A large single-nucleus RNA-seq atlas to study transcriptomic changes of brain myeloid cells along the spatiotemporal progression of AD

Brain tissue samples from 32 donors at varying stages of tau pathology were split into 4 groups based on prior neuropathological characterization (Fig. S1a, Table S1). To capture myeloid cells along the stereotypical progression of tau pathology, we selected 5 brain regions that included allocortex and neocortex, from expected high to low pathology: entorhinal cortex (EC), inferior temporal gyrus (ITG), prefrontal cortex (PFC), visual association cortex (V2), and primary visual cortex (V1) (Fig. 1a). Compared to pioneering snRNA-seq studies [17, 34], we captured more than 150 times the number of brain myeloid cell nuclei per tissue sample with our enrichment protocol (337,475 total). Further, separation of nuclei by their corresponding cell types combined with deeper sequencing led to increased numbers of genes (more than three times the number of median UMIs) detected per myeloid cell nucleus compared to any published study (Fig. 1b). With our enrichment protocol (Fig. S1b), brain myeloid cells amounted to 24–34% of total nuclei per region (Fig. 1c, d).

Fig. 1

Study design and identification of brain myeloid cells across brain regions. a Study design. Samples from 5 brain regions of in total 32 donors along four stages of AD pathology progression were snRNA-seq profiled and characterized by quantitative readouts of tau as well as Aβ 3D6 IHC. Samples were divided into 4 pathology groups, according to their Braak and Thal stage. b Comparison of dataset size and median UMIs per microglia/brain myeloid nucleus vs. public microglial studies. c Per region microglia/brain myeloid cell numbers as proportion of all NeuN-/Olig2- cells per region. d UMAP representation of NeuN-/Olig2- cells, brain myeloid cells are colored in blue. Grey and blue numbers correspond to the absolute NeuN-/Olig2- sorted non-myeloid cells (e.g., astrocytes, endothelial cells, and pericytes) and myeloid cells, respectively. e pTau/Total Tau, HT7 Aggregated Tau, HEK seeding, and 3D6 Amyloid-β measurements for each pathology group (CTRL → AD), across brain regions. f Quantification of CD11c and CD68 immunohistochemistry across brain regions and pathology groups. g Representative CD11c and CD68 IHC (EC, grey matter) of pathology group 4 samples, with CD11c in brown and plaques (3D6) in red, and CD68 in brown and plaques (D54D2) in red, respectively (scale bar 100 µm). IHC across pathology groups in Fig S1d/e

Biochemical and neuropathological characterization reveals brain myeloid cell responses associated with spatiotemporal aspects of AD pathology

To correlate transcriptomic changes in myeloid cells with local levels of pTau and Aβ pathology, we conducted an extensive biochemical and immunohistochemical quantitative analysis in samples adjacent to those used for snRNA-seq. Tau protein becomes hyperphosphorylated early in disease, which contributes to its aggregation [2]. In line with prior histological and biochemical studies (e.g., [11]), we observed a pattern of pTau/Tau levels of EC > ITG > PFC > V2 > V1 (Fig. 1e, Fig. S1c). In a given brain region pTau/Tau levels reflected the pathology groups, with donors of pathology group 4 (dark blue, Braak VI) and 1 (yellow, Braak 0/I/II) showing highest and lowest levels of pTau, respectively. A similar pattern was seen for HT7 aggregated tau and the propensity of lysate material for tau seeding by HEK biosensor cells (HEK seeding). As expected, Aβ pathology, as measured by 3D6 immunoreactivity, was highest in neocortex (PFC and ITG) (Fig. 1e). These pathology readouts confirmed earlier studies and revealed the expected pathology levels in the brain samples selected for this study.

To determine the dynamics of reactive microglia with respect to disease progression, we stained FFPE tissue sections from the contralateral hemisphere to that used for snRNA-seq with antibodies for reactive microglia (CD11c and CD68) and plaques (3D6), and quantified the area covered by microglia markers, plaques, and co-localized area. CD11c (encoded by ITGAX) increased from pathology group 1 to pathology group 4, and showed highest expression in EC (high-tau) and PFC (high-tau & Aβ), suggesting first a tau-associated (path. group 3) and later tau & Aβ associated (path. group 4) changes (Fig. 1f, g, Fig. S1d). On the other hand, CD68 protein expression increased from pathology group 1 to group 4 in EC, ITG, and V1, and had highest immunoreactivity in EC followed by PFC within all pathology groups, suggesting an association with early tau and early Aβ pathology (Fig. 1f, g, Fig. S1e). Thus, these typical reactive microglia markers are both associated with tau and Aβ pathology progression, but demonstrate unique spatial and temporal patterns.

Comparison of myeloid cells across brain regions reveals an EC-specific signature

Clustering of brain myeloid cell nuclei based on their transcriptomes showed few donor-specific clusters after mapping across all donors (Fig. 2a, Fig. S2a). When integrating brain myeloid cells across different brain regions, these appeared to be very similar (Fig. 2b), with < 1% of detected genes being differentially regulated in any given region (Fig. S2b, Table S2) and with high correlation of clusters between regions (Fig. S2c, Table S3). Although most brain myeloid cells were highly similar across brain regions, one group of brain myeloid cells from EC clustered separately from those in other regions and showed differentially expressed genes (DEGs) associated with vesicles and potassium transport (Fig. 2c, Fig. S2d). Notably, this cluster was observed across all donors, and was neither specific to donors with high or low pathology, nor enriched for Aβ or tau pathology readouts (Fig. S2e, f). Although the relative proportion of DEGs in any given region was small, EC also showed the most DEGs of any of the 5 regions [75 genes, 62 up—including IFNGR1 (encoding the interferon gamma receptor 1)—and 13 down], followed by V1 (67 DEGs, 26 up and 37 down), while V2 showed the least DEGs (zero genes) (Fig. 2d, e). In summary, this analysis revealed a unique transcriptomic signature of myeloid cells in EC, highlighting allocortical vs. neocortical differences that might contribute to differences in vulnerability to tau.

Fig. 2

Brain myeloid cell similarity across brain regions. a Brain myeloid cell subclustering per brain region. Macrophage cluster numbers are indicated in bold (based on LYVE1, MRC1, CD163, and F13A1 marker genes). b Cross-region integration of subsampled brain myeloid cells across brain regions shows alignment between regions for most cells, except for one EC-enriched population of cells (highlighted in black circles). Shown are combined and per region UMAP plots. c EC enriched population (indicated as green cells in UMAP plot) was compared to all other brain myeloid cells across regions. Biological process GO term enrichment indicates upregulated synapse vesicle cycle changes and ion transport differences. d Up- and downregulated differentially expressed gene (DEG) numbers per region, filtered for microglial genes (average log2FC > 0.25). e Top 5 upregulated microglia genes per region (no significantly upregulated V2 markers identified)

Correlation with global and local AD neuropathology reveals distinct homeostatic and AD-associated brain myeloid states

To identify brain myeloid cell subsets associated with AD pathology, we analyzed the percentage of myeloid cell nuclei per cluster from high and low-pathology donors, including microglia and perivascular macrophages (PvMs) as identified by marker genes LYVE1, MRC1, F13A1, and CD163. We reasoned that AD-associated brain myeloid cell clusters should have a significantly higher proportion of nuclei from high vs. low-pathology donors and/or correlate positively with any of the local tau and Aβ pathology readouts. By contrast, homeostatic microglia clusters should have a higher proportion of nuclei from low vs. high pathology donors and/or correlate negatively with the local tau and Aβ pathology readouts.

Screening for AD-associated microglia, we detected several clusters with a significantly higher number of high- and low-pathology donor myeloid cell nuclei than expected by chance. For example, ITG microglia clusters 3 and 4 showed significantly more high pathology (pathology group 3 and/or 4) donor nuclei (Fig. 3a, ITG upper heatmap; cluster 3: adj. p values 5.3e-7 and 8.5e-8, respectively; cluster 4: adj. p value 2.2e-14). Further, we identified several clusters for each brain region for which the percentage of nuclei per donor was correlated with its pathology readout. For example, the proportion of microglia in ITG cluster 3 showed a significant positive correlation with all tau and Aβ pathology readouts (Fig. 3a, ITG lower heatmap; 3D6 p value < 0.01, pTau231/Total Tau p value < 0.05, HEK seeding p value < 0.001, HT7 Aggregated Tau p value < 0.01). Importantly, the gene signature of this cluster (i.e., DEGs as compared to cluster 0, which was equally contributed by all four pathology groups and did not correlate with any pathology readout) also positively correlated with the “AD1” human AD microglia described by Gerrits et al. [16] and negatively correlated with human homeostatic microglia reported by Mancuso et al. [32] (Fig. S3a, ITG heatmap, “Gerrits_AD1” p value < 2.22e-16, “Mancuso_HM” p value 1.67e-10), reinforcing the identity of this microglia cluster as the AD-associated microglia. Overall, microglia clusters that positively correlated with tau or with tau & Aβ pathology were mainly observed in early tau regions (e.g., EC cluster 4, ITG clusters 3, 4 and 9) (Fig. 3a, solid black boxes), and these clusters were characterized by genes in pathways including “Scavenging by Class A Receptors” (EC cluster 4, ITG clusters 3 and 4), and “Cell recruitment (pro-inflammatory response)” (ITG cluster 3) (Table S4, Table S5). The proportion of PvM clusters did not show any significant correlations with any of the pathology readouts, but did show increased proportion of nuclei from pathology group 4 donors in EC, ITG, V2, and V1 (proportions EC—7.3%; ITG—3.1%; PFC—7.2%; V2—3.2%; V1—1.4%; adj. p values EC 4.7e-15, ITG 2.2e-14, V2 5.4e-12, and V1 4.77e-5, respectively).

Fig. 3

Identification of tau- and Aβ -associated microglia and brain macrophage subpopulations. a Per cluster pathology group enrichment shown as observed over expected ratios (scaled to 1) (upper panels of heatmaps) and Spearman correlation of 3D6 and tau readouts with proportion of brain myeloid cells per cluster (lower panels of heatmaps). ‘*’corresponds to significant enrichment >  = 10% (binomial test, adj. p value < 0.001), and significant Spearman correlation (p value < 0.05), respectively. Solid black boxes denote clusters positively correlated with pathology; dashed black boxes denote clusters negatively correlated with pathology. Bold cluster numbers indicate macrophage clusters, characterized by increased expression of LYVE1, MRC1, F13A1, and CD163. b Mapping of disease-associated clusters per region (right) to cross-region integrated data (left) confirms similarity of disease-associated clusters across brain regions, albeit indicating expression differences between primarily tau- (clusters 2/4 in integrated data) and tau + Aβ -associated clusters (clusters 5/8 in integrated data). c 3D6 IHC, pTau/Total tau, and HT7 aggregated tau readouts were binned into 5 equally spaced categories, representing no-to-late pathology. For simplicity, integrated microglia are shown for no, early, and late pathology only (first, middle, last bin), based on their cellular density in individual clusters (UMAP representation). Grey plots beneath visually summarize shifts of brain myeloid cells into clusters stratified for early (lightblue) and late (darkblue) pathology. For HT7 aggregated tau, bin #4 (not #5) is shown at late stage, as last bin (#5) only contained data from one donor. d Spearman correlation of cross-region brain myeloid cell clusters (using DEGs per cluster vs. cluster 0) with public genelists. Significant correlation indicated by ***p value < 0.001, **p value < 0.01, *p value < 0.05, grey boxes indicate insufficient data (number of overlapping genes between data sets < 10). AD1 and AD2 human microglia signatures from [16]; laser capture microdissected samples from [9] with signatures “Das_LCM_Plaque” (ThioflavinS + plaques), “Das_LCM_Peri_Plaque” (50 µm area around plaques), “Das_LCM_NFT” (neurofibrillary tangles with the 50µm area around them), “Das_LCM_Distant” (area > 50µm away from plaques), “Das_LCM_Plaque_vs_NFT” (ThioflavinS + plaques vs. neurofibrillary tangles); human iPSC-derived microglia-like cells transplanted into mice, with signatures CRM2 cytokine response 2, CYT/CRM1 cytokine response 1, DAM (disease associated), HLA antigen-presenting response, HM homeostatic, IRM (Interferon response), RM (ribosomal response), TRANS transitioning CRM from [32]; and primary mouse microglia tau fibril response genes from [51]

Regarding homeostatic microglia, microglia clusters negatively correlated with pathology were mainly observed in later tau regions, as expected (V2 cluster 0: 3D6 p value < 0.01, HEK Seeding p value < 0.05, HT7 Aggregated Tau p value < 0.01; V1 cluster 0: 3D6 p value < 0.05) (Fig. 3a, dashed black boxes). These microglia clusters showed typical markers of microglia homeostasis, e.g., P2RY12, and newly identified homeostatic microglia genes, such as SYNDIG1, FOXP2, OXR1, and LINC02232 (Table S4). Among the microglia clusters negatively correlated with pathology were ITG cluster 1 and PFC cluster 2 (Fig. 3a, dashed black boxes). Both showed a significant negative correlation with the pTau/total Tau ratio (Fig. 3a; p values < 0.05 and 0.01, respectively) and were characterized by an increased expression of ribosomal genes associated with translation and viral transcription, as well as iron uptake and storage genes FTL and FTH1, encoding ferritin protein light and heavy chains, respectively (Table S5). Thus, in this study, FTL + microglia did not increase with pTau or Aβ load, unlike previously reported [26], but rather showed significantly decreased proportions with increasing tau progression. Furthermore, ferritin-positive microglia have previously been described as “dystrophic” and “senescent” (e.g., [30]); however, we did not observe any enrichment of genes or pathways associated with apoptosis or senescence within these clusters (Table S5). These microglia clusters (ITG cluster 1 and PFC cluster 2) showed a high similarity to ribosomal response microglia recently described in a human microglia transplantation model [32] (Fig. S3a, “Mancuso_RM”).

To confirm the observed microglia and PvM phenotypes at the protein level and their localization with respect to pathology, we performed immunohistochemistry in the ITG region across all donors using antibodies for markers of pathology-associated (CPM) and homeostatic (TMEM119) microglia, and PvMs (CD163), with nearly adjacent sections (between 30 and 100 µm away) stained for amyloid plaque and tau pathology (Fig. S3b-e). We did not observe a difference in TMEM119 immunoreactivity between pathology groups, suggesting that homeostatic microglia are not correlated with pathology. However, we did observe an increase in CPM immunoreactivity with respect to pathology groups, and CPM positive cells were observed adjacent to plaques and dystrophic neurites, according to nearly adjacent sections stained with 3D6 and AT100, respectively. These results are in line with the observed transcriptional changes: TMEM119 is a homeostatic marker of ITG cluster 0, whose proportions are not significantly different with respect to pathology group or pathology readouts. CPM is a marker of ITG cluster 3, whose proportion was significantly positively associated with pathology group 3 and 4 donors as well as with all 4 pathology readouts in ITG. CD163 immunoreactivity was observed in cells with a monocyte/ macrophage-like morphology in the brain vasculature (Fig. S3d, arrow). Total levels of CD163 immunoreactivity did not change with respect to pathology, with no significant differences observed between pathology groups. This trend is in line with our observation of increased CD163 gene expression in cluster 6 of ITG, which matches the transcriptomic profile of PvMs and does not show any correlation with pathology groups. We also observed scattered CD163-positive cells in the brain parenchyma (Fig. S3d, arrowhead), and an increase in this parenchymal CD163 in 2 donors of pathology group 4, a finding which would need to be confirmed in a larger cohort. Interestingly, ITG microglia cluster 4 showed significantly increased CD163 as compared to all other clusters, and a significant correlation with higher pathology groups. This cluster did not have an overtly PvM-like phenotype based on transcriptomic profile, suggesting that it may align with the CD163-positive parenchymal microglia-like cells that we observed by IHC that increased in 2 pathology group 4 donors. The parenchymal CD163-positive cells were not abundant enough to determine the exact localization with respect to amyloid or tau pathology in nearly adjacent sections in these 2 donors, despite their trend toward increase in relation to overall pathology load.

In summary, correlations between local biochemical and neuropathological measures of tau and Aβ pathology and microglia transcriptomic clusters enabled us to discern between homeostatic and AD-associated microglia in multiple brain regions.

Correlation with local tau vs. Aβ measures reveals distinct subsets of AD-associated microglia

Once established that the existence of AD-associated microglia is distinct from homeostatic microglia, we aimed to identify associations between microglia clusters and local tau vs. Aβ measures that may indicate specialized responses to one or the other pathology. Prior research suggested that microglia show unique responses to tau vs. tau & Aβ [16]. Thus, we investigated marker genes and pathways in microglia clusters positively or negatively correlated with tau measures, but with no significant association to Aβ pathology, like EC cluster 4 and PFC cluster 6. EC cluster 4 showed upregulated markers of hypoxia and inflammatory response (HIF1A, DUSP1, FOS) and was represented by pathways including “response to decreased oxygen levels” (Table S5). Moreover, it correlated with cytokine response (“Mancuso_CYT_CRM1” p value 3.33e-3, “Mancuso_CRM2” p = 1.26e-4), HLA (“Mancuso_HLA” p = 2.75e-3), and tau fibril-treated microglia (“Wang_Tau_Fibril” p = 2.81e-6) as identified in published studies (Fig. S3a). PFC cluster 6 showed markers and pathways similar to those of EC cluster 4, including “response to decreased oxygen levels”, and exhibited a positive correlation with cytokine response microglia identified by Mancuso et al., 2022 (“Mancuso_CRM2” p = 3.43e-5) (Fig. S3a, Table S5).

Only V2 cluster 1 showed a significant positive correlation with Aβ (3D6 p < 0.05) but not tau, and only two clusters (EC cluster 7 and V1 cluster 0) had a significant negative correlation with Aβ (both 3D6 p values < 0.05) but not tau. Finally, ITG clusters 3 and 9 correlated positively and V1 cluster 0 correlated negatively with both Aβ and tau readouts (ITG cluster 3: p values see above; ITG cluster 9: 3D6 p < 2e-5, pTau231/Total Tau p < 0.001, HEK seeding p < 0.001, HT7 Aggregated Tau p < 5.53–5; V1 cluster 0: 3D6 p < 0.05). While several tau and tau & Aβ-associated clusters showed significant correlation with Gerrits et al. “AD1” tau & Aβ signature (e.g., Fig. S3a, ITG microglia clusters 3 and 9, p values < 2.22e-16 and < 2.3e-6, respectively), none of the tau-only associated clusters (e.g., EC cluster 4, ITG cluster 4, PFC cluster 6) showed a positive correlation with Gerrits et al. “AD2” tau-only signature, based on fold change comparisons against homeostatic microglia (Fig. S3a).

These data suggest that there are distinct transcriptomic responses of AD-associated microglia to tau vs. Aβ pathology as well as a signature common to both pathologies. To better understand the similarities and differences in homeostatic, tau and Aβ-associated clusters between regions, we mapped the homeostatic and pathology-associated clusters from individual regions to our cross-region clusters. Individual region microglia clusters showing a negative correlation with tau pathology (Fig. 3a, Fig. S3a, S3f, dashed boxes/ circles) aligned into cross-region clusters 0 and 3 (Fig. 3b, Fig. S3f). Moreover, AD pathology-associated EC cluster 4 and PFC cluster 6 aligned with cross-region cluster 4, while ITG clusters 3 and 4 mapped to cross-region clusters 5, 8, and 2, respectively (Fig. 3b, S3f). Genes characterizing these cross-region pathology-associated clusters 2, 4, 5, and 8 include the top regulated genes CD163 (a typical marker of brain macrophages) and RGS1, PTPRG, and CPM, respectively (Fig. S3g, Table S6, Fig. S3h, Tables S7), and pathways such as “CDC42 GTPase cycle”, and “binding and uptake of ligands by scavenger receptors” (Fig. S3i, Table S8). PvMs (CD163, LYVE1, MRC1, and F13A1) were mainly found in cross-region cluster 6, while cluster 10 was marked by increased expression of CCR2, suggesting that cells in this cluster constitute myeloid cells of a peripheral origin, e.g., monocytes (Fig. S3j).

Identification of shifts in microglia states from homeostatic to AD-associated

Next, we asked whether homeostatic microglia transition to an AD-associated state along the disease course. We first investigated shifts in microglial density from homeostatic to AD-associated along the accrual of pTau/Total tau, HT7 aggregated tau, and Aβ plaques. We binned these readouts into 5 classes and plotted the density of microglia in each pathology bin and cross-region cluster (Fig. 3c; 3 of 5 bins [no/early/late] shown for simplicity). Interestingly, of the cross-region AD-associated clusters 2, 4, 5, and 8, clusters 2 and 4 showed the highest density of microglia nuclei in the early and late tau bins, while clusters 5 and 8 showed a high density of microglia nuclei in the late Aβ bin (Fig. 3c, Table S9, Table S10), suggesting differential early vs. late responses of these microglial clusters to tau and Aβ pathologies. Notably, the tau progression-associated clusters 2 and 4 showed correlations with Gerrits et al. “AD1” and tau fibril-treated microglia (Fig. 3d, Gerrits et al. “AD1” p values < 0.01 and < 2e-5, respectively, and “Wang_Tau_Fibril” p values < 2.8e-6 and < 0.05, respectively). Furthermore, the Aβ-associated cross-region clusters 5 and 8 showed the strongest correlation with “AD1” signatures (rho = 0.8 and 0.7, respectively; p values < 2.22e-16 for both), positive correlations with laser capture microdissected plaques and peri-plaque signatures (“Das_LCM_Plaque”: rho = 0.61 and 0.55, with p < 8.3e-6 and < 1.9e-6, respectively, and “Das_LCM_Peri_Plaque”: rho = 0.54 and 0.56, with p < 0.05 and < 0.01, respectively) [9], and a negative correlation with “Mancuso_HM” (homeostatic microglia, rho =  – 0.48 and – 0.49, with both p values < 0.05), and cluster 5 additionally showed a positive correlation with “Mancuso_DAM” (p < 0.05) (Fig. 3d). Interestingly, the tau-associated cross-region clusters 2 and 4 did not show significant positive correlations with any of the Mancuso et al. signatures or the Gerrits et al. “AD2” tau signature (Fig. 2d), but cluster 4 did show positive correlations with “Das_LCM_Plaque”, “Das_LCM_Peri_Plaque”, and “Das_LCM_Plaque_vs_NFT” (p < 2.13–5, < 0.05 and p < 0.001, respectively). This suggests that the Gerrits et al.’s study, with limited brain regions, and the Mancuso et al.’s study, which used human iPSC-derived microglia transplanted into mouse brain, incompletely describe the microglia and macrophage signatures that we were able to detect in human AD across all 5 brain regions.

In silico modeling identifies “phasic” genes as potential regulators of microglia transition during disease

The shifts in proportion of homeostatic and AD-associated microglia clusters with increasing levels of pathology supported a transition from the former to the latter. To model human microglia transition along disease progression, we calculated trajectories and identified transitionally upregulated genes (‘phasic’ genes) in the conversion from the main homeostatic microglia cluster (cluster 0) to the AD-associated clusters identified in Fig. 3c (Fig. 4a). These suggest that human AD microglia can transition from a homeostatic (cluster 0, HOM) to either a ribosomal activation state (cluster 3, ribosomal response or RR) or AD-associated states that correlate with increases in AD pathology (clusters 4, 5) (Fig. 4b). Cluster 3 was marked by upregulation of ribosomal response-associated genes, while clusters 4 and 5 showed separate trajectories, and were designated as early Aβ/late tau (EALT), and late Aβ response (LAR), respectively, based on mapping to pathology readouts in Fig. 3c. Phasic genes from homeostatic (cluster 0) to AD-associated clusters 3, 4, and 5 included genes implicated in CDC24 GTPase cycle, RHO GTPase cycle, fibrin clot formation, IRAK1 recruitment of IKK complex (Fig. 4b, top, cluster 0 to 3), IL-4/IL-13 signaling, response of EIF2AK1 to heme deficiency, signaling by interleukins, IFNγ signaling (Fig. 4b, middle, cluster 0–4), and genes involved in axon guidance, fibrin clot formation, semaphorin interactions, and IL-4/IL-13 signaling (Fig. 4b, bottom, cluster 0 to 5). Of note, we also were able to identify trajectories from homeostatic microglia to PvM and monocyte clusters 6 and 10, respectively, indicating that microglia, PvMs and monocytes may exist on a continuum once these cells are localized within the brain parenchyma (Fig. S4a).

Fig. 4

Microglia subtype conversion in human. a Trajectories to disease-associated clusters were identified with monocle3 [50]. b For individual trajectories 0 (HOM)–3 (RR), 0–4 (EALT), and 0–5 (LAR), transitionally upregulated genes were identified by splitting pseudotime into quartiles and filtering for genes expressed at a higher level in middle quartiles (transitionally higher expressed) compared to 1st and 4th one, the latter representing cluster-enriched expression in HOM, or disease-associated cluster, respectively. c Expression of top 30 genes per microglial phenotype (HOM, DAM1, DAM2, EADAM, LADAM) identified in [28], averaged per cluster in cross-region integrated brain myeloid cells. Red indicates high average expression levels; blue indicates low average expression levels. Clear upregulation of HOM (cluster 0) and DAM1 (cluster 3) phenotypes are observed, as well as enriched expression of LADAM and EADAM genes across clusters 4, 5, and 10, indicated by black boxes. Clusters 6 and 10 show strong relative downregulation of homeostatic microglia markers, as indicated by blue boxes. d,e Volcano plots showing genes differentially expressed between cluster 4 and 10 (d), and cluster 5 and 10 (e)

Comparison with prior mouse single-cell transcriptomics studies highlights differences between microglial responses in human and mice

To determine whether microglia AD progression signatures are shared between human disease and mouse models, we compared our cross-region brain myeloid dataset to several previously reported disease-associated microglia (DAM) mouse signatures. Prior mouse scRNA-seq studies have identified a Trem2-dependent and a subsequent Trem2-independent stages of AD progression in the 5xFAD model, termed DAM1 and DAM2, respectively [27], as well as two additional DAM phenotypes, early DAM (EADAM), increased in dual Aβ and tau pathology mice, compared to single pathology mice, and late DAM (LADAM) [28]. We observed high expression of mouse homeostatic genes in our cluster 0 and of DAM1 genes in our cluster 3, and moderate expression of DAM2 genes across clusters with some expression in our clusters 5 and 8 (Fig. 4e, Fig. S4b, black boxes). Interestingly, in our data, DAM1-like cluster 3 did not precede later stage DAM2-like clusters in pseudotime (i.e., the positioning of cells along the trajectory that quantifies the relative progression of the underlying biological process), suggesting a different pathology-associated microglia transcriptional program in mouse vs. human. Remarkably, our cross-region cluster 3 contains individual region clusters with significant negative correlation with AD pathology (ITG cluster 1, PFC cluster 2), yet showed the strongest DAM1 signature (mainly ribosomal response-associated genes), suggesting that these microglia disappear with increasing pathology in human disease, which contrasts with microglial DAM1 phenotype observations in mouse models. Furthermore, expressions of EADAM and LADAM genes were not clearly delineated across human myeloid clusters, with clusters 4, 5, and 10 showing both EADAM and LADAM gene upregulation (Fig. 4c, black boxes), suggesting that this mouse early pathology-specific response may not be clearly identifiable within human donors even when analyzing multiple brain regions across a spectrum of pathology severity.

We further observed strong downregulation of mouse homeostatic microglia markers in clusters 6 and 10 (Fig. 4c, blue boxes). While cluster 6 corresponded to macrophages identified in our individual region analysis, characterized by increased expression of LYVE1, MRC1, F13A1, and CD163, cluster 10 showed increased expression of the peripheral monocyte marker CCR2 [37]. Recent studies have identified microglia/macrophage-like cells expressing both the microglial marker TMEM119 and the macrophage marker CD163 surrounding Aβ plaques in human AD brains but not in control brains [40, 44]. Although microglia clusters 4 and 5 and monocyte cluster 10 were associated with EADAM and LADAM mouse microglia signatures, clusters 4 and 5 showed comparatively higher expression of canonical microglia genes P2RY12 and TMEM119, and of the “AD1” gene SPP1 (Fig. 4d, e, Table S11). Cluster 6, the PvM cluster, showed elevated F13A1, MRC1, LYVE1, and CD163, macrophage marker expression, and increased P2RY12 as compared to monocyte cluster 10 (Fig. S4c).

In summary, some aspects of microglia transcriptomic responses to AD pathology are shared between human and mouse models, but not others.

Pseudobulk analysis reveals genes impacted by tau, Aβ, or both, and confirms early tau dysregulation of the transcriptional regulators BACH1 and PRR5

While analysis at the single-cell level is a powerful tool to characterize microglial phenotypes based on cell-to-cell variation, single cells from the same tissue sample cannot be considered truly independent sample. Leveraging the cohort size, we were interested in expanding our analysis to also identify disease-associated genes at a population level. To confirm tau vs. tau & Aβ driven changes in late-stage AD brain myeloid cells, we compared pseudobulk gene expression in high-tau/low Aβ vs. low-tau/low-Aβ (EC vs. V1, tau-driven), high-tau/high-Aβ vs. high-tau/low-Aβ (PFC vs. EC, Aβ-driven), and high-tau/high-Aβ vs. low-tau/low-Aβ (PFC vs. V1, tau & Aβ-driven) regions within high pathology group 4 donors, controlling for regional changes observed in low-pathology group 1 donors (Fig. 5a, Table S12). Tau-driven changes included interferon-related genes (IFITM10, IFI44L, and IFG20), which were also increased in tau-associated clusters from our single-cell level analysis (e.g., IFI44L in cross-region tau-associated cluster 2 vs. cluster 0), as well as the previously “AD2” identified gene GRID2. Tau & Aβ-driven changes included genes related to cytokine (TNFRSF21, TGFBI) signaling. Clustering of genes across regions per pathology group demonstrated that the majority of pathology group 3 and 4 genes spike in PFC, suggesting mainly Aβ-influenced signatures later in disease progression. These included pathology group 3 gene cluster 2, represented by pathways such as “antigen processing: ubiquitination & proteasome degradation” (Fig. 5b, Table S13, S5a, Table S14). On the other hand, pathology groups 1 and 2 had gene clusters following tau progression (high EC–low V1 expression or low EC–high V1 expression), e.g., pathology group 1 gene cluster 4, mainly corresponding to pathology group 2 gene cluster 2 and represented by pathways such as “extracellular matrix organization” (Fig. 5b, Fig. S5a). Correlation of each gene cluster with biochemical readouts indicated overall highest correlation in pathology group 4, across gene clusters (Fig. 5c). Additionally, pathology group 2, gene cluster 3 showed positive correlations with all pathology readouts, while pathology group 3 gene cluster 2 showed positive correlations with Aβ but not tau, as expected based on gene expression patterns of EC > ITG > PFC > V2/V1 and EC < ITG < PFC > V2/V1, respectively. We further sought to identify genes that showed opposite expression patterns in early disease stages (path group 1 compared to path group 2), indicating early tau-associated dysregulation, which included the transcriptional regulators BACH1 and PRR5 (Fig. S5b, Table S15). Notably, genes previously identified as transitionally upregulated in the conversion from cluster 0 (HOM) to 4 (EALT) showed significant overlap (p value 1.74e-5) with genes showing early tau pathology-driven dysregulation, e.g., BACH1 and PRR5, thus supporting the validity of our trajectory results.

Fig. 5

Tau- and Aβ-associated brain myeloid cell signatures. a Heatmap of differentially expressed genes (DEGs) up- (⇧) or down- (⇩) regulated in EC vs. V1, PFC vs. EC and PFC vs. V1 regions, within pathology group 4. Tau-driven changes (EC vs. V1) include interferon-related genes, while Aβ driven and tau & Aβ-driven changes (PFC vs. EC, PFC vs. V1) include growth factor, and cytokine signaling related genes, respectively. Results were adjusted for gender and respective pathology group 1 DEGs were filtered out. Color-coding of aggregated expression per sample (column) and gene (row), annotation shows pathology group, 3D6 Aβ IHC, pTau/Total Tau, HEK seeding, and HT7 Aggregated Tau. Filtering for DEGs based on nominal p value < 0.01 and logFC > 1.2. Expression patterns included in respective comparisons are indicated by black boxes; expression patterns of other regions are shown for completeness. b K-means gene clustering across regions, per pathology group. Gene numbers are color-coded. Sankey diagrams, color-coded according to pathology group, show percentage change of genes from given gene clusters in one pathology group to gene clusters in next pathology group. Pathology group 3 and 4 gene clusters spike in PFC region, suggesting Aβ influenced expression, while pathology group 1 and 2 contain also gene clusters showing linear correlation along regions. c Spearman correlation per pathology group of each gene cluster with biochemical readouts; overall highest correlation is observed in pathology group 4, across gene clusters. High correlation is indicated by red and low correlation by blue color