A Pathway-Based Machine Learning Approach Identifies Region-Specific Markers and Patterns in Alzheimer’s Disease Patients Based on Spatial and Severity Metadata

Alzheimer’s disease (AD) exhibits profound spatial heterogeneity in its molecular and pathological features, yet the basis of this regional selectivity remains obscure. Here, we analyze transcriptomic profiles from publicly available bulk RNA-sequencing datasets collected from two cortical regions implicated in AD vulnerability: the insular cortex and Brodmann area 32 (BA32) of the anterior cingulate cortex. Following rigorous covariate correction and region-specific differential expression analysis, we observe extensive transcriptional disruption in the insula, comprising 4,883 significantly dysregulated genes, while BA32 exhibits no differentially expressed genes at the same statistical threshold. Pathway-level modeling using donor-aware, nested elastic-net logistic regression accurately classified AD versus control cortical samples (pooled BA32 and insula; mean AUROC = 0.813 ± 0.127; AP = 0.938 ± 0.040). SHAP-based interpretability identified a concise set of biologically coherent pathways underlying predictive performance. Cross-regional generalization revealed partial transferability of insular signatures to BA32, indicating shared yet muted molecular changes in the latter. Comparisons across Braak stages suggested stage-associated shifts in cholesterol, complement, and oxidative phosphorylation pathways, consistent with progressive remodeling of metabolic and immune programs. These findings are consistent with AD being a regionally patterned, rather than transcriptionally uniform, disease. They provide a framework for prioritizing region-specific molecular pathways that may inform future therapeutic exploration.