Balanced deep learning on multi-omics networks identifies molecular subgroups of pathological brain aging

Area of Science:

• Neuroscience
• Genomics
• Computational Biology

Background:

• Neurodegenerative diseases like Alzheimer's disease (AD) present significant heterogeneity, complicating diagnosis and treatment.
• High-dimensional, imbalanced multi-omics data integration with biological networks is a key methodological challenge.

Purpose of the Study:

• To develop and validate a network-informed multi-omics integration framework for identifying molecular subgroups in brain aging.
• To characterize the molecular and neuropathological differences across identified subgroups.

Main Methods:

• Developed a framework integrating multi-omics data (transcriptomics, proteomics, metabolomics) with brain networks from ROS/MAP cohorts.
• Utilized graph embedding (AD Atlas) and autoencoders to derive multi-omics expression scores for hierarchical clustering.
• Validated subgroup robustness and classifier performance in an independent cohort.

Main Results:

• Identified five molecular subgroups with distinct cognitive performance and neuropathological profiles.
• Demonstrated reliable subgroup discrimination using cross-validated classification of transcriptomic and proteomic data.
• Revealed stage-dependent biological patterns, including synaptic/immune activation, mitochondrial dysfunction, and proteostatic impairment.

Conclusions:

• The framework successfully identified molecular subgroups of brain aging, offering insights beyond clinical diagnosis.
• A spectrum of disease progression was observed, from at-risk controls to typical Alzheimer's disease, differentiated by tau pathology.
• Identified subgroups include a reference control, a mixed pathology group, and stages of AD progression.