Integrated postmortem brain transcriptomics identifies a five-gene Alzheimer’s disease-associated signature with a reproducible whole-blood CEBPB correlate

Transcriptomic profiling provides a systems-level view of Alzheimer’s disease (AD), capturing coordinated neuronal, glial, vascular, and immune-related changes. Yet the same breadth that makes transcriptomics informative also complicates interpretation because technical heterogeneity, anatomical sampling, cellular composition, and outcome-guided feature selection can amplify apparent disease signals. We therefore sought to derive a compact AD-associated transcriptomic signature and, critically, to define the biological and analytical conditions under which that signal remains interpretable. We integrated 593 array-level profiles from three postmortem brain microarray series (410 AD -group and 183 controls), performed differential-expression analysis, and applied a sequential LASSO/XGBoost/SVM-RFE prioritization workflow. GSE33000 and GSE44770 were used for external candidate evaluation and final panel prioritization. The resulting five-gene signature was then interrogated across donor-aware cortical snRNA-seq, CNS-oriented and exploratory leukocyte-related transcriptional-context analyses, cross-disease testing, multi-cohort covariate and CNS composition sensitivity analyses, Braak neuropathology, retrospective development-sample model characterization, two PAXgene whole-blood cohorts, and a same-subject benchmark against established CSF biomarkers. The integrated discovery matrix yielded 109 differentially expressed genes, of which 25 remained in the sequential feature-prioritization intersection. External candidate evaluation converged on NCALD, GFAP, CD200, CEBPB, and AEBP1, and the resulting five-gene model showed strong observed separation in GSE33000 (AUC 0.934, 95% CI 0.910–0.957) and GSE44770 (0.949, 0.922–0.975). Cross-disease testing showed heterogeneous non-AD performance, supporting an AD-associated rather than AD-exclusive interpretation. Positive score effects persisted under several available covariate models but attenuated to different degrees after CNS composition adjustment. Donor-aware snRNA-seq identified region-robust cell-class-specific effects, including lower NCALD in oligodendrocytes and higher CEBPB and CD200 in excitatory neurons. In GSE125583, 75.4% of the age/sex-adjusted AD-associated score effect remained after Braak adjustment. CEBPB was also reproducibly higher in AD whole blood in both GSE63060 and GSE63061. Together, these analyses define a five-gene AD-associated transcriptomic signature with strong observed brain-cohort separation, donor-aware cellular resolution, context-dependent persistence under covariate and composition adjustment, a Braak-associated neuropathological relationship, and replicated peripheral CEBPB expression. The study therefore provides a biologically anchored framework for interpreting the signature while setting explicit boundaries on disease specificity, mechanistic inference, and clinical translation.