alzheimer’s disease risk and vulnerable brain networks. Sparse Canonical Correlation analysis (SCCA) revealed relationships between brain subgraphs and AD risk, with bootstrap based confidence intervals. When constructing a composite AD risk factor based on sex, age, genotype, the highest weight was associated with genotype. Next, we mapped networks associated with auditory, visual, and olfactory memory, and identified networks extending beyond the main nodes known to be involved in these functions. The inclusion of cognitive metrics in a composite risk factor pointed to vulnerable networks, and associated with the specific memory tests. These regions with the highest cumulative degree of connectivity in our studies were the pericalcarine, insula, banks of the superior sulcus and cerebellum. To help scale up our approach, we extended Tensor Network Principal Component Analysis (TNPCA) to evaluate AD risk related subgraphs, introducing CCA components and sparsity. When constructing a composite AD risk factor based on sex, age, and genotype, and family risk factor the most significant risk was associated with age. Our sparse regression based predictive models revealed vulnerable networks associated with known risk factors. The prediction error was 17% for genotype, 24% for family risk factor, and 5 years for age. Age prediction in groups including MCI and AD subjects involved several regions that were not prominent for age prediction otherwise. These regions included the middle and transverse temporal, paracentral and superior banks of temporal sulcus, as well as the amygdala and parahippocampal gyrus. The joint estimation of AD risk and connectome based mappings involved the cuneus, temporal, and cingulate cortices known to be associated with AD, and add new candidates, such as the cerebellum, whose role in AD is to be understood. Our predictive modeling approaches for AD risk factors represent a stepping stone towards single subject prediction, based on distances from normative graphs.