Epigenomic subtypes of late-onset Alzheimer's disease reveal distinct microglial signatures

Abstract

Growing evidence suggests that clinical, pathological, and genetic heterogeneity in late onset Alzheimer's disease (LOAD) contributes to variable therapeutic outcomes, potentially explaining many trial failures. Advances in molecular subtyping through proteomic and transcriptomic profiling reveal distinct patient subgroups, highlighting disease complexity beyond amyloid-beta plaques and tau tangles. This underscores the need to expand subtyping across new molecular layers, to identify novel drug targets for different patient subgroups. In this study, we analyzed genome-wide DNA methylation (DNAm) data from three independent postmortem brain cohorts (N = 831) to identify epigenetic subtypes of LOAD. Unsupervised clustering approaches were employed to identify distinct DNAm patterns, with subsequent cross-cohort validation. We assessed how subtype-specific methylation signatures map onto individual brain cell types by comparing them with DNAm profiles from purified cells. Next, we integrated bulk and single-cell RNA-seq data to determine each subtype's functional impact on gene expression. Finally, we explored clinical and neuropathological correlates of the identified subtypes to elucidate biological and clinical significance. We identified two distinct epigenomic subtypes of LOAD, consistently observed across three cohorts. Both subtypes exhibit significant yet distinct microglial methylation enrichment. Bulk transcriptomic analyses further highlighted distinct biological mechanisms underlying these subtypes: subtype 1 was enriched for immune-related processes, while subtype 2 was characterized by neuronal and synaptic pathways. Single-cell transcriptional profiling of microglia revealed subtype-specific inflammatory states: subtype 1 displayed chronic innate immune hyperactivation with impaired resolution, whereas subtype 2 exhibited a more dynamic inflammatory profile, balancing pro-inflammatory signaling with reparative and regulatory mechanisms. These findings reveal distinct epigenetic and functional microglial states underlying LOAD subtypes, advancing our understanding of disease heterogeneity. This work lays the groundwork for targeted therapeutic strategies tailored to specific molecular and cellular disease profiles.

Keywords: Alzheimer’s disease; DNA methylation; Epigenetics; Microglia; Subtyping.

Figure 1

Cross-cohort generalizability of DNAm-based clusters. (a) Visual representation of the correlation-based replication of the identified clusters across three cohorts using Hierarchical (Left), and K-means (Right) methods. The first block (LOADS1) includes UKBBN cluster A, PITT-ADRC cluster B, and ROSMAP cluster A, while the second block (LOAD-S2) includes UKBBN cluster B, PITT-ADRC cluster C, and ROSMAP cluster B. (b) Spatial overlap analysis showing cluster assignments across the three cohorts. In this example, the clusters were confirmed through iterative projections, where UKBBN cohort’s first two latent spaces were projected onto the PTT-ADRC and ROSMAP to verify spatial overlap across datasets.

Figure 2. Characterization of the methylomic signatures of predicted LOAD subtypes.

Visualization of the first two PCs for LOAD subtypes across all cohorts. Plot of the first two PCs (PC1 and PC2) for DNAm profiles from all samples in the UKBBN, PITT-ADRC, and ROSMAP cohorts. Samples are labeled according to their subtype assignments: Subtype 1 (LOAD-S1) and Subtype 2 (LOAD-S2). ‘Unassigned’ samples are also indicated, representing clusters without matching correlations across cohorts.

Figure 3. Transcriptomic characterization of LOAD epigenomic Subtypes.

(a) Heatmap showing Z-scores for the top 20 enriched Gene Ontology (GO) terms in LOAD-S1 and LOAD-S2 subtypes. Z-scores indicate the deviation of observed gene counts from expected values, normalized by standard deviation. LOAD-S1 is enriched in immune-related pathways, reflecting a dominant immune/inflammatory signature, while LOAD-S2 shows enrichment in synaptic and neuronal pathways. Immune pathways prevalent in LOAD-S1 are demonstrating weak enrichment in LOAD-S2, and vice versa. (b) Comparative number of DEGs across seven microglial (MG) states (MG 0–6) highlights distinct gene expression profiles for LOAD-S1 and LOAD-S2, with minimal overlap, underscoring subtype-specific transcriptional landscapes. (c) shows GO term enrichment for the 269 DEGs identified across MG 0–MG 6 in ROSMAP single-cell pseudobulk profiles.