We analyzed DNA methylation in eight brain regions from a unique cohort of participants aged 90 years and older in order to identify DNA methylation differences related to the endophenotypes of the three hallmark neuropathologic lesions of AD (Fig. 1a, b). Using the Illumina 850k platform, we assayed methylation of 853,307 CpGs in eight regions of the human brain: middle frontal gyrus (MFG), cingulate gyrus (CG), entorhinal cortex (EC), hippocampus dentate gyrus (DG), hippocampus CA1 (CA1), substantia nigra (SN), locus coeruleus (LC) and cerebellar cortex (CBM). Computational cell type deconvolution was used to investigate differential methylation in CTS data (Fig. 1d, e).

Characteristics of study participants

After removing samples from individuals with failed methylation array processing (see Methods), our final cohort consisted of 321 samples from 47 individuals (Fig. 1a). Eight brain regions were processed from each individual. Characteristics of study participants are shown in Table 1. The mean age was 97.4 ± 3.5 years, with little difference in age observed between sexes. The majority of individuals were females (n = 35), reflecting the overall demographics of people aged over 90 [41]. Amongst all individuals, 17.32% are heterozygous carriers of the APOE ɛ4 allele; data from 10 individuals were missing due to problems in sequencing. Concerning clinical diagnosis, 40.43% of our participants were diagnosed as cognitively normal, 31.9% had cognitive impairment, no dementia (CIND) and 27.66% were diagnosed with dementia. Only 4 individuals (8.51%) had an overall low neuropathological burden of AD as measured by the AD severity score (see Methods), 15 were classified as intermediate (31.91%) and 28 individuals had a high burden of AD neuropathology (59.57%). Detailed demographics and neuropathological data for each case can be found in Additional File 4.

Computationally deconvoluted cell type proportions unveil regional differences

We estimated cell type proportions of our samples using a cell type decomposition method that utilizes an existing reference panel from EpiScore [12] (see Methods). As expected, neurons were the most abundant cell type in most brain regions (Additional File 1: Fig. S2) and predicted proportions were consistent with observations from previously published epigenome wide association studies [2, 7, 42,43,44,45]. Cell composition varied between brain regions, with the highest proportion of neuronal signal in the cerebellar cortex, followed by regions in the forebrain, midbrain, and finally pons. Similar to previous findings, the proportion of neurons also varied within the same brain region between individuals (Additional File 1: Fig. S2) [46, 47]. Microglia were the least abundant glial cell type, only being detected consistently in SN and LC. Microglia account for 0.5%-16.6% of cells across brain regions, with the pons and basal ganglia showing higher amounts of microglia than cerebral cortical regions [48], but our average proportions of 3.2% ± 4.4 in the SN and 6.8% ± 5.1 in LC were modestly lower than previously reported proportions greater than 10% [48]. Our predicted proportions of oligodendrocytes/OPCs varied from 0% (in the cerebellum) to up to ~ 40% and were consistent with proportions found in the adult mouse brain (0–40%) [49]. The low estimated proportion of glial cells in the cerebellum also matched previous findings [50]. Similar to existing literature, we did not find any significant correlations of cell composition with neuropathological traits or age at death [42,43,44,45].

Cell-type-specific methylation profiles differ across brain regions

We recovered CTS methylation signals using cell type deconvolution with TCA [11], which relies on the cell type proportions estimated with EpiScore [12] (see Methods). Although these CTS profiles were imputed from bulk data and not obtained through sorting, for simplicity, we will refer to them as data from their respective cell types. In addition to analyzing the CpG level data, we also aggregated our deconvolved data by averaging methylation across individual CpGs within promoters, gene bodies and CpG islands. We used principal component analysis (PCA) and random matrix theory to identify CTS methylation profiles with sufficiently non-random signals. We computed the full set of eigenvectors and eigenvalues for each cell type and aggregation type, and excluded CTS profiles where the number of eigenvalues which exceeded the theoretical limit expected under the Marcenko-Pastur distribution [37] for a matrix of random noise was less than or equal to 5 (Additional File 3). Most microglial profiles were excluded, except in the LC and the CpG and gene level data in the SN. We further excluded all datasets of oligodendrocytes/OPCs and astrocytes in the CBM, and the CpG islands dataset of astrocytes in the EC. Except for astrocytes in the EC, these profiles corresponded to cell types with very low average estimated proportions in their respective regions, ranging between 0 and 1.2%. We used the UMAP algorithm [51] to visualize the latent space in two dimensions across CpGs for all cell types and brain regions (Additional File 1: Fig. S3) and found that the bulk profiles from different brain regions embedded closely together relative to the CTS profiles, indicating high homogeneity across regions. With the exception of nigral neurons, neuronal profiles were embedded in proximity to bulk data, reflecting the predominance of the neuronal signal in bulk data. The clear separation of CTS methylation profiles from bulk profiles shows that cell type deconvolution is an effective strategy for disentangling signals from distinct cell types in tissue homogenates (Additional File 1: Fig S3). Figure 2 displays a UMAP plot of CTS profiles across brain regions and cell types and visualizes that cell type and not brain region is the main driver of variance in our data. Neurons and astrocytes of the SN were embedded further from similar cell types in other brain regions. Among the non-neuronal CTS profiles (astrocytes, oligodendrocytes/OPCs, microglia and endothelial cells), we found that profiles from the same cell type were usually embedded more closely to each other than to other samples from the same region (Fig. 2). This observation is broadly consistent with findings reported in single cell ATAC-Seq data, where non-neuronal cell types showed greater homogeneity across regions than neurons [52].

Fig. 2

UMAP plot displaying clustering of brain region and cell-type-specific methylation data. We used the Uniform Manifold Approximation and Projection (UMAP) technique for dimension reduction to visualize similarities across cell type and brain-region-specific methylation data. Dimensionality reduction is performed with all brain regions and cell types combined. Each dot represents one individual sample. Colors reflect the cell types. Olig Oligodendrocytes, OPC Oligodendrocyte Precursor Cells, MFG Middle Frontal Gyrus, CG Cingulate Gyrus, CA1 Hippocampus CA1, DG Dentate Gyrus, EC Entorhinal cortex, SN Substantia nigra, LC Locus coeruleus, CBM Cerebellar cortex

Cell type deconvolution uncovers ADNC related methylation differences not present in bulk data

We used differential methylation to associate bulk and CTS methylation profiles with ADNC endophenotypes in all brain regions and clinical measures (see Methods). We chose to highlight results for associations with averaged promoter methylation in protein-coding genes because these associations are more readily interpretable. Generally, promoter methylation is negatively correlated with gene expression, and limiting the analysis to protein-coding genes avoids concerns about poorly-annotated gene models for non-coding genes. We found no association between ADSS and bulk methylation in any brain region (Additional File 1: Fig. S4). In contrast, numerous associations were found between ADSS and CTS profiles, especially in the dentate gyrus (DG) where we identified 911 differentially methylated promoters (DMPTs) in neurons, all of which were unique to the DG (Fig. 3).

Fig. 3

Overview of the number of differentially methylated promoter-associated protein-coding regions for dentate gyrus and cingulate gyrus by cell type. Barplots display the number of significant (FDR p-value < 0.05) differentially methylated promoter regions of protein-coding genes. Each barplot shows the results for one brain region and neuropathological score combination. Color coding of the bars reflects the different cell types. The dentate gyrus (DG) and cingulate gyrus (CG) were the two regions with the highest amount of differentially methylated promoters (DM promoters, DMPTs) across neuropathological scores. We did not discover any DM promoters across different NIA-AA C scores, and there were no differentially methylated sites found in bulk data. FDR False discovery rate, Olig Oligodendrocytes, OPCs Oligodendrocyte Precursor Cells, CG Cingulate Gyrus, DG Dentate Gyrus, NIA-AA National institute of Aging Alzheimer's Association, AD Alzheimer’s Disease

We wanted to perform a more granular analysis of the association of ADNC endophenotypes with methylation by examining differential methylation for Aβ plaques (NIA-AA A score), neurofibrillary tangles (NIA-AA B score) and neuritic plaques (NIA-AA C score). No associations were seen between the individual scores and bulk methylation profiles in any region except for 2 DMPTs associated with NIA-AA C score in the cerebellum. In the CTS profiles, the largest numbers of DMPTs were found in neurons of the CG in association with neurofibrillary tangles and in neurons of the DG in association with amyloid burden (Fig. 3). The DG was the only brain region where NIA-AA A score was associated with a substantial number of DMPTs in neurons (n = 5897 DMPTs), astrocytes (n = 1096 DMPTs), endothelial cells (n = 255 DMPTs), or oligodendrocytes/OPCs (n = 85 DMPTs). Another 4 DMPTs associated with NIA-AA A score were found in neurons of the MFG.

The NIA-AA B score, a measure of neurofibrillary tangle (NFT) burden, is most commonly used as a measure of AD neuropathology in epigenome-wide association studies (EWAS) of AD. In our study, the CG showed the highest number of DMPTs in association with neurofibrillary tangles, found mainly in neurons (n = 556) but also in astrocytes (n = 11) and oligodendrocytes/OPCs (n = 211) of the CG (Fig. 3). Only a few DMPTs were found in neurons of the DG (n = 55) and no other brain regions showed significant DMPTs in association with neurofibrillary tangles. The DMPTs of neurons in the GG are visualized in Fig. 4a, with the top 20 hyper-and hypomethylated genes labeled accordingly. Most differentially methylated promoter regions were hypomethylated with increasing NIA-AA B score.

Fig.4

Manhattan mirror plot of differentially methylated promoter regions (DMPTs) in the dentate gyrus. Manhattan plots visualizing a the results from differential methylation analysis in promoter associated regions in neuronal signals of the dentate gyrus across NIA-AA A scores (Aβ plaque burden) and b the results from differential methylation analysis in promoter associated regions in neuronal signals of the dentate gyrus comparing individuals with different NIA-AA B scores (neurofibrillary tangle burden). Each dot represents the averaged methylation across all CpGs within a specific promoter region. The top part of each plot contains all promoters that are hypermethylated with a higher Aβ plaque burden or b higher burden of neurofibrillary tangles, and the bottom plot respectively shows all promoters that are hypomethylated. The x-axis displays chromosomes from 1 to 22 from the left to the right. The y-axis is displaying the-log10 FDR p-value as a significance measure for the methylation difference across neuropathological scores. The red dotted line marks the significance threshold of p < 0.05. The blue dots highlight the top 20 significant protein-coding promoters as ranked by log fold change of the methylation beta value. Labels display the name of the associated protein-coding gene for each of the top 20 differentially methylated promoters. DMPT differentially methylated promoter region, NIA-AA National institute of Aging Alzheimer's Association

For the NIA-AA C score, only a small number of DMPTs were found in microglia of the LC (n = 3, Additional File 1: Fig. S4), and no significant DMPTs were identified in the other brain regions (CBM, EC, SN, CA1).

We examined several copathologies and did not find any significant DMPTs associated with TDP-43 burden or Braak staging for Lewy bodies in any of the brain regions in the CTS profiles (Additional File 1: Fig. S5). We found very few significant (FDR p < 0.05) DMPTs associated with microvascular lesions in neurons of the locus coeruleus (n = 5, F2, CISD2, WFDC6, LYPLAL1, GPR33) and substantia nigra (n = 1, KLHL14 hypomethylation, logFC = 0.70) and in astrocytes of the locus coeruleus (n = 1, PPP1RC hypermethylation, logFC = 0.58) and entorhinal cortex (n = 1, SCML4 hypomethylation, logFC = 0.60). One DMPT was associated with microvascular lesions in oligodendrocytes/OPCs of the substantia nigra (MYL6B hypermethylation, logFC = 0.84).

Hypomethylation at promoter regions of known AD risk loci with increasing burden of AD neuropathology is mainly found in neurons of the dentate gyrus

We identified several significantly hypomethylated promoters among the top 33 AD risk loci previously identified in AD GWAS studies [53]. In neurons of the CG, the promoter region of the UNC5CL gene was hypomethylated with increasing burden of neurofibrillary tangles (logFC = 0.71, FDR p < 0.05). In astrocytes of the CG, promoter regions of the SPI1 and CR1 genes were hypomethylated with increasing burden of amyloid plaques (SPI1 logFC = 0.56, CR1 logFC = 0.55, FDR p < 0.05). In neurons of the DG, 12 out of the top 33 AD GWAS risk loci were hypomethylated with increasing amyloid plaque burden (SPI1 logFC = 0.53, WNT3 logFC = 0.64, CLNK logFC = 0.37, CLU logFC = 0.46, UNC5CL logFC = 0.31, BIN1 logFC = 0.36, SORL1 logFC = 0.37, IL34 logFC = 0.52, ACE logFC = 0.45, INPP5D logFC = 0.59, PLCG2 logFC = 0.49, CD2AP logFC = 0.27; FDR p < 0.05). CLU and ACE promoter hypomethylation was also seen in association with the overall AD Severity Score in neurons of the DG (CLU logFC = 0.42, ACE logFC = 0.38, FDR p < 0.05).

Promoter hypomethylation of PEN-2 with increasing burden of amyloid plaques is unique to neurons of the dentate gyrus

We identified the largest number of DMPTs in the DG, predominantly in neurons, but also in astrocytes, endothelial cells and oligodendrocytes/OPCs (Fig. 3). All 911 DMPTs found in neurons were specific to the DG. The DG has been underrepresented in methylation studies of the human brain, prompting us to explore further the DMPTs associated with NIA-AA A score in this region. The results are visualized in Fig. 4b, with the top 20 hyper- and hypomethylated promoter regions labeled accordingly. Similar to DMPTs associated with NIA-AA B score in the CG (Fig. 4a), DMPTs associated with increased Aβ plaque load in promoter regions of neurons of the DG were mostly hypomethylated (Fig. 4b). This also holds true for global methylation at overall non-averaged CpGs. We saw a negative correlation of global methylation levels with NIA-AA A score in neurons (spearman rho =  − 0.58, FDR p = 0.005), astrocytes (rho =  − 0.49, FDR p = 0.019) and microglia (rho =  − 0.55, FDR p = 0.007) in DG. These results are consistent with previous findings where overall neuronal DNA-methylation in the hippocampus was negatively correlated with AD burden [54]. The top 20 DMPTs (based on log fold change (logFC)) found in the DG included several genomic regions that are known to be altered in AD [55,56,57,58] like presenilin enhancer 2 (PEN-2, logFC =  − 0.80, FDR p = 0.001), solute carrier family 22 member 6 (SLC22A6, logFC = -0.94, FDR p < 0.001), lipopolysaccharide binding protein (LBP, logFC =  − 0.81, FDR p < 0.001), and S100 calcium binding protein A13 (S100A13, logFC =  − 0.81, FDR p < 0.001). A complete summary of DMPT statistics is available in Additional File 5. Across these four genes, promoter hypomethylation in neurons with increasing burden of Aβ plaques was unique to the dentate gyrus (Additional File 1: Fig. S6). Focusing on PEN-2, hypomethylation was also significantly associated with ADSS (FDR p = 0.016, Additional File 1: Fig. S7), reflecting the influence of the amyloid burden in this aggregated score. In the DG, the association of NIA-AA A score with PEN-2 hypomethylation was only found in neurons, and, crucially, could not be identified in the bulk methylation profile (Fig. 5a). As promoter methylation is often expected to reduce gene expression [59,60,61], we could expect the hypomethylation we observed to lead to higher expression of PEN-2 and perhaps contribute to proteolytic processing of amyloid precursor protein along the gamma(γ)-secretase pathway [62, 63]. Indeed, downregulation of the PEN-2 gene directly impairs γ-secretase activity, and overexpression has been found to increase activity of the γ-secretase [64, 65]. Other subunits of the γ-secretase complex are the presenilins (PSEN1 and PSEN2), aph-1 homolog A (APH-1A) and nicastrin (NCSTN). Numerous mutations in PSEN1 and PSEN2 have been identified in cases of familial early onset AD [66], but less is known about the regulation of the γ-secretase complex in late onset AD. Among the four subunits of the γ-secretase complex, in addition to PEN-2, hypomethylation of NCSTN was also associated with higher Aβ plaque burden in the DG neuron profile (Additional File 1: Fig. S8).

Fig. 5

Cell-type-specific promoter methylation of the PEN-2 gene in dentate gyrus (DG) across individuals with different NIA-AA A scores. a Scatterplots with smoothers showing the relationship between neuronal methylation of the promoter region of the PEN-2 gene (y-axis) across the five different cell types and bulk data from individuals with different Aβ plaque burden (NIA-AA A scores, x-axis). Methylation beta values are displayed on the y-axis and the categories of the NIA-AA A score on the x-axis. Each individual plot shows data from the dentate gyrus (DG) for different cell types. Each dot represents one individual sample. The standard linear regression was plotted as smoothers on top of the data: Smoothers curves are showing the relationship (solid line) between the NIA-AA A score and the methylation beta-value. Shaded areas indicate the 95% confidence interval of the smooth curve. We saw significant hypomethylation (**FDR p = 0.001, logFC = 0.80) in the promoter region of the PEN-2 gene with increasing Aβ plaque burden in the dentate gyrus. b Immunohistochemistry (IHC) of PEN-2 in the hippocampal region Cornu Amonis 3 (CA3) of an individual with high Alzheimer’s disease neuropathological changes (AD severity score = 3). IHC Scoring 3+ . The respective participant was Participant 44; for extended phenotype data of this individual see Additional File 4. Olig/OPCs Oligodendrocytes/Oligodendrocyte Precursor Cells. DG Dentate gyrus, Aβ Amyloid beta

Immunohistochemistry staining for PEN-2 shows regional expression in the dentate gyrus and substantia nigra

In our cohort, the hypomethylation of PEN-2 with an increasing burden of Aβ plaques was unique to neurons of the dentate gyrus. As there is little known about the regional expression of PEN-2 in the human AD brain, we conducted immunohistochemistry (IHC) staining on one case with high ADNC (AD severity score of 3) across all eight brain regions that were prior sampled for methylation analysis. The positivity for PEN-2 by IHC was strongest in the hippocampus (Fig. 5b) and substantia nigra with robust positivity and very specific staining, followed by the periaqueductal gray with intermediate positivity, and cingulate and middle frontal gyrus both showed the weakest positivity with diffuse staining in the gray matter. Cerebellum and pons were both negative (Figures not shown). IHC is a localizing and not a robustly quantitative technique because of pre-analytic and analytical variability [67]. Furthermore, the correlation between gene methylation and protein abundance is poor, reflecting the multiple layers of regulation between these two events. With these cautions in mind, we performed PEN-2 IHC on the hippocampus of two cases with high and two cases with low ADNC and observed variation across cases that was independent of ADNC status (Additional File 1: Fig. S9). Nevertheless, our small IHC experiment showed that in a brain with ADNC, PEN-2 was highly expressed in the hippocampus and substantia nigra. In contrast to findings in mouse models of AD, we did not see expression in cerebellum and pons [68].

Differential promoter methylation of clathrin-mediated endocytosis genes associates with increasing burden of Aβ plaques in neurons of the dentate gyrus

In addition to findings in genes of the γ-secretase complex, amongst the top 20 DMPTs in neurons of the dentate gyrus, the promoter region of clathrin light chain A (CLTA) was hypermethylated with increasing burden of Aβ plaques (Fig. 4b). CLTA encodes one of two clathrin light chain proteins, which form part of the regulatory function of the clathrin protein, an important protein for endocytotic processes in synaptic trafficking [69]. Clathrin-mediated endocytosis (CME) of amyloid precursor protein is of great relevance to AD pathology because it impacts the production of Aβ [70, 71]. Although not amongst the top 20 DMPTs, we found two more genes that are involved in CME to be significantly (FDR p < 0.05) hypomethylated with increasing amyloid burden: BIN1 and CD2AP. These genes were previously found to be associated with AD in both epigenome and genome wide association studies [2, 72,73,74,75]. In our study, the differential methylation of the promoter regions of CLTA, CD2AP and BIN1 with increasing amyloid burden was unique to neurons of the dentate gyrus and could not be found in any other brain region or cell type (Additional File 1: Figs. S10, S11). Notably, we did not see any association with AD neuropathological scores other than amyloid plaque burden.

Differential methylation in the dentate gyrus is related to neuropathology but not cognitive performance

In addition to neuropathologic changes, we assessed whether methylation differences were also associated with cognitive performance in old age. We tested differential methylation at promoters of protein-coding genes across individuals with one of three cognitive statuses: normal, cognitive impaired not dementia (CIND), or dementia. In contrast to the results from neuropathologic features, we did not detect any DMPTs in the dentate gyrus. The ERC was the only region showing cell-type-specific methylation differences across clinical groups, with a small number of DMPTs found in neurons (n = 6) distinguishing cognitively normal from demented individuals and in astrocytes (n = 6, Additional File 1: Fig. S12) distinguishing individuals with CIND and normal cognitive status, none of which overlapped. Further, treating the clinical status as a continuous rather than a categorical variable yielded no significant associations that might indicate a continuous methylation change related to cognitive decline.