Spatial proteomics reveals human microglial states shaped by anatomy and neuropathology

Microglia are implicated in aging, neurodegeneration, and Alzheimer’s disease (AD). Traditional, low-plex, imaging methods fall short of capturing in situ cellular states and interactions in the human brain. We utilized Multiplexed Ion Beam Imaging (MIBI) and data-driven analysis to spatially map proteomic cellular states and niches in healthy human brain, identifying a spectrum of microglial profiles, called the microglial state continuum (MSC). The MSC ranged from senescent-like to active proteomic states that were skewed across large brain regions and compartmentalized locally according to their immediate microenvironment. While more active microglial states were proximal to amyloid plaques, globally, microglia significantly shifted towards a, presumably, dysfunctional low MSC in the AD hippocampus, as confirmed in an independent cohort (n=26). This provides an in situ single cell framework for mapping human microglial states along a continuous, shifting existence that is differentially enriched between healthy brain regions and disease, reinforcing differential microglial functions overall.


Introduc$on
Microglia, the resident macrophages of the brain, are highly dynamic cells that adapt to their ever-changing environment and perform a vast array of funcSons spanning brain development, adulthood, aging, and neurodegeneraSve and neuroinflammatory diseases 1 . In aging and disease microglia can assume hypo-or hyperacSve states that fail to perform their funcSons or become pathogenic and aberrantly engulf synapses, thus contribuSng to cogniSve decline [2][3][4] . Corresponding to their diverse roles, microglia can assume a wide range of phenotypes across contexts and brain regions. Given the relaSve ease of targeSng microglia 5 , a reprogrammable, renewable and replaceable cell type versus neurons themselves, their cellular state diversity is key to idenSfying their spectrum of funcSon. Historically, microglia were defined by their morphology and later by expression of a few immune markers that defined a "bad" (M1) or a "good" (M2) state 6 . In recent years these definiSons have expanded enormously. A myriad of single-cell proteomic [7][8][9] and transcriptomic phenotypes of microglia have been categorized and named including in disease: DAMs 10 , MGdN 11 , GAMs 12 , ARMs and IRMs 13 , HAMs 14 , MIMs 15 , LDAMs 16 , and in development and aging: WAMs 17 , ATMs 18 and PAMs 19 . While they have unique names, these cell state descripSons can overlap significantly, demonstraSng the plasScity of microglial funcSons with clearly discriminaSng phenotypes yet to be determined. At the same Sme, it is accepted that microglia exhibit variability between healthy brain regions and even between layers of neurons within a region [20][21][22][23][24][25][26] . ParScularly in the human, the binning of microglia into disparate contextual subtypes, even when captured from different brain regions under different contexts, has led to confusion in the field, and a more unifying approach to describing these cellular states is required 27 .
ComputaSonal tools typically used to separate microglial states include clustering algorithms that were developed to idenSfy discrete cell lineages 7,28 . Much as the case for myeloid cells in the periphery, these tools fail to capture the range of overlapping microglial phenotypes presumably linked to funcSonal idenSty. This is further compounded by most studies of microglia being performed in the mouse where there is growing evidence that their diversity does not overlap with that of humans 29 . ExisSng human brain samples have been limited in the anatomical regions analyzed, cell number acquired, or methods targeSng nuclear rather than cytoplasmic transcripts [30][31][32] . Analysis of single-cell suspensions leaves spaSal informaSon behind as they require extracSon of cells instead of in situ analysis within the microenvironment which further informs funcSon. SpaSal transcriptomics 33 a n d immunohistochemistry studies 34,35 of microglia to date have been restricted by relaSvely liqle human data, low resoluSon, low mulSplexing capabiliSes, and very few cells with the purpose to validate specific genes or proteins of interest found through scRNA-seq or CyTOF 12,36 , and have largely lacked quanStaSve metrics of markers and mulS-cellular niche analysis.
Here we invesSgate human microglial phenotypic diversity at a quanStaSve, spaSal proteomic, and morphological level within different regions of the human brain in post-mortem formalin-fixed paraffin embedded (FFPE) Sssue from individuals enrolled in a research study who were repeatedly and carefully annotated by neuropsychological and clinical assessments, and ulSmately by comprehensive neuropathologic evaluaSon. We made use of a single-cell, spaSal proteomic imaging tool called mulSplexed ion beam imaging (MIBI) 37,38 to capture microglia within intricate mulS-cellular microenvironments and pathologic niches of the brain. MIBI has been used previously to study various archival human Sssue types from tumors [39][40][41] to placenta 42 and granulomas 43 in infecSous disease and, recently, the brain 44 . With this we created a 41-plex imaging analysis to capture major brain resident cell types, and importantly, simultaneously quanSfy the expression of 17 microglial immunophenotyping proteins. From our iniSal imaging study we extracted 203,717 myeloid cells -199,204 were microglia, 430 were monocytes and 4,080 were vessel-associated macrophages (BAMs 7 ) -across 5 brain regions from individuals with healthy brains. Instead of organizing cells into clusters, we recognized a conSnuous progression of cell phenotypes that we defined here as the microglial state conSnuum (MSC). The distribuSon of microglia on the MSC was anatomically restricted between healthy brain regions, and conversely, compartmentalized within local microenvironments. ContrasSng the healthy brain profiles to Alzheimer's disease (AD) hippocampus, we idenSfied sparse DAM phenotypes enriched around amyloid plaques, and more significantly, revealed that all hippocampal microglia were polarized towards a less acSve MSC compared to control Sssue. This shij in the AD related MSC was validated using an independent analysis of hippocampal Sssue from individuals exhibiSng a spectrum normal cogniSon and AD-related cogniSve decline. Using our high parameter data as a guide, we used the combinatorial expression of HLA-DR and Iba1 to recapitulate the high dimensional MSC on an addiSonal ~60,000 microglia sampled from 26 individuals. Here, we showed a significant shij in AD as well as an overall correlaSon of MSC progression with cogniSve decline. Taken together, this study is the first highly mulSplexed, quanStaSve, spaSal proteomic analysis of microglial phenotypes in the human brain, creaSng a new, evidenced-based definiSon of a conSnuum of cell states that varies between brain regions and is skewed towards an impaired state in neurodegeneraSve disease.

Mul$plexed spa$al proteomics organizes macro-and microenvironments in the human brain.
Most human brain samples exist in archival FFPE blocks. Immunohistochemistry (IHC) of archival FFPE Sssue is challenging when mulSplexing is introduced. Fluorescent IHC, parScularly the brain, is confounded by high background autofluorescence that becomes even more problemaSc with mulSplexed probes. In order to reliably study complex microglial states in situ across the human brain, we implemented MIBI technology that uses metal labelled probes and does not have the same constraints as fluorescent or dichromaSc IHC. We started by targeSng five brain regions from one healthy aged donor (90-year-old male without demenSa, APOE ε3/ ε3, and a post-mortem interval of 3.4 hours), including the hippocampus (HIP), cerebellum, substanSa nigra (SN), caudate nucleus, and middle frontal gyrus (MFG) (Figure 1A, leB). SecSons were stained with a panel of 38 metal-tagged anSbodies (Extended Data Table S1) capturing neuronal structures including soma, dendrites, axons, synapses, astrocytes, myelin, vasculature, cellular processes including metabolism, proteolysis and translaSon, the hallmark AD neuropathologic changes (ADNC) including amyloid-β (Aβ) plaques and tau tangles, and a sub panel of 17 microglial phenotyping proteins. AddiSonally, we stained nucleic acids (DNA & RNA) using free indium 115 3+ ions 45 and captured endogenous iron (Fe), an important marker of brain health 46 and microglial state 47 . Areas outlined in black within each Hematoxylin-Eosin (H&E) and Luxol Fast Blue (LFB) stained secSon were imaged by MIBI by Sling mulSple images of 700 µm x 700 µm each, including larges area of mostly grey maqer (labelled "Grey") as well as smaller areas within the deep white maqer (labelled "White"). The data analysis was a mulSpronged approach including quanSficaSon of protein expression paqerns in large demarcated sub-regions, segmentaSon of microglial cells, in silico s o r Sng of morphological subtypes, phenotyping of microglial single cells and evaluaSng changes across large brain regions, pixel clustering of local microenvironments and spaSal analysis of niche enrichment of microglial states, and comparison to an AD demenSa case (Figure 1A, right).
Approximately 6.5 mm x 5.5 mm of the hippocampus was imaged at ~470 nm pixel resoluSon ( Figure 1B). Exemplary sStched overlays of the hippocampus show markers VGLUT1 (excitatory synapSc density), MAP2 (neuronal dendrites and soma), MCNPase (myelin), GFAP (astrocytes), DNA & RNA (nuclei and neuronal soma as stained by free indium 115 3+ metal ions) and CD31 & CD105 (vasculature). The boxed FOV1 in larger format overlay ( Figure 1C) of the same image highlights the large areas of synapSc density, myelin, astrocytes as well as the fine structures of dendrites, soma of the hippocampal pyramidal neurons and granule cells of the dentate gyrus (DG), nuclei and vasculature. Next, in order to assess the biological validity of pan-brain proteomic targets, separate from those used for microglial phenotyping, we correlated expression at the pixel level within the enSre hippocampus ( Figure 1D). As anScipated, nuclear markers correlated highest with other nuclear markers (HH3, Indium115 3+ , 8-OH-Guanosine (a marker of DNA & RNA damage), p-S6 (a surrogate marker for mTOR pathway acSvity), see image panel (D i)), astrocyte markers correlated posiSvely with other astrocyte markers (GFAP, S100β, CD44, (D ii)) although not highly since these proteins are expressed in different sub-cellular locaSons of astrocytes, and VGLUT1 correlated well with other widely dispersed "carpet" markers like CD56 (aka NCAM1, expressed by neurons and glia) and VDAC1 (a mitochondrial protein) and anS-correlated with nuclear markers. ApoE is produced primarily by astrocytes and microglia and by some neurons in response to injury, and contributes to integrity of the blood brain barrier (BBB) 48 . In our images, ApoE was widely dispersed and concentrated at blood vessels (iii). Axonal proteins, neurofilament (NF) high (H) and low (L) molecular mass, correlated posiSvely with each other, NFH with MAP2 as well and NFL with MCNpase (iv), although only in specific sub-regions of the hippocampus while in other subregions the proteins were not co-expressed. Fe correlated highly with the myelin enzyme MCNPase, also in specific sub-regions (v). PolyUK48, a protein upregulated during the proteolysis, correlated posiSvely with CPT2 (carniSne palmitoyltransferase 2) and Iba1 corelated posiSvely with Fe and nuclear markers. Overall, our pan-brain markers correlated with biologically expected sets of proteins and built confidence that our method for interpreSng the brain through MIBI is sound and not confounded by technical arSfacts.
To further analyze the regional distribuSon of these markers in the hippocampus, given its familiar differenSal protein expression across its sub anatomy, we demarcated the subregions based on well documented neuropathological annotaSons including the subiculum, cornu ammonis (CA1-4), DG, outer molecular layer, stratum lacunosum moleculare, stratum radiatum, alveus and glia limitans ( Figure 1F). The density of expression of brain and microglial immunophenotyping proteins in these local sub-regions was quanSfied as average pixel expression per mm 2 , with the inner regions of the hippocampus highlighted. The subiculum and CA1-4 were enriched for neuronal and synapSc markers like VGLUT1, NFH, MAP2 and CD56, VDAC1 and ApoE. InteresSngly, CA2, a subregion of hippocampus relaSvely resistant to AD-type changes, had a high level of CPT2. The DG is compacted with granule cell neurons, as evidenced by the high expression of HH3, 8-OH-Guanosine and the binding of indium115 3+ . Moving from the DG through the outer molecular layer, there is a disSnct diminishing of nuclear markers and pan-brain marker expression. The stratum lacunosum moleculare was defined by a high density of astrocyte soma and blood vessels as seen by the strong expression of S100β and vascular markers CD31 & CD105. While the stratum radiatum that harbors radiaSng dendrites from the CA1 area displayed the highest expression of MAP2. The alveus which is compacted with axons is highlighted by dense MCNpase protein. Finally, the glia limitants, the astrocySc barrier covering the brain, is enriched in GFAP and CD44 as well as pS6 expression. InteresSngly, microglial phenotyping marker density was lower in CA1 compared to neuronal and synapSc markers. DifferenSal expression of microglial markers was observed between subiculum (P2RY12, Tmem119, TREM2), stratum lacunosum moleculare (Iba1, CD68, GPNMB, FerriSn-L, CD45, HLA-DR, CD11c, MRP14, MerTK) and alveus (CD74, TSPO). The DG and glia limitans had the highest expression of CD16 and CD14, respecSvely.
Beyond our ability to capture expected neurobiology and co-localizaSon, this mulSplexed, quanStaSve spaSal proteomic analysis within hippocampal subregions uniquely captures compartmentalizaSon of different neuronal, structural, and glial features including differenSal densiSes of microglial phenotyping proteins across this important brain region. Further to this, spaSal proteomic analysis in the MFG, cerebellum, SN, and caudate (Figure S1A-D) was able to obtain similar correlaSons of subregional compartmentalizaSon to quanSfied features.

Quan$fying proteomic expression profiles on microglia with spa$al-anatomical context.
With the faithful quanSficaSon of mulSplexed protein expression in different neuroanatomical spaces and subregions at a pixel and bulk level, we sought to idenSfy discrete microglial cells and define their distribuSon within the different subregions. To this end, we used classical microglial markers Iba1 and CD45 and, leveraging the high resoluSon of our images, we were able to capture microglial cells and parts of cells. Figure 2A shows a single 700 µm x 700 µm image from the hippocampus with larger format in Figure 2B highlighSng a single microglial cell.
In order to phenotype microglia at a single cell level, we used an intensity thresholdingbased segmentaSon approach as a part of the MIBI image processing toolkit (see methods). Microglia were defined as Iba1 + and/or CD45 + and a composite channel mask was created with which a total of 203,714 microglia (with BAMs and monocytes) were extracted from our images and pre-processed for single cell analysis (Figure 2C, S2A). Microglial cell extracSon largely excluded pixels from other glia, for example astrocytes (Figure 2D), although these pixels were present to some extent as cells appear to overlap in the collapsed Z plane. Microglial cell density across all the large brain regions was quanSfied as the number of cells per mm 2 per FOV, within grey and deep white maqers ( Figure 2E, F, S2B) which showed the highest density of microglia in the hippocampus, parScularly in the grey maqer (here defined as the enSre area shown in Figure 2E, including the white maqer tracts, and separate from the deep white maqer areas imaged separately); the cerebellum had the lowest microglial density and the SN, MFG and Caudate showed intermediate expression. Striking differences in microglial density between the hippocampus and cerebellum is beqer appreciated in the overlaid images shown in Figure  2E, where each dot represents a microglial cell colored by its local density in a 4.9 µm x 4.9 µm area. Notably, microglial density was parScularly compartmentalized to sub-regions within the hippocampus and cerebellum, both allocorScal areas (Figure 2E, S2B). For instance, the microglial density in the cerebellum was highest in the local myelin layer, corresponding to the high pixel density of glial markers in our annotated sub-region analysis of the cerebellum ( Figure  S1B), and lowest in the molecular layer and granular where there are relaSvely sparse VGLUT1 pos synapses and dense HH3 pos CD56 pos neuronal soma ( Figure S1B). In the hippocampus, microglial density was highest in CA4 and CA2, where MAP2 pos axons are sparse and synapses are very dense, and lowest in CA1, outer molecular layer, and subiculum where MAP2 pos axons are dense, and synapses are sparser. Microglial density was also relaSvely high in the hippocampal myelin and astrocyte rich areas, as in the cerebellum. These paqerns of microglial density being enriched in areas of dense synapses and dense myelin are in accordance with their roles in synapSc pruning 49 and maintaining myelin integrity 50,51 in concert with astrocytes.
InteresSngly, expression of our microglial phenotyping proteins within our segmented cells varied and was localized to different intracellular niches within microglia ( Figure 2G). For instance, where Iba1 and CD45 were observed within the enSre cytoplasm of the cell, Tmem119 (considered a bona fide microglial marker in mice 52 ) was only present in some areas of the cell, leaving whole segments unmarked. P2RY12, also an exclusively microglial marker in the brain, had a more homogenous intracellular distribuSon but its expression was generally low. HLA-DR and CD11c were highly expressed throughout the cell, while MRP14 and GPNMB tended to be found in bright, dense intracellular spots, as were MerTK and CD74. CD44 was expressed within microglia and outside of microglia in astrocytes. The product of free radical damage to nucleic acid, 8-OH-Guanosine, was also not restricted to microglia. Endogenous Fe was present in large areas that we imaged, parScularly in the hippocampus and SN. Amongst other cell types, Fe was localized to microglia and ojen co-localized with FerriSn light chain (FerriSn-L, Figure 2H). CorrelaSon of these features across all microglia extracted from all brain regions on a cell-level revealed programs of co-expression ( Figure 2I). Iba1, however, was not highly correlated with any other marker in the panel; in fact, it was negaSvely correlated with CD45 and CD14 and only slightly posiSvely correlated with Tmem119. The two "homeostaSc" markers for microglia, P2RY12 and Tmem119, correlated most highly with each other. Fe had a moderately high correlaSon with FerriSn-L, consistent with it being a marker of iron storage in microglia 53 . Amongst the posiSvely correlated set of markers, CD16 (also known as FcγRIII), CD45, CD11c and MerTK formed a subset, and HLA-DR, GPNMB, CD68, MRP14 (all implicated in phagocytosis), CD14 and FerriSn-L formed another subset. MerTK, also a protein involved in phagocytosis, was promiscuously correlated with both subsets. TSPO, a marker rouSnely used in PET imaging to evaluate microglial acSvaSon in people and animal models which has been shown rather to reflect microglial density 54,55 , was not highly correlated with any other proteins in this panel; and polyUK48 even less so. Taken together, MIBI imaging of microglia across the human brain allows for robust, mulSplexed, single cell phenotyping and local cellular density analysis where the quanStaSve nature of the data allows us to reveal dispariSes in microglial diversity not previously realized.

Classifica$on of microglia based on mul$plexed morphological features separates nucleated cells and fragments, but not anatomical loca$on
In the absence of disSnct markers for microglial states, morphology has been historically used in an aqempt to classify these cells as resSng/homeostaSc (ramified with long, fine processes), acSvated (rounded, amoeboid, large, thickened), dystrophic (fragmented, distorted) and many states in between 56,57 . Now that we were able to segment microglia and assess their intracellular and intercellular protein expression, we set out to define their morphological states and determine whether this morphological classificaSon could inform protein expression, or vice versa. We extracted a set of morphological features from our 2D segmented microglia ( Figure 3A), including size (area, perimeter), circularity, elongaSon (major and minor axis length, aspect raSo), cell border undulaSon (perimeter to area raSo (P/A raSo); number concaviSes), and symmetry (radial asymmetry, solidity), as well as the presence of a nucleus in the plane of view. Many morphological features were highly correlated with each other (posiSvely and negaSvely) (Figure S3A), shown in representaSve cells at low, medium and high values for each feature ( Figure 3C). Cells tended to be similar in size, many had zero concaviSes, others ranged from one to seven concaviSes; circularity was also widely distributed; HH3 nuclear signal had three major peaks likely corresponding to how much nucleus was in the plane ( Figure 3B).
Cell objects that did not have any HH3 signal and were thus anucleate, were likely pieces of cytoplasm. In order to separate very small cell objects that were likely parts of microglial processes from larger cellular areas, we postulated that transversely secSoned processes would be small, round and anucleate. Indeed, size and circularity were negaSvely correlated. However, circularity was not highly correlated with HH3 and there were four disSnct populaSons of HH3 signal and circularity (Figure 3D), suggesSng that we captured parts of processes (PR), irregularly shaped anucleate cell objects, rounded nucleate cells with their cell body (CB) in the plane, and irregularly shaped nucleate cells ( Figure 3E). All cell objects had Iba1 and/or CD45 expression throughout their cytoplasm but some markers, like CD74 and HLA-DR, were very low in small round processes. We clustered our microglia by morphological features using FlowSOM 58 and performed manual meta-clustering into 4 main morpho clusters ( Figure 3F, G): (1) anucleate, small, rounded and transverse processes high in circularity and solidity; (2) anucleate, branched cells that were larger and had some concavity; (3) nucleate, rounded cells; and (4) nucleate, large, even more branched cells. The four 'morpho' clusters differed in protein marker expression for a few of our phenotyping proteins, where parScularly the anucleate small, rounded processes and, to a lesser extent the anucleate branched cells, had lower expression of CD11c, CD74, MerTK, TSPO and GPNMB, and HLA-DR ( Figure 3H, S3B). Since we had seen that an overlapping set of proteins with these tended to be found in very localized intracellular niches (Figure 2C), when the cell body of a given microglia was largely beyond our rastered plane we therefore would not be able to capture that protein reliably.
We also calculated frequency of morpho clusters per FOV acquired in each brain region (grey and white maqer) but found no significant enrichment across brain regions ( Figure 3I). This seems to indicate that healthy brain Sssue is relaSvely uniform in its microglial morphological composiSon, and that the variance we see here is a combinaSon of normal biology with the technical limitaSon of our shallow (< 1 µm) imaging plane. Furthermore, our analysis of microglial morphological features allowed us to in silico sort cells and cell parts to select reliable morpho clusters for protein detecSon. Therefore, we excluded morpho cluster 1small, rounded, transverse processes -from subsequent cellular phenotyping.

Microglia align along a con$nuum of phenotypic cell states which are enriched in specific brain regions.
Phenotyping microglia through clustering methods has been employed across disease contexts in transcriptomic and proteomic data, resulSng in the idenSficaSon of someSmes disSnct populaSons in murine models 7,10 but ojen, especially in human studies, seemingly arbitrary boundaries are drawn between clusters that are very similar to each other in the high dimensional space 9,12,31 . We sought to organize these microglial states based on protein expression across brain regions without forcing cells into staSc, somewhat arbitrary, clusters. Excluding small cell pieces by excluding morpho cluster 1 ( Figure S4A) lej 93,679 cells. We removed sparse monocytes found in blood vessels that would otherwise contaminate our analysis. We idenSfied intravascular monocytes as MRP14 high , round, nucleate cells found close to CD31 + & CD105 + blood vessels ( Figure S4B) which we gated out, leaving a total of 93,325 microglia from all Sssues for analysis.
In our iniSal aqempt to organize these cells, we used FlowSOM to idenSfy 12 clusters of microglia. However, while different in magnitude, these clusters had similar protein expression profiles overall, resembling a conSnuum of states, where cluster 1 had overall low expression and cluster 12 had overall high expression across many proteins associated with microglial funcSon ( Figure S4C). Most cells landed in the mid-expression clusters 5, 2 and 10 ( Figure S4D). Further, dimensionality reducSon showed a progression of clusters from 1 to 12 in embedded UMAP space, again resembling a conSnuum of states rather than disSnct populaSons ( Figure  S4E). As the boarders of cell clusters seemed inseparable in high dimensional space, there was concern it would create hazard in mapping meaningful changes between clusters and brain regions or disease.
As such, we sought to embrace the conSnuous nature of these microglial states for quanSficaSon and comparison rather than binning into disSnct populaSons. To this end, we employed SCORPIUS 59 , a computaSonal method for trajectory modeling to generate a pseudoSme calculated using protein features only which aligned our cells along a trajectory we termed the microglial state conSnuum (MSC) ( Figure 4A). The MSC followed the original cluster progression along the UMAP embedding. Since the hippocampus region had the largest imaged area, and therefore greatest number of cells ( Figure S4F), we wanted to confirm that the MSC was not skewed by this region and that cell states from less abundant regions were sSll captured. Therefore, we compared single, separately calculated, brain-region MSCs to that calculated form combined regions and found that they had strong agreement with each other ( Figure S4G). To further assess whether subsets of the MSC calculated with all cells is biased towards a specific brain region, we generated a simple linear regression model of the MSC using a 70-30 train-test split and compared the test error (squared residuals) of cells in each region along the MSC. There was no region-specific difference in error for any region along the MSC, and the smoothed error lines for each region were overlapping ( Figure S4H). From this we concluded that our approach of calculaSng an MSC trajectory as a quanStaSve summary of cell idenSty was a robust approach to capture and compare high dimensional microglial phenotypes from both abundant and sparsely distributed Sssues.
InterrogaSng protein expression, which was normalized to extracted cell size ( Figure 4B, S4J), along the MSC showed that expression of many molecules increased from low to high levels as the MSC progressed, parScularly HLA-DR, MerTK, CD11c, MRP14, TREM2, etc. Other proteins like P2RY12 showed parabolic expression, high in the middle and low in the ends of the MSC. Iba1, on the other hand, fluctuated more than other markers, decreasing with increasing MSC. RepresentaSve overlay of marker expression is shown for individual cells within low, middle and high MSC ( Figure 4C). Overall, low MSC cells appeared to be expressing less protein and late MSC cells appeared to be expressing more protein overall, suggesSng that low MSC represents a quiescent, potenSally senescent, cell state while middle to high MSC represents more acSve, but not necessarily acSvated, cellular states.
We also assessed MSC calculaSon with morphological features (MSC-morph). The MSCmorph progression ( Figure S4I-J) failed to organize proteomic features and resulted in high variaSon making the correlaSon to the clear proteomic spectrum low ( Figure S4K). This discrepancy shows that morphology alone do not organize microglia into a unified spectrum of cell states. One caveat to this is the shallow depth of our imaging plane (< 1 µm), thus we may not have sufficient cell area alone to capture their morphology robustly. While the MSC we moved forward with was calculated using only the protein features of cells, morphological features were somewhat enriched where the low MSC cells were on average small with fewer concaviSes and potenSally dystrophic, and late MSC cells were larger with more concaviSes and potenSally more acSve cells ( Figure 4D, S4J), aligning with our interpretaSon of the proteomic expression features overall.
Next, we sought to reveal whether microglia differed in their cell states along the MSC between different brain regions. Cell density along the MSC was generally highest towards the middle of the trajectory and lowest at the extremes ( Figure 4E). Microglia from the MFG and caudate were skewed towards the low MSC, suggesSng less acSvaSon for microglia in those brain regions. Cerebellar microglia had the highest enrichment in the middle compared to other brain regions, with fewer highly acSvated or highly inacSvate microglia. The hippocampus and SN were skewed towards the high MSC, suggesSng that microglia in those regions are on average more acSve. Given these brain region-specific differences in cell distribuSons, we next asked whether there were significant differences between the protein markers along the MSC of cells from different brain regions. To this end, we designed a non-parametric method for calculaSng differences between trajectories from single-cell data which we termed Pairwise Comparison of Trajectories by Binned PermutaSons (PCTBP, Figure S4J, see methods). To show these differences, we summed the number of significant feature comparisons for each brain region pair in equally divided low, middle, and high MSC phases ( Figure 4F). The hippocampus and MFG had the most differences between features while the caudate and SN had the fewest. Feature differences between brain regions were concentrated in the middle MSC with fewer differences in the low and high MSC suggesSng that the phenotypic state of intermediate acSvaSon of microglia was most highly variable among brain regions. InteresSngly, even though the hippocampus and SN had a similar distribuSon of cell density along the MSC, they sSll had the second highest number of phenotypic feature differences, with a large fracSon of those being in high MSC, suggesSng that while hippocampal and SN microglia may be skewed towards a more acSve state, this state is not necessarily of the same high dimensional phenotype. The top five variable features between brain regions were protein markers ( Figure 4H), namely VDAC1, APOE, CD44, Iba1 and P2RY12, and are shown individually as z-scaled progressions along the MSC (Figure 4G, top) as well as their PCTBP significance scores for each brain region pair ( Figure 4G, bo1om), and all unscaled inter-regional levels of protein and morphological features along the MSC in order of staSsScal significance ( Figure S4J). Overall, protein features were more varied between brain regions while morphological features were very similar for all brain region pairs across the MSC (Figure 4H). Since microglia from white and grey maqer perform different funcSons and have been shown to assume different states 36 , we also compared the MSC from areas of exclusively white maqer vs. grey maqer that were in the same Sssue blocks as the regions of interest and found that, again, only protein features varied between regions, including APOE, CD68, Iba1, polyUK48, Tmem119 and VDAC1 (Figure 4I, S4K). Grey maqer microglia had higher APOE, Tmem119 and VDAC1 expression while white maqer microglia had higher CD68 and poyUK48 expression, proteins linked to phagocytosis and protein degradaSon, respecSvely.
Overall, phenotypic differences of microglia between brain regions and white or grey maqer along the MSC were subtle but significant. Our data allowed us to not only detect marker expression but quanSfy relaSve expression between cells and thus revealed changes across many proteins that may have gone unnoSced by more binary methods. The differences between low and high MSC microglia may not be picked up by standard clustering methods since all cellular states between low and high MSC are also present. We observe that a conSnuum approach to assessing microglial states is key to understanding microglial heterogeneity within the human brain where a spectrum of microglial states exists together.

Microglial states are compartmentalized within local microenvironments of larger brain regions.
To appreciate the distribuSon and cellular niches of the microglial phenotypic states, we next spaSally mapped them back onto Sssue. Brain textures surrounding microglia are composed of astrocytes, neurons, synapses, and vasculature and form subregions within larger brain regions. Due to the highly irregular shapes of these features, lack of robust membrane markers and seemingly covering large unbounded spaces, they are difficult to segment using standard nuclear expansion segmentaSon approaches used in immune Sssue. Thus, to overcome these limitaSons and uncover common signatures of microenvironments across the brain, we used unbiased, automated pixel clustering 60 across images from brain regions (including grey and white maqer) and idenSfied 20 pixel clusters corresponding to unique textures in the Sssue using neuropathology lineage features ( Figure 5A). These pixel clusters included regions dominated by proteins that were expressed widely in the extracellular space, with intermediate to high excitatory synapSc density (cluster 1: CD56 high VGLUT1 pos , cluster 2: CD56 pos V G L U T 1 high ), APOE and CD56 expression (cluster 3: APOE pos CD56 pos , cluster 11: APOE high ), dense white maqer areas (cluster 9: MCNPase high ), mixed white and grey maqer areas (cluster 5: MCNPase pos CD56 pos ), and areas with dense mitochondrial protein and free iron (cluster 7: VDAC1 hi Fe hi ). Pixel clustering also idenSfied nuclei (cluster 4) and cells, including neuronal soma (cluster 10: RNA high HH3 low ), microglia (cluster 8), intravascular immune cells (cluster 12), astrocyte cell bodies (cluster 19), as well as vasculature (cluster 13) and astrocyte endfeet (cluster 6). Neuronal structures like dendrites (cluster 20), white maqer axons (cluster 15) and grey maqer axons (cluster 16) were also clearly idenSfiable, despite having elongated and irregular shapes. Tau tangles (cluster 18) were found in the hippocampus as well as intracellular Ab1-42, a common occurrence in the aged brain.
The proporSon of each pixel cluster for each brain region grey and white maqer showed that the white maqer was indeed homogenously dominated by myelin (MCNPase high cluster 9) while the grey maqer areas from each brain region varied in their composiSon ( Figure 5B). All grey maqer regions had synapSc textures and the hippocampus, MFG and caudate had higher proporSons of VGLUT1 high areas (cluster 2). The SN had the highest proporSon of VDAC1 high Fe high areas while the hippocampus had a small proporSon and other brain regions did not have any, suggesSng that, in this case, the SN was a large reservoir of accumulated free iron. The cerebellum had the largest proporSon of neuronal soma (cluster 10), which corresponded to the densely packed granular layer ( Figure S1B). The hippocampus, MFG and SN had MAP2 pos dendrites (cluster 20) and some (< 25%) white maqer textures (clusters 5 and 9). All 20 pixel clusters are shown in the hippocampus in a composite image ( Figure 5C) with an enlarged area (FOV1) corresponding to the one shown in Figure 1C, as well as enlarged areas depicSng a large blood vessel with intravascular immune cells (FOV2) and the white maqer axons and myelin of the alveus (FOV3). Individual pixel clusters are visualized within FOV1 ( Figure 5D) and within the enSre hippocampus ( Figure S5A). The combined clusters across the SN, MFG, cerebellum, and caudate were also visualized and showed the varying microenvironments of the grey maqer ( Figure S5B).
We next asked whether microglial states along the MSC are preferenSally restricted to certain microenvironments across larger brain regions. To do this, we visualized the MSC of all microglia across the enSre hippocampus (and other brain regions, Figure S5B) colored by their MSC value and saw that low MSC cells appeared to be enriched within specific hippocampal sub-regions like CA1 (Figure 5E). Middle phase MSC cells appeared to be concentrated in CA4 and DG, while high MSC cells were disSnctly compartmentalized within areas like the alveus and stratum radiatum which are both myelin and glial rich areas of the hippocampus. This yet unappreciated organized distribuSon of conSnuous microglial states within the healthy adult hippocampus suggests that different microglial states might perform different roles in specific local microenvironments related to the composiSon of that niche.
To fully untangle the niche composiSon, we sought to computaSonally and spaSally analyze the texture of the microenvironments directly around each microglia. To that end, we quanSfied the proporSon of pixel clusters within a 20 µm radius of the centroid of each microglial cell (Figure 5F). We then aligned the proporSon of pixel clustered brain textures and microglia for each of 180 bins along the MSC (Figure 5F). Low MSC cells were surrounded by a high proporSon of dense synapSc VGLUT1 high CD56 pos textures (cluster 2) and MAP2 pos dendrites (cluster 20), features that were predominantly present in the CA1 and parts of the stratum radiatum, but were also found in CA2, CA3 and CA4 sub-regions. Middle MSC cells tended be surrounded by a higher proporSon of the less synapScally dense CD56 high VGLUT1 pos texture (cluster 1), which was predominant in the subiculum and parts of CA4, DG and stratum lacunosum moleculare. As microglia transiSoned from middle to high MSC, there was a disSnct enrichment of white maqer (cluster 9: MCNPase high myelin and cluster 5: MCNPase high CD56 pos ) surrounding them and a loss of dense synapSc textures. The cerebellum also showed a disSnct enrichment of white maqer, white maqer axons as well as vasculature surrounding high MSC cells ( Figure S5B). If low MSC microglia are in a hypoacSve, possibly dystrophic state, their associaSon with dense synapses in the CA1 in this healthy aged hippocampus implies that they may not be performing their normal homeostaSc roles there. This is in line with the CA1 area being highly suscepSble to degeneraSon in aging 61 and the gradual increase of senescent microglia with age 57 .
The MFG, a less compartmentalized isocorScal brain region, had a higher density of low MSC cells in general and middle phase MSC cells were modestly enriched with surrounding white maqer texture. The SN, like the hippocampus, had a higher overall density of high MSC cells and in the SN these had more astrocyte-rich microenvironments around them. The caudate showed the least compartmentalizaSon of microglial states which may be a result of the homogenous nature of the acquired FOVs in this area. Within the deep white maqer tracts in the hippocampus that were acquired separately from the large grey maqer area, microglia were also more homogenously surrounded by the same microenvironments which were, unsurprisingly, dominated by myelin ( Figure 5G) and similar to other deep white maqer brain regions ( Figure S5C). However, high MSC cells in the white maqer had a higher proporSon of vasculature in their direct vicinity, suggesSng that vessel-proximate microglia were more highly acSvated than microglia further away from vasculature, possibly due to cytokines and other factors that penetrate the blood brain barrier and acSvate microglia. We calculated the correlaSon (both posiSve and negaSve) of pixel cluster frequency to progression of MSC and saw that the more varied grey maqer brain regions had a higher correlaSon of MSC to specific pixel clusters overall than the homogeneous white maqer regions (Figure 5H). Following this paqern, the hippocampus had the highest correlaSon of MSC progression with pixel cluster frequencies and was the most organized and compartmentalized brain region we analyzed. Pixel cluster changes that correlated with MSC progression supported what we saw previously, namely VGLUT high/pos , MAP2 pos and MCNPase high rich areas which represent high synapSc, neuronal and myelinated axon density respecSvely, correlated with changing microglial states between them.
This data driven pixel clustering analysis of local brain microenvironments provides a new model of brain regional cellular composiSon and localizaSon based on intracellular and extracellular protein expression 60 , aspects that are not captured by tradiSonal cellular segmentaSon. We uncovered a compartmentalizaSon of microglial cell states within specific microenvironments where the more structured a brain area is, like in the allocorScal hippocampus, the more enrichment of polarized microglial cell states are found within them. Microglia had higher MSC in white maqer areas and were more likely hypoacSve or dystrophic in grey maqer areas with high synapSc and neuronal density. Vasculature, regardless of locaSon, usually implied surrounding microglia were more acSve than those more distant from vasculature. Notably, the CA1 subregion of the hippocampus had the most hypoacSve, low MSC cells across the enSre hippocampus, a finding that may be relevant for the high suscepSbility of this subregion to degeneraSon in AD and several other diseases of brain.

Alzheimer's disease skews microglial states locally around plaques and globally towards low MSC throughout the hippocampus.
Having observed that microglia are beqer analyzed through a conSnuum of states, varying between brain regions and within local microenvironments, we sought to assess microglial changes along the MSC within a neurological disease context. We imaged sub-regions of the hippocampus (CA1, CA4, DG) from an AD donor (82-year-old male with AD demenSa, Braak score of V, MMSE score of 19, APOE ε3/ε3, and a post-mortem interval of 2.95 hours) which included part of the subiculum, CA1, CA4, and DG subfields ( Figure S6A). Previous studies of microglia in mouse models of AD showed that microglia proximal to Aβ plaques downregulate homeostaSc markers like P2RY12 and upregulate acSvaSon markers and have been called disease associated microglia (DAMs) 11 . To idenSfy similar niches in human AD, we carried out proximity analysis around plaques that we stained for with pan-Aβ, Aβ1-40, Aβ1-42 and tau tangles with PHF-tau. We segmented out the plaques and tangles from our images using masks created by pan-Aβ and PHF-tau expression, respecSvely ( Figure 6A) and idenSfied different types of plaques through FlowSOM (Figure 6B, S6C). We categorized two major types as dense or sparse plaques by their pan-Aβ intensity. These can be further divided into plaque subtypes with or without PHF-tau + starburst neuriSc processes: dense non-neuriSc plaques (type 1), sparse non-neuriSc plaques with Aβ1-42 (type 2), sparse neuriSc plaques with Aβ1-40 and Aβ1-42 (type 3), and dense neuriSc plaques with Aβ1-42 (type 4). Plaque density was much higher in the grey maqer than in the white mater, and plaque types 3 and 4 were most abundant (Figure 6B, 6D). We next made a disSncSon between PHF-tau pos processes that were present outside of plaques and those that were present within plaques by gaSng on pan-Aβ pos and PHF-tau neg features (Figure S6B, S6C). PHF-tau + tangles were present in both the healthy and AD hippocampus although there was a higher density in the AD case in both the grey and white maqer.
In order to determine the phenotype of microglia directly around plaques and tangles, we selected those cells which had a plaque or a tangle nearby (within a specific radius, see methods) and quanSfied the number of cells proximal to each plaque and tangle type ( Figure  6C). Since microglia are much more abundant than plaques and tangles, most microglia were not proximal to plaques or tangles (grey bar). Of the microglia that were proximal to plaques, most were nearby neuriSc plaques (types 3 and 4). RelaSvely few microglia were proximal to plaques generally (less than 100 cells in total) compared to tangles (close to 10,000 cells), likely due to the relaSve abundance of tangles over plaques in the hippocampus of this AD case. We looked to whether microglial phenotype was influenced by proximity to each pathologic feature. We observed that many proteins were changed in comparison to microglia distant from plaques and tangles, including upregulaSon of HLA-DR, APOE, CD74, CD11c, FerriSn-L, GPNMB, poyUK48 and CD68, and downregulaSon of CD16, CD14 and P2RY12, (Figure 6D, S6D), suggesSng a DAM phenotype and therefore supporSng observaSons from previous studies in mice 10,11 . Microglia proximal to dense plaques were shijed towards high MSC states, and therefore more acSvated, while microglia proximal to sparse plaques were predominantly in middle phase MSC ( Figure  S6E). Microglia that were proximal to tangles were not significantly different in their protein phenotype to microglia that were distant from tangles ( Figure S6F). Pixel clustering of the AD hippocampus idenSfied, in addiSon to textures we saw in other healthy brain regions, two types of plaques, namely Aβ1-40 pos APOE high and Aβ1-42 pos plaques (Figure S6E). High MSC microglia had an enrichment of plaques and vasculature in their direct microenvironment in the grey maqer ( Figure S6H) and an enrichment of vasculature in the white maqer (Figure S6I), as in healthy brain Sssue.
Although we found DAMs around plaques in AD hippocampus, these were comparaSvely few cells (~100) compared to the total number of cells imaged in the AD hippocampus (~20,000). We asked whether microglia from the AD hippocampus exhibit global changes compared to microglia from the healthy hippocampus. We found that AD hippocampal microglia were globally skewed towards low MSC states while healthy hippocampal microglia were skewed towards middle-high MSC states in the grey maqer areas (Figure 6E, top), suggesSng that microglia from AD hippocampus grey maqer were generally less acSvated and presumably senescent/dysfuncSonal compared to healthy hippocampal grey maqer microglia in this single case. As in other comparisons between healthy brain regions, the staSsScally significant feature differences were found within the protein rather than morphologic features (Figure 6E, boYom). Once again, APOE and VDAC1 were highly staSsScally significant where AD hippocampal microglia had higher APOE expression and lower VDAC1 expression than healthy hippocampal microglia (Figure 6F, 6G). AD hippocampal microglia had higher levels of p-S6 and lower levels of CD45 than healthy hippocampal microglia. Unique to the AD versus healthy comparison, the trajectory of marker progression changed dramaScally at high MSC for several features (Figure 6G). Iba1 decreased suddenly in high MSC in AD microglia, while polyUK48 increased, compared to healthy microglia. Morphological features also changed in high MSC where healthy microglia were the largest and had the highest number of projecSons in the control hippocampus, but AD microglia were smaller and had fewer projecSons. Overall, there was a lower density of presumably healthy funcSonal and acSve microglia in AD, and their state of healthy acSvaSon was perturbed. The same was not true for AD and healthy hippocampal white maqer microglia (Figure S6J), which supports the pathological changes in AD being found predominantly in the grey maqer. Indeed, microglia in the grey maqer AD hippocampus were skewed towards low MSC compared to microglia in the white maqer AD hippocampus ( Figure  S6K).
Since this iniSal screen was based on a single AD case, we sought to validate our findings of MSC changes in a larger cohort of healthy and AD hippocampal samples. While the mulSplexing capacity of MIBI allowed us to uncover the coordinated MSC across mulSple proteins and morphologies, we now needed to run many samples with only a few quanSfied proteins. To do this, we used two-color immunohistochemistry and chose two markers as proxies for the MSC: HLA-DR, which increased from low to high MSC, and Iba1, which decreased from low to high MSC. Thus HLA-DR low Iba1 high cells would represent low MSC cells and HLA-DR high Iba1 low cells would represent high MSC cells. We stained a cohort of 13 healthy and 13 AD demenSa age-matched hippocampal cores ( Figure 6H, top le6) with Iba1 and HLA-DR on blue and brown chromogens, respecSvely, and extracted FOVs from CA1 of each core. We then separated the brown and blue colors in silico (Figure 6H, top right) and segmented out the microglia using the same strategy as for our MIBI data, i.e., creaSng a composite microglial mask using Iba1 and HLA-DR expression. We extracted approximately 30,000 cells from each of the healthy and AD groups (Figure 6H, bo1om le6). Based on a simple biaxial plot, healthy hippocampal microglia expressed higher levels of HLA-DR than AD hippocampal microglia. We then computed a proxy MSC using Iba1 and HLA-DR which, like our highly mulSplexed MIBI MSC, also progressed from high Iba1 and low HLA-DR to low Iba1 and high HLA-DR expression ( Figure 6H, bo1om middle). Indeed, the AD hippocampal microglia were skewed towards low MSC compared to healthy hippocampal microglia which were enriched in the middle MSC, presumably reflecSng a healthy acSve state compared to a potenSally senescent inacSve state in AD. We calculated progressing p-values along MSC bins ( Figure 6H, bo1om right) and found that differences between healthy and AD MSCs were staSsScally significant in the low and middle MSC, and not in high MSC, suggesSng that AD microglia were indeed globally skewed towards having more inacSve HLA-DR low microglia and fewer healthy acSve HLA-DR +/high middle MSC microglia. However, more acSvated, high MSC microglia were sSll present in both healthy and AD hippocampi, which may represent DAMs around plaques in AD and/or microglia in myelin rich niches without pathology.
Finally, we assessed correlaSon between MSC and MMSE (Mini Mental State Exam), which is a cogniSve assessment test taken by all the donors of our cohort towards the end of life, where a score of 24 or less suggests a likely diagnosis of demenSa. Rather than grouping samples into two groups, MMSE allowed for a conSnuous variable associaSon with MSC and correlated posiSvely (R = 0.31, p < 2.2 e-16) where low MSC, corresponding to impaired microglial states, associated with low MMSE score, and higher MSC correlated with higher MMSE score ( Figure S6L). Overall, our highly mulSplexed spaSal proteomic MIBI data and chromogenic validaSon show that microglia from the AD hippocampus, specifically the CA1 subregion, are globally skewed towards a less expressive, less acSve, potenSally senescent state microglia. These deficient microglia are likely not performing their crucial roles in brain homeostasis and may be contribuSng to synapSc loss and cogniSve decline in this highly suscepSble and vulnerable brain area. Microglia from the healthy hippocampus CA1 more acSvely expressed many proteins and are likely performing compensatory funcSons that help maintain brain health and cogniSon in old age. Finally, our data support that differenSal expression of HLA-DR and Iba1 by microglia can be used to understand cellular state in health and disease, thus providing easy-to-use tools for tracking microglial states in future studies using any human brain Sssue.

Discussion
We employed spaSal proteomics and endogenous metal quanSficaSon by MIBI to deeply phenotype human microglia across five healthy brain regions and AD hippocampi ( Figure  1). We segmented individual microglia (Figure 2-3) and implemented a pseudoSme trajectory analysis to reveal a conSnuum of phenotypic states we termed the Microglial State ConSnuum (MSC), characterized by a high dynamic range of protein expression (Figure 4). Low MSC microglia had low expression levels for many proteins associated with myeloid cell acSvaSon, were smaller and had dystrophic morphology while middle to late microglia had higher to very high expression levels for many proteins and were larger and thicker, resembling healthy acSvated to highly acSvated states. Besides density of microglial contribuSon, inter-brain regional variaSon of MSC distribuSon was also seen, where the healthy MFG and caudate were skewed towards low MSC, the cerebellum was enriched in middle MSC and the hippocampus and substanSa nigra skewed towards high MSC (Figure 4). Furthermore, we demonstrated that this mulS-cellular brain environment's relaSonship to the MSC posiSon was not just driven by macro-anatomical features, but also micron-scale, subregional niches using pixel clustering and neighborhood analysis (Figure 5). For instance, we showed enrichment of high MSC microglia within niches of high myelin and axon density as well as close to vasculature, and of low MSC microglia in niches that had high synapSc and soma density, parScularly in the hippocampus. We also localized DAMs in direct contact with amyloid plaques in AD hippocampus, which were enriched in middle to high MSC, away from the most acSvated states we observed (Figure 6). These changes in MSC-based localizaSon as a funcSon of both anatomical space and proteopathy suggest a link towards cell (dys)funcSon, and reinforce the idea of uniquely diverse and plasSc human microglia with specifically tuned roles in brain homeostasis 62 .
Notably, in the AD hippocampus, microglia spaSally uninvolved with plaques and tangles vastly outnumbered DAMs (i.e., >> 99:1) and exhibited a low, presumably inacSve, MSC phenotype, suggesSng a dystrophic and perhaps senescent and impaired funcSonal state. This overall low-shijed MSC persisted when we validated our MIBI findings across an independent cohort of 13 AD and 13 control hippocampi by targeted low-plex IHC, quanSfying the coexpression of Iba1 and HLA-DR (Figure 6). Like other aspects of neurodegeneraSon, it will be important to understand if shijs in these microglial states are causal to the disease process or a response to the degeneraSon process. If causal, the expectaSon is that similar shijs may be present of other degeneraSve processes with associated proteopathy, like Parkinson's disease. Moreover, the anatomical influence on MSC suggests that cauSon should be taken when assessing glial cells from one brain region and making conclusion on their influence in another.
While most proteomic imaging and single cell analysis techniques tend to binarize molecules of interest into posiSve and negaSve cells and pixels, this approach would miss important, but less obvious, changes of microglial protein expression. Our applicaSon of MIBI allowed the quanSficaSon of proteins across a dynamic range on a cell/pixel basis allowing us to capture a progression of microglial states based on gradual expression changes along a spectrum which corresponds both to anatomical biases as well as healthy vs. impaired cogniSve states. Microglial molecular phenotypes, whether proteomic or transcripSonal, are no longer viewed as discrete populaSons but rather as overlapping and progressively changing in response to an onslaught of sSmuli 63,64 . Furthermore, spaSal informaSon is criScal to understand the diversity of microglia present in the brain. Mouse studies focus on specific DAM states around amyloid plaques which are ojen not found in human studies, largely due to the lack of single-cell and spaSal resoluSon. Our data demonstrated that human microglia are indeed able to transiSon to the DAM state, which is considered protecSve in AD, however, it also shows this represents only a small fracSon of microglia that are in direct contact with plaques undergo this transiSon. Instead, we see that most microglia in the human AD hippocampus are not in contact with plaques or tangles but appear to instead be globally shijed to a dystrophic or impaired state. Other studies of human aged and AD brain regions have shown microglia to have an accelerated aging profile in AD (HAMs, Human AD Microglia) 14 rather than a DAM profile, although the DAM signal was likely lost in the bulk analysis, so this likely reflects the global microglial switch to an impaired response. Altogether, this study framework provides a single cell lens through which to view human microglial diversity in situ that seems not to be limited to those cells perturbed by disease pathology interacSon, but also global, anatomically specific shijs that could be used to understand dysfuncSon in disease and are a potenSal point of therapeuSc intervenSon.

Human brain $ssue
Human FFPE brain samples that was imaged by MIBI were acquired from the Arizona Study of Aging and NeurodegeneraSve Disorders and Brain and Body DonaSon Program at Banner Sun Health Research InsStute (brainandbodydonaSonprogram.org) 68 . Brain regions including the hippocampus, cerebellum, substanSa nigra, caudate nucleus and middle frontal gyrus were from a healthy donor who was a 90-year-old male with no demenSa, an MMSE score of 27, Braak stage I, APOE genotype ε3/ε3, and a post-mortem interval of 3.4 hours. The AD demenSa case imaged by MIBI was the hippocampus from an 82-year-old male with MMSE score of 19, Braak stage V, APOE genotype ε3/ε3, and a post-mortem interval of 2.95 hours. The Sssue microarray used to validate MIBI findings comprised 26 hippocampal human FFPE brain cores selected from paSents of The 90+ Study cohort (hqps://doi.org/10.1002/alz.12981) whose pathologic evaluaSon was performed at Stanford Pathology Department. Donors were agematched and selected based on their clinical diagnosis and neuropathological scores evaluated by NIA-AA guidelines 65 . The donor details are listed in Extended Data Table S2.
Tissue staining FFPE brain Sssues were secSoned (5μm secSon thickness) from Sssue blocks on gold and tantalum-spuqered microscope slides. Slides were baked at 70°C overnight, followed by deparaffinizaSon and rehydraSon with washes in xylene (3×), 100% ethanol (2×), 95% ethanol (2×), 80% ethanol (1×), 70% ethanol (1×) and ddH2O with a Leica ST4020 Linear Stainer (Leica Biosystems). Slides next underwent anSgen retrieval by submerging sides in 3-in-1 Target Retrieval SoluSon (pH 9, DAKO Agilent) and incubaSng at 97°C for 40 min in a Lab Vision PT Module (Thermo Fisher ScienSfic). Ajer cooling to room temperature for 1 h, slides were washed in wash buffer (1× PBS IHC Washer Buffer with Tween 20 (Cell Marque) with 0.1% (w/v) bovine serum albumin (Thermo Fisher)). Next, all slides underwent two rounds of blocking, the first to block endogenous bioSn and avidin with an Avidin/BioSn Blocking Kit (BioLegend). Slides were then washed with wash buffer and blocked for 1 h at room temperature with 1× TBS IHC Wash Buffer with Tween 20 with 3% (v/v) normal donkey serum (Sigma-Aldrich), 0.1% (v/v) cold fish skin gelaSn (Sigma-Aldrich), 0.1% (v/v) Triton X-100, and 0.05% (v/v) Sodium Azide. The first round of staining was done with free indium 115 3+ (8 mM diluted in PBS, Fluidigm) in staining buffer (1x TBS IHC Wash Buffer with Tween 20 with 3% (v/v) normal donkey serum) and incubated overnight at 4°C in a humidity chamber. The following day, slides were washed twice for 5 min on a shaker in wash buffer. The second round of staining was done using the cocktail of metal conjugated anSbodies prepared in staining buffer at their respecSve concentraSons and filtered through a 0.1 μm centrifugal filter (Millipore) prior to incubaSon with Sssue overnight at 4°C in a humidity chamber. Following the overnight incubaSon with the anSbody cocktail, slides were washed twice for 5 min in wash buffer. On the third day, anS-bioSn 152 Eu was prepared as described and incubated with the Sssues for 1 h at 4°C in a humidity chamber. Following staining, slides were washed twice for 5 min in wash buffer and fixed in a soluSon of 2% glutaraldehyde (Electron Microscopy Sciences) soluSon in low-barium PBS for 5 min. Slides were then washed in PBS (1×), 0.1 M Tris at pH 8.5 (3×) and ddH2O (2×) and then dehydrated by washing in 70% ethanol (1×), 80% ethanol (1×), 95% ethanol (2×) and 100% ethanol (2×). Slides were dried under vacuum overnight prior to imaging. For the detailed staining protocol see dx.doi.org/10.17504/protocols.io.dm6gprk2dvzp/v5.

Image acquisi$on on MIBI
Prior to imaging, slides were spuqer coated with 10 nm of gold over the enSre stained Sssue secSon on each slide in order to ensure no charging effects of the Sssue which impact on field uniformity during imaging. Imaging was performed using a MIBI instrument with a Hyperion ion source by sequenSal rastering: pre-rastering at half the ion dose to remove the gold coaSng on FOVs of interest, followed by final rastering and image collecSon at full ion dose. Xe + primary ions were used to sequenSally spuqer pixels for a given FOV. The following imaging parameters were used: aperture se‚ng, 300 µm; acquisiSon se‚ng of 100 kHz; FOV size, 700 μm x 700 μm at 1,024 ×1,024 pixels per FOV; sample bias, 20 V (pre-raster, ions not funneled into the TOF chamber to preserve the detector) and 50 V (final raster, ions funneled into the TOF chamber); dwell Sme, 0.5 ms (pre-raster) and 1 ms (final raster for image acquisiSon); median gun current on Sssue, 13.6 nA; an ion dose of ~15 nAmp.ms.nm -2 (pre-raster) and ~30 nAmp.ms.nm -2 (final raster). AcquisiSon Sme was 8.7 minutes per FOV (pre-raster) and 17.33 minutes per FOV (final raster). FOVs were Sled across each brain region with an overlap of 50 µm in the x and 20 µm in the y direcSon. A total of 420 FOVs were acquired across all brain regions, including grey and deep white maqer.

Low-level image processing
MulSplexed image sets were extracted, slide background-subtracted, denoised and aggregate filtered as previously described 40,44 , using Matlab scripts and in-house GUIs for MIBI image processing. AddiSonally, non-specific binding to charged neurons (due to over fixaSon of Sssue in formalin prior to embedding in paraffin blocks) was subtracted by masking neurons using the indium 115 3+ signal and removing signal from channels that contained non-specific binding. For visualizaSon purposes, individual FOVs were sStched together to reconstruct large areas of each brain region using the Fiji/ImageJ image processing environment and exisSng plugins.

Microglial and pathology segmenta$on
To segment microglia, amyloid plaques and PHF tau tangles we used EZSegmenter, a MATLAB regionprops thresholding-based segmentaSon GUI developed in-house and available as a part of the MIBI image processing toolkit here: hqps://github.com/angelolab/MAUI, as previously described 44 . Briefly, mulSplexed TIF images from mulSple FOVs are loaded into the GUI. Microglial masks were created using the combined Iba1 and CD45 protein expression, amyloid plaque masks were created using pan-Aβ protein expression and PHF tau tangle masks were created using PHF tau protein expression. Parameters for masking are adjusted for each mask separately (e.g. Gaussian Blur, minimum and maximum object pixel size) and fixed across all FOVs. Masks are then used to extract pixel-level signal intensiSes across each channel and then cell or object size normalized before import into an output cell table csv file. Single cell or object data was then imported into R for arcsin h transformaSon and normalizaSon.

Microglial single-cell computa$onal analysis Clustering
Protein and morphology features of segmented microglial cells were mean-centered and scaled, and unsupervised clustering was performed on protein-only features using FlowSOM 58 v2.2.0 in R version 4.1.3.

Computa2on of the microglial state con2nuum
The microglial state conSnuum (MSC) was generated using the SCORPIUS 59 algorithm with a lowess smoother and 1000 iteraSons. Trajectory inference was performed using either proteinonly, or morphology-only features. Microglia from the HIP, Caudate, Cerebellum, SN, and MFG brain regions were included (grey and white maqer), and features were mean-centered and scaled prior to trajectory inference. Trajectory inference was also performed on microglia from individual brain regions and a spearman rank correlaSon was used to compare the pseudoSme esSmate of each individual region, to the all-region MSC, which had strong agreement with each other.

Es2ma2ng microglia state in the AD hippocampus using machine learning
To esSmate the MSC of microglia in the AD hippocampus, we used kernel support vector machine (KSVM) with epsilon regression and a radial basis funcSon Gaussian kernel. KSVM was done using the ksvm() funcSon in the R package kernlab 66 v0.9.30 with parameters type = 'epssvr' and kernel = 'rbfdot'. To validate predicSve accuracy of the model, we used 5-fold cross-validaSon. Microglia data from all regions except AD hippocampus was randomly split into 5 folds and SCORPIUS was computed on the training set for each fold. The KSVM model was trained with protein features and pseudoSme as the response variable, and then predicSon was performed on the hold-out set. We compared the predicted pseudoSme to the pseudoSme computed separately on the hold-out set. Model performance was assessed using correlaSons and residuals. Predicted and true pseudoSme esSmates were significantly correlated with values between 0.99 and 1.0, residuals were randomly distributed along the fiqed line, and absolute mean and standard deviaSon of residuals across all folds was less than 1.8% and 0.02%, respecSvely.

Pairwise Comparison of Trajectories by Binned Permuta2ons (PCTBP)
To compare protein and morphology expression between microglia from different brain regions along the MSC in a pairwise manner, we developed a non-parametric staSsScal method using binned permutaSons along the trajectory with 3 major steps.
1. Calculate the opSmal bin width using Freedman-Diaconis rule 67 . 2. Perform permutaSon analysis: Randomly scramble, without replacement, the pair of labels to compare. Within each bin, calculate the mean difference between the scrambled labels. Compare the absolute permutaSon mean difference from the scrambled labels to the absolute true difference (test staSsSc) between the true labels in the same bin. Repeat this step 1000x Smes. The permutaSon score is the proporSon of samples that have a test staSsSc at least 1.5x greater than the permutaSon mean difference. A minimum 1.5x difference is required to improve robustness of the method by incorporaSng a fold change threshold.
3. The permutaSon score of each bin along the trajectory is corrected for mulSple comparisons and smoothed using a locally esSmated scaqerplot smoothing and the final PCTBP score is the average smooth permutaSon score along the trajectory. The bin-based permutaSon approach to generaSng the PCTBP score allows for flexible analysis that can capture both local and global differences along the trajectory. The PCTBP score is generated using permutaSon scores along the enSre binned conSnuum, unless otherwise noted. When comparing low, middle, and high parts of the conSnuum, smoothing is performed prior to spli‚ng the low, middle, and high PCTBP scores.

Pixel clustering and spa$al enrichment analysis
Pixel clusters were idenSfied using the Pixie pipeline 60 . Briefly, single-pixel expression profiles were extracted from pre-processed MIBI images from all brain regions combined. A Gaussian blur was applied using a standard deviaSon of 2 for the Gaussian kernel. Pixels were normalized by their total expression, such that the total expression of each pixel was equal to 1. A 99.9% normalizaSon was applied for each marker. Pixels were clustered into 100 clusters using a selforganizing map (SOM) based on the expression of 17 markers: HH3, CD56, Indium, APOE, CD45, CD105, NEFL, NEFH, VDAC1, MCNPase, VGLUT1, Iba1, MAP2, S100β, GFAP, pan-Aβ and PHFtau. The average expression of each of the 100 SOM clusters was found and the z-score for each marker across the 100 SOM clusters was computed. All z-scores were capped at 3, such that the maximum z-score was 3. Using these z-scored expression values, the 100 SOM clusters were hierarchically clustered using Euclidean distance into metaclusters. These metaclusters were manually adjusted and mapped back to the original images such that each pixel was assigned to one pixel cluster. To characterize the microenvironment in direct vicinity of each microglial cell, we defined a radius of 40 pixels (20 µm) from the centroid of each microglial mask and calculated the proporSon of pixel clusters within. Microglia were binned into 180 bins along the MSC for the hippocampus, and 80 bins for the other brain regions. Average pixel cluster proporSons were calculated for the cells within each bin and ploqed in ascending order along the MSC.

Pathology and DAM analysis
Plaques were clustered using FlowSOM based on pan-Aβ, Aβ1-40, Aβ1-42 and PHF-tau expression within them. To idenSfy microglia that localized directly around plaques and tangles (i.e. DAMs), we bucketed microglia containing masked plaques or tangles within a set radius from each cell centroid. The radius was determined by taking the major axis length of each cell and adding a 10-pixel buffer to ensure only those cells in relaSvely close proximity to pathology would be considered DAMs. AddiSonal buffers, including 50 and 100 pixels were tested (data not shown), but lost the ability to separate out the majority of microglia into DAM or not.

Immunohistochemistry
For IHC screening of panel anSbodies, FFPE human hippocampal Sssue was secSoned onto standard glass slide at 5 µm thickness. Slides containing Sssue were baked at 70 °C overnight. Tissue secSons were then processed and stained using the sequenza method with single primary anSbody. The IHC protocol mirrors the MIBI protocol, with the addiSon of blocking endogenous peroxidase acSvity with 3% (v/v) H2O2 (Sigma-Aldrich) in ddH2O ajer epitope retrieval. On the second day of staining, instead of proceeding with the MIBI protocol, Sssues were washed twice for 5 min in wash buffer and stained using ImmPRESS universal (AnS-Mouse/AnS-Rabbit horse radish peroxidase) kit (Vector Laboratories). For detailed IHC staining techniques see dx.doi.org/10.17504/protocols.io.bf6ajrae and dx.doi.org/10.17504/ protocols.io.bmc6k2ze.

SoBware, data and code availability
Sojware for running the MIBI equipment was developed by SAI (MiniSIMS 2 Data Systems). The code for the analysis can be downloaded at hqps://github.com/bryjcannon/ MIBI_Brain_Analysis. All the informaSon required for cell and object segmentaSon are available at Ark-Analysis hqps://github.com/angelolab/ark-analysis.
All imaging data and analysis annotaSons will be made available in a public repository with a fixed DOI # upon peer reviewed publicaSon.  (C) An enlarged view of FOV1, the CA2 and parSal dentate gyrus, shown in (B), highlighSng fine cellular features including nuclei, pyramidal neuronal soma and dendrites, synapSc densiSes, myelin, astrocytes and vasculature. (D) Pixel correlaSon of pan-brain proteins used to idenSfy mulS-cellular niches in the brain, highlighSng strong posiSvely and negaSvely correlated protein programs. (E) Images of proteins depicted in (D) demonstraSng pixel signal overlap and exclusion for (i) nuclear proteins, (ii) astrocytes proteins, (iii) vascular proteins, (iv) axonal and dendrite proteins in neurons, and (v) endogenous iron (Fe). (F) Sub-regional organizaSon of the human hippocampus from expert neuropathological annotaSon (lej) and the protein abundance in each sub-region (right) calculated as average pixel value per mm 2 , with the inner cornu ammonis, subiculum and dentate gyrus areas highlighted.        (A) Local brain textures idenSfied through pixel clustering of all FOVs from all healthy human brain regions from a healthy donor using pan-brain protein markers, with expression of all panel protein makers for each of twenty pixel clusters. (B) DistribuSon of pixel clusters for each large brain region, quanSfied as average frequency per FOV for grey and deep white maqer. (C) SpaSal compartmentalizaSon of overlaid pixel clustered brain textures in the hippocampal local sub-regions with enlarged FOVs (right): FOV1 depicSng grey maqer with high synapSc density (pixel clusters 1 and 2), nuclei (pixel cluster 4), pyramidal neurons (pixel cluster 10), dendrites (pixel cluster 20), areas of grey and white maqer mixtures (pixel cluster 5); FOV2: astrocyte endfeet alone vasculature (pixel cluster 6), intravascular immune cells (pixel cluster 12); FOV3: white maqer axons (pixel cluster 15) and myelin (pixel cluster 9).   (A) SegmentaSon of amyloid plaques (lej) and tau tangles (right) from one human AD hippocampus by eZsegmenter through masking (blue) Pan-Aβ (magenta) and PHF-tau (yellow) expression. Human donor details: 82-year-old male with AD demenSa, Braak score of V, MMSE score of 19, APOE ε3/ε3, and a post-mortem interval of 2.95 hours. (B) Phenotyping of plaque types with FlowSOM clustering (lej, top) results in four plaque types with varying levels of Pan-Aβ, Aβ42 and Aβ40. Density of each plaque type quanSfied as number of plaques per mm 2 in the grey and deep white maqer of the human AD hippocampus (lej, boqom). RelaSve levels of each plaque phenotyping protein for each plaque type (right). (C) SpaSal enrichment analysis of plaques and tangles proximal to single microglia (lej) through masking each feature and quanSfying the presence of a plaque or tangle within a specified radius (r = major axis length of each cell + 10 pixel buffer) around each microglial mask. Total number of microglia with a plaque or tangle of each type within their direct vicinity (right). (D) QuanSficaSon of differenSally expressed proteins in disease associated microglia (DAMs) as defined by their proximity to each plaque type, with representaSve plaque types 3 and 4 shown with microglia overlaid as dashed magenta lines. (E) Microglial cellular density along the MSC between the healthy (green line) and AD (navy blue line) human hippocampus (top) with the total number of staSsScally significant and not significant pairwise differences within protein and morphology features. (F) StaSsScally significant feature differences (only proteins) between the healthy and AD human hippocampus as volcano plot, with their PCTBP scores and area between the curves. (G) All feature comparisons (protein and morphology) shown along the MSC for the healthy and AD human hippocampi. (H) ValidaSon of MIBI findings through an independent cohort of 13 healthy and 13 AD hippocampi through low-plex immunohistochemistry: (top) staining the Sssue microarray cores of the CA1 region of the hippocampus with Iba1 (blue) and HLA-DR (brown) as a dual stain and in silico separaSon of each protein signal; (boqom) segmentaSon of microglia through masking both Iba1 and HLA-DR signals together in eZsegmenter and quanSficaSon of signal intensity at a single-cell level in each FOV acquired in healthy and AD hippocampus groups. A proxy MSC was calculated with Iba1 and HLA-DR expression from extracted cells (boqom, middle) and the average microglial cellular density ploqed for the healthy (green line) and AD (navy blue line) hippocampus (with individual donors in dashed lines) and Iba1 (light blue) and HLA-DR (peach) expression along the MSC. StaSsScal significance (p-value) was calculated along binned MSCs (boqom, right) with PCTBP scores > 1 interpreted as staSsScally significant.