Assessing polyomic risk to predict Alzheimer's disease using a machine learning model.

Area of Science:

• Biomarkers and diagnostics
• Computational biology and bioinformatics
• Neurodegenerative disease research

Background:

• Alzheimer's disease (AD) is the leading cause of dementia in older adults.
• Neuropathology of AD begins decades prior to symptom manifestation, highlighting the need for early detection tools.
• Early intervention strategies for AD can be facilitated by timely and accurate screening methods.

Purpose of the Study:

• To develop and evaluate polyomic prediction models for Alzheimer's disease (AD) affection status and age at onset.
• To identify key predictive features and informative data modalities for early AD detection.
• To assess the utility of "AD-by-proxy" cases in enhancing prediction models.

Main Methods:

• Utilized tree-based and deep learning algorithms to train polyomic prediction models.
• Integrated genomic, proteomic, metabolomic, and drug use data from UK Biobank.
• Employed SHAP analysis to determine feature importance and identify key predictors.

Main Results:

• The best polyomic model achieved an AUROC of 0.87 for AD prediction.
• Glial fibrillary acidic protein (GFAP) and CXCL17 proteins were identified as the strongest predictors, alongside apolipoprotein E (APOE) alleles.
• Incorporating "AD-by-proxy" cases did not significantly improve model performance.

Conclusions:

• Genomics and proteomics emerged as the most informative data modalities for AD prediction based on AUROC.
• Blood-based biomarkers GFAP and CXCL17 show potential for early, presymptomatic AD prediction.
• The developed polyomic model effectively predicts AD and age at onset using omics and EHR data.