A deep-learning microglial response score integrates single-cell and spatial transcriptomics for Alzheimer’s disease stratification

Background Microglia exhibit diverse responses in Alzheimer’s disease (AD), yet linking these responses to tissue pathology and disease heterogeneity remains challenging. We aimed to develop a deep-learning microglial response score (DL-MRS) integrating single-cell, spatial and bulk transcriptomics to characterize AD-associated transcriptional patterns. Methods We analyzed human single-cell, spatial and bulk transcriptomic data to characterize microglial clearance and inflammatory programs. These programs and tissue-context features formed eight module-level inputs to a multilayer perceptron that generated DL-MRS. Model performance and feature contributions were assessed using five repeats of five-fold cross-validation, cross-cohort testing, feature-set comparisons and SHapley Additive exPlanations (SHAP) analysis. Spatial localization, individual-level prediction using spatial transcriptomic data, and molecular heterogeneity within AD were also examined. Results Single-cell analyses characterized clearance-enriched and inflammation-enriched microglial groups. Descriptive spatial projection visualized regional variation in the weighted microglial and tissue-context programs. Across 697 AD and control bulk samples, DL-MRS achieved a repeated out-of-fold area under the receiver-operating characteristic curve (AUC) of 0.939 and a mean cross-cohort transfer AUC of 0.947. SHAP analysis of the expanded model identified complement as the largest contributor, alongside contributions from microglial and tissue-context modules. Among 129 AD samples, median DL-MRS was higher in the immune-enriched than in the neuronal/synaptic-enriched molecular group (0.970 versus 0.800). These findings support DL-MRS as an interpretable research tool for integrating microglial and tissue-context signals to characterize AD-associated transcriptional patterns and within-disease molecular heterogeneity. Conclusion DL-MRS integrates microglial and tissue-context signals into an interpretable measure of AD-associated transcriptional remodeling. By capturing coordinated variation across immune, glial, vascular, neuronal and synaptic programs, DL-MRS reveals molecular heterogeneity among AD samples and provides a framework for investigating interindividual differences and advancing molecular stratification in AD research.