Abstract
Human Alzheimer’s disease (AD) astrocytes are transcriptionally heterogeneous across brain regions, pathological stages and cohorts, complicating the distinction between reproducible disease signals and dataset-specific effects. We asked whether CREB5, identified by donor-level pseudobulk analysis in SEA-AD middle temporal gyrus (MTG), retained a positive astrocyte association across sequential regional, independent-cohort, confirmatory and prespecified external-validation analyses, and then examined the neuropathological context of this replicated signal. CREB5 was increased in SEA-AD MTG (log2FC 1.203, FDR 0.019), remained positive in SEA-AD dorsolateral prefrontal cortex and both regions of GSE160936, and was independently confirmed in a diagnosis- and target-insulated reanalysis of GSE157827 (log2FC 0.857, P = 3.78×10−5). Three additional accessions were then analyzed under a fixed internal validation plan. GSE188545 supported the association after Holm correction, whereas GSE138852 and GSE174367 yielded smaller, nonsignificant positive estimates. Their prespecified fixed-effect meta-analysis gave a pooled log2FC of 0.678 (95% CI 0.294–1.062; P = 5.36×10−4), with moderate heterogeneity (I² = 58.6%); random-effects sensitivity remained positive. Across six cohorts and eight primary regional or cohort analyses, all eight estimates were positive. After the primary validation sequence was frozen, author-provided sample-level astrocyte data from the five-region AD Progression Atlas (GSE268599; 32 donors) were analyzed with donor as the biological unit. CREB5 was higher in Pathology Groups 3–4 than in Group 1 (donor-level Welch P = 6.19×10−5), and a region-adjusted analysis with donor-clustered standard errors supported the same association (β = 0.379, 95% CI 0.209–0.548; P = 1.16×10−5). Portal-reported correlations with local Aβ plaque or pTau/total-tau burden remained nonsignificant after correction. A separate atlas-scale Malva audit showed heterogeneous CREB5 behavior across additional AD astrocyte studies and was retained as contextual rather than inferential evidence. Together, these findings support CREB5 as a reproducible human AD astrocyte-associated signal with broader neuropathological-stage context, while arguing against a universal effect or a simple linear relationship with local plaque or tau burden.