Enhancing Alzheimer's disease prediction using genetic-guided brain volumetric phenotype network

1401 

Abstract Submission 

Yonghyun Nam¹, Jakob Woerner¹, Sang-Hyuk Jung¹, Erica Suh¹, Haochang Shou¹, Li Shen¹, Dokyoon Kim¹

¹Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania, Philadelphia, PA

Yonghyun Nam  
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
Philadelphia, PA

Jakob Woerner  
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
Philadelphia, PA

Sang-Hyuk Jung  
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
Philadelphia, PA

Erica Suh  
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
Philadelphia, PA

Haochang Shou  
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
Philadelphia, PA

Li Shen  
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
Philadelphia, PA

Dokyoon Kim  
Department of Biostatistics, Epidemiology and Informatics, University of Pennsylvania
Philadelphia, PA

Genetic predispositions related to brain volumetric phenotypes are known to be associated with complex brain-related traits, including Alzheimer's disease (AD), which often involves significant brain volume reduction. Recent genome-wide association studies (GWASs) have shown that brain imaging-derived phenotypes (IDP) are phenotypically and genetically associated with AD. However, existing AD risk prediction models, primarily based on conventional polygenic risk scores (PRSs), are limited in capturing the complex relationships between IDPs and AD. To address these limitations, we have developed a network-based risk scoring model to enhance AD risk prediction ability.

We propose the BrainNetScore, a novel AD risk scoring model that quantifies genetic impacts of associations among multiple brain IDPs and AD incidence. First, a brain connectivity network was constructed using genetic correlations from the UK Biobank GWAS summary statistics across 96 regional brain volume IDPs. Subsequently, this network was expanded into a heterogenous BrainNet graph, comprising 96 IDPs and their associated 12,043 common variants (SNPs), by attaching significant SNPs (P-value < 5e-8) to each brain IDP. The set of SNPs in BrainNet were then used to obtain individual genotype information from independent cohorts, such as the Alzheimer's Disease Neuroimaging Initiative (ADNI). Next, label propagation algorithms were applied to generate individualized predicted scores for each IDP, leveraging genetic correlations and SNP effect directions. To provide a comprehensive view of risk factors with respect to AD, logistic regression was utilized to aggregate the predicted scores for the IDPs, denoting these aggregated scores as BrainNetScore.

To build the BrainNet, we obtained the GWAS summary statistics of 96 brain volume IDPs for UK biobank European participants downloaded from the Brain Imaging Genetics Knowledge Portal (Big-KP). Individual genotype data for 914 samples (550 AD cases and 364 cognitive normal controls) were collected from ADNI and ADNI-WGS-2 with ADSP Follow-Up Study after excluding samples involved in International Genomics of Alzheimer's Project (IGAP) consortium. To demonstrate the utility of BrainNetScore in predicting AD status compared to conventional risk models, PRSs, as a baseline model, were calculated using pruning and thresholding (PRS P+T, P-value threshold as 1e-6) with GWAS summary statistics from IGAP. We assessed the predictive performance of three models: a singleton scoring model, an additive model incorporating covariates (sex, age, and APOE), and a combined model integrating PRS and BrainNetScore. Across 10-fold cross-validation, the combined PRS + BrainNetScore model yielded an average AUC of 0.684 ± 0.034, surpassing both the PRS only model (0.595 ± 0.075) and the BrainNetScore only model (0.666 ± 0.029). Models with sex as a covariate further enhanced the predictive robustness of the combined model (0.684 ± 0.030). Additionally, when integrating APOE genotypes into the combined model, the predictive accuracy further increased (0.778 ± 0.043), suggesting the significant role of genetic factors in AD risk assessment (Figure 1).

The study demonstrates the enhanced predictive ability of the BrainNetScore model for AD risk, particularly when combined with conventional PRSs. These findings underscore the value of incorporating network-based approaches and comprehensive genetic information in improving AD risk prediction models. The enhanced predictive capability of the combined model shows promise for personalized medicine and early intervention strategies in AD and brain-related traits. Future research should focus on expanding the model's applicability to a wider range of populations.

Genetics Other ²

Classification and Predictive Modeling ¹

Connectivity (eg. functional, effective, structural)

Methods Development

Aging

Data analysis

Degenerative Disease

Machine Learning

Neurological

Phenotype-Genotype