Robust transcriptomic signatures of Alzheimer's disease progression: validated explainable AI approach.

The lack of validated stage-specific biomarkers hampers the understanding of Alzheimer's disease (AD) progression and clinical translation. Current transcriptomic methods often produce unstable results with limited stage discrimination. We aimed to develop an explainable machine learning pipeline to identify robust, interpretable gene signatures linked to distinct AD neuropathological stages. We analyzed multi-region transcriptomic data from the MSBB cohort using a multi-class XGBoost-SHAP-driven framework. Class imbalance was addressed using SMOTE, and model robustness was assessed through cross-validation and permutation-based validation. Our pipeline accurately classified Early, Mid, and Late Braak stages and achieved accurate Braak stage classification (regional ROC AUCs up to 0.76). The method identified a concise set of high-confidence, stage-specific genes, showing minimal signature overlap (~ 1.7%). Key validated novel candidate biomarkers included ARX (Early), MKNK2 (Mid), and SLC25A16/NEURL1B (Late), linked to GABAergic, inflammatory, mitochondrial, and synaptic pathways. This explainable framework overcomes key limitations of conventional analyses, providing a stable, interpretable gene signature for AD staging. It establishes a robust method for transcriptional biomarker discovery and offers new biological insight into AD progression, highlighting potential stage-specific therapeutic targets.