Interpretable Machine Learning Framework Annotates Brain Cells and Classifies Alzheimer’s Disease

A computational framework using single-cell RNA sequencing data can simultaneously annotate major brain cell subpopulations and classify Alzheimer’s disease status, according to a study published in BMC Biology. Developed by a research team including Hao Du, Yi-Yang Wang, Lu Jia, and colleagues, the machine learning methodology directly addresses a recognized gap in neurodegenerative disease research. While Alzheimer’s disease is acknowledged as a multicellular disorder involving cell-type-specific dysfunctions, computational tools capable of resolving individual cell identities while simultaneously discriminating affected patients from healthy individuals using single-cell transcriptomic profiles have remained limited.

Alzheimer’s disease represents the most prevalent neurodegenerative disorder and continues to pose a major public health challenge across aging populations worldwide. To bridge the gap in existing computational frameworks, the investigators developed an interpretable learning approach designed to process single-cell RNA sequencing data. Their framework specifically links the identification of distinct cellular subpopulations with the clinical classification of Alzheimer’s disease, providing a unified analytical strategy for transcriptomic datasets.

To identify the optimal feature set for cellular annotation, the researchers systematically integrated six feature selection methods and four base classifiers. Through this comparative analysis, combining Tuned ReliefF, or TURF, feature selection with an XGBoost classifier yielded superior performance in isolating a critical 229-gene set. This model configuration proved highly effective at annotating six major brain cell subpopulations. In benchmark evaluations, the TURF and XGBoost combination achieved an overall accuracy of 97.42 percent, a precision of 96.19 percent, a recall of 94.76 percent, and an F1-measure of 95.44 percent across the cellular subpopulations.

Biological interrogation of the identified 229-gene panel revealed significant functional enrichment in neuronal signaling pathways, reinforcing the biological relevance of the selected transcriptomic signatures. Furthermore, this computational interrogation uncovered potential cell-type-specific markers, explicitly highlighting ST18, LHFPL3, GPM6A, and SYT1.

Building upon the foundational 229-gene set, the researchers constructed and tested five distinct machine learning models to distinguish Alzheimer’s disease patients from healthy individuals. Among the evaluated algorithms, an ensemble machine learning model achieved the highest diagnostic performance. The ensemble model attained an overall classification accuracy of 92.08 percent and an area under the receiver operating characteristic curve, or AUC, of 0.96. These experimental results demonstrated that the gene set identified through cellular subpopulation annotation also retains powerful predictive utility for discriminating disease status in patient cohorts.

To enhance practical applicability and reduce computational overhead, the researchers implemented four dimensionality reduction strategies to identify a more compact biomarker panel. This process yielded 53 candidate Alzheimer’s-associated genes, encompassing several well-documented genes such as SNAP25, NLGN1, GPM6A, and LRRTM4. Notably, these 53 candidate genes retained strong classification performance, achieving an accuracy of 86.13 percent and an AUC of 0.90. To facilitate broader scientific adoption and practical usage, the team also developed a dedicated webserver enabling researchers to perform brain cell annotation and Alzheimer’s disease status classification.

The investigators concluded that their computational system provides a robust and interpretable framework for exploring the cell-type-specific molecular characteristics of Alzheimer’s disease. By simultaneously offering cell annotation capabilities and disease status discrimination, the approach delivers practical computational tools for single-cell neurobiology.

Subject of Research: Machine learning framework for single-cell brain annotation and Alzheimer's classification

Article Title: Interpretable learning for brain cell annotation and Alzheimer’s classification

Article References: Du, H., Wang, Y.-Y., Jia, L., Wang, J., Ta, L., Lv, H., & Zhang, L.-Q. (2026). Interpretable learning for brain cell annotation and Alzheimer’s classification. BMC Biology. https://doi.org/10.1186/s12915-026-02789-z

Image Credits: AI Generated

DOI: 10.1186/s12915-026-02789-z

Keywords: Alzheimer's disease, single-cell RNA sequencing, transcriptomics, brain cell annotation, machine learning, TURF feature selection, XGBoost, ensemble models, neuronal signaling, ST18, LHFPL3, GPM6A, SYT1, SNAP25, NLGN1, LRRTM4, dimensionality reduction