Neural pathway–informed unsupervised learning for cross-resolution transcriptomic profiling of Alzheimer’s disease heterogeneity

Alzheimer’s disease (AD) is a heterogeneous neurodegenerative disorder, yet the molecular and cellular mechanisms underlying its heterogeneity remain poorly understood. Previous studies have largely relied on a single molecular modality, limiting comprehensive characterization of AD heterogeneity. We aimed to characterize AD heterogeneity across bulk, single-nucleus, and spatial transcriptomic data using a common neural pathway–informed analytical framework. We developed a neural pathway-informed unsupervised learning framework based on the activities of 25 curated neural pathways. The framework was applied separately to six bulk brain transcriptomic datasets, a single-nucleus RNA sequencing dataset, and a spatial transcriptomic dataset to identify neural pathway–associated subgroups and compare their biological characteristics across resolutions. We comprehensively characterized subgroup-specific clinical features, DNA repair capacity, metabolic activity, immune signatures, cellular heterogeneity, cell-cell communication, and spatial distributions. A common three-subgroup partition characterized relatively high, intermediate, and low neural enrichment across the analyzed bulk transcriptomic datasets. Neu-H exhibited enhanced metabolic function and DNA repair capacity, lower immune signatures, and reduced tissue-level cellular heterogeneity, whereas Neu-L showed opposite characteristics and was associated with more severe clinical pathology. Compared with controls, AD samples showed lower neural gene-set enrichment in four of six bulk datasets, no significant difference in ROSMAP, and higher enrichment in GSE5281. Lower DNA-repair and metabolic gene-set enrichment and higher immune signatures were observed in most bulk datasets. Single-nucleus analyses identified neuronal subgroups with differing neural pathway enrichment and associated molecular characteristics, providing a cellular perspective on the patterns observed in bulk tissues. Neu-H excitatory and inhibitory neurons were more diverse, differentiated, metabolically active, and DNA repair–competent than Neu-L neurons, but AD neurons in all three subgroups exhibited lower DNA repair capacity, increased heterogeneity, and weakened cell–cell communication than those of controls. Spatial transcriptomics recapitulated these patterns in situ: Neu-H spots were enriched in neurons and gray matter, whereas Neu-L spots predominated in white matter and had increased microglial and oligodendrocyte precursor cell representation. Compared with controls, AD tissues showed greater representation of Neu-L spots and lower neural, DNA-repair, and metabolic gene-set enrichment scores, suggesting an association between these transcriptional patterns and AD pathology. Our neural pathway–informed cross-resolution framework characterizes AD heterogeneity at tissue, cellular, and spatial levels, highlighting associations among neural pathway expression, metabolic and DNA repair signatures, immune signatures, and tissue organization.