Basic Science and Pathogenesis.

Area of Science:

• Genomics
• Computational Biology
• Neuroscience

Background:

• Early prediction of Alzheimer's disease (AD) is crucial for timely intervention.
• Genomic data offers a promising avenue for AD prediction, but its high dimensionality presents challenges.
• Deep learning (DL) models are well-suited for analyzing complex genomic datasets.

Purpose of the Study:

• To develop a novel deep learning model utilizing a transformer architecture.
• To predict Alzheimer's disease (AD) dementia solely from whole genome sequencing (WGS) data.
• To investigate the model's ability to integrate local and distant genetic relationships for enhanced prediction accuracy.

Main Methods:

• Analysis of 1,050 whole genome sequencing (WGS) datasets from ADNI and ADSP-FUS1-ADNI-WGS-2.
• Development of a three-step DL model: window-based attention, annotation-enhanced analysis, and graph convolutional network (GCN) integration.
• Focus on a 1,000,000 base pair segment of WGS data surrounding the APOE gene on chromosome 19.

Main Results:

• The complete DL model achieved a significant prediction accuracy (AUC = 0.68) compared to the window-based step (AUC = 0.53).
• The model identified key single nucleotide polymorphisms (SNPs) associated with AD dementia, including novel discoveries.
• Top SNPs rs429358, rs120074114, rs120074111, and rs5122, located in or near APOE/APOC genes, were highlighted.

Conclusions:

• The developed DL model effectively performs multi-scale analysis of WGS data for improved AD dementia prediction.
• The transformer architecture facilitates the integration of local and distant genetic relationships, enhancing predictive power.
• This study establishes a foundation for utilizing DL models with WGS data in early AD prediction and efficient data processing.