Disentangling Context-Specific Alzheimer's Disease Progression Signatures through an Explainable AI Framework

Alzheimer's disease (AD) progression involves stage-dependent transcriptional changes that are often masked by dominant cellular and biological variation. Here, we present ADVISE (Alzheimer's Disease Variational Isolation of Signal Embeddings), an interpretable deep generative framework designed to disentangle Braak stage-associated transcriptional signals from background biological variation through separate latent representations. Using explainable AI, ADVISE prioritizes genes and pathways contributing to the disease-specific representation. Applying ADVISE to large-scale single-nucleus RNA sequencing data from the human prefrontal cortex, we identify shared and context-specific signatures of AD progression across global, sex-specific, and cell-type-specific analyses, revealing convergent molecular programs involving synaptic transmission, HSF1-associated proteostatic stress responses, and SLIT/ROBO axon guidance. Notably, ADVISE prioritizes ROBO2 as a progression-associated candidate in excitatory neurons. Experimental validation using CRISPR activation in hiPSC-derived excitatory neurons demonstrates that ROBO2 activation reduces the phosphorylated Tau-to-total Tau ratio. ADVISE thus reveals context-dependent molecular signatures of AD progression and bridges human single-cell discoveries with experimental validation of candidate regulators of AD pathology.