Epigenome-wide study uncovers large-scale changes in histone acetylation driven by tau pathology in the aging and Alzheimer human brain

Abstract

Accumulation of tau and amyloid-β are two pathologic hallmarks of Alzheimer’s disease (AD). We conducted an epigenome-wide association study using the H3K9 acetylation (H3K9ac) mark in 669 aged human prefrontal cortices: in contrast to amyloid-β, tau protein burden had a broad effect on the epigenome, affecting 5,990 out of 26,384 H3K9ac domains. Tau-related alterations aggregated in large genomic segments reflecting spatial chromatin organization, and the magnitude of these effects correlated with the segment’s nuclear lamina association. Functional relevance of these chromatin changes was demonstrated by (1) consistent transcriptional changes in three independent datasets and (2) similar findings in two AD mouse models. Finally, we found that tau overexpression in iPSC-derived neurons altered chromatin structure and that these effects could be blocked by a small molecule predicted to reverse the tau effect. Thus, we report broad tau-driven chromatin rearrangements in the aging human brain that may be reversible with Hsp90 inhibitors.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Introduction

Alzheimer’s disease (AD) is a chronic neurodegenerative disease characterized pathologically by the accumulation of amyloid-β plaques and tau tangles which leads to neuronal cell death, cognitive impairment, and, ultimately, a diagnosis of dementia. Although loci harboring genetic risk factors have been identified in genome-wide studies¹, a large portion of late onset AD dementia risk remains unexplained, indicating the need for complementary approaches such as exploring the aging brain’s epigenome. Epigenomic alterations can be caused by genetic and non-genetic risk factors such as life experiences and environmental exposures but can also occur as a consequence of AD pathologies². Hence, studying the AD epigenome may also be helpful for understanding molecular events resulting from the toxicity of AD pathologies. Evidence for epigenomic perturbations in AD was found in smaller human studies³. For example, a recent study characterized the redistribution of histone 4 lysine 16 acetylation (H4K16ac) in the temporal cortex of 12 AD subjects compared to controls⁴, but, so far, large AD-related epigenome-wide association studies of the human cortex have been limited to DNA methylation^5,6. These studies demonstrated reproducible DNA methylation changes in AD subjects, but did not distinguish explicitly between amyloid-β- and tau-related alterations. Notably, tau pathology has recently been associated with epigenetic changes in model systems. In Drosophila, tau overexpression has been shown to relax heterochromatin^7,8. On the other hand, there is also evidence for a physiological function of nuclear tau in maintaining and regulating heterochromatin which may be lost by pathologically phosphorylated tau⁹. However, whether these mechanisms cause major chromatin alterations in the human AD brain, translate to transcriptional alterations, and are restricted to heterochromatic regions remains unknown.

Here, we studied the acetylation of the ninth lysine of histone 3 (H3K9ac), which marks transcriptionally active open chromatin, genome-wide in the dorsolateral prefrontal cortex (DLPFC) of 669 subjects. Our data support the hypothesis that tau but not amyloid-β causes widespread chromatin remodeling. We mapped the location of these alterations genome-wide, characterized the architecture of these large areas of coordinated chromatin remodeling, replicated them in transcriptomic data, demonstrated that tau is sufficient to cause such chromatin rearrangements prior to tangle formation, and identified a compound that may attenuate this chromatin perturbation.

Results

Tau but not amyloid-β pathology has a broad effect on histone acetylation in the human brain

We studied H3K9ac in the DLPFC of 669 participants enrolled in either the Religious Order Study (ROS) or the Rush Memory and Aging Project (MAP), two longitudinal studies of aging and dementia^10,11. Participants were not demented upon study entry (Fig. 1a). At autopsy, neuropathologic examination was performed and quantitative measurements of the density of phosphorylated tau tangles and the burden of amyloid-β were obtained (Table 1). A wide spectrum of tau tangles and amyloid-β burdens with a moderate correlation of ρ=0.48 between these two AD pathologies was observed in our subjects (Fig. 1b). ChIP-seq was performed in isolated gray matter from frozen DLPFC samples to generate genome-wide H3K9ac profiles. A median of 55 million 36 bp single-end reads were sequenced per sample (Supplementary Table 1).

Figure 1: The active histone mark H3K9ac is broadly associated with tau pathology in the human cortex.

a, H3K9ac was studied in the prefrontal cortex of 669 subjects. b, Pearson correlation between tau and amyloid-β levels was 0.48 in this cohort. Amyloid-β peptide levels were below the detection threshold for 103 samples. c, Horizontal bars depict the relative frequency of each chromatin state within the H3K9ac domains (97 Mbp) and the whole genome (3.1 Gbp). Chromatin states were obtained from a DLPFC sample with minimal neuropathology (E073) included in the Roadmap Epigenomics Project. For better clarity, the respective upper bars depict the relative frequencies without the quiescent state. d, Median H3K9ac ChIP-seq read coverage is shown on the y-axis for highly transcribed, lowly transcribed and not detectable RefSeq transcripts within ±3 kb around the transcription start site. Transcripts were grouped by median transcription values (high: x > 22 rpkm (0.95 quantile), low: 1 rpkm < x < 1.5 rpkm (0.6 – 0.65 quantiles), not detectable: x = 0). The median of the H3K9ac read density across all samples was calculated for each transcript. Then, the median of all transcripts within each of the three groups was calculated and plotted as solid line. The 25% and 75% quantiles for each group are depicted by the transparent bands. e, Bars depict the number of H3K9ac domains that were significantly associated at an FDR of 0.05 with age, gender, tau levels and amyloid-β levels. f, Bars depict the number of H3K9ac domains stratified by promoter, enhancer and other domains. Dark shading indicates the number of domains whose H3K9ac levels were significantly associated with tau or amyloid-β.

Table 1: Summary characteristics of the ROS/MAP data set.

Mean values and standard deviation (in brackets) are shown for quantitative variables. Number of cases and percentages (in brackets) are shown for categorical variables. H3K9ac and transcription data were available for 452 samples, H3K9ac and DNA methylation data for 645 samples, and transcription and DNA methylation data for 496. All three data types were available for 449 samples.

	H3K9ac (n=669; ChIP-seq)	Transcription (n=500; RNA-seq)	DNA methylation (n=729; BeadChip)
Age (years)	88.3 (±6.5)	88.5 (±6.6)	88.0 (±6.6)
Male gender	233 (35%)	189 (38%)	267 (37%)
Tau protein load (%−1/2)	2.2 (±1.4)	2.1 (±1.3)	2.1 (±1.4)
Amyloid-β load (%−1/2)	1.5 (±1.1)	1.5 (±1.1)	1.5 (±1.1)
Pathologic AD diagnosis	412 (62%)	290 (58%)	441 (60%)
Post mortem interval (hours)	7.5 (±5.9)	7.1 (±4.9)	7.4 (±5.8)

We found 26,384 H3K9ac peaks or “domains” which showed a distinct signal above the genome-wide background in at least 100 of the 669 subjects (Supplementary Table 2). As with DNA methylation⁵, there is limited inter-individual variability: the average correlation of domains between two individuals was 0.98, despite the vast differences in life experiences between these older individuals. Nearly half (41%) of the genomic regions covered by H3K9ac domains in our data were annotated as containing an active transcription start site, and more than a quarter (27%) were annotated as being in an enhancer site (Fig. 1c) based on the chromatin state annotation generated for this cortical region by the Roadmap Epigenomics Project (sample E073; a MAP subject with minimal neuropathology)¹². We thus binned H3K9ac peaks into “promoter” (n=15,225), “enhancer” (n=8,071) and “other” (n=3,088) domains (Supplementary Fig. 1a–d) for subsequent analyses. Using RNA-seq data from the same region in a subset of subjects (n=500), we verified the expected positive correlation between H3K9ac and transcriptional activity (Fig. 1d).

To distinguish between tau- and amyloid-β-related epigenomic changes, we modeled both pathologies simultaneously as explanatory variables in a regression model with the H3K9ac levels as the outcome. At a false discovery rate of 0.05, 23% of H3K9ac domains showed an association with tau whereas only 2% were associated with amyloid-β (Fig. 1e, Supplementary Table 3). Only a few domains (n=88) had acetylation levels that were significantly associated with both pathologies. This is a striking and unexpected difference in the impact of tau and amyloid-β, and it may be related to the fact that tau pathology initially accumulates intracellularly and may directly affect neuronal chromatin organization, whereas amyloid-β is secreted and affects cells extracellularly. For tau, the largest proportion of associated domains was observed among promoter regions (Fig. 1f), but the estimated effect sizes were similar for domains found in promoters, enhancers and other regions (Supplementary Fig. 1e–f), indicating that the mechanism driving these associations is not specific for a type of domain. The greater average read depth at promoters probably explains the larger number of significant domains in these regions. Interestingly, known AD risk loci¹ were not enriched in tau-associated H3K9ac domains: out of 39 H3K9ac domains within 50 kb of 19 known AD loci, 9 were associated with tau, corresponding to the genome-wide relative frequency of 23% (Supplementary Table 4).

Next, we investigated whether tau-related changes in H3K9ac led to functional consequences by analyzing our RNA-seq data: we repeated the simultaneous evaluation of amyloid and tau pathology for 18,257 out of 24,594 active transcripts that could be mapped to an H3K9ac domain. We observed a positive but weak correlation of 0.14 between the coefficients for tau (0.12 for amyloid) from the transcription and the H3K9ac data: in general, epigenomic and transcriptomic data showed the same direction of effect (Supplementary Fig. 1g–h) (Supplementary Table 5), with notable exceptions that indicate the presence of other regulatory mechanisms. The modest extent of the correlation is probably attributable to the fact that H3K9ac is not sufficient to define a transcriptionally active region (Supplementary Fig. 1i).

Spatial pattern of tau-related alterations in H3K9ac reflects higher-order chromatin organization

Because of the large number of tau-associated H3K9ac domains, we evaluated the spatial distribution of tau-associated domains using a “Chicago Plot” that presents the physical location, significance, and directionality of each tested domain’s association with tau (Fig. 2a, Supplementary Fig. 2a). Throughout the genome, we find large-scale genomic segments whose H3K9ac domains were coordinately enriched for associations with tau. Unlike genome-wide SNP studies where we see very small genomic regions being associated with disease because linkage disequilibrium is observed only over ~18 kilobase pairs (kbp) in most regions of the human genome¹³, we see a clustering of disease-related epigenomic associations in certain segments of the genome that cover several megabase pairs (Mbp). These results suggest that a physical aspect of the chromosome may be implicated in disease. To define the boundaries of these large-scale segments we applied a segmentation algorithm¹⁴ to our data, and this method divided the genome into 178 segments within which domains display a similar tau association (Fig. 2a). The median size of these segments was 5.2 Mbp, and they contain a median of 79.5 H3K9ac domains per segment (Supplementary Table 6). In contrast to tau, no similar spatial pattern was observed for domains associated with amyloid-β pathology (Supplementary Fig. 2b), suggesting a fundamental difference in the relation of these pathologies to the human cortex’ epigenome.

Figure 2: Tau-related chromatin alterations show spatial patterns.

a, Manhattan plot depicts log-transformed unadjusted p-values for the tau coefficients of all H3K9ac domains along chromosome 1 (two-sided Wald tests; n=669 subjects). Transformed p-values were plotted with the sign of the respective coefficient to distinguish between positive and negative associations. Red dashed lines indicate FDR threshold of 0.05 when adjusting for all 26,384 domains. Broad genomic segments covering H3K9ac domains showing associations with tau similar in direction and strength are depicted as orange line segments. The gray polygon represents the chromsome’s centromere. Two ribbons at the bottom show chromatin structure related annotation. The first ribbon indicates A and B compartments. The second ribbon indicates lamina-associated domains. b, Boxplot shows larger mean tau-related changes in H3K9ac levels in A compartments (n=284) than in B compartments (n=223) (Mann-Whitney-Wilcoxon test, p=8×10⁻¹³). The center line of the boxplot shows the median, the outer hinges correspond to the lower and upper quartiles, and the whiskers extend to the most extreme observed value within 1.5×IQR from the hinges. Compartments with no or a very few H3K9ac domains (< 5) were discarded. c-d, Scatter plots show the association between segments’ mean tau-related changes in H3K9ac levels and nuclear lamina association (c), and mean tau-related changes in mRNA levels (d). Each dot represents a segment (n=138, sex chromosomes were excluded). Weighted linear regression line (blue line), 95% confidence interval of the mean change (shaded region), and weighted Pearson correlation are depicted in the plots. Segments were weighted by the number of H3K9ac domains. e, Columns of the heatmap represent segments sorted by lamina association first and then by chromatin accessibility to break ties since multiple segments were completely free of LADs. Chromatin accessibility was calculated using DLPFC specific annotation from the Roadmap Epigenomics Project (see Methods). Column widths reflect segment sizes. In rows one to six, color indicates the mean tau effect size observed in H3K9ac (n=669 subjects), RNA transcription (n=500 subjects), DNA methylation (n=729 subjects) at active TSSs, weakly transcribed regions, enhancers and quiescent regions. In the last two rows, color indicates the mean difference between AD and control samples observed in published transcription data from temporal cortex tissue (row seven, n=292 subjects) and dissected neurons (row eight, n=34 subjects). Weighted Pearson correlations between the mean tau effect in the segments and histone accessibility (left column) and nuclear lamina association (right column) are shown on the right side. Segments were weighted by the number of H3K9ac domains in the segments.

At the megabase pair scale, chromatin is spatially organized into type A and type B compartments¹⁵. Type A compartments are primarily characterized by open and type B compartments by closed chromatin. To study whether the spatial chromatin architecture underlies the observed spatial pattern of H3K9ac alterations, we derived a compartment map from previously published Hi-C data of the cortical and subcortical plate of a human fetal brain¹⁶. The genome was divided into 636 compartments and H3K9ac domains were mapped to compartments. Figure 2a shows the locations of H3K9ac domains and A/B compartments for chromosome 1. Genome-wide, the majority of H3K9ac domains (70%) were located in type A compartments, but we also observed 223 type B compartments out of 507 compartments that contained at least five H3K9ac domains. For these 507 compartments, we calculated the mean tau effect on H3K9ac and found a significantly larger tau effect in type A compartments (Fig. 2b) indicating that the tau effect on H3K9ac is associated with the spatial chromatin organization.

A critical element of chromatin structure is the nuclear lamina because it provides anchor points that couple chromatin to the lamina. Nuclear lamina is associated with repressive chromatin and type B compartments. Recent work in Drosophila suggests that tau can disrupt lamina function and thereby induce heterochromatin relaxation⁷. To elucidate the role of the lamina in our observations that center on euchromatin, we calculated the lamina association of our tau-defined segments as the proportion of lamina-associated domains (LADs) within a segment using previously published genome-wide DamID lamin B1 data from the human HT1080 fibrosacoma cell line¹⁷. Even though some LADs are cell type specific, the majority of LADs are strongly conserved, so that LADs derived from human fibroblasts are still useful for annotating whole brain tissue^17,18. We plotted the segments’ lamina association versus the average tau effect of all H3K9ac domains within a segment (Fig. 2c). The negative Pearson correlation (ρ=−0.86) indicates that the effect of tau pathology on single H3K9ac domains varies depending on the lamina association of the surrounding genomic region. In fact, the 178 genomic segments within which tau associations are correlated explain 29% of the variance of the tau effects observed across all 26,384 H3K9ac domains. Thus, a large portion of the tau effect can be explained by chromatin organization.

We next evaluated our RNA-seq data and found that each segments’ mean tau effects obtained from the RNA-seq data was correlated (ρ=0.83) with the mean tau effects from the H3K9ac data (Fig. 2d), reflecting a shift in average gene transcription that is related to the burden of tau pathology and is concordant with the segments’ chromatin alterations. However, in line with the weak correlation of alterations in H3K9ac and RNA levels at the single transcript level, only 2% of the variance of tau-related effects in RNA-seq data could be explained by the segments defined with the chromatin data. This suggests that, while large-scale epigenomic changes explain some of the effect of tau on the transcriptome, other regulatory mechanisms must also be influenced by this neuropathologic process.

We then speculated that other epigenetic marks may show similar tau-related patterns: in evaluating our DNA methylation data from the same brain region⁵, we found consistent evidence for a perturbation of the epigenomic architecture. Specifically, we binned CpG’s into four groups using the tissue specific reference chromatin state map from the Roadmap Epigenomics Project¹² (sample E073): active transcription start sites, enhancers, weakly transcribed regions and quiescent regions. These four chromatin states cover most of the CpGs in our data set, and we repeated the segment-based analysis for each group of CpG’s. As shown in Fig. 2e, a negative correlation between tau effects on DNA methylation and nuclear lamina association was observed in enhancers, quiescent regions and weakly transcribed regions, which is in agreement with the paradigm that high methylation of gene bodies is a feature of active genes¹⁹. The variance of tau effects in DNA methylation data that could be explained by the H3K9ac derived segments ranged between 6% for enhancer, 4% for transcribed regions, and 1% for quiescent regions (Supplementary Table 7). By contrast, no distinct correlation was observed at active transcription start sites, in line with previous work reporting that promoter regions are spared from altered methylation in other diseases²⁰.

To validate our findings in an independent set of samples, we assessed a published transcriptomic data set from the temporal cortex of subjects from a late onset AD (LOAD) case-control study²¹. Because quantitative amyloid-β and tau burdens were not available, we compared samples with AD diagnosis (n=202) to controls (n=90; non-AD and non-progressive supranuclear palsy diagnosis) and calculated the mean AD-related change in transcription over the genes within each of the 178 genomic segments defined using our H3K9ac data. Although we could not distinguish the tau from the amyloid-β effects in these data, we did confirm the presence of a negative correlation (ρ=−0.71) between the segments’ nuclear lamina association and AD-related transcriptional changes (Fig. 2e). Similar to our RNA-seq data, 4% of the variance of AD effects in the transcriptomic data of the LOAD study could be explained by the segments. Nevertheless, as in our RNA-seq data, the mean tau-related changes in the H3K9ac data were positively correlated (ρ=0.65) with the mean AD-related changes in the transcription data from this LOAD study, validating our observation.

Since the data analyzed so far were generated from bulk tissue, it is critical to assess whether the observed alterations may be driven by changing cell type compositions. We estimated the proportion of neurons, astrocytes, oligodendrocytes, myeloid cells and endothelial cells from our RNA-seq data using a set of five marker genes per cell type (Supplementary Table 8)²². The proportion of neurons was modestly decreased (p=0.043, two-sided t-test, n=452 subjects) in subjects with AD, indicating that cell type composition might be a confounder. However, the coefficients for tau derived from a model adjusted for cell type proportions fitted on a subset of n=452 subjects that had both RNA-seq and H3K9ac data were highly correlated (ρ>0.99) with the tau coefficients from the unadjusted model used in our primary analysis (Supplementary Fig. 3). Similarly, p-values did not change noticeably after adjustment for cell type proportions. While we cannot exclude confounding by changing cell type compositions completely²³, these results indicate that varying cell type compositions are probably not the main driver of our observations and we decided to further investigate which cell type(s) may be affected by tau pathology.

Altered transcription due to tau-related chromatin remodeling occurs in neurons

A priori, neurons are the most likely candidate cell type since intracellular phosphorylated tau is known to accumulate and form tangles in neurons. To explore this hypothesis, we repurposed a gene transcription data set generated from laser-capture microdissected neurons from the superior frontal gyrus of individuals with AD (n=23) and control subjects (n=11)²⁴. Since tau and amyloid-β loads were not reported in these subjects, we estimated the AD effect instead of the tau effect for each gene. The AD effects were then averaged within each of the 178 genomic segments defined in our H3K9ac data, and the correlation with the segment’s nuclear lamina association was calculated. Again, we observed a negative correlation of ρ=−0.64 (Fig. 2e), and the correlation of these neuronal data with the average tau effect in our H3K9ac data was ρ=0.67. Similar to the tissue transcriptomic data, 2% of the AD effect variance were explained by the segments. Interestingly, the original report of these neuronal data specifies that neurons lacking neurofibrillary tangles were selected for laser capture; this indicates that the observed AD-related transcriptional changes are occurring early in pathogenesis, prior to the accumulation of tangles²⁴. Since the transcriptional changes consistent with our epigenomic changes are found in purified neurons, they are likely to be cell-autonomous for neurons; however, we cannot rule out that other cell types may also be affected.

H3K9ac data from tau mouse models suggest similar structural changes in the murine brain epigenome that involve the nuclear lamina

To explore whether the epigenomic changes that we found in human brain are recapitulated in mouse models known to accumulate tau, we generated hippocampal H3K9ac ChIP-seq profiles from two different mouse models at (1) an early time point and (2) a late stage of neurodegeneration (Supplementary Table 9). Specifically, we studied 6 and 11 months old mutant tau mice (MAPT P301S), which start to accumulate phosphorylated tau in neurons by 6 months²⁵. Wild-type mice of the same age were used as controls. The second mouse model was the CK-p25 model, which is characterized by increased amyloid-β levels early after p25 induction followed by increased tau phosphorylation and neuronal loss at later stages^26,27. Three months old CK-p25 mice were studied 2 weeks and 6 weeks after p25 induction and were compared to CK littermate controls. In total, we found 44,165 H3K9ac domains across the epigenome of both mouse models. Tau mice showed less significantly altered H3K9ac domains compared to the CK-p25 mice, especially at the early stage of 6 months (Fig. 3a). Chromosome-wide plots did not show the explicit spatial pattern observed in human cortex, possibly because of the low sample size of 3 mice per time point. We then classified the mouse H3K9ac domains into two bins: “close to lamina” (if domain <50 kb from a LAD) or “distant to lamina” based on published DamID lamin B1 data from mouse embryonic fibroblasts¹⁷. As anticipated, we observed smaller differences between AD mice and control mice in H3K9ac domains that were located in the proximity of LADs (Fig. 3b–c). This observation was consistent for both mouse models and more distinct at the later time points during which the pathology has accumulated more extensively, confirming our observations from the human cortex where the effect of tau was smaller in segments enriched with LADs.

Figure 3: H3K9ac alterations in AD mouse models reflect spatial pattern observed in the human cortex.

a, Bar plot shows the number of domains with significant different H3K9ac levels observed in the tau mouse model (blue bars) and the CK-p25 mouse model (green bars) compared to respective control mice for different time points. CK-p25 mice were 3 months old when p25 was induced. b,c, Boxplots depict differences in H3K9ac levels between tau mice (b) or CK-p25 mice (c) and respective control mice separately for H3K9ac domains distant to and close to LADs. The center line shows the median, the outer hinges correspond to the lower and upper quartiles, and the whiskers extend to the most extreme observed value within 1.5×IQR from the hinges. The median observed log fold changes at lamina-free and lamina-associated H3K9ac domains differed by 0.20 (p ≤ 10⁻¹⁶, Mann-Whitney-Wilcoxon test) at 11 months (0.01 at 6 months) for the tau model, and by 0.13 (p ≤ 10⁻¹⁶, Mann-Whitney-Wilcoxon test) at 4.5 months (0.04 at 3.5 months) for the CK-p25 model.

Induced neurons overexpressing MAPT demonstrate chromatin structure alterations

Given the evidence from neuronal expression data that the observed changes in chromatin modifications are likely to be occurring in neurons, we assessed whether overexpression of the 4R isoform of MAPT, which encodes the tau protein, in forebrain neurons derived from human induced pluripotent stem cells (iNs) could recapitulate the epigenomic changes that we found in our cortical H3K9ac data. In iNs bearing a familial AD mutation, this model system induces AD-related features such as the intracellular accumulation of phosphorylated tau²⁸. To characterize the epigenomic changes in this in vitro model system, we used the Assay for Transposase-Accessible Chromatin with sequencing (ATAC-seq) to map genome accessibility²⁹. In contrast to anti-H3K9ac ChIP-seq, ATAC-seq generates reads at all open regulatory elements where the Tn5 transposase can access exposed DNA. We therefore expect to observe ATAC-seq peaks in the vicinity of nucleosomes carrying the H3K9ac mark and additional peaks at enhancers and insulators that might not be marked by H3K9ac.

All iNs were derived from the same iPSC line, and we observed 40,637 ATAC domains that have a distinct peak in at least 4 of our 18 samples. Each sample consisted of 2–3 pooled individual wells of transfected iNs; 9 samples represented independent transductions of MAPT-overexpressing iNs. The remaining 9 samples were control iNs transduced with a GFP overexpressing construct. Samples were generated in 3 experimental batches of independent differentiations. ATAC domains were annotated with chromatin states of hESC-derived neurons (sample E010) from the Roadmap Epigenomics Project¹² (Fig. 4a) and clustered into four groups: transcription start site (TSS), TSS flanking region, enhancer and other domains (Supplementary Fig. 4a–c). As expected, most of the TSS domains (91%) overlapped one of the H3K9ac domains from our whole brain tissues, but only 17% of the enhancer domains were included in our set of cortical H3K9ac enhancer domains (Fig. 4b) indicating that H3K9ac is not an explicit enhancer marker. Assessing these domains in the other direction, the majority of cortical H3K9ac domains were overlapped by at least one ATAC domain (Fig. 4c). Thus, while there are important differences between the two sets of profiles, there is good overlap, particularly in TSS where the bulk of the tau effect is observed in our H3K9ac tissue-level data.

Figure 4: Tau induces chromatin alterations in iPSC-derived neurons.

a, Horizontal bars depict the relative frequency of each chromatin state within the ATAC domains (31 Mbp) and the whole genome (3.1 Gbp). Chromatin states were obtained from H9 derived cultured neurons (E010) included in the Roadmap Epigenomics Project. For better clarity, the respective upper bars depict the relative frequencies without the quiescent state. b, Bars depict proportion of ATAC domains in neurons that were overlapped by an H3K9ac domain in the DLPFC. c, Bars depict proportion of H3K9ac domains in the DLPFC that were overlapped by an ATAC domain in neurons. d, Bars depict the number of ATAC domains stratified by TSS, TSS flanking, enhancer and other domains. Dark shading indicates the number of domains whose chromatin accessibility differed significantly between MAPT overexpressing neurons and controls at an FDR of 0.05. e, Manhattan plot depicts the log-transformed unadjusted p-values for differences between MAPT overexpressing neurons and controls of all ATAC domains along chromosome 1 (two-sided t-tests, n=18 experiments). Transformed p-values were plotted with the sign of the test statistic to distinguish between increased and decreased chromatin accessibility. Red dashed lines indicate FDR threshold of 0.05 when adjusting for all 40,637 domains. Broad genomic segments covering ATAC domains showing similar changes in chromatin accessibility are depicted as purple line segments. The gray polygon represents the chromsome’s centromere. Two ribbons at the bottom show chromatin structure related annotation. The first ribbon indicates A and B compartments. The second ribbon indicates lamina-associated domains. f, Boxplot shows different mean tau-related changes in chromatin accessibility in A compartments (n=306) and B compartments (n=283) (p=4×10⁻⁷, Mann-Whitney-Wilcoxon test). The center line of the boxplot shows the median, the outer hinges correspond to the lower and upper quartiles, and the whiskers extend to the most extreme observed value within 1.5×IQR from the hinges. Compartments with no or a very few ATAC domains (< 5) were discarded. g, Scatter plot shows association between segments’ nuclear lamina association and mean change in chromatin accessibility. Each dot represents a segment (n=89, sex chromosomes were excluded). Weighted linear regression line (blue line), 95% confidence interval (shaded region), and weighted Pearson correlation are depicted in the plot. Segments were weighted by the number of ATAC domains. h, Each dot represents either an A (n=284) or B (n=221) compartment derived from fetal human brain Hi-C data. For each compartment, the mean tau-related change in H3K9ac levels in the human DLPFC on the x-axis is plotted versus the mean change in chromatin accessibility between MAPT overexpressing neurons and controls on the y-axis (Pearson correlation ρ=−0.51). Compartments with less than 5 H3K9ac domains or less than 5 ATAC domains were discarded. i, Columns of the heat map represent the 138 genomic segments (sex chromosomes were excluded) derived from the DLPFC H3K9ac data and were first sorted by lamina association and then by chromatin accessibility. The color in the first row indicates the mean tau effect size observed in the DLPFC H3K9ac data (n=669 subjects) as shown in Fig. 2e. The second and third rows depict the differences in H3K9ac (n=3 experiments) and chromatin accessibility (n=18 experiments) observed between MAPT overexpressing neurons and control neurons. Weighted Pearson correlations between the mean tau effect in the segments and histone accessibility (left column) and nuclear lamina association (right column) is shown on the right side. Segments were weighted by the number of DLPFC H3K9ac domains in the segments. j, Bars depict the variance of differences between MAPT overexpressing and respective control iNs that can be explained by segments (see Supplementary Table 13). Blue bars indicate that the MAPT overexpressing iNs were treated with 17-DMAG.

In total, 747 out of 40,637 ATAC domains showed a difference in their chromatin accessibility at an FDR < 0.05 when comparing the MAPT overexpressing iNs to control iNs infected with a non-relevant vector (Fig. 4d). The plots in Fig. 4e and Supplementary Fig. 2c depict large genomic segments of predominantly concordant tau-related changes; segments were defined using the same segmentation algorithm that we used for the cortical H3K9ac data and are highlighted in purple. In iNs, we found 99 segments that explained 11% of the variance of tau effects observed for single ATAC domains (Supplementary Table 10). As seen in the H3K9ac data, the tau effect differed significantly between type A and type B compartments (Fig. 4f), and the average tau effect of the segments correlated (ρ=0.92) with the proportion of LADs in the segment (Fig. 4g). These results were found consistently in each of the three batches (Supplementary Fig. 4d–j). Interestingly, in contrast to the H3K9ac changes observed in human and in mouse brain tissue, segments containing many LADs showed a large positive average tau effect indicating that chromatin relaxation preferentially occurred in lamina-associated segments in this in vitro system. Concordantly, when looking at the human fetal brain A/B compartments, the tau effect observed in human brain H3K9ac data was negatively correlated with the tau effect in iNs (Fig. 4h). To verify that the model system presents a tau effect that is the inverse of what we see in the human cortex, we repeated the iN experiment with anti-H3K9ac ChIP-seq and estimated the effect size of tau for each of the 34,988 detected H3K9ac domains. Then, using the 178 genomic segments derived from the cortical H3K9ac data and calculated the segments’ average tau effect in (i) iN H3K9ac data and (ii) iN ATAC data to compare it to the brain tissue H3K9ac data (Fig. 4i). The H3K9ac and ATAC data from iNs consistently indicate that tau is associated with chromatin opening in genomic segments with a high fraction of lamina association. Thus, segments of the genome are significantly affected by tau pathology in vivo and in vitro, but the direction of effect is reversed in relation to tau pathology, which accumulates over decades in the human brain and over 14 days in our MAPT overexpressing iN model system.

The Hsp90 inhibitor 17-DMAG is a candidate drug for attenuating tau-related alterations of chromatin structure

To find putative drugs with the potential to protect neurons from tau toxicity, we scanned the Connectivity Map data base³⁰ for compounds whose gene expression signature was negatively correlated with our H3K9ac signature for tau pathology derived from ROS/MAP cortical profiles. H3K9ac domains were mapped to genes based on transcription start sites inside the domains. Domains without a transcription start site were discarded. Out of 1309 compounds in the database, the N-terminal Heat Shock Protein 90 (Hsp90) inhibitor 17-(dimethylaminoethylamino)-17-demethoxygeldanamycin (17-DMAG, alvespimycin) achieved the smallest p-value according to the database’s search algorithm and the fifth smallest correlation coefficient (Supplementary Table 11). Hsp90 and other heat shock proteins are involved in protein folding, and Hsp90 has been previously evaluated in the context of Alzheimer’s disease^31,32.

To assess the effect of 17-DMAG, we treated three independent cultures of MAPT overexpressing iNs with 17-DMAG at a concentration of 3 nM for 24 hr before harvesting (on day 31), based on published studies³³. The differences between the treated MAPT overexpressing iNs and the respective controls were calculated for each of the 40,637 ATAC domains. The variance of these tau-induced differences was then decomposed into between-segment and within-segment variance using the 99 tau-associated chromosomal segments derived for the iNs using ATAC-seq data. The prediction is that if 17-DMAG attenuates the tau effect on chromatin organization, the variance explained by the segments (i.e. the between-segment variance) should decrease for the 17-DMAG treated iNs. In our experiment, we observed such a protective effect of 17-DMAG (Fig. 4j). We then confirmed these results in a second experiment and a dose-response curve: triplicates of MAPT-overexpressing iN cell cultures with 17-DMAG at concentrations of 1 nM, 3 nM or 10 nM. As controls, we cultured MAPT overexpressing iNs and control iNs in DMSO solution. While our experiments indicate that 17-DMAG may protect neurons from chromatin alterations related to tau (Fig. 4j, Supplementary Tables 12–13), further studies with different lines of iNs are necessary to replicate the effect of 17-DMAG and to explore the interplay between 17-DMAG and tau pathology.

Discussion

Our epigenome-wide association study revealed that tau pathology is associated with broad changes in the brain’s epigenome. We observed large genomic segments of several megabases within which H3K9ac domains showed similar tau-associated gains or losses of histone acetylation. At the megabase pair scale, the genome is organized into two major types of compartments, A and B, defined by patterns observed in Hi-C interaction maps¹⁵. Type A compartments are characterized by euchromatin which is transcriptionally active whereas the inactive type B compartments exhibit high chromatin density. These two major types of compartments are essential for chromatin organization³⁴ and cell identity³⁵, and they impose a cis- and trans-correlation structure on epigenetic marks³⁶. We demonstrated that the effect of tau on H3K9ac differed between type A and type B compartments and that the segments observed in the Figure 2a reflect higher-order chromatin structure. Thus, tau appears to be preferentially affecting certain elements of nuclear architecture and chromosomal organization.

The nuclear lamina is a key element in the spatial organization of chromatin and is associated with inactive B compartments. Previous work in Drosophila suggested that the lamin nucleoskeleton is disrupted in tauopathies, which causes heterochromatin relaxation and mediates neuronal cell death^7,8. In line with this hypothesis, we observed a strong correlation between the effect of tau in a genomic segment and the nuclear lamina association of that segment. However, we measured the euchromatic mark H3K9ac, and it remains unclear how this mark relates to heterochromatin relaxation. This intriguing association between the effect of tau on H3K9ac and key elements of spatial chromatin organization needs to be explored further using techniques such as those that map chromatin contact points.

The association of tau pathology with alterations in chromatin structure is not specific to the H3K9ac chromatin mark and has functional consequences: we replicated the presence of the segments with coordinated alterations in relation to tau pathology in both DNA methylation and transcriptomic data from the same persons. Further, we find the same pattern of results in publically available transcriptomic data from a different cohort, and a second dataset from laser-captured neurons further refines the observation by implicating a specific cell population. Even though the pattern was less distinct in the various transcriptomic data sets, demonstrating the presence of these transcriptional perturbations in relation to tau is important in understanding the consequences of these alterations. Not all epigenetic changes directly translate to altered transcription. In addition, if chromatin alterations induce the transcription of epigenetically silenced genes, transcription may occur at a low level in a tissue sample, and we may thus be under-estimating the extent of transcriptional alterations in tissue-level RNA-seq data.

We employed AD mouse models and a MAPT overexpressing iNs in vitro model to further explore our findings and to address some of the limitations of our human cortex study. Looking at the results from our iN experiments, we demonstrated that MAPT overexpression is sufficient to induce chromatin reorganization in neurons. Further, this event occurs prior to tangle formation and neuronal cell death in our model system as well as in the repurposed RNA dataset from laser-captured neurons²⁴: perturbation of epigenomic architecture may therefore be an early event in tau pathology. This is an important point that we cannot address with our cross-sectional human H3K9ac brain data. In the longitudinal CK-p25 mouse model data, we detected altered H3K9ac levels depending on the proximity of the H3K9ac domain to the nuclear lamina two weeks after p25 induction. For the same mouse model at the same time point, a study reported altered localization of lamin in hippocampal neurons indicating a dispersion of the nuclear lamina membrane that could affect chromatin organization³⁷. This effect was stronger over time and preceded apoptosis³⁷. Overall, we hypothesize that tau-induced epigenomic changes are early events that occur downstream of pathologic tau accumulation but before neurofibrillary tangles develop.

Interestingly, we observed large positive tau effects on H3K9ac in lamina free regions of the human cortex and mouse brain, but the direction of the association was reversed for iNs. This is an intriguing result and suggests a complex interaction between tau and chromatin structure. iNs grow in vitro in the absence of the three dimensional context of the cerebral cortex and do not interact with non-neuronal cells. The inverse directionality that we see at an early time point after MAPT overexpression could be an initial response of iNs to a rapid accumulation of tau (compared to accumulation over decades in the human brain), in line with recent work suggesting that nuclear tau also has a physiological function in assembling and maintaining heterochromatin⁹. Importantly, the segmental pattern and the association of the tau effect with the proportion of LADs in the segments is seen in both the human cortex and iNs, suggesting the involvement of the same mechanism. More work is required to delineate the molecular events happening in vivo and in vitro, and to study whether the mechanism is specific for tau or can occur with other stimuli. Further, we used publicly available data from different cell types to annotate our results, and therefore, the extent to which spatial organization changes occur with tau pathology needs to be interrogated directly. iNs overexpressing tau may be a useful model system with which to address these questions. This is illustrated by our testing of an Hsp90 inhibitor that is predicted to block the effects of tau based on our brain data: 17-DMAG attenuates the chromatin alterations induced by tau in iNs.

Overall, we have identified large-scale changes throughout the epigenome of the human AD brain and shown that tau-induced alterations of chromatin structure are much more profound than the changes that are attributable to amyloid pathology. The effect of these alterations propagate into the transcriptome. Our data supports the hypothesis that the nuclear lamina is a key element in mediating tau toxicity in vivo and in vitro, and we demonstrated that iNs are a useful model system to further study tau-associated chromatin changes and to screen for drugs such as 17-DMAG that might prevent or reverse these chromatin alterations.

Methods

ROS/MAP cohort and pathologic characterization

All participants were enrolled in the Religious Orders Study (ROS)¹⁰ or the Memory and Aging Project (MAP)¹¹. Studies were approved by the Institutional Review Boards of Rush University Medical Center and Partners Healthcare. Participants were not demented at the time of enrollment and agreed to donate their brain upon death. Amyloid-β and tau tangles were assessed postmortem as previously described^38,39. Briefly, brains were cut into 1-cm-thick coronal slabs and immersion fixed in 4% paraformaldehyde. Tissue blocks from 8 brain regions (hippocampus -CA1/subiculum-, angular gyrus, entorhinal, superior frontal, dorsolateral prefrontal, inferior temporal, anterior cingulate, and calcarine cortices) were embedded in paraffin and sectioned at 20μm. Paraffin-embedded sections were immunostained for amyloid-β using 1 of 3 monoclonal anti-human antibodies: 4G8 (1:9000; Covance Labs, Madison, WI), 6F/3D (1:50; Dako North America Inc., Carpinteria, CA), and 10D5 (1:600; Elan Pharmaceuticals, San Francisco, CA). Paired helical filament (PHF) tau tangles were labeled with an antibody specific to phosphorylated tau (AT8, Thermoscientific, Waltham, MA, USA). A computerized sampling procedure combined with image analysis software was used to calculate the percentage area occupied with amyloid-β and the density of PHFtau tangles. Composite scores were computed for overall amyloid-β burden and PHFtau tangle density by averaging the scores obtained from the eight brain regions. Amyloid-β and tau tangle scores were square root transformed for better statistical properties. Brain samples were randomly selected and sent to ChIP-seq processing in an arbitrary order. Individuals generating ChIP-seq data were blinded to the outcome measures. No statistical methods were used to pre-determine sample sizes but our sample sizes are similar to those reported in previous studies^5,40.

Human brain H3K9ac ChIP-seq

We identified the Millipore anti-H3K9ac mAb (catalog # 06–942, lot: 31636) as a robust mAb validated for ChIP-seq (validation data from Millipore available at http://www.emdmillipore.com/US/en/product/Anti-acetyl-Histone-H3-Lys9-Antibody,MM_NF-06–942). 50mg of gray matter was dissected on ice from biopsies of the DLPFC of the ROS/MAP cohorts. The tissue was minced and crosslinked with 1% formaldehyde at room temperature for 15mins and quenched with 0.125M Glycine. The tissue was then homogenized in cell lysis buffer using the Tissue Lyser and a 5mm stainless steel bead. Then the nuclei were lysed in cell lysis buffer and chromatin was sheared by sonication. Samples were incubated overnight at 4C with 2.5μl of the H3K9ac antibody with a final volume of 3ml using the ChIP Dilution Buffer. Chromatin bound to the antibody was purified with protein A sepharose beads. The final DNA was extracted and used for Illumina library construction following usual methods of end repair, adapter ligation and gel size selection. Samples were pooled and sequenced on the Illumina HiSeq (36 bp single-end reads).

Human brain H3K9ac ChIP-seq data pre-processing, peak detection and annotation

Single-end reads were aligned by the BWA algorithm against the human reference genome GRCh37 and reads with a mapping quality of 0 were removed⁴¹. Peaks were detected for each sample individually by MACS2 using the broad peak option, a stringent q-value cutoff of 0.001, and pooled genomic DNA of 7 samples as negative control library⁴². A combination of different ChIP-seq quality measures were employed to remove low quality samples⁴³: samples that did not reach (i) ≥ 15×10⁶ unique reads, (ii) non-redundant fraction ≥ 0.3, (iii) cross-correlation ≥ 0.03, (iv) fraction of reads in peaks ≥ 0.05 and (v) ≥ 6000 peaks were removed. After quality control, 669 out of 712 samples remained (Supplementary Table 1). We defined our H3K9ac domains by calculating all genomic regions that were detected as a peak in at least 100 (15%) of our 669 samples. Regions neighbored within 100 bp were merged and very small regions of less than 100 bp were removed, resulting in 26,384 H3K9ac domains. H3K9ac levels were quantified by counting the number of fragments overlapping the H3K9ac domains in each sample. Fragments were estimated by extending the short sequence reads towards the 3’-end by the optimal shift obtained by maximizing the cross-correlation for each sample. The mean optimal shift was 271 bp. Only non-duplicated uniquely mapped reads were used to quantify H3K9ac levels. H3K9ac domains were annotated using chromatin states (core 15-state model) from an unaffected DLPFC sample included in the Roadmap Epigenomics project¹² (E073). Proportion of chromatin states were calculated for each H3K9ac domain. Then, k-means clustering was applied to partition the H3K9ac domains according to their chromatin states into k=5 clusters that explained 61% of the variance of the domain annotation. Chromatin states of the five cluster centroids (Supplementary Fig. 1b) revealed that two clusters represented H3K9ac domains at promoters and another two clusters represented H3K9ac domains at enhancers. Promoter and enhancer clusters were merged resulting in three classes of H3K9ac domains: promoter, enhancer and other.

Mouse models

All mouse work was approved by the Committee on Animal Care of the Division of Comparative Medicine at MIT and complied with the relevant ethical regulations. Adult (3 months old) female double-transgenic CK-p25 mice²⁶ (crosses of p25 Tg mice, see https://www.jax.org/strain/005706, and Camk2a-tTA mice, see https://www.jax.org/strain/003010) and their respective control littermates were used for the experiments. Brain tissue was collected at either 2 or 6 weeks after p25 induction. For the tau (P301S) mouse model²⁵ (see https://www.jax.org/strain/008169), adult female 6 months old and male 11 months old transgenic mice with age- and gender-matched wild-type mice as controls were used. Hippocampal tissue was extracted for ChIP-seq and flash-frozen in liquid nitrogen. No animals were excluded from the study. Supplementary Table 9 summarizes the number of animals available at each time point.

Mouse brain H3K9ac ChIP-seq and data pre-processing

H3K9ac ChIP of mouse hippocampal tissue was performed as described for the human brain tissue using the Abcam antibody ab#4441 (lot numbers GR196840–1 and GR207546–1) at a concentration of 1.5 μl (= 1.5 μg) of antibody / 1 ml of chromatin sample. Validation data is available from the manufacturer at https://www.abcam.com/histone-h3-acetyl-k9-antibody-chip-grade-ab4441.html. ChIP-seq data was pre-processed similar to the human DLPFC ChIP-seq data: BWA was used to map reads against the reference assembly NCBI m37. Same quality measures were calculated (Supplementary Table 14) as for the human data, but we did not require samples to have ≥ 15×10⁶ unique reads, because CKp25 samples were sequenced at a lower coverage. MACS2 was used for peak detection and H3K9ac domains were defined as regions that were covered by a peak in at least 4 (17%) out of the 23 samples (n=12 CKp25; n=11 tau) resulting in 44,165 H3K9ac domains.

Induction of human neuronal cells overexpressing MAPT

The iPSC line used in this study (YZ1) was originally generated from the IMR-90 cell line (ATCC) and characterization of this line was described previously⁴⁴. Due to a karyotype abnormality in a small subset of cells, monoclonal isolates were obtained and confirmed to be karyotypically normal and pathogen free prior to initiating this study. The iPSC line used was confirmed to be of the correct identity prior to and at the conclusion of the study using short tandem repeat (STR) profiling (Genetica Cell Line Testing). Neurons were generated from the direct conversion of induced pluripotent stem cells by transduction with Neurogenin 2 as previously described⁴⁵. Neurons were plated on Matrigel coated 96-well plates on DIV4 and maintained in media consisting of 485 ml neurobasal medium (Gibco), 5 ml Glutamax, 7.5 ml 20% Dextrose, 2.5 ml MEM NEAA with 1:50 B27, BDNF, CNTF, GDNF and doxcylcine. On DIV17, cells were transfected with a 1:1 dilution of lentivirus-packaged open reading frames (ORF) expressing human MAPT (titer 6.2×10⁶), resulting in overexpression of tau protein, or a GFP expressing construct (pRosetta). Following 18-hour incubation, virus-containing media was removed and replaced with fresh media to incubate until DIV31 when harvested for ATAC-seq. RNA-sequencing revealed a 10-fold increase of MAPT mRNA in cells transfected with MAPT cDNA compared to controls. In total, 9 MAPT OE and 9 control iNs were generated in 3 batches (differentiation rounds) of triplicate wells.

17-DMAG treatment of induced neurons

The second and third batch of experiments contained cell cultures of induced neurons treated with 1, 3, or 10 nM 17-DMAG for 24 hours prior to collection. Concentrations were chosen based on previous work³³. To account for vehicle specific effects, MAPT overexpressing cells and control cells cultured in concentration-matched DMSO solution for 24 hours prior to collection. Experimental design of the iN experiments is shown in Supplementary Table 13.

Assay for transposase-accessible chromatin using sequencing (ATAC-seq)

ATAC-seq was performed as previously described²⁹ with three modifications: (i) To reduce mitochondrial DNA contamination, we substituted 0.1% Ipegal with 0.01% Tween in the nuclear lysis buffer. (ii) Additional tagmentation buffer (TD; 2x) was prepared as follows: 20 mM Tris(hydroxymethyl)aminomethane and 10 mM MgCl2; adjusted to pH 7.6 with 100% acetic acid before addition of 20% (vol/vol) dimethylformamide⁴⁶. (iii) We performed the PCR amplification in a 25μl total volume rather than 50μl to save half of our tagmented DNA for backup. After 31 days in vitro, iPSC neurons were trypsinised and counted to provide 50,000 live cells for ATAC-seq. iPSC-derived neurons were lysed with 0.01% Tween, 10 mM Tris-Cl, pH 7.4, 10 mM NaCl, 3 mM MgCl2 and nuclei pelleted, washed and transposed upon addition of 2.5 μl Nextera Tn5 Transposase in 1x TD buffer (Nextera Kit, Illumina) and incubation at 37°C for 30 mins. Immediately following transposition, DNA fragments were purified using a Qiagen MinElute PCR Purification Kit (Qiagen) following the manufacturer’s instructions. Transposed DNA was amplified with dual indexed Nextera PCR primers for 5 cycles prior to removing a 5 μl aliquot from each sample to test further amplification requirements by qPCR. Rounding up the cycle threshold from each sample decides how many additional cycles each sample needs to be amplified. For these samples, we needed to re-array and continue amplifying the remaining 20 μl for 5–8 cycles. Final libraries were cleaned twice with 1.5X Ampure XP SPRI beads and eluted in 12 μl of water. We quantified with Qubit HS DNA assay and assessed for quality using a Bioanalyzer High-Sensitivity DNA analysis Kit. 20 ng of each sample were pooled and diluted to 4nM using an estimated library length of 300 bp for all and submitted to the Broad Institute’s Genomics Platform for Illumina HiSeq2500 25 bp paired-end sequencing.

ATAC-seq data pre-processing, peak detection and annotation

Paired-end reads from ATAC-seq were aligned by the BWA algorithm against the human reference genome GRCh37⁴¹. MACS2 was applied to each sample using fragment sizes obtained from the paired alignments and a q-value cutoff of 0.001⁴². Our ChIP-seq quality control pipeline was adapted for ATAC-seq. Instead of cross-correlation we calculated the median insert size and verified that the distribution of insert sizes showed a periodicity equal to the helical pitch of DNA (Supplementary Table 15). No experiments were excluded. ATAC domains were defined as genomic regions covered by a peak in at least 4 (22%) of our 18 samples (n=9 MAPT OE; n=9 controls). Domains less than 50 bp away from each other were merged. In total, we obtained 40,637 ATAC domains. ATAC domains were annotated with chromatin states obtained from hESC-derived neurons included in the Roadmap Epigenomics project (sample E010, core 15-state model)¹². k-means clustering was applied to partition the ATAC domains according to their chromatin states into k=4 clusters that explained 65% of the variance of the domain annotation. Chromatin states of the four cluster centroids (Supplementary Fig. 4a) revealed that the first three clusters represent ATAC domains at TSS, TSS flanking regions and enhancers. The fourth cluster represents ATAC domains with various other chromatin states.

H3K9ac Chip-seq of induced neurons

H3K9ac ChIP-seq data from iNs were generated and processed as described for the DLPFC tissue samples. No experiments were excluded based on quality measures (Supplementary Table 15). Pooled input DNA from all three samples (n=2 MAPT OE, n=1 control) were used as negative control library during peak detection. Peaks that were detected in at least two of the three samples were defined as neuronal H3K9ac domains (n=34,988) and used for subsequent analyses.

Statistical analysis

The numbers of reads Y within each of the m=26,384 H3K9ac domains and n=669 individuals were calculated and assumed to follow a negative binomial distribution Y ~ NB(μ,θ) with mean μ and dispersion parameter θ ⁴⁷. A log-linear regression model

$$log\left(E\left(Y\right)\right)=log\left(n\right)+{\beta }_0+x_t{\beta }_t+x_a{\beta }_a+z\gamma$$ (1)

was fitted for each H3K9ac domain separately using the R package MASS. The offset log(n) accounts for different total numbers of reads n (summed over all domains) per sample. Coefficients β_t and β_a model the effect of the observed square root-transformed tau and amyloid-β levels on the H3K9ac levels. Coefficients were tested by Wald tests and the false discovery rate (FDR) was calculated to account for multiple testing. Vector γ models the biological covariates age and gender as well as the technical covariates post mortem interval, batch and cross-correlation.

Normalized ROS/MAP transcription data from the DLPFC were downloaded from Synapse (Synapse: syn3388564). RNA-seq data generation and quantification with RSEM have been described elsewhere^40,48. Pathologic variables were available for n=500 individuals. Only transcripts with rpkm values larger or equal to 2 in at least 25% of the samples were considered as active in the human brain and kept in the data set (m=24,594 transcripts). A linear regression model was fitted for each transcript with the log₂-transformed rpkm values as outcome variable and tau and amyloid-β as explanatory variables. Models were adjusted for age, gender, RIN score (RNA integrity number), post mortem interval, study index (ROS or MAP) and logarithmized total number of aligned reads. Coefficients were tested by t-tests and the FDR was calculated to account for multiple testing.

Mouse H3K9ac data were analyzed using the voom method implemented in the R package limma⁴⁹. For each mouse model and time point, we compared n=3 AD mice (n=2 for the tau mouse model at t=6 months) to n=3 respective control mice. Voom has been designed for small sample sizes and provides reliable domain-specific estimates of the variance by estimating the mean-variance trend using information from all H3K9ac domains. The voom method is incorporated in a weighted linear regression framework with log₂-transformed reads per million (log₂ rpm) as outcome. A model was fitted for each of the m=44,165 domains and the cross correlation was added as technical covariate in addition the AD/control indicator variable. Significance of the AD/control variable was assessed by moderated t-tests and the FDR was calculated to adjust for multiple testing⁵⁰.

Neuronal ATAC-seq data of MAPT OE iNs (n=9 experiments) and control iNs (n=9 experiments) from three batches were analyzed using voom⁴⁹. For each of the m=40,637 ATAC domains, log₂-transformed rpm values were regressed on the tau indicator variable (MAPT OE or control) adjusted for batch and logarithmized total number of reads in all domains of the sample (library size). Significance of the tau variable was assessed by moderated t-tests and the FDR was calculated to adjust for multiple testing⁵⁰. Hierarchical clustering (Euclidian distance, average linkage) shown on Supplementary Fig. 4d was applied to logarithmized rpm values after subtracting the estimated effect of the total number of reads (library size). For the comparison of batches (Supplementary Fig. 4e–j), the same regression models but without batch variable were applied to each batch (n=3 MAPT OE iNs, n=3 control iNs) separately. The effect of 17-DMAG treatment was analyzed separately for each batch/concentration combination (batch 2: 3 nM, batch 3: 1, 3 and 10 nM) using linear regression models within the voom framework. A binary variable encoding treatment or control condition (n=3 MAPT OE iNs + 17-DMAG, n=3 control iNs) and the logarithmized total number of reads (library size) were used as explanatory variables.

Accounting for altered cell type proportions

The Digital Sorting Algorithm (DSA) was applied to estimate the proportions of neurons, astrocytes, oligodendrocytes, myeloid cells and endothelial cells in the cortical tissue from RNA-seq data of n=452 individuals that had both H3K9ac and RNA-seq data²². A set of five marker genes was derived for each cell type (Supplementary Table 8) from published RNA-seq data of purified cells⁵¹. We calculated the mean rpkm value for each gene in each cell type and required a marker gene for a specific cell type to have a mean rpkm value ≥ 25 in that cell type and a mean rpkm value ≤ 3 in the other four cell types. Next, we selected the top five marker genes per cell type with the smallest coefficients of variation in our tissue-level RNA-seq data to avoid selecting genes with highly variable transcription which are likely to be affected by various cellular processes resulting in imprecise estimations. For cell type proportion estimations, we used gene-level transcription values derived from the RNA-seq data available at Synapse (Synapse: syn3505732) instead of the isoform-level values used in the main analysis. Estimated cell type proportions are given in Supplementary Table 1. The proportion of neurons was decreased (p=0.043) and proportions of oligodendrocytes (p=0.003) and myeloid cells (p=0.016) were increased in AD subjects (two-sided t-tests, n=452 individuals, unadjusted p-values). The main analysis of the H3K9ac data for the subset of 452 samples with RNA-seq data was repeated using (i) no adjustment, (ii) the proportion of one of the five cell types as additional covariate (iii) the proportions of all five cell types as additional covariates in the regression model used for the main analysis given by equation (1). Pearson correlation between the tau coefficients of the unadjusted model and the tau coefficients of any adjusted model was > 0.99 (Supplementary Fig. 3a–f). In total, 2,514 H3K9ac domains were significantly associated with tau (FDR ≤ 0.05) in the unadjusted model with the reduced sample size of n=452. None of the significant H3K9ac domains lost significance (FDR > 0.1) in any of the adjusted models. When adjusting for all cell types, an additional 11 H3K9ac domains were significantly (FDR ≤ 0.05) associated with tau that were not associated (FDR > 0.1) in the unadjusted model. Although cell type proportion affected the tissue-level H3K9ac levels, these results indicate that changing cell type proportions did not severely confound the tau effect and we therefore decided to not adjust our primary analysis for cell type proportions so that we can the full sample size of n=669 individuals.

Determining genomic segments of concordant tau-related changes

Genomic segments for the DLPFC H3K9ac data (Fig. 2a, Supplementary Table 6) were defined on the p-values for the tau coefficient β_t from model (1). P-values were log-transformed and the signs of the respective tau coefficients β_t were assigned to the transformed p-values. The circular binary segmentation algorithm implemented in the R package DNAcopy was then applied to the transformed p-values with the corresponding genomic positions of the H3K9ac domains¹⁴. We identified 178 genomic segments using the default value of α=0.01 for acceptance of new change-points. The same approach was applied to the transformed p-values from the neuronal ATAC-seq data (comparing MAPT-OE to control iNs) to identify 99 genomic segments in iNs (Fig. 4e, Supplementary Table 10).

Calculating mean tau effects of genomic segments

The mean tau effect of a segment was defined as the weighted mean of the tau coefficients β_t of all H3K9ac (or ATAC) domains located within the segment. The tau coefficients were weighted by their inverse standard deviation. Similarly, for transcription data, the TSS was used to map genes to segments. Coefficients and their standard deviations for DLPFC H3K9ac, DLPFC RNA-seq and iN ATAC-seq data were obtained from the models described in the section Statistical analysis. In addition, Fig. 2e and 4i depict segments’ mean tau effects from DLPFC methylation data, Mayo LOAD study transcription data, neuronal transcription data, and iN H3K9ac data. Mean effects were calculated in the same manner using coefficients and standard deviations obtain from the data sets and models described as follows.

Normalized ROS/MAP DNA methylation data from the DLPFC were downloaded from Synapse (Synapse: syn3157275). Data generation and processing have been described elsewhere⁵. Pathologic variables were available for n=729 individuals. Relative methylation levels (“beta values”) were assumed to follow a beta distribution⁵². A beta regression model with probit link function was fitted for each CpG with the methylation level as outcome using the R package betareg⁵³. Tau, amyloid-β, age, gender, post mortem interval, batch and bisulfite conversion rate were used as explanatory variables. Most CpG’s were located in the chromatin states (Sample E073, 15 states model) quiescent (24%), active transcriptional start site (21%), weakly transcribed (15%) and enhancer (7%). Segments’ mean tau effects in DNA methylation data were calculated for these four chromatin states separately since DNA methylation levels vary between different functional regions⁵⁴.

Normalized gene expression data from the Mayo LOAD study were downloaded from Synapse (Synapse: syn3157225)^21,55. Individuals with progressive supranuclear palsy (PSP) were excluded resulting in n=202 samples with AD and n=90 no-AD/PSP samples. Array probes with a low call rate (<50%) were excluded reducing the number of probes from 24,526 to 17,910. For each probe, we fitted a linear regression model with the log-transformed expression value as outcome. Since quantitative measurements of AD pathologies were not available, a binary variable encoding the AD diagnosis was used as explanatory variable to model the effect of AD on transcription. Models were adjusted for age, gender, PCR plate and RIN score.

Raw neuronal expression data were downloaded from Gene Expression Omnibus (GEO: GSE5281)²⁴. The RMA method implemented in the R package affy was applied to normalize data and summarize intensity values at the probeset level⁵⁶. Only samples collected from the superior frontal gyrus were used (n=23 AD, n=11 non-AD samples). The R package limma was used to fit a linear regression model for each of the 54,675 probesets with the log-transformed transcription level as outcome⁵⁰. AD diagnosis and gender were used as explanatory variables.

H3K9ac ChIP-seq data from iNs (n=2 MAPT OE iNs, n=1 control iNs) were generated and processed as described for the DLPFC tissue samples. No experiments were excluded after quality control (Supplementary Table 15). Pooled input DNA from all three samples were used as negative control library during peak detection. Peaks that were detected in at least two of the three samples were defined as neuronal H3K9ac domains (m=34,988) and used for subsequent analyses. The voom method was applied to test for differences between MAPT OE and control neurons⁴⁹. Cross-correlation was calculated and added to the models to adjust for technical variability.

A/B compartment and lamina-associated domain annotation

Human fetal brain A/B compartments were derived from a published Hi-C data set (GEO: GSE77565) as described in the original publication¹⁶. Briefly, Pearson correlation matrices were calculated from the intra-chromosomal observed/expected 40 kbp binned interaction matrices. For each chromosome, the sign of the first principal component obtained from the correlation matrix was used to categorize bins as type A or B compartment. Mean tau effects in compartments (Fig. 2b and 4h) were calculated as described for genomic segments.

The 178 genomic segments derived from DLPFC H3K9ac data were annotated with nuclear lamina association using published DamID data for Lamin B1 in human fibroblasts (GEO: GSE22428)¹⁷. Probes of the DamID array were classified into nuclear lamina bound or unbound in the original work. We re-mapped the probes against the genome version GRCh37 and defined the lamina association of a segment as the ratio of nuclear lamina bound probes to all probes in the segment. To sort segments in Figures 2e and 4i, chromatin accessibility was calculated from chromatin state annotation of sample E073 (DLPFC) from the Roadmap Epigenomics Project¹². Chromatin accessibility was defined as the fraction of the segment that was annotated with any chromatin state other than heterochromatin, repressed polycomb, weak repressed polycomb or quiescent/low. These four chromatin states exhibited the lowest DNA accessibility measured by DNase-seq¹². As expected, chromatin accessibility was negatively correlated with lamina association in the DLPFC (Spearman correlation ρ=−0.86, n=178 segments): that is, segments of more open chromatin were less likely to include a LAD. Nuclear lamina association for the 99 genomic segments derived from iN ATAC-seq data was calculated in the same manner using the same DamID data for lamin B1 from human fibroblasts. Published lamin B1 data from mouse embryonic fibroblasts were used to annotate our mouse H3K9ac data¹⁷. Genomic segments and their lamina association are given in Supplementary Tables 6 (n=178 DLPFC H3K9ac segments) and 10 (n=99 iN ATAC-seq segments). Small segments with < 20 H3K9ac domains or < 20 active transcripts (<20 ATAC domains) and segments located at sex chromosomes were removed when calculating the correlation between lamina association and mean tau effects shown in Fig. 2c–e (Fig 4g, i) reducing the number of DLPFC H3K9ac (iN ATAC-seq) segments to 138 (89).

Analysis of the overall effect of MAPT OE

We overexpressed MAPT in iNs to replicate the spatial pattern of tau-related H3K9ac alterations in the human brain data (Fig. 2a). We found a similar pattern of tau-related alterations in the Manhattan plot derived from the ATAC-seq data of the iNs (Fig. 4e), but the p-values in the Manhattan were not independent since all tests were conducted on the same set of samples, and a few samples could drive the pattern. To address this question, we defined a segmentation score which quantifies whether the ATAC domain values of a single sample follow the spatial pattern that we observed in the human brain H3K9ac data. To calculate the score, we first regressed the log rpm values of each ATAC domain on the technical variables batch and logarithmized total number of reads. This is the same model as used for the main analysis, but we excluded the variable that indicates whether the sample overexpresses MAPT. The residuals r_ij for sample i and domain j from these domain-wise regression models were extracted and still contain the MAPT overexpression effect. Next, the linear model r_ij = a_i + b_i S_ι(j) + e_ij was fitted for each sample i. S_k denotes the mean tau effect observed for segment k in the human brain H3K9ac data, and ι is a function mapping ATAC domain j to the segment k where the domain is located. Factor b_i is the segmentation score for sample i. The rationale is to model a sample’s residuals using the segments and the segments’ mean tau effects from the H3K9ac data. Segmentation scores for the 9 MAPT OE iNs and the 9 control iNs are shown in Supplementary Fig. 4k. Since residuals are centered at 0, the score is also centered at 0. The score was significantly different (p = 0.02; Wilcox-Mann-Whitney test) between the two groups. In line with the inverse tau effect in the Manhattan plot (Fig. 4e), the score was smaller for MAPT overexpressing iNs.

Search for drug candidates in the Connectivity Map data base

Only H3K9ac domains spanning a TSS of an active transcript (≥ 100 samples with rpkm ≥ 2 in the RNA-seq data) were considered. Gene annotation of the TSS were used to map H3K9ac domains to genes in the Connectivity Map data base³⁰. The search algorithm provided by the Connectivity Map 2.0 requires a set of up- and downregulated genes as input. As suggested in the Connectivity Map documentation, we selected the top 250 positively tau-associated genes and the top 250 negatively tau-associated genes that had the smallest p-values in our H3K9ac data and calculated the enrichment score and p-values on the web server (https://www.broadinstitute.org/cmap). In addition, we calculated the Spearman rank correlation coefficient between the vector of standardized coefficients β_t/σ_βt for tau obtained from our H3K9ac data and the differential expression statistic from the Connectivity Map for each compound (n=1309) in the data base. A negative correlation indicates that genes showing increased H3K9ac levels with tau were repressed by the respective compound.