An interpretable deep learning framework identifies proteomic drivers of Alzheimer's disease

Area of Science:

• Neuroscience
• Computational Biology
• Proteomics

Background:

• Alzheimer's disease (AD) is a leading neurodegenerative disorder with unclear pathogenesis and no cure.
• Multi-omics data from healthy and AD individuals are available, but machine learning models often lack interpretability.
• The proteomic landscape of the AD brain is less understood than its genetic landscape.

Purpose of the Study:

• To develop an interpretable deep learning method for analyzing complex proteomics data in Alzheimer's disease.
• To identify novel molecular drivers and signaling pathways implicated in AD pathogenesis.
• To explore the relationship between proteomic alterations and patient age at death.

Main Methods:

• Developed EnsembleOmicsAE, a deep learning ensemble of autoencoders, to reduce proteomics data complexity into stable latent features.
• Combined brain proteomic data from 559 individuals across three AD cohorts.
• Implemented an iterative feature scrambling algorithm to calculate protein feature importance and identify signaling modules enriched in protein-protein interactions.

Main Results:

• EnsembleOmicsAE generated stable latent features suitable for biological interpretation.
• Identified novel AD molecular drivers, including integrin signaling and cell adhesion, missed by linear methods.
• Characterized relationships between identified signaling modules and age of death, revealing differential regulation of vimentin and MAPK signaling in younger versus older AD patients.

Conclusions:

• EnsembleOmicsAE provides an interpretable deep learning framework for Alzheimer's disease proteomics research.
• The method successfully identified previously unrecognized molecular pathways contributing to AD.
• Findings suggest distinct proteomic profiles associated with age in AD patients, highlighting vimentin and MAPK signaling.