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

Area of Science:

• Neuroscience
• Genomics
• Computational Biology

Background:

• Alzheimer's disease (AD) progression understanding is limited by a lack of validated, stage-specific biomarkers.
• Current transcriptomic methods yield unstable results, hindering accurate AD staging and clinical translation.

Purpose of the Study:

• To develop an explainable machine learning pipeline for identifying robust and interpretable gene signatures associated with distinct AD neuropathological stages.
• To overcome limitations of conventional transcriptomic analyses for biomarker discovery in AD.

Main Methods:

• Utilized a multi-class XGBoost-SHAP framework on multi-region transcriptomic data from the MSBB cohort.
• Addressed class imbalance with SMOTE and ensured model robustness via cross-validation and permutation-based validation.
• Applied an explainable AI approach for biomarker identification and stage discrimination.

Main Results:

• The pipeline accurately classified Early, Mid, and Late Braak stages with high regional ROC AUCs (up to 0.76).
• Identified a concise set of high-confidence, stage-specific genes with minimal overlap (approx. 1.7%).
• Validated novel candidate biomarkers: ARX (Early), MKNK2 (Mid), and SLC25A16/NEURL1B (Late), linked to key biological pathways.

Conclusions:

• The explainable framework provides a stable, interpretable gene signature for AD staging, surpassing conventional methods.
• Establishes a robust methodology for transcriptional biomarker discovery in neurodegenerative diseases.
• Offers novel biological insights into AD progression and identifies potential stage-specific therapeutic targets.