Interpretable learning for brain cell annotation and Alzheimer’s classification

Alzheimer’s disease (AD), the most prevalent neurodegenerative disorder, poses a major public health challenge in aging populations. Although AD is recognized as a multicellular disease with cell-type-specific dysfunctions, machine learning frameworks that simultaneously annotate brain cell subpopulations and discriminate AD from healthy controls using single-cell transcriptomic profiles remain limited. We present a computational framework bridging this gap using single-cell RNA sequencing data. Integrating six feature selection methods and four base classifiers, we identified a 229-gene set via TURF feature selection combined with the XGBoost classifier, achieving superior performance in the annotation of six major brain cell subpopulations (accuracy, 97.42%; precision, 96.19%; recall, 94.76%; F1-measure, 95.44%). Biological interrogation of these genes revealed significant enrichment in neuronal signaling pathways and identified potential cell-type-specific markers, including ST18, LHFPL3, GPM6A, and SYT1. Building upon the 229-gene set, five machine learning models were constructed to distinguish AD patients from healthy individuals. The ensemble model achieved the best performance, with an accuracy of 92.08% and an AUC of 0.96. To enhance applicability, we implemented four dimensionality reduction strategies and identified 53 candidate AD-associated genes, including several well-documented genes such as SNAP25, NLGN1, GPM6A, and LRRTM4. Notably, these 53 genes retained strong classification performance, with an accuracy of 86.13% and an AUC of 0.90. To facilitate applications, we developed a webserver enabling brain cell annotation and AD status classification. This study provides a robust framework for exploring cell-type-specific molecular characteristics of AD and offers practical tools for brain single-cell annotation and AD status classification.