Spatial proteomic analysis in human Alzheimer’s disease brains enables identification of microenvironment-dependent microglial cell states

Paula Sanchez-Molina; Dennis-Dominik Rosmus; Dillon Brownell; Mert Meral; Cavanagh Gohlich; Aditya Pratapa; Yaser Peymanfar; Alyssa Whitley; Yue Hou; Nadezhda Nikulina; Alina Bogachuk; Ellen Lara Bouchard; Aude Chiot; Heidrun Kuhrt; Peter Wieghofer; Randall Woltjer; Fabian Svara; Oliver Braubach; Bahareh Ajami

doi:10.1038/s41593-026-02267-3

. 2026 May 18;29(7):1599–1614. doi: 10.1038/s41593-026-02267-3

Spatial proteomic analysis in human Alzheimer’s disease brains enables identification of microenvironment-dependent microglial cell states

Paula Sanchez-Molina ^1,^2,^#, Dennis-Dominik Rosmus ^3,^4,^5,^#, Dillon Brownell ^1,^2,^#, Mert Meral ⁶, Cavanagh Gohlich ^1,², Aditya Pratapa ⁷, Yaser Peymanfar ⁷, Alyssa Whitley ⁷, Yue Hou ⁷, Nadezhda Nikulina ⁷, Alina Bogachuk ^1,², Ellen Lara Bouchard ^1,², Aude Chiot ^1,², Heidrun Kuhrt ³, Peter Wieghofer ^3,⁴, Randall Woltjer ⁸, Fabian Svara ⁶, Oliver Braubach ^7,^✉, Bahareh Ajami ^1,^2,^9,^✉

PMCID: PMC13337490 PMID: 42151483

Abstract

Disease-associated microglial states are thought to contribute to Alzheimer’s disease (AD) progression, but characterizing them and their relationships to pathology remains challenging. Here we introduce CODEX-CNS—a multiplexed protein imaging technology with a custom data analysis pipeline for use in human brain samples. We profiled 704,706 cells in samples from the frontal cortex of 8 people with AD and 8 healthy controls and mapped features including blood–brain barrier, meningeal components and cell–cell interactions within the same tissue sections. Amongst the myeloid cell populations we identified, we found a border-associated macrophage-like microglial subset associated with aging. Further classifying myeloid cell subsets based on their spatial neighborhood, we identified a border-associated macrophage-like microglial subpopulation that was associated significantly with dense amyloid-β plaques, which we termed human plaque-associated microglia. This work offers insights into myeloid cell heterogeneity in AD and provides a new spatial approach to characterizing brain cells at the single-cell protein level.

Subject terms: Neuroimmunology, Microglia, Fluorescence imaging, Alzheimer's disease

Myeloid cells show marked heterogeneity in Alzheimer’s disease. This study introduces CODEX-CNS, a single-cell spatial proteomics pipeline, and identifies a human microglial subpopulation enriched in Alzheimer’s disease brains that associates with dense amyloid-β plaques.

Main

Understanding the complexity of microenvironmental, cellular and molecular properties associated with brain disorders is key to identifying the drivers of neurological diseases. Alzheimer’s disease (AD) is a heterogeneous disease characterized by the presence of extracellular amyloid-β (Aβ) deposition as neuritic plaques, intracellular accumulation of hyperphosphorylated tau as neurofibrillary tangles, and inflammation, which culminate in neurodegeneration¹. AD pathobiology features complex interactions that unfold across molecular, cellular and tissue-level scales, and involve virtually every principal brain cell type¹. Previous studies comparing patients with AD to age and sex-matched controls revealed that disease-associated transcriptional changes at early AD stages are cell-type specific^2–4. Therefore, understanding the characteristics of individual cell types, such as their morphology, molecular phenotype and communication with other cells, is essential for the development of both experimental models of AD and therapeutics.

Recent single-cell genomic and proteomic analyses have revealed extensive diversity within brain cell types^5–8. Several microglial subpopulations or activation states have been identified in AD by single-cell RNA sequencing (scRNA-seq) approaches^9–11. Two studies using scRNA-seq in mouse models of AD-associated amyloidosis identified a microglial phenotype, which was termed microglial neurodegenerative phenotype or disease-associated microglia (DAM)^9,11, associated with Aβ plaques. However, although transcriptomics data suggest that such microglial subpopulations exist in animal models, there is a lack of consensus regarding similar AD-associated microglia populations in humans^2,12–17. Furthermore, cross-species single-cell analyses revealed that risk genes for neurodegenerative conditions are significantly overexpressed in human microglia¹⁸. In line with this, a recent study applying scRNA-seq to characterize human microglial states across numerous neurodegenerative disorders points to microglial subpopulations that are present solely in the diseased human brain¹⁹. Therefore, comprehensive characterization beyond gene expression, such as spatially resolved analysis and proteomics, is needed to gain new insights into the role of microglial subsets in human AD pathophysiology.

Our understanding of microglial cell diversity in AD is limited by currently available technologies. Conventional single-cell methods that require tissue dissociation, such as flow cytometry, mass cytometry by time of flight (CyTOF) or scRNA-seq^20,21, cannot assess cell–cell interactions or interactions of cells with pathological features of AD, such as abnormal extracellular protein aggregates. Although valuable, next generation spatial transcriptomics technologies²² do not provide in-depth proteomic information at the single-cell level. As protein aggregation is a key feature of AD, identifying protein-specific signatures at single-cell resolution and simultaneously integrating these properties with the diseased brain microenvironment is of particular interest. Although conventional immunohistological analysis remains the standard method for combining protein expression analysis with spatial context, it detects only a limited number of markers within a tissue section. As a result, the diversity of cells, their interactions and how they contribute to AD pathology cannot be fully explored. A multiplexed protein tissue imaging technology with single-cell resolution could improve the characterization of the cellular relationships in healthy and pathological brain tissue samples.

The codetection by indexing (CODEX) method combines the high-parameter capabilities of single-cell technologies with spatial resolution²³. CODEX is a multiplex immunofluorescence (IF) imaging technology that relies on the cyclic addition and removal of complementary fluorescently labeled DNA probes against DNA-conjugated antibodies and has been used to visualize up to 100 proteins simultaneously in situ^23–25. This method enables labeling of all main cell types and substructures within a single tissue section. Until now, the CODEX technology has been deployed mainly in cancer research, whereas its use for neuroscience applications has been limited²⁶. This is due to the challenges presented by brain tissue, such as high autofluorescence and complex cell morphologies that are difficult to segment^27–29.

Here we defined the heterogeneity of myeloid cells in human brain formalin-fixed paraffin-embedded (FFPE) tissue samples from patients with AD and healthy subjects by adapting the existing CODEX technology to study central nervous system (CNS) tissue. This CODEX-CNS method consists of a brain-specific histological preparation protocol, a cell segmentation approach that enables the detection of complex cell morphologies, as well as a custom analysis pipeline that is suitable for cellular phenotyping and spatial analytics. Applying CODEX-CNS, we characterized myeloid cell heterogeneity based on morphology, protein expression and features in the spatial neighborhood. We demonstrated that the brain microenvironment correlates with the microglial cell phenotype, which is undetectable through conventional clustering analysis based solely on protein expression, and identified a microglial subpopulation that accumulates in the brains of patients with AD and is closely associated with dense Aβ plaques. These findings provide insights into the role of microglial subpopulations in human AD pathology, implying that the cell identity could be determined by the local neighborhood, and showcase the application of CODEX-CNS to enable diverse spatial proteomics studies of human brain tissue.

Results

Development of CODEX-CNS to study the human brain

To overcome autofluorescence and enable precise imaging and spatial analysis of FFPE tissue samples—the most common type of specimens available in brain banks—we developed the CODEX-CNS workflow (Fig. 1a). We modified the standard CODEX protocol by adding a pretreatment step to remove autofluorescence, which involved incubating FFPE tissue sections in a 4.5% H₂O₂ bath while exposing it to a broad-spectrum LED light (Supplementary Fig. 1a). The intense autofluorescence and visible lipofuscin particles that untreated FFPE brain tissues present (Supplementary Fig. 1b, upper panel) were considerably reduced after employing the CODEX-CNS autofluorescence reduction treatment (Supplementary Fig. 1b, lower panel). Thus, an increase in the signal-to-noise ratio was detected in treated samples compared to untreated adjacent sections, using the same antibodies, imaging channels and exposure times (Supplementary Fig. 1c). All data subsequently presented in this study were acquired using the CODEX-CNS protocol on FFPE frontal cerebral cortex samples (Supplementary Tables 1 and 2). In this study, a total of 16 samples were analyzed using the CODEX-CNS protocol. Initially, we profiled eight samples (four patients with AD versus four healthy controls), followed by validation of our findings in an independent cohort of eight additional samples (four patients with AD versus four healthy controls) using CODEX-CNS. To further confirm our results, we performed conventional IF on 35 samples across two independent laboratories. One of these IF cohorts closely matched the original CODEX-CNS samples, whereas the other represented a distinct, unrelated cohort. Altogether, the study included 51 samples, comprising 26 AD and 25 healthy control samples (Supplementary Tables 1, 2 and 3).

Fig. 1 — a, Schematic overview of CODEX technology and computational analyses performed in FFPE human brain samples. b, Custom 32-plex oligonucleotide-conjugated antibody panel designed to detect protein markers for distinct modules: immune cells, astrocytes, oligodendrocytes/OPCs, neurons and blood vessels, as well as proteins associated with neuropathology. c, Representative fluorescent images showing markers from each antibody module in the same brain tissue section of the human frontal cerebral cortex from a patient with AD. Image is representative of n = 4 AD samples included in the initial analysis. d, Individual fluorescent staining of different brain cell types (left) and their corresponding cell masks (right) using a machine learning segmentation approach. Images are representative of n = 4 AD samples and n = 4 healthy control samples analyzed in 7 independent experiments. e, Representative tissue image showing DAPI, NeuN (neurons), OLIG2 (oligodendrocytes/OPCs), GFAP (astrocytes) and IBA1 (microglia) fluorescent staining (top panel) and its corresponding spatial map showing the localization of segmented neurons, oligodendrocytes/OPCs, astrocytes and microglia (bottom panel) in the healthy human frontal cerebral cortex. Each dot represents one segmented cell. Note the distinction between WM and GM areas, as well as different cortical layers. f, Proportions of the main brain cell types in the GM and WM areas of the human frontal cortex (n = 4). Box heights indicate mean values per region and cell type group; error bars: 95% confidence intervals. Panel a created in BioRender; Rosmus, D. https://biorender.com/eihz1yz (2026).

We designed a panel of 32 antibodies to target neurons, glia, cerebral vasculature, peripheral immune cells and pathological features related to AD for use in CODEX-CNS (Fig. 1b, Supplementary Fig. 2 and Supplementary Table 4). As a proof of concept, neurons, astrocytes, oligodendrocytes, oligodendrocyte precursor cells (OPCs), microglia, blood vessels and Aβ plaques were identified in the same tissue section, confirming that CODEX-CNS can detect several cell populations and noncellular structures simultaneously in archival FFPE brain tissue samples (Fig. 1c). Consistent with the known anatomy of the human cerebral cortex, our imaging data clearly distinguished between gray matter (GM) and white matter (WM) regions (Extended Data Fig. 1a–l), as well as stratifications between different cortical layers (Extended Data Fig. 1a–d). We were also able to visualize components of the blood–brain barrier in detail by the presence of astrocytic end-feet in contact with the vasculature (Extended Data Fig. 1m, arrowheads), collagen IV in the basement membrane (Extended Data Fig. 1n), pericytes or vascular smooth muscle cells (Extended Data Fig. 1o, arrows) and endothelial cells containing tight junctions marked by claudin-5 (Extended Data Fig. 1p). Furthermore, simultaneous imaging of markers specific to blood vessels, immune cells and astrocytes allowed for the visualization of the meningeal membranes: the pia and the arachnoid mater, as well as the glia limitans superficialis, which is in contact with the pial surface³⁰ (Extended Data Fig. 1q–t). These data provide a striking example of how CODEX-CNS offers sufficient imaging resolution to visualize the CNS cytoarchitecture. However, current nuclear segmentation approaches are insufficient to capture morphologically complex brain cells, whose processes are not registered properly within the circumference of the nucleus segmentation mask. Therefore, the large number of markers that can be measured simultaneously by the CODEX technology is accompanied by the need for new bioinformatic methods to accurately annotate cell types, quantify marker expression and identify spatial relationships.

Extended Data Fig. 1 — a, Anatomical distinction between gray matter (GM) and white matter (WM) areas. Red inset in a delineates the region shown in higher magnification **(b-d)** demonstrating distinction of cortical layers. **e-l**, Representative images at higher magnifications showing specific GM **(e-h)** and WM **(i-l)** cellular properties using oligodendrocyte and neuronal markers. **m-p**, Representative images showing different components of the blood-brain barrier, such as astrocytic end-feet (white arrowheads) **(m)**, basal lamina **(n)**, pericytes/smooth muscle cells (white arrows) **(o)**, and endothelial cells **(p)**. **q-t**, Characterization of human meninges including the pia, arachnoid matter as well as the glia limitans superficialis (white arrow) using antibodies that correspond to the immune cells, vasculature, and astrocytes modules. Abbreviations: BV (blood vessel).

In this study, we used a machine learning segmentation approach that can detect the cytoplasm and processes of each brain cell. Cell instance segmentation was performed using the pretrained, convolutional neural network-based deep-learning models available in the ariadne.ai SPATIAL platform. This segmentation method captures neurons, astrocytes, microglia/macrophages and oligodendrocytes/OPCs based on neuronal nuclei (NeuN), glial fibrillary acidic protein (GFAP), ionized calcium-binding adapter molecule 1 (IBA1) and OLIG2 expression, respectively (Fig. 1d–f and Extended Data Fig. 2a,b). A total of 471,181 cells were segmented in the initial CODEX-CNS dataset, of which 267,970 were in the GM, 192,545 were in the WM and 10,666 were in the meninges. Cells segmented from the GM and WM were used for subsequent analyses. Different cell types exhibited distinct regional distributions, with neurons present mainly in the GM (39.35% of all cells) and oligodendrocytes in the WM (78.41% of all cells) (Fig. 1f). For each brain area, no significant differences in density of any cell type were observed between AD and age-matched healthy samples (Extended Data Fig. 2c).

Extended Data Fig. 2 — a, Fluorescent images showing the four main cell types of the brain (upper panel) and spatial maps showing their corresponding segmentations in tissue sections (bottom panel). b, UMAP plots of all cells in the original CODEX-CNS dataset (n = 4 AD samples and n = 4 healthy controls). Corresponding cell clusters (left panel) or tissue samples from which the cells originate (right panel) are color-coded, respectively. c, Quantification of IBA1⁺, GFAP⁺, NeuN⁺, and OLIG2⁺ cell density in the white matter (WM) and gray matter (GM) of healthy and AD brains (n = 4 per group, except for OLIG2⁺ cells where n = 2 per group). Error bars represent the standard error of the mean (SEM). Statistical analysis was performed by post-hoc Tukey test.

AD brains exhibit specific cell–cell interactions

Using CODEX-CNS, we captured pathological features of the AD brain microenvironment by simultaneously visualizing Aβ and principal brain cell types at single-cell resolution. Diffuse and dense-core Aβ plaques have been described in AD¹. We observed colocalization of apolipoprotein E (ApoE) and Aβ staining in dense (Fig. 2a–c, red arrows) and diffuse (Fig. 2a–c, dashed circles) Aβ plaques, consistent with the reported role of ApoE in Aβ plaque formation³¹. Reactive astrocytes with high levels of GFAP and Vimentin expression were localized primarily around dense Aβ plaques (Fig. 2d,e). In addition to astrocytes, microglia and potentially infiltrating macrophages were detected surrounding Aβ plaques (Fig. 2f–i). As previously reported in AD human brains^32–34, some of the IBA1⁺ cells (IBA1 is a marker of macrophages and microglia) around the Aβ plaques expressed CD163 (Fig. 2i)—a marker that is expressed predominantly by border-associated macrophages (BAMs), suggesting that either BAMs have infiltrated the brain parenchyma or some microglia have obtained BAM phenotype in response to the local microenvironment³⁵. In conjunction with glial reactivity, we found that the immediate vicinity of some Aβ plaques lacked normal postsynaptic density protein 95 (PSD95) and synaptophysin punctate staining but contained aberrant spheroid accumulations of synaptophysin and neurofilament (Fig. 2j–l, arrowheads). These observations are consistent with synaptic and neuritic dystrophy near dense Aβ plaques^1,36,37.

Fig. 2 — a–c, Representative images showing diffuse (dashed circles) and dense (red arrows) Aβ plaques (a) as well as ApoE depositions (b), demonstrating a high degree of colocalization (c). d,e, Representative images of Vimentin⁺ astrocytes (d) expressing GFAP (e) in close proximity to Aβ plaques. Note that Vimentin is also expressed in blood vessels. f–i, Representative images of astrocytes (GFAP (f) and Vimentin (g)) and microglia/macrophages (IBA1 (h) and TMEM119 and CD63 (i) surrounding an Aβ plaque. j–l, Representative images showing neuronal markers in relation to an Aβ plaque, showing a merged image (j) and individual staining for NeuN and neurofilament (k) and PSD95 and synaptophysin (l). Note the abnormal synaptic protein accumulation in the dense Aβ plaque site (arrowheads). Images are representative of n = 4 samples of frontal cerebral cortex from patients with AD.

Next, we leveraged the simultaneous detection of different cell types to analyze cell–cell interactions in the healthy human brain and their changes in AD (Fig. 3). Preferential cell contact was explored by comparing the frequency of cell interactions within a 15-µm radius from each cell centroid to a permuted background (Fig. 3a and Methods). This analysis revealed preferential contacts between neurons and oligodendrocytes/OPCs (Fig. 3b,c), as well as between microglia and astrocytes (Fig. 3b,c), in the GM of the human frontal cerebral cortex of healthy people. Oligodendrocytes in close vicinity to the neuronal soma, also known as perineuronal oligodendrocytes, provide metabolic support to neurons^38,39. However, the biological role of interactions between astrocytes and microglia under physiological conditions remains less explored. Microglia and astrocytes also exhibited frequent interactions with blood vessels (Fig. 3b,c). In WM, oligodendrocytes/OPCs exhibited a higher frequency of interactions with each other compared to interactions with other cell types (Extended Data Fig. 3a,b). This is consistent with the characteristic linear arrangement of oligodendrocytes along neuronal axons as well as the known axon-supportive functions of oligodendrocyte-formed myelin⁴⁰.

Fig. 3 — a, Representative image showing the cell interaction analysis approach used in samples of human frontal cerebral cortex from healthy people (n = 4) and patients with AD (n = 4). Cell and blood vessel frequencies are quantified within a 15-µm radius (red circle) from the centroids (blue dots) of each cell. b, Network plots showing the likelihood of close interactions between the main brain cell types in the GM of the human frontal cerebral cortex of healthy people (n = 4). Interaction likelihoods were generated by comparing observed interaction frequencies to a permuted background. Red lines represent high cell-to-cell vasculature interactions, whereas blue lines represent low cell-to-cell vasculature interactions. Line thickness represents deviation of the observed interaction, either positive or negative, from what would be expected under the null hypothesis that cells show no spatial patterning. c, Representative fluorescent images showing the most predominant cell interactions quantitatively detected in the GM of the human frontal cerebral cortex of healthy people (n = 4). d, Comparison of conditional cell interaction likelihoods in GM between AD and healthy brains (HC) (n = 2 per group) by two-sided Student’s t-test. Bars represent obtained t-statistics. *P < 0.1. Exact P values and t-test statistics can be found in the corresponding source data file. e, Representative fluorescent images showing cell interactions detected more frequently in the GM of patients with AD (n = 4) compared with healthy controls (n = 4). f, Comparison of cell protein expression between microglia interacting with other microglia and noninteracting microglia, as well as between neurons interacting with astrocytes and noninteracting neurons in the GM of AD brains (n = 4). Statistical analysis was performed by two-sided Mann–Whitney U test. Significant differences were considered as P < 0.05 and log₂FC > 0.5 or < −0.5. Each box extends from the first quartile (Q1) to the third quartile (Q3) of the data, with a line at the median. The whiskers extend from the box to the farthest datapoint. Exact P values and t-test statistics can be found in the corresponding source data file.

Extended Data Fig. 3 — a, Network plots showing the likelihood of close interactions between the main brain cell types in the white matter of the human frontal cortex (n = 4 healthy controls). Interaction likelihoods were generated by comparing observed interaction frequencies to a permuted background. Lines in red represent high cell-to-cell/vasculature interactions, whereas lines in blue represent low cell-to-cell/vasculature interactions. Line thickness represents deviation of the observed interaction, either positive or negative, from what would be expected under the null hypothesis that cells show no spatial patterning. b, Representative fluorescent image showing high interaction between OLIG2⁺ cells. c, Comparison of conditional cell interaction likelihoods in white matter between AD and healthy brains (n = 2 per group) via two-sided Student’s t-test. Bars represent obtained t-statistics. T-statistics, p-values, and pre-clipped cell values can be found in the corresponding source data file. Neurons were excluded from this analysis due to the low expression of NeuN in white matter areas.

We next explored the frequencies of different cell interactions in AD brains compared to healthy brains. Microglia–microglia contacts were more abundant in the GM of AD brains than in control brains (Fig. 3d,e, upper panel), probably due to their tendency to aggregate around Aβ plaques. No significant differences in microglia–microglia contacts between AD and healthy brains were observed in WM (Extended Data Fig. 3c). In the GM of AD brains, we detected a higher expression of CD163 and human leukocyte antigen DR isotype (HLA-DR) in microglia interacting with other microglia compared to noninteracting microglia (Fig. 3f). Significant differences in proliferating cell nuclear antigen (PCNA) expression were not observed (Fig. 3f), suggesting that microglial accumulation in the GM of AD brains results from cell migration probably prompted by pathological cues rather than from cell proliferation⁴¹. We also observed a tendency towards an increased interaction between astrocytes and both neurons and oligodendrocytes/OPCs, as well as vice versa, in the GM of AD brains compared to healthy controls (Fig. 3d,e, bottom panel). At the same time, we found that neurons interacting with astrocytes present higher levels of γH2A.X—a marker of DNA damage—than neurons not interacting with astrocytes (Fig. 3f). These findings suggest that astrocytes communicate with damaged neurons in the context of AD.

Taken together, this analysis provides a deeper understanding of cellular communication in the human frontal cerebral cortex and identifies differences in specific cell–cell interactions between AD and healthy human brains.

Distinct microglial morphologies across tissue microenvironments

Although extensive work employing sophisticated computational pipelines has been conducted to analyze microglial morphology in mouse models^42–45, a detailed morphometric characterization of human microglia is needed. To profile morphological phenotypes amongst myeloid cells in the human brain (using healthy controls and AD samples, n = 4 per group), we conducted unsupervised clustering on all IBA1⁺ segmented cells based on 12 morphological features (Methods), which were computed using masks for cells, somas, processes, branches and the cell convex hull (which measures cell solidity) (Fig. 4a and Extended Data Fig. 4a). Our initial analysis identified three main clusters that we annotated as Rounded (high cell circularity and absence of processes), Intermediate (few and thick processes) and Ramified (high number of processes and branches) (Fig. 4b,c and Extended Data Fig. 4b,c). To explore heterogeneity within the Ramified cluster, we applied a subclustering analysis that further stratified these microglia into three subclusters (Ramified 1–3), distinguished mainly by cell and soma size, soma circularity, number of processes and branches and process length (Fig. 4b,c and Extended Data Fig. 4b). Furthermore, we observed a correlation between morphological features and protein expression (Supplementary Fig. 3). Of note, cell circularity and solidity were associated highly with the expression of monocyte and macrophage markers CD14 and CD163, whereas features related to cell ramification were associated with the expression of the microglia marker transmembrane protein 119 (TMEM119) (Supplementary Fig. 3). The proportion of these morphologically distinct cell clusters varied across different compartments of the frontal cerebral cortex, with the Intermediate cluster being more prevalent in the WM, and the Ramified 2 and Ramified 3 clusters being more prevalent in the GM (Fig. 4d,e and Extended Data Fig. 4d). No statistically significant differences in the proportions of morphological clusters were observed between AD and control brains.

Fig. 4 — a, Representative schematic image showing cell, soma, process and branch detections used for the obtention of morphological parameters. b, t-SNE plot visualizing 63,873 myeloid cells colored by supercluster (left: Rounded, Intermediate and Ramified) and subclusters (right: Ramified 1, 2 and 3) identified after applying Leiden clustering to IBA1⁺ segmented cells of the human frontal cerebral cortex (n = 8) based on morphological parameters. Heatmaps show the mean morphological feature value within each cluster scaled by the minimum and maximum value. c, Representative images of cell masks displaying the characteristic morphological features observed in the five identified clusters. d, Representative tissue map showing the spatial distribution of the morphological clusters identified in the human frontal cerebral cortex from a healthy person. Each dot represents one cell. Dashed lines delimit the WM area. e, Pie charts showing the percentage of each morphological subcluster in GM (top panels) and WM (bottom panels) areas from AD (n = 4) and healthy control (n = 4) brains. f, Schematics illustrating the method used for the neighborhood analysis applied to all myeloid cells identified. The expression of different proteins is quantified in a 30-µm radius from the cell centroid of each segmented IBA1⁺ cell. g, Heatmaps showing the association between the identified morphological clusters and the proteins expressed in their neighborhood within the GM of the human frontal cerebral cortex from donors with AD (n = 4) and healthy control donors (n = 4). Asterisks denote significance (*P < 0.05, **P < 0.01, ***P < 0.001) as obtained from two-sided Student’s t-tests comparing values within each cell cluster to values from all other cell types for a given protein. Scale bar: rescaled average z-scored protein percentages in each group. Exact P values and t-test statistics can be found in the corresponding source data file.

Extended Data Fig. 4 — a, Representative images showing the cellular elements employed to obtain the morphometrics used in the myeloid morphological analysis. b, tSNE plots showing IBA1⁺ segmented cells colored by morphometric values. Scale bar represents raw values. c, tSNE plots colored by cell distribution between AD and control samples (upper), cell distribution in gray matter (GM) vs white matter (WM) areas (middle), and cell distribution between individual samples (bottom). d, Spatial maps of each sample showing the distribution of the morphology-based myeloid clusters in control (upper) and AD (bottom) brain tissue sections. Each dot represents one IBA1⁺ segmented cell.

As changes in microglial morphology could arise from alterations in their microenvironment, we further examined the cellular neighborhood of the five morphologically defined myeloid cell clusters throughout the human frontal cerebral cortex of all eight samples. We performed a custom neighborhood analysis quantifying the abundance of different markers present within a radius of 30 µm from each myeloid cell centroid (Fig. 4f and Methods). In both GM and WM, the rounded morphology was associated significantly with blood vessel markers and PCNA (Fig. 4g and Supplementary Fig. 4). This morphology was also associated with ApoE except in the GM of patients with AD, where high levels of ApoE colocalize with Aβ plaques. In addition, a positive association between the Ramified 3 morphology and microtubule-associated protein 2 (MAP2) expression was observed regardless of the disease condition, suggesting that a high degree of branching in microglia is associated with dendrites (Fig. 4g). Ramified 3 morphology was also associated with NeuN in GM of control brains, but not in AD brains (Fig. 4g). Notably, a significant spatial association between the Ramified 2 morphology and its proximity to Aβ, ApoE and NeuN was observed exclusively in the GM of patients with AD (Fig. 4g). By contrast, in the GM of control brains, but also in the WM of AD brains (where Aβ plaques are less prevalent), this morphology was associated with neurofilament (Fig. 4g and Supplementary Fig. 4). This result suggests that highly ramified cells with large somas may be associated with pathological features of AD. Overall, these data indicate that the aged human brain has well-defined morphological microglial phenotypes, which are distributed differentially across GM and WM as well as the surrounding tissue microenvironment.

Brain myeloid cells form five protein-based clusters characterized by their spatial distribution

Beyond morphological features, discerning protein phenotypes of myeloid cells and their spatial organization in brain tissue is essential for understanding their diversity and function in disease. Using CODEX-CNS, we performed Leiden unsupervised clustering in a total of 63,873 IBA1⁺ segmented cells (Fig. 5a) across eight frontal cerebral cortex samples from patients with AD and aged-matched controls.

Two myeloid populations (A and B) were identified based on the expression of IBA1, TMEM119, CD74, CD14, CD163, CD68, CD11b, CD11c, HLA-DR and CD45 (Fig. 5b and Extended Data Fig. 5a–c). Population A demonstrated a monocyte/macrophage phenotype and population B demonstrated a microglial phenotype based on the presence or absence of CD163 (BAM marker) and TMEM119 (microglia marker; Fig. 5b). To further explore the heterogeneity of these two populations, we performed a subclustering analysis that resulted in five myeloid cell populations (Fig. 5b). The five clusters were annotated based on the expression of TMEM119 and/or CD163, as well as their location in the tissue (Methods). Annotated clusters included: monocytes (MO), perivascular macrophages (PVMs), BAM-like microglia (BLM), microglia 1 (MG1) and microglia 2 (MG2). Most of the cells from the BLM subpopulation expressed both CD163 and TMEM119 proteins, indicating a unique phenotype that crosses that of microglia and PVMs (Extended Data Fig. 5d). These five myeloid clusters were mapped back onto the tissue samples and their phenotypes were validated in situ (Fig. 5c and Supplementary Fig. 5). These protein clusters were identifiable by both the ariadne.ai machine learning segmentation approach and our custom Otsu’s Thresholding-based Segmentation and Merge open-source algorithm (Methods), using reference mapping (Supplementary Fig. 6).

Extended Data Fig. 5 — a, tSNE plots showing IBA1⁺ segmented cells colored by protein expression levels. Scale bar represents raw values. b, tSNE plot showing cells colored by the Leiden clusters (A-C) and heatmap showing the mean protein expression in each cluster. Cluster C, which represents poorly stained and/or segmented cells, was removed from all downstream analyses. Scale bar represents the mean protein expression levels within each cluster divided by the maximum expression value. c, tSNE plots colored by cell distribution between AD and control samples (left), cell distribution in gray matter (GM) vs white matter (WM) areas (middle), and cell distribution between individual samples (right). d, Proportion of TMEM119⁺ (upper panel) and CD163⁺ (bottom panel) cells in the different protein-based myeloid clusters. e, Spatial maps of each sample showing the myeloid clusters distribution in control (upper panel) and AD (bottom panel) brain tissue sections. Each dot represents one IBA1⁺ segmented cell.

We next assessed the correspondence between clusters defined by morphology and those defined by protein expression. Whereas monocytes and PVMs belonged mostly to the rounded morphology cluster, BLM, MG1 and MG2 displayed a greater heterogeneity of morphologies (Fig. 5d), indicating morphological heterogeneity in each protein-based microglial subpopulation. Spatial mapping of segmented myeloid cells revealed different proportions of these subpopulations in GM and WM (Fig. 5e,f and Extended Data Fig. 5e). For all samples, MG1 and MG2 clusters were the most abundant, whereas monocytes were the least abundant (Fig. 5f). However, MG2 was the most predominant cluster observed in GM and MG1 was the most predominant cluster observed in WM (Fig. 5f). The expression of the markers CD74, CD68, CD11c and HLA-DR in MG1 is consistent with the activated microglial phenotype reported in WM areas during physiological aging of mice and humans^46,47. Moreover, we observed a significantly higher proportion of PVMs and monocytes in GM compared to WM. No statistically significant differences in cluster proportions were detected between AD and healthy brains.

We then explored whether the identified protein-based subpopulations were associated with specific microenvironmental features, or neighborhoods. As expected, PVMs were associated positively with vascular markers and PCNA (Fig. 5g,h and Supplementary Fig. 7a,b). Of note, we detected a significant association between monocytes and vascular markers specifically when considering their cell masks as their neighborhood (Supplementary Fig. 7a). This is due to the colocalization of monocytes with blood vessels. In GM areas, MG1 was associated positively with neuronal markers (Fig. 5g,h). The BLM subpopulation was associated significantly with Aβ, ApoE and GFAP in GM of AD brains (Fig. 5g,h). This result suggests that BLM migrate to Aβ plaques in AD brains. Of note, PVMs rather than BLM were associated with ApoE in control brains (Fig. 5g and Supplementary Fig. 7b). This could be explained by the expression of ApoE by vascular smooth muscle cells and pericytes under physiological conditions^48,49. These data indicate distinct neighborhood preferences of myeloid subpopulations throughout brain tissue, highlighting variations between AD and control subjects.

The BLM subpopulation forms the strongest association with dense Aβ plaques

Observing that certain morphological (Ramified 2 subpopulation) and protein-based microglial subpopulations (BLM subpopulation) are associated significantly with Aβ expression in their neighborhood, we next used CODEX-CNS to understand their relationship with different Aβ plaque types. We trained a machine learning model to identify Aβ plaques and differentiate them into dense and diffuse types (Fig. 6a,b). For each IBA1⁺ segmented cell, we measured the distance from its centroid to the nearest Aβ plaque (Fig. 6c). Since myeloid populations differ between GM and WM areas based on their morphology and protein expression (Figs. 4e and 5f), we analyzed their distance to Aβ plaques only in GM, where the plaques are predominant. Our results demonstrated heterogeneity in morphological clusters next to both dense and diffuse Aβ plaques, with slight variation in cluster proportions based on their distance from Aβ plaques (Fig. 6d, left panels). This indicates a weak association between morphological clusters and Aβ plaques.

Fig. 6 — a, Representative fluorescent image showing Aβ staining in the frontal cerebral cortex from an AD sample. Image is representative of n = 4 AD samples analyzed in 4 independent experiments. b, Representative images showing dense and diffuse Aβ plaque segmentations. c, Spatial map showing myeloid cells colored by their distance to the nearest Aβ plaque. d, Stacked barplots showing the average proportion of each myeloid morphology-based (left panels) and protein-based (right panels) clusters in relation to dense (top panels) and diffuse (lower panels) Aβ plaque proximity in the GM of AD brains (n = 4). e, Representative fluorescent images showing staining of IBA1, CD163 and HLA-DR markers in dense (upper panel) and diffuse (lower panel) Aβ plaques. Image is representative of n = 4 AD samples analyzed in 4 independent experiments. f, Representative images of multicolor IF labeling validating the presence of HLA-DR⁺CD68⁺ myeloid cells in proximity to dense Aβ plaques (white arrowheads). Images are representative of n = 10 AD brains. g, Protein clustering t-SNE plot with pseudotime visualization of myeloid cells. Line color represents the pseudotime values. h, Neighborhood analysis of the pseudotime values in AD (n = 4) and healthy control (n = 4) samples showing markers for vasculature (claudin-5, SMA, collagen IV), neurons (NeuN) and Aβ plaques. i, Box and whisker plots showing the mean z-scores of vascular markers (claudin-5, collagen IV) expressed in the neighborhood of each myeloid protein-based cluster identified within the GM of the human frontal cortex (n = 4 AD samples, n = 4 healthy controls). Differences were first confirmed by one-way ANOVA Benjamini–Hochberg-adjusted P value < 0.05, followed by post hoc Tukey test. Asterisks denote per-comparison significances (*P < 0.05, **P < 0.01, ***P < 0.001) obtained from the post hoc Tukey test. Each box extends from the first quartile (Q1) to the third quartile (Q3) of the data, with a line at the median. The whiskers extend from the box to the farthest data point lying within 1.5× the interquartile range from the box. Flier points are those past the end of the whiskers, although these were not excluded from statistical analysis. ANOVA P values, post hoc Tukey P values, confidence intervals and Q test statistics can be found in the corresponding source data file. j, Distance analysis of the myeloid protein-based clusters with respect to their nearest segmented blood vessel. Overall significance assessed by one-way ANOVA upon raw cell distances split by cluster, followed by Benjamini–Hochberg adjustment. Pairwise significances were calculated by post hoc Tukey test and Tukey’s Q critical values were used to determine universal 95% confidence interval widths around each group mean (see Methods for plotting details). Comparisons to BLM cluster (represented in blue) are highlighted as either significant in red (P < 0.05) or insignificant in gray (P ≥ 0.05). ANOVA P values, post hoc Tukey P values, confidence intervals and Q test statistics can be found in the corresponding source data file.

Examining protein-based myeloid cell subpopulations and distance from Aβ plaques, we found that the BLM subpopulation comprises the highest percentage of cells (~40%) in close proximity (0–5 µm) to dense Aβ plaques, with this percentage decreasing at greater distances from the plaques (Fig. 6d,e, upper right panels). Of the cells in close proximity (0–5 µm) to diffuse Aβ plaques, only ~20% were BLMs, and this proportion showed little variation based on distance from the plaques (Fig. 6d,e, bottom right panels). To further corroborate these findings, we performed conventional IF labeling on FFPE sections from ten AD and ten healthy control brains. In line with the results obtained from CODEX-CNS analysis, CD68⁺HLA-DR⁺ myeloid cells, which recapitulate cardinal features of the proteomic BLM subpopulation (Fig. 5b), are found frequently in close vicinity to dense Aβ plaques (Fig. 6f). Furthermore, CD68⁺HLA-DR⁺ myeloid cells were also present in the brain parenchyma of healthy controls (Extended Data Fig. 6a), as already indicated by CODEX-CNS analysis (Fig. 5f). Therefore, we have identified a strong spatial association between dense Aβ plaques and the BLM cluster in the brains of patients with AD.

Extended Data Fig. 6 — a, Representative images of immunofluorescence labeling depicting HLA-DR⁺CD68⁺ myeloid cells (white arrowheads) in the human brain parenchyma. Images are representative for n = 10 healthy control brains from two independent experiments. Scale Bar = 100 µm. b, Representative images of immunofluorescence labeling visualizing CD163⁺ myeloid cells in close proximity to dense Aβ plaques in AD brains (white arrowheads) as well as in contact with blood vessels (white asterisks), whereas in the healthy human brain CD163⁺ myeloid cells only colocalize with blood vessels. Images are representative for n = 10 brains per group (AD and healthy controls) from two independent experiments. Scale Bar = 100 µm. c, *CD163* expression level in human iPSC-derived microglia untreated and exposed to synthetic Aβ fibrils *in vitro*. Data obtained from Dolan et al.⁵⁰. d, *CD163* expression level in human iPSC-derived microglia transplanted in the brain of 5xFAD (a model of amyloid pathology) and wild type (WT) mice (9 months old), both in a MITRG background. Data obtained from Hasselmann et al.¹⁷. e, *CD163* expression level in human iPSC-derived microglia transplanted in the brain of *App*^NL-G-F (a model of amyloid pathology), *App*^hu/hu (the genotype control), and WT mice (6 months old), all expressing human *CSF1* on a *Rag2*/*Il2rg* deficient background. Data obtained from Mancuso et al.⁵¹.

Due to the finding that microglial and macrophage markers are coexpressed in the BLM subpopulation, we performed a pseudotime analysis on the five protein-based myeloid clusters to predict possible developmental trajectories. Our results showed a trajectory from PVM to MG2, suggesting that the PVM population might represent a lineage that develops earlier in the differentiation trajectory of the BLM subpopulation, or vice versa (Fig. 6g). In concordance with their blood-derived origin, our data indicated that the monocyte population was not part of the obtained trajectory between PVMs and MG2 (Fig. 6g). Next, we plotted neighborhood analysis scores against pseudotime scores to interrogate how cellular neighborhoods changed across the inferred trajectory. This analysis showed that as the cells transition from PVM to MG2 there is a decrease in their association with blood vessels (claudin-5, smooth muscle actin (SMA), collagen IV) concurrent with an increase in their association with neurons (NeuN) and with Aβ in the case of AD samples (Fig. 6h). These results suggest a potential phenotypic transition between PVMs and microglia depending on the brain microenvironment, whereas monocytes are separate from all others.

The neighborhood analysis performed on the protein-based clusters revealed that BLM cells were surrounded by blood vessels to a greater extent than the other two microglial populations (Fig. 6i). Moreover, distance analysis showed BLM are the closest parenchymal myeloid subpopulation to blood vessels (Fig. 6j). This is further validated by IF labeling experiments in which CD163⁺ myeloid cells were associated with dense Aβ plaques and blood vessels in AD, and with blood vessels in healthy human brain (Extended Data Fig. 6b). Altogether, our data indicate that BLM display an intimate relationship with the brain vasculature, in addition to dense Aβ plaques.

One possible explanation for this observation is that PVMs infiltrate the brain parenchyma and acquire a microglia-like phenotype. To explore this hypothesis, we evaluated whether CD163 expression is induced in microglia derived from human induced-pluripotent stem cells (iPSCs) when they are exposed to Aβ. Analyzing several published scRNA-seq datasets^50,51, we did not detect upregulation of the CD163 gene in iPSC-derived microglia either stimulated with Aβ in vitro or transplanted into the brain of AD mouse models expressing Aβ plaques (Extended Data Fig. 6c–e). Thus, given that human iPSC-derived microglia do not upregulate CD163 in response to Aβ, these results suggest that one possible origin for the BLM subpopulation in humans is infiltrating PVMs, which should be explored in the future.

Brain microenvironments drive the accumulation of specific myeloid phenotypes in the human AD cortex

Given that the microenvironment is a potent determinant of microglia and macrophage phenotypes³⁵, we next used CODEX-CNS to cluster myeloid cells based on their surrounding microenvironment. Using 13,444 IBA1⁺ segmented cells in the GM of AD brains (n = 4), we employed Leiden clustering based on the percentage of different proteins expressed within a radius of 30 μm from each cell centroid. We obtained 16 clusters (0–15 clusters) based on their neighborhood (Fig. 7a,b, Extended Data Fig. 7 and Supplementary Fig. 8a). Notably, applying different radii ranging from 20 μm to 50 μm for neighborhood analysis provided comparable results for BLM and microglia clusters (Supplementary Fig. 8b). Conversely, monocytes and PVMs showed a decreased association with blood vessels, whereas monocytes increasingly displayed an association with neuronal markers when larger radii were applied (Supplementary Fig. 8b). We further explored the protein and morphological characteristics of the cells composing these neighborhood-based clusters (Fig. 7c,d).

Fig. 7 — a, UMAP plot visualizing 13,444 myeloid cells colored by the clusters identified after applying Leiden clustering to IBA1⁺ segmented cells from the GM of human frontal cerebral cortex samples from patients with AD (n = 4) based on their neighborhood. b–d, Heatmaps show the average of z-scored median expression of the neighborhood percentage (b), protein expression (c) and morphological feature (d) values. Red rectangles highlight Aβ-associated clusters. e, Distance analysis of the myeloid neighborhood-based clusters with respect to their nearest segmented dense (right panels) and diffuse (left panels) Aβ plaque in human frontal cortex samples from human patients with AD (n = 4). Images show colored neighborhood cluster cell masks and their association with Aβ plaques. Overall significance assessed by one-way ANOVA upon raw cell distances split by cluster, followed by Benjamini–Hochberg adjustment. Pairwise significances were calculated by post hoc Tukey test and Tukey’s Q critical values were used to determine universal 95% confidence interval widths around each group mean (see Methods for plotting details). Comparisons to group of lowest mean distance (represented in blue) are highlighted as either significant in red (P < 0.05) or insignificant in gray (P ≥ 0.05). ANOVA P values, pairwise post hoc Tukey P values, confidence intervals and Q test statistics can be found in the corresponding source data file. f, Protein-based t-SNE plot visualizing cells from the GM of AD frontal cerebral cortex samples colored by the Aβ-associated clusters obtained by neighborhood-based clustering. Cells belonging to the BLM subpopulation are colored in orange, HPAM are encircled by a dashed line. g, Protein-based t-SNE plots showing a selected Aβ-associated subpopulation composed mostly of Cluster 15 in AD (left panel, n = 4) and mostly absent in control (right panel, n = 4) samples. This population was termed HPAM. Cells belonging to the HPAM subpopulation are colored in yellow. h, Representative visualization by fluorescence microscopy of CD11c⁺ (upper panel, white arrowhead) and TMEM119^low (lower panel, white arrow) myeloid cells in association with dense Aβ plaques. Of note, CD11c⁻ (upper panel, black arrowhead) and TMEM119^high (lower panel, black arrow) myeloid cells without contact to dense Aβ can be identified. Images are representative of n = 5 brains per group (AD and healthy controls).

Extended Data Fig. 7 — a, Spatial maps of the AD samples of our initial CODEX-CNS analysis (n = 4). Each dot represents a cell and is color-coded according to the neighborhood analysis-defined cell cluster to which it belongs. b, Barplot depicting the relative contribution of each AD sample to the corresponding neighborhood-defined cell cluster.

Most of the clusters had neuronal components in their neighborhood, and the myeloid cells corresponding to these clusters exhibited similar protein and morphological patterns (Fig. 7b–d). However, Clusters 3 and 11 were associated with a neighborhood significantly enriched in MAP2 and neurofilament (Fig. 7b). Clusters 4, 8, and 10 showed elevated expression of GFAP and/or Vimentin in their neighborhood, indicating that these myeloid cell subpopulations associate with astrocytes (Fig. 7b). Of note, cells from Cluster 4 were surrounded almost exclusively by GFAP and characterized by high expression of CD74, suggesting a specific interaction between astrocytes and CD74⁺ microglia (Fig. 7b,c and Supplementary Fig. 8c). This result aligns with the reported crosstalk between pSTAT3⁺ reactive astrocytes and CD74⁺ microglia/macrophages in brain metastasis⁵². Cells from Cluster 4 were located at the border of the brain tissue, which could indicate the presentation of antigens derived from the meninges by microglia to astrocytes as CD74 plays a crucial role in antigen presentation (Supplementary Fig. 8c). Among the clusters highly associated with vascular markers (Clusters 1 and 12), Cluster 12 was characterized by high expression of macrophage-related proteins (CD14, CD163 and CD11b), absence of TMEM119 and a round morphology (Fig. 7b–d), which correlates with the PVM population.

We found three neighborhood-based myeloid cell subpopulations associated significantly with Aβ: Clusters 0, 9 and 15. Cells in Cluster 0 were characterized by a neighborhood with lower Aβ and higher ApoE expression in comparison to the other Aβ-related cell clusters (Fig. 7b). By contrast, cells in Cluster 9 were in proximity to Aβ deposition that lacked ApoE (Fig. 7b). Cluster 15 comprised the cell population with the highest Aβ content in its neighborhood. In addition to Aβ, cells from Cluster 15 were surrounded by significant levels of GFAP and Vimentin and, to a lesser extent, ApoE (Fig. 7b). These data indicate an association of Cluster 15 with ApoE⁺ Aβ plaques and reactive Vimentin⁺ astrocytes. Cells from this cluster exhibited a specific phenotype characterized by high expression of CD68, CD11c, HLA-DR and CD45, as well as the highest CD163 expression after the vasculature-associated Cluster 12 (Fig. 7c). Moreover, the percentage of CD163⁺ cells in Cluster 15 was comparable to that of Cluster 12 (Supplementary Fig. 8d). In addition, cells from Cluster 15 presented the largest cell and soma sizes and thickest processes compared to the rest of the clusters (Fig. 7d). The heterogeneity of microglial cells in relation to Aβ deposition led us to further explore their relationship with different Aβ plaque types. Notably, our distance analysis revealed that Cluster 9 was the closest to diffuse Aβ plaques, whereas Cluster 15 was the closest to dense Aβ plaques (Fig. 7e).

We next asked whether the 16 neighborhood-based clusters were associated with clusters obtained previously based on protein expression (Fig. 5) or morphology (Fig. 4). Although there is heterogeneity between neighborhood-based clusters and protein-based clusters, we found that the neighborhood Cluster 4 (astrocyte-associated) was made up mostly of cells from the MG2 cluster, the neighborhood Cluster 12 (vasculature-associated) is comprised mainly of PVMs, and the neighborhood Cluster 15 (Aβ-associated) consists predominantly of BLMs (Supplementary Fig. 9a). The key difference between the proteomic signature of BLMs and cells belonging to Cluster 15 is that BLMs are TMEM119⁺CD163⁺ cells, whereas Cluster 15 cells are characterized by lower TMEM119 expression levels. Regarding morphology-based clusters, we observed that cells from the 16 neighborhood-defined clusters were distributed across all morphology-based clusters, with the rounded cluster being represented particularly in Cluster 12 (Supplementary Fig. 9b). Further exploring the Aβ-associated neighborhood-based clusters (0, 9 and 15), we found that Cluster 15, which is the cluster associated with dense Aβ plaques, represents a subcluster within the BLM population (Fig. 7f) and is enriched in human frontal cerebral cortex samples from patients with AD (n = 4) in comparison to healthy controls (n = 4) (Fig. 7g). Due to its distinct phenotype (CD163⁺CD68⁺HLA-DR⁺CD11c⁺CD45^high TMEM119^low) in comparison to BLMs (CD163⁺CD68⁺HLA- DR⁺CD11c⁺CD45^intermediate TMEM119^high) and its association with Aβ plaques in human brains, we termed this cluster human plaque-associated microglia (HPAM) (Fig. 7f,g). Of note, when comparing myeloid cells belonging to the HPAM cluster in AD cases and the very few detected in healthy controls, we observed similar morphological characteristics in both conditions, whereas HPAM in AD displayed an enrichment of Aβ and reactive astrocyte-related markers such as GFAP in their neighborhood (Extended Data Fig. 8). Similarly, HPAM in AD exhibited lower TMEM119 and higher CD68 expression levels than HPAM in healthy controls, thereby indicating nuanced differences in their proteomic signature (Extended Data Fig. 8).

Extended Data Fig. 8 — Heatmap displaying the average of z-scored median expression of the neighborhood percentage (left), protein expression (middle), and morphological feature (right) values for HPAMs in AD cases (n = 4) and healthy controls (n = 4).

To validate our findings, we employed CODEX-CNS to analyze an additional independent cohort consisting of frontal cerebral cortex samples from another four patients with AD and four healthy controls (Extended Data Fig. 9). In line with our findings in the original CODEX-CNS analysis, Leiden-based clustering of myeloid cells according to protein expression recapitulated the myeloid cell clusters identified previously, including BLM (Extended Data Fig. 9a). These myeloid cell subpopulations also displayed protein expression signatures comparable to those of the original clusters (Extended Data Fig. 9b). In addition, the BLM cluster displayed an intimate relationship with Aβ-associated markers in their neighborhood (Extended Data Fig. 9c). Similarly, combining neighborhood-based clustering with protein expression and morphological features in the resulting cell clusters provided comparable results to the original CODEX-CNS analysis, including the identification of HPAM (Cluster 15) (Extended Data Fig. 9d). This is further supported by the observation that HPAM-associated markers, such as CD11c, CD163 and HLA-DR, displayed the strongest inverse correlation with their distance to Aβ, thereby indicating that cells expressing those markers are associated closely with Aβ (Extended Data Fig. 9e).

Extended Data Fig. 9 — a, t-SNE plots of Leiden clustering based on protein expression using an independent CODEX-CNS validation cohort consisting of frontal cortex samples from n = 4 AD and n = 4 healthy controls (left plot contains cells from both conditions, right panel displays data separated by condition). b, t-SNE plots depicting IBA1⁺ segmented cells colored by protein expression. c, Heatmaps showing the association between the identified protein clusters and the proteins expressed in their neighborhood within the gray matter of the human frontal cortex from AD (n = 4) and control (n = 4) donors. Asterisks denote significances (*p < 0.05, **p < 0.01, ***p < 0.001) as obtained from two-sided Student’s t-tests comparing values within each cell cluster to values from all other cell types for a given protein, followed by Benjamini-Hochberg adjustment. Cell values derived by averaging scaled mean neighborhood metrics within each cell type-protein-sample group across samples, Z-transformation within each protein group, then clipping between -2 and +2. T-statistics, p-values, and pre-clipped cell values can be found in the corresponding source data file. Scale bar represents rescaled average z-scored protein percentages in each group. d, Heatmaps show the average of z-scored median expression of the neighborhood percentage (left), protein expression (middle), and morphological feature (right) values. Red rectangles highlight Aβ-associated clusters. e, Spearman correlation (left) and Pearson correlation test ranking the association between myeloid cell protein expression features and dense Aβ plaques in their proximity.

Finally, we sought to validate our findings on the HPAM population obtained by employing CODEX-CNS by using IF labeling on another independent cohort of eight AD and seven healthy control brain samples. We found that myeloid cells near dense Aβ plaques exhibited upregulated CD11c expression, whereas myeloid cells without contact to these plaques were CD11c⁻ in the same samples (Fig. 7h). Similarly, dense Aβ plaque-associated-myeloid cells were also characterized by lower expression levels of the homeostatic microglia marker TMEM119 (IBA1⁺TMEM119^low), whereas myeloid cells that did not interact with dense Aβ plaques were IBA1⁺TMEM119^high cells (Fig. 7h). By contrast, myeloid cells in the frontal cerebral cortex of healthy controls did not express CD11c and displayed a more homogenous TMEM119 expression pattern and morphology (Extended Data Fig. 10a).

Extended Data Fig. 10 — a, Representative images of immunofluorescence labeling on FFPE sections from healthy human brain samples. In the brain parenchyma, IBA1⁺ myeloid cells do not exhibit immunoreactivity for CD11c (white arrowheads) and homogenously express TMEM119 (white arrows). Images are representative for n = 5 body donors per staining obtained from two independent experiments. Scale Bar = 100 µm. b, Machine-learning prediction of dense (upper panel) and diffuse (bottom panel) Aβ plaques based on myeloid cells features (neighborhood, protein expression, morphology). Actual Aβ plaques values are represented in blue and predicted Aβ plaques values in green.

We next sought to empirically determine the power of myeloid cell features (morphology, protein expression and neighborhood) to predict the presence of Aβ plaques using machine learning tools. Our findings indicated that dense Aβ plaques can be predicted by the nearby myeloid cells, whereas diffuse Aβ plaques were harder to predict (Extended Data Fig. 10b). The neighborhood of myeloid cells was the best predictor of dense Aβ plaques, followed closely by their protein expression profiles (Extended Data Fig. 10b and Supplementary Fig. 10). However, the morphology of myeloid cells was a poor predictor of Aβ plaques (Extended Data Fig. 10b). The most powerful prediction was obtained combining morphology, protein expression and neighborhood (Extended Data Fig. 10b). To lend some insight into which specific myeloid cell characteristics were driving predictive power of dense Aβ plaques, we ranked all individual myeloid cell features by their correlation with dense Aβ plaques and found CD11c and CD68 expression as the strongest correlating features (Supplementary Fig. 10). Furthermore, neighborhood features such as GFAP or Vimentin were among the strongest correlating features for the prediction of dense Aβ plaques (Supplementary Fig. 10).

Therefore, our analyses using CODEX-CNS to identify morphology, protein expression profile and neighborhood of myeloid cells in frontal cerebral cortex samples from healthy controls and people with AD have identified a potentially AD-relevant microglial subpopulation that is associated with dense Aβ plaques.

Discussion

In this study, we report cell–cell interactions and characterization of myeloid cell subpopulations based on their morphology, protein expression and microenvironment in frontal cerebral cortex samples of patients with AD compared with healthy controls. Most previous studies, largely in mice, defined myeloid cell populations by transcriptional, protein or morphological data. However, CODEX-CNS enabled the integration of protein, morphology and microenvironmental features to identify spatially segregated, disease-enriched subpopulations in AD FFPE brain samples. Brain cells reside in a delicate ecosystem in which complementary and interdependent relationships between different cell types are essential for tissue functions. Current single-cell technologies do not provide critical information about the brain microenvironment under pathological conditions, which is essential for gaining new insights into cell–cell signaling cascades that orchestrate pathobiological processes^20,21. Therefore, a toolkit to study the brain microenvironment and cell–cell interactions at single-cell resolution is of particular interest. In this context, high-plex, single-cell, spatial proteomic approaches such as CODEX can address key limitations of previously established single-cell technologies in neuroscience. However, their application in brain research is complicated greatly by endogenous autofluorescence in human, aged brain samples^27,29. By modifying the existing CODEX protocol, our CODEX-CNS approach overcomes this barrier and enables high-quality, imaging-based protein profiling in FFPE brain samples. Leveraging the ability of the CODEX technology to measure more than 100 proteins at single-cell resolution, our optimized platform provides a fundament for future studies disentangling the molecular and context-dependent cell heterogeneity in the human CNS.

By applying CODEX-CNS on frontal cerebral cortex samples from patients with AD and healthy controls, our protein-based clustering analysis of myeloid cells identified a microglial subpopulation (BLM cluster) characterized by a BAM-like phenotype. One intriguing observation was a specific association of the BLM cluster with astrocytes, ApoE and dense Aβ plaques in AD brains. However, a similar proportion of the BLM cluster was observed in both patients with AD and healthy controls, indicating that BLM is an aging-associated microglial subpopulation that migrates to the Aβ plaque sites in AD or, potentially, that Aβ plaques develop around them. The detection of neurodegenerative disease-associated microglial populations such as DAM has also been reported previously in the healthy brains of mice^9,53. However, similar to what we observed with the BLM population, these signatures are specifically localized around Aβ plaques when they are present.

Our data indicated that the BLM subpopulation expresses CD163, which was previously identified as a marker of BAMs in CNS-associated tissues^54–56. Furthermore, several studies have reported a CD163⁺ microglial subpopulation near Aβ plaques in AD brains^32,33. Muñoz-Castro and colleagues recently suggested that these CD163⁺ parenchymal cells are infiltrating monocytes rather than microglia³⁴. This contrasts with the traditional view that blood-derived cells do not infiltrate AD brain, as shown in parabiosis studies of 5xFAD mice⁵⁷. However, to our knowledge, CD163⁺ cells have not been reported in aged or AD mouse brains, suggesting species differences in myeloid infiltration or microglial states. Because this CD163⁺ myeloid population seems human-specific, its origin cannot be investigated with parabiosis or fate mapping. The cell trajectory observed in our pseudotime analysis indicates that BLM represent a transitional state between a PVM- and a microglia-like phenotype, which is supported by our analysis of published scRNA-seq datasets demonstrating the absence of CD163 upregulation in human iPSC-derived microglia exposed to Aβ. Nevertheless, given their shared embryonic origin⁵⁸ and their similar transcriptional and proteomic profiles, PVMs and microglia might respond similarly to local vascular or parenchymal Aβ accumulation, thereby limiting the resolution of our pseudotime analysis using our current CODEX-CNS panel. Further research is needed to decipher the source of parenchymal CD163⁺ cells in the human brain and their role in Aβ pathology.

A key strength of our CODEX-CNS protocol was integrating microenvironmental context into brain myeloid cell clustering. Unlike scRNA-seq, CODEX-CNS enables simultaneous single-cell analysis of protein expression, morphology and local neighborhood. Our analysis revealed a cell cluster associated with Aβ, ApoE and Vimentin⁺ astrocytes, which is defined by a specific protein signature and morphology and which we termed HPAM. The big soma size and thick processes observed in HPAM could reflect a high phagocytic activity in the Aβ plaque site. In terms of protein expression, HPAM combined DAM (CD11c), antigen presentation (HLA-DR), phagocytic (CD68) and macrophage (CD163) markers, along with high expression of the pan-leukocyte marker CD45. The simultaneous expression of markers related to different microglial signatures reinforces the lack of consensus on an AD-associated microglial signature in humans. Notably, HPAM represent a subpopulation within the BLM cluster, suggesting that cell populations traditionally identified by a limited number of markers could represent a yet unknown combination of different subpopulations. Although the BLM cluster was found in both healthy donors and donors with AD, the HPAM subcluster, which is molecularly distinct from the BLM core signature, was largely enriched in AD brains and associated with Aβ plaques. Thus, these data demonstrate that microenvironmental context can reveal disease-associated cell subpopulations missed by conventional protein-based clustering analysis.

The association of cells with a certain neighborhood, along with their protein expression and morphology, suggests their functionality. For example, the high expression of CD68 in HPAM indicates that these cells may play an important role in Aβ phagocytosis. Conversely, HLA-DR expression in cells surrounded by Aβ might suggest antigen representation to infiltrating T cells. Our results demonstrated that HPAM cells were associated with dense, but not diffuse, Aβ plaques, which are also associated with abnormal neurites. Thus, the putative phagocytic and antigen-presenting functions of HPAM cells could be elicited in response to neuritic pathology rather than Aβ deposition itself. The potential importance of HPAM cells to AD pathology is supported further by the findings of a multiomic study characterizing microglial Aβ-clearance in the brain of a patient with AD treated with the Aβ-targeting antibody Lecanemab, in which a pronounced upregulation of HPAM-associated markers in Aβ-enriched areas was observed⁵⁹. This could point to a role of HPAM in Aβ-clearance, which might be exploited therapeutically in the future. Therefore, CODEX-CNS could facilitate considerably the investigation of the intimate relationship between myeloid cells and hallmarks of AD, thereby promoting the development of new therapeutics for this still incurable neurodegenerative disease.

Overall, CODEX-CNS, which bridges single-cell signatures at the protein level with their spatial location and morphology, represents a powerful method for investigating the cytoarchitecture of the human brain that will provide new insights into the cellular mechanisms underlying neurological diseases with unprecedented resolution.

Methods

Postmortem brain samples

This study used brain frontal lobe sections, containing both GM and WM areas, from patients with AD (n = 4 for initial analysis, n = 4 for CODEX-CNS validation experiments) and people without manifested neurological pathology (n = 4 for initial analysis, n = 4 for CODEX-CNS validation experiments) (Supplementary Tables 1 and 2). All subjects were donors to the Oregon Brain Bank following consent of next of kin and deidentification of tissues as stipulated by the Institutional Research Board (IRB) to conform to the requirements of nonhuman subject research on human postmortem tissue. Samples were allocated to the experimental groups (healthy controls versus AD) based on their clinical and neuropathological diagnosis. No statistical methods were used to predetermine sample sizes, but our sample sizes were similar to those reported in previous publications^8,46,60. Investigators were blinded during staining, image acquisition and part of the data analysis. Brain fixation was performed by immersion in 10% neutral buffered formaldehyde solution at room temperature (RT) for a minimum of 2 weeks with periodic inspection until complete. Fixed tissue was dissected and processed into paraffin blocks. Tissue blocks were then sectioned at 7-μm thickness with a microtome and placed directly on poly-l-lysine coated coverslips. Neuropathologic diagnosis was performed as described previously⁶¹. Our data did not show any variance in staining quality or myeloid subpopulations correlating with the postmortem interval time of the samples (Supplementary Fig. 11).

For IF validation experiments focusing on the BLM population (Fig. 6f and Extended Data Fig. 6a,b), frontal cerebral cortex sections from ten patients with AD and ten healthy people were provided by the Oregon Brain Bank. Furthermore, IF-based validation of HPAMs in human AD (Fig. 7h and Extended Data Fig. 10a) was carried out in frontal cerebral cortex samples that were obtained from eight patients with AD and seven healthy donors provided by the brain bank of the Institute of Anatomy at Leipzig University (Supplementary Table 3). All body donors provided written, informed consent, which was secured by contract during lifetime, and no other body donor information then sex, age and cause of death were disclosed (Supplementary Table 3). After tissue removal, human brain samples were fixed in 4% paraformaldehyde, dehydrated and embedded in paraffin.

Design of CODEX-CNS antibody panel

Our antibody panel is listed in Supplementary Table 4. The panel consisted of 32 antibodies that were organized into modules designed to label neurons, glial cells, immune cells, vasculature and pathological features of AD (Fig. 1b). Individual antibody staining data are provided in Supplementary Fig. 2.

For detection of brain-resident and peripheral immune cells, we used antibodies against the leukocyte common antigen (CD45) and, more categorically, against general markers of myeloid-lineage cells (CD11b) and T cells (CD3e). Antibodies against CD4 and CD8 were also included to distinguish between T helper and T cytotoxic cells, respectively. In addition to CD45 and CD11b markers, IBA1, TMEM119, CD163, CD74, CD68, CD11c, CD14 and HLA-DR antibodies were used for the characterization of the macrophage/microglia population. IBA1 is a general marker for both microglia and macrophages, whereas TMEM119 is a highly expressed microglia-specific marker⁶². CD163 is a scavenger receptor specifically expressed in monocytes/macrophages. Increased expression of CD74, CD68, CD11c, CD14 and HLA-DR markers has been associated with microglial activation in different neurological pathologies such as AD⁹. Briefly, CD74 acts as a chaperone for major histocompatability complex class II molecules as well as macrophage migration inhibitory factor receptor⁶³, CD68 is a lysosomal marker expressed highly in phagocytic cells and expression of CD11c in microglia has been linked to phagocytosis and insulin-like growth factor 1 induction⁶², CD14 is involved in microglial responses to pathogen- and damage-associated molecular patterns acting as a TLR4 coreceptor. HLA-DR complex present antigenic peptides to T helper cells; however, their function in microglia remains undefined⁶⁴.

To identify neurons, we used antibodies against NeuN, MAP2 and neurofilament. Whereas NeuN staining is restricted to the nucleus, MAP2 and neurofilament proteins form part of the neuronal cytoskeleton and are expressed in dendrites and axons. These markers allowed us to differentiate easily between GM and WM areas. Moreover, we used synaptophysin (also known as the main synaptic vesicle protein p38) and PSD95 antibodies to detect presynaptic and postsynaptic components, respectively.

Oligodendrocytes/OPCs were identified using an antibody against OLIG2—a transcription factor presents in precursors and mature oligodendrocytes.

For the detection of astrocytes, GFAP was used due to its canonical expression in the astrocytic cytoskeleton. Aquaporin-4 (AQP4), S100B and Vimentin were additional markers included in the antibody panel for astrocytic visualization and characterization. S100B is a calcium-binding protein expressed predominantly in astrocytes and associated with Aβ and tau aggregates in the AD brain⁶⁵. AQP4—the main water channel expressed in the CNS—is expressed densely in astrocyte end-feet and visualized in meninges. Vimentin is a type III intermediate filament that in the adult brain is expressed in astrocytes, ependymal cells and endothelial cells. Thus, this marker allowed for the detection of both astrocytes and vascular cells.

CD31 (a platelet endothelial cell adhesion molecule), claudin-5 (a component of tight junction strands), collagen IV (a component of the basal lamina) and SMA were used as vascular markers. SMA staining was observed only in blood vessels containing smooth muscle cells or pericytes.

Our antibody panel also included markers whose expression is related to pathological features, such as Aβ, ApoE, PCNA, P2RX7 (P2X purinergic receptor for ATP) and γH2A.X (H2A histone family member X).

Tissue preparation and staining

CODEX-CNS

All data in this study were generated following the commercially available CODEX protocol (https://www.akoyabio.com/wp-content/uploads/2021/01/CODEX-User-Manual.pdf). Below, we provide an abbreviated version of this protocol with an emphasis on modified steps that were taken to establish CODEX-CNS.

Poly-l-lysine coated coverslips with FFPE sections were baked for 40 min at 55 °C on a slide warmer and dewaxed by incubating twice in HistoChoice Clearing Agent for 5 min. The FFPE sections were then rehydrated by sequentially incubating them in 100%, 90%, 70%, 50% and 30% ethanol each time for 5 min, followed by three washes in distilled water for 5 min each. Heat-induced epitope retrieval was then performed by incubating coverslips in a beaker with Tris-EDTA buffer (pH 9) in a pressure cooker for 20 min. Following heat-induced epitope retrieval, the beaker with coverslips was removed from the pressure cooker and equilibrated at RT for about 30 min. Coverslips were then incubated in phosphate-buffered saline (PBS) three times for 5 min.

We used a slightly modified protocol from the one published by Du and colleagues⁶⁶ to reduce tissue autofluorescence in our samples (Supplementary Fig. 1a). Specifically, coverslips were placed in a freshly prepared bleaching solution (4.5% (w/v) H₂O₂ and 20 mM NaOH in PBS) in a six-well plate and exposed to an ultrabright LED light panel (Aibecy A4 Ultra Bright 25,000 Lux LED Light Box-Tracing Pads) positioned 2 cm away from the plate for 45 min at RT. Samples were then transferred to a fresh well of bleaching solution and exposed to the LED light for another 45 min at RT. Coverslips were then washed in PBS three times for 5 min. Finally, samples were equilibrated by first placing them in Akoya Hydration buffer twice for 2 min and once more in Akoya Staining buffer for 20 min at RT.

For staining, a cocktail with the antibodies listed in Supplementary Table 4 was prepared as recommended by the Akoya protocol. Staining was performed in a humidity chamber overnight at 4 °C before coverslips were washed twice in Akoya Staining buffer for 2 min and postfixed in 1.6% paraformaldehyde in Akoya Storage for 10 min. After three washes in PBS, coverslips were incubated in ice-cold methanol for 5 min at 4 °C, and washed again in PBS, three times for 5 min. Finally, coverslips were placed in Akoya fixative reagent for 20 min in a humidity chamber, followed by three washes in PBS.

Immunofluorescence

For IF labeling, FFPE sections of human frontal cerebral cortex were deparaffinized following a standard protocol. Afterwards, sections were blocked with blocking buffer containing 2% bovine serum albumin (BSA), 0.1% Triton X-100 and 2% normal goat serum (NGS) or normal donkey serum (NDS), respectively, in PBS for 60 min at RT. Subsequently, sections were incubated with primary antibodies diluted in PBS containing 0.1% Triton X-100, 2% BSA, and 2% NGS or NDS, overnight at 4 °C. The following primary antibodies were used for IF labeling in this study: anti-CD11c (Abcam, catalog number ab52632; dilution: 1:500), anti-IBA1 (Abcam, catalog number ab5076; dilution: 1:500), anti-Aβ (Sigma–Aldrich, catalog number A8354; dilution: 1:5,000; BioLegend, catalog number 856501; dilution: 1:200), anti-TMEM119 (Abcam, catalog number ab185333; dilution; 1:500), anti-CD68 (ThermoFisher Scientific, catalog number 14-0688-80; dilution: 1:200), anti-HLA-DR (Abcam, catalog number ab92511; 1:200), anti-CD163 (Cell Signaling Technology, catalog number 93498; dilution: 1:200) and anti-collagen IV (Synaptic Systems, catalog number 462 004; dilution: 1:200). Primary antibodies were omitted for negative controls. Following extensive washing with 0.1% Triton X-100, 2% BSA and 0.2% NGS/NDS in PBS, sections were incubated with fluorophore-labeled secondary antibodies diluted 1:500 in 0.1% Triton X-100, 2% BSA and 0.2% NGS/NDS in PBS at RT for 90 min in the dark. The following secondary antibodies were used for IF labeling in this study: Donkey anti-guinea pig CF770 (Biotium, catalog number 20242-1; dilution: 1:500) Donkey anti-rat DyLight 755 (Invitrogen, catalog number SA5-10031; dilution: 1:500), Donkey anti-goat Alexa Fluor 647 (Invitrogen, catalog number A21447; dilution: 1:500), Donkey anti-rabbit Alexa Fluor 568 (Invitrogen, catalog number A10042; dilution: 1:500), Donkey anti-mouse Alexa Fluor 555 (Invitrogen, catalog number A31570; dilution: 1:500), Donkey anti-rabbit Alexa Fluor 488 (Invitrogen, catalog number A21206; dilution: 1:500) and Donkey anti-mouse Alexa Fluor 488 (Invitrogen, catalog number A21202; dilution: 1:500). After washing at least three times with 0.1% Triton X-100, 2% BSA and 0.2% NGS/NDS in PBS, slides were counterstained with 4′,6-diamidino-2-phenylindole (DAPI) 1:10,000 for 10 min, washed three times with PBS followed by autofluorescence quenching with TrueBlack Lipofuscin Autofluorescence Quencher (Biotium, catalog number 23007) according to the manufacturers’ instructions.

Tissue imaging

CODEX

Imaging was performed with an Akoya Biosciences PhenoCycler connected to a Keyence BZ-X800 epifluorescence microscope. Before imaging, 96-well plates were prepared using the cycle reporter configuration as outlined in the CODEX user manual. Imaging was performed with a ×20 objective (Nikon CFI Plan Apo ×20/0.75 numerical aperture (NA)); an 8 ×8 mm large region was identified initially using the stitching option in BZ-X800 viewer software (v.1.1.1) and z-stacks were then acquired throughout this region with a total depth of up to 10 μm each. The multiplex cycles were set up using Akoya’s CODEX Instrument Manager (v.1.29.3.6) and acquired images were then processed with the CODEX Processor to perform cycle alignment, background subtraction, deconvolution, extended depth of field, shading correction, tile registration and stitching. The resulting Qptiff image files were inspected manually for quality and uniform staining using QuPath software and then analyzed using a custom workflow developed around QuPath, ariadne.ai and Scanpy as described below in ‘Myeloid cell clustering based on protein expression.’

Microscopy

Slides were imaged using a confocal Fluoview FV1000 (Olympus) equipped with 0.75 NA U Plan S Apo ×20 and ×40 0.95 NA U Plan S Apo ×40 2 (Olympus, Japan) or a Zeiss Axioscan 7 slide scanner (Zeiss) equipped with a ×20 0.80 NA Plan-Apochromat (Zeiss). For confocal images, z-stacks were used subsequently to generate projection images using maximum intensity projections over the z-axis.

Image analysis

The first step of CODEX-CNS image analysis included manual inspection and quality control of processed imaging data, followed by annotation of tissue regions and segmentation of single cells.

Tissue annotation

Regional annotations of GM and WM areas were performed manually using the QuPath annotation tool following antibody staining patterns that were readily visible in QuPath (v.0.3.2). The segmented cells, whose centroids fell within these expert-annotated regions, were assigned to the respective brain areas. Cells not assigned to any region were excluded from further analysis (that is, cells in the meninges).

Cell segmentation

Cell instance segmentation was performed using the ariadne.ai SPATIAL platform (https://ariadne.ai/spatial), which offers deep-learning-based cell segmentation models for glia and neurons, and calculates morphological metrics for each cell (‘Morphological analysis’). NeuN, GFAP, OLIG2 and IBA1 markers were selected to segment neurons, astrocytes, oligodendrocytes/OPCs and microglia/macrophages, respectively. The accuracy of all segmented cell types was assessed meticulously by a neurohistological expert during the segmentation process. With the purpose of capturing cells with soma and processes located in different imaging planes from the cell nucleus, DAPI expression was not considered as a necessary parameter in the cell segmentation algorithm. For microglia/macrophages, we also implemented a custom Otsu’s Thresholding-based Segmentation and Merge open-source algorithm, which consists of the following main steps: (1) adjusting the brightness and contrast for each myeloid marker manually to account for any background or tissue autofluorescence signal; (2) creating a composite single-channel grayscale image of the cell containing the maximum pixel value (signal intensity) across the markers obtained in the previous step; (3) applying a background cutoff of 30 (that is, any pixel value less than 30 is considered to be background); (4) applying Otsu’s thresholding method to convert the grayscale image into a binary image, where 0 corresponds to background and 1 corresponds to the microglial cell mask; (5) computing areas of connected regions of 1 s (detected objects) with a single mask, and discard any objects consisting of less than 50 µm² in connected area; (6) merging objects whose centroids are closer than 10 µm using a line, which often corresponds to cell processes that are in different focal regions; (7) dilating the binary mask by 1 pixel and fill any holes to obtain a contour around the detected objects and (8) assigning a unique cell label for each contiguous detection. To exclude monocytes, IBA1⁺ segmented cells within vasculature were filtered by removing cell masks with a collagen IV signal intensity value of greater than 50. Cell masks lacking DAPI staining were also removed.

Blood vessel segmentation

Blood vessels were segmented by the following steps: (1) establishing a minimum intensity threshold for collagen IV positive staining per-sample, (2) binarizing the collagen IV staining, (3) segmenting connected pixels using skimage’s measure.label function (https://scikit-image.org/docs/stable/api/skimage.measure.html#skimage.measure.label) and (4) establishing a minimum threshold of 100 pixels of collagen IV⁺ segmented area.

Aβ plaque segmentation

Aβ plaques are categorized in different types, being dense-core (fibrillar) and diffuse (nonfibrillar) the most common classification¹. In diffuse plaques, Aβ accumulation is spread or scattered widely and they lack an amyloid core. On the other hand, dense plaques have a condensed core of Aβ surrounded by an optically clear region with little Aβ, and then an outer corona of more diffuse Aβ. Using a machine learning approach, we were able to discriminate between dense and diffuse Aβ plaques (Fig. 6b). We filtered the Aβ fluorescent signal with a 5 × 5 Gaussian kernel and created plaque masks by thresholding at the 95th pixel intensity percentile. We manually labeled 393 plaques as ‘dense’ or ‘diffuse.’ Using the labeled masks, we trained a random forest classifier for binary classification from the mean and standard deviation of Gabor-filtered masked image patches (frequencies: 0.2, 0.4, 0.6, 0.8; angles: 0, π/4, π/2, 3π/4 radians).

Cell interaction analysis

With the aim to capture cell–cell interactions in close proximity, we selected the xy coordinates corresponding to the centroids of each segmented cell and measured the frequency of other cell centroids in a small radius of 15 μm (the average radius of a prefrontal neuronal soma is 7–8 μm⁶⁷) (Fig. 3a). In addition to that, we included in the analysis the xy centroid coordinates of blood vessels to see interactions between cells and vasculature (Fig. 3a). Cell centroid distances and neighborhood frequencies were calculated using the Scimap python module (https://github.com/labsyspharm/scimap), and the sm.tl.spatial_interaction function was used to calculate the likelihood of cell types being adjacent to each other by comparing to a permuted background. Cell interaction results were represented by network chord diagrams (Fig. 3b), where the line color represents greater than expected (red) or lesser than expected (blue) likelihood of interaction assuming a random spatial distribution of cell types, and the line thickness represents the deviation of the likelihood from random. Positive interactions were defined as interactions occurring more frequently than expected given a random spatial distribution of cells, and negative interactions were defined as interactions occurring less frequently than expected given a random spatial distribution of cells. To assess differences in cell interaction frequencies between control and AD samples, likelihood ratios were calculated for each sample group according to the method established by Schürch and colleagues²⁵ and compared by a two-tailed Student’s t-test (n = 2), with resulting t statistics and significances plotted in Fig. 3d.

Myeloid cell clustering based on protein expression

To identify myeloid cell types, we considered the protein expression per cell of the following ten markers: IBA1, TMEM119, CD74, CD14, CD163, CD68, CD11b, CD11c, HLA-DR and CD45, and performed unsupervised clustering. This workflow involved the following steps:

Protein quantification: within each segmented IBA1⁺ cell, the average intensity of the 10 markers listed above was computed.
Normalization: for each sample and each marker, the expression values in each cell were first log-transformed using log(1+p). For all samples, we then performed batch-effect removal using Scanpy’s Combat function (https://github.com/brentp/combat.py), which uses an empirical Bayes framework to account for batch-effects^68–70. Here the parameter we used to remove batch-effects was the sample ID. We then normalized the batch-effect corrected expression values using z-score normalization with a maximum value of 10.
Unsupervised clustering: to obtain cluster labels, we started by computing the cell neighbor graph based on expression similarity, and then ran Leiden clustering based on the cell neighbor graph. Silhoutte and Carlinski scores, both calculated using the scikit-learn version 0.24.2 metrics toolkit (https://github.com/scikit-learn/scikit-learn), were considered to determine the resolution used for the clustering analysis. Attending to the highest scores, initial clusters were obtained using a resolution of 0.2, whereas subclusters from the two main clusters (A and B) were obtained using a resolution of 0.2 and 0.3, respectively. Clusters were visualized with a lower dimensional embedding of cells using t-distributed stochastic neighbor embedding (t-SNE).
Cluster annotation: for each cluster, we assigned a cell type label based on their protein expression and tissue location as follows:

If a cluster was positive for TMEM119 and located in the brain parenchyma, it was assigned the label ‘microglia.’
If a cluster was negative for TMEM119 and positive for CD163, and located in the perivascular space, it was assigned the label ‘perivascular macrophage.’
If none of the mentioned two markers were expressed in the cluster and the cells were located within blood vessels, it was assigned the label ‘monocyte.’

(5)
Biological validation: all single-cell analyses and visualizations were performed using the Python-based Scanpy tool⁶⁸ v.1.9.1, with RapidsAI backend for GPU-accelerated computing.

Clustering of all cells was accomplished in a similar fashion to clustering of myeloid cells; however, datasets were integrated with Scanpy’s bbknn function (https://github.com/Teichlab/bbknn), rather than Combat, and visualized with uniform manifold approximation and projection (UMAP), also implemented through scanpy.

Spatial neighborhood analysis

We implemented a custom neighborhood analysis measuring protein expression percentage of different markers in a 30-μm radius from the centroid of each IBA1⁺ segmented cell (Fig. 4f). The process consists of the following steps: (1) for each image protein channel, a manual threshold was set for protein expression intended to reduce background noise; (2) all pixels below this value were set to 0, the resulting matrix was then binarized by setting all nonzero pixel expression levels to 1; (3) for each myeloid cell, a 30-μm radius was drawn around the cell segmentation’s centroid; (4) for each radius drawn, the number of pixels assigned as positive for a given marker was divided by the total number of pixels considered: those pixels which were both within the radius and not belonging to the mask of the segmentation were considered. Although we opted to exclude any pixels belonging to the segmentation mask of the centroid in question, we separately explored the protein expression of pixels lying directly beneath each cell mask in Supplementary Fig. 7a and (5) we applied a z-score transformation across all cells within each sample, to reduce discrepancies in remaining post-threshold noise.

Following this process, we aggregated data from all samples and calculated per-sample mean z-scores for each profiled protein. These values were compared by one-way analysis of variance (ANOVA) and post hoc Tukey’s HSD, both implemented in scipy.stats v.1.9.1 (https://github.com/scipy/scipy). We also compared mean cell type scores by performing one-to-all two-sided Student’s t-tests, which were also implemented through scipy.stats, between per-sample protein scores within a given cell type and per-sample protein scores averaged among other cell types. Significances, as well as rescaled per cell type mean z-scores were plotted on heatmaps generated using seaborn v.0.12.2 (https://github.com/mwaskom/seaborn).

Morphological analysis

In the IBA1⁺ segmented cells, we analyzed the following 12 cell morphological features/metrics: cell perimeter (μm), cell area (μm²), cell circularity, cell solidity (cell area to convex hull area ratio), soma area (μm²), soma circularity, number of processes (arise from the cell soma), number of branches (ramifications of the processes; arise from the processes), branching degree (branches to processes ratio), branch proximity: distance from the center of the cell soma to the closest branch (μm), average length of processes (μm) and average thickness of processes (μm). Extended Data Fig. 4a shows the features needed (cell segmentation, soma segmentation, convex hull, processes detections and branches detections) to compute these morphological metrics.

Myeloid cell clustering based on morphology

IBA1⁺ segmented cells were clustered initially using the Leiden algorithm based on the above explained 12 morphological parameters (resolution = 0.2), obtaining three superclusters (Rounded, Intermediate and Ramified). Subclustering analysis was performed in the Ramified cluster (resolution = 0.3). Clusters were visualized with a lower dimensional embedding of cells using t-SNE.

Distance analysis

The distances between the IBA1⁺ cell centroids to the closest dense and diffuse Aβ plaque centroids were computed by supplying segmentation cell centroids to scipy’s cdist function. Distances to plaques were first compared by separating cells into bins based on distance interval, with distance intervals increasing in size. Distances were compared statistically between cell types using a one-way ANOVA test and post hoc Tukey comparison, offered by scipy (https://github.com/scipy/scipy) and statsmodels v.0.13.2 (https://github.com/statsmodels/statsmodels), respectively. Results were displayed using a the TukeyHSDResults.plot_simultaneous method offered through statsmodels, which plots a single confidence interval per group informed by the Tukey’s Q critical value. Comparisons between the interest, which was colored blue, and all other groups were highlighted; overlapping groups (insignificant differences) were colored gray, and nonoverlapping groups (significant differences) were colored in red.

Pseudotime analysis

scFates v.1.0.7 (https://github.com/LouisFaure/scFates) was employed for pseudotime analysis on the five myeloid subpopulations (MO, PVM, BLM, MG1 and MG2) identified after applying clustering analysis based on protein expression. Untransformed protein expression data was fed into scFates’ tl.curve function. A principal curve was drawn specifying 15 nodes and an epg_lambda of 0.5. The resulting curve was rooted arbitrarily in the PVM cluster, as opposed to the MG2 cluster (Fig. 6g). Cells were binned by resulting pseudotime values in intervals of 100, and neighborhood feature percentages were averaged across cells in each bin and plotted (Fig. 6h).

Myeloid cell clustering based on neighborhood

IBA1⁺ segmented cells were clustered using the Leiden algorithm (resolution = 1) based on their neighborhood. The neighborhood of the myeloid cells was determined as explained in ‘Spatial neighborhood analysis.’ Clusters were visualized with a lower dimensional embedding of cells using UMAP-learn v.0.5.2. Of note, all neighborhood-defined cell clusters could be detected in each brain sample used in the original CODEX-CNS analysis (Extended Data Fig. 7b). Extended Data Fig. 7a provides color-coded images of the distribution of the obtained neighborhood clusters in all samples from the original CODEX-CNS dataset.

Aβ plaque prediction based on myeloid cell features

The variable we sought to predict was minimum–maximum normalized dense and diffuse Aβ plaque segmentations within a 30-μm radius from the centroid of each IBA1⁺ segmented cell. This metric was calculated much in the same way protein expression in myeloid cell neighborhoods was, exchanging binarized protein expression for the actual dense and diffuse Aβ segmentation masks. The feature data was calculated previously, and could be split into three broad categories: cell protein expression, cell morphology and cell neighborhood (protein expression within a 30-μm radius from the cell centroid). From this last category, we excluded the features Aβ and ApoE, as they would essentially act as near-perfect predictors of the response variable. We chose to limit this exploration to only the GM region of patients with AD, where Aβ plaques are found most commonly.

The first step we took to assess feature predictive power was to rank all features by their absolute Spearman correlation with the response variable. This was accomplished by first generating correlation scores with the pandas.DataFrame.corrwith method—specifying ‘spearman’ as the method—then ranking said correlations by their absolute values (https://github.com/pandas-dev/pandas).

The next step we took to assess feature predictive power was to determine empirically which feature categories conferred the greatest predictive power using a machine learning-powered analysis. We used a Random Forest Regressor model available through scikit-learn to predict either dense or diffuse Aβ plaques using each individual feature category as well as all three together (https://github.com/scikit-learn/scikit-learn). After obtaining predicted scores, we assessed feature category predictive strength by calculating and comparing the coefficient of determination (R squared) between runs, which was calculated using scikit-learn’s r2_score function. Finally, to unclutter the predicted score plots, we averaged each predicted score with the ten preceding and ten following predicted scores, ordered by actual Aβ plaques. These averaged scores were used only for visualization purposes, and not in the calculation of the coefficient of determination.

Statistics

Statistical analyses between two groups were performed using unpaired Student’s t-tests. One-way or two-way ANOVA with Tukey’s multiple comparisons test was performed when required. Assumption of normality was assessed by Shapiro–Wilk test as implemented in Scipy’s shapiro function (https://github.com/scipy/scipy). Specific statistical tests performed are explained in each corresponding Methods subsection. In addition, statistical test details are reported in the figure legends. Statistical significance was considered when P < 0.05 for all of the tests performed, unless otherwise indicated in the figure legend. No datapoints were excluded from the analyses.

Reporting summary

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

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41593-026-02267-3.

Supplementary information

Supplementary Information^{(10.8MB, pdf)}

Supplementary Tables 1–4 and Figs. 1–11.

Reporting Summary^{(93.9KB, pdf)}

Supplementary Data 1^{(23.9KB, xlsx)}

Statistical source data for Supplementary figures.

Source data

Source Data^{(63.8KB, xlsx)}

Statistical source data for main and extended data figures.

Acknowledgements

We express our gratitude to all tissue donors and their families for their kind and selfless gift. We would like to thank the Layton Aging and Alzheimer’s Disease Center for providing us with brain tissue samples collected with the support of grant no. P30AG008017. Furthermore, we thank I. Bechmann and the brain bank of the Institute of Anatomy at Leipzig University for providing brain tissue samples for IF validation experiments. We thank members of the J. Wherry laboratory at UPENN for input and constructive advice related to the autofluorescence quenching protocol as well as our colleagues at Synaptic Systems and ABCAM for supporting our antibody panel development. We thank B. Ben Cheikh, J. Singh, A. Crotti and A. Dhawan for their contribution to experimental design and data analysis. We thank A. Bardacke, D. P. Wolf and A. Bayat for their generous donation to the Ajami lab. Funding for this research was provided by the Alzheimer’s Association (AARGD-22-928829), Collins Medical Trust in Portland, Oregon, Oregon Alzheimer’s Disease Research Center grant no. P30-AG066518 and Oregon citizens through the Alzheimer’s Disease Research Fund of the Oregon Charitable Checkoff Program to B.A. and by Akoya Biosciences to O.B. D.-D.R. and P.W. were supported by the Medical Faculty of the University of Augsburg. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Extended data

Author contributions

B.A. conceived and oversaw this study. O.B. designed and oversaw the CODEX optimization and operation. A.C. provided input in design of microglial markers in the panel. D.-D.R., H.K., P.S.-M., A.P., R.W., P.W., O.B. and B.A. provided experimental and data analysis design. D.-D.R., C.G., Y.P., A.W., Y.H., H.K., N.N. and A.B. performed experiments. D.B., M.M., A.P., E.L.B. and F.S. designed and implemented bioinformatic analysis. D.-D.R., P.S.-M., P.W., O.B., R.W. and B.A. interpreted the data. D.B., D.-D.R., P.S.-M., F.S., P.W., O.B. and B.A. wrote the paper. All authors contributed to editing and reviewing the paper.

Peer review

Peer review information

Nature Neuroscience thanks Rong Fan, Ajit Johnson Nirmal, Marta Olah and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Data availability

The raw datasets generated for this study are available via Zenodo at 10.5281/zenodo.18903415 (ref. ⁷¹), 10.5281/zenodo.18878506 (ref. ⁷²), 10.5281/zenodo.20725082 (ref. ⁷³), 10.5281/zenodo.20706335 (ref. ⁷⁴) and 10.5281/zenodo.20726162 (ref. ⁷⁵). Source data are provided with this paper.

Code availability

Computer code for this project can be provided upon reasonable request and found at https://github.com/BaharehAjami/CODEX-CNS.

Competing interests

M.M. and F.S. were paid employees and/or shareholders of Ariadne.ai at the time that this work was conducted. N.N., A.P. and O.B. were paid employees and/or shareholders of Akoya Biosciences at the time that this work was conducted. The other authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Paula Sanchez-Molina, Dennis-Dominik Rosmus, Dillon Brownell.

Change history

7/8/2026

In the version of this article initially published, only two of five datasets were listed in the Data availability section, as is now amended in the HTML and PDF versions of the article.

Contributor Information

Oliver Braubach, Email: obraubach@icloud.com.

Bahareh Ajami, Email: ajami@ohsu.edu, Email: Bahareh.Ajami@cshs.org.

Extended data

is available for this paper at 10.1038/s41593-026-02267-3.

Supplementary information

The online version contains supplementary material available at 10.1038/s41593-026-02267-3.

References

1.Serrano-Pozo, A., Frosch, M. P., Masliah, E. & Hyman, B. T. Neuropathological alterations in Alzheimer disease. Cold Spring Harb Perspect Med1, a006189 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Mathys, H. et al. Single-cell transcriptomic analysis of Alzheimer’s disease. Nature570, 332–337 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Wang, Q. et al. Single cell transcriptomes and multiscale networks from persons with and without Alzheimer’s disease. Nat. Commun.15, 5815 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Mathys, H. et al. Single-cell multiregion dissection of Alzheimer’s disease. Nature632, 858–868 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Ajami, B. et al. Single-cell mass cytometry reveals distinct populations of brain myeloid cells in mouse neuroinflammation and neurodegeneration models. Nat. Neurosci.21, 541–551 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Siletti, K. et al. Transcriptomic diversity of cell types across the adult human brain. Science382, eadd7046 (2023). [DOI] [PubMed] [Google Scholar]
7.Batiuk, M. Y. et al. Identification of region-specific astrocyte subtypes at single cell resolution. Nat. Commun.11, 1220 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Bottcher, C. et al. Human microglia regional heterogeneity and phenotypes determined by multiplexed single-cell mass cytometry. Nat. Neurosci.22, 78–90 (2019). [DOI] [PubMed] [Google Scholar]
9.Keren-Shaul, H. et al. A unique microglia type associated with restricting development of Alzheimer’s disease. Cell169, 1276–1290 (2017). [DOI] [PubMed] [Google Scholar]
10.Olah, M. et al. Single cell RNA sequencing of human microglia uncovers a subset associated with Alzheimer’s disease. Nat. Commun.11, 6129 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Krasemann, S. et al. The TREM2-APOE pathway drives the transcriptional phenotype of dysfunctional microglia in neurodegenerative diseases. Immunity47, 566–581 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Chen, Y. & Colonna, M. Microglia in Alzheimer’s disease at single-cell level. Are there common patterns in humans and mice? J. Exp. Med.218, e20202717 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Srinivasan, K. et al. Alzheimer’s patient microglia exhibit enhanced aging and unique transcriptional activation. Cell Rep.31, 107843 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Friedman, B. A. et al. Diverse brain myeloid expression profiles reveal distinct microglial activation states and aspects of Alzheimer’s disease not evident in mouse models. Cell Rep.22, 832–847 (2018). [DOI] [PubMed] [Google Scholar]
15.Zhou, Y. et al. Human and mouse single-nucleus transcriptomics reveal TREM2-dependent and TREM2-independent cellular responses in Alzheimer’s disease. Nat. Med.26, 131–142 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Del-Aguila, J. L. et al. A single-nuclei RNA sequencing study of Mendelian and sporadic AD in the human brain. Alzheimers Res. Ther.11, 71 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Hasselmann, J. et al. Development of a chimeric model to study and manipulate human microglia in vivo. Neuron103, 1016–1033 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Geirsdottir, L. et al. Cross-species single-cell analysis reveals divergence of the primate microglia program. Cell179, 1609–1622 (2019). [DOI] [PubMed] [Google Scholar]
19.Tuddenham, J. F. et al. A cross-disease resource of living human microglia identifies disease-enriched subsets and tool compounds recapitulating microglial states. Nat. Neurosci.27, 2521–2537 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Lee, J., Hyeon, D. Y. & Hwang, D. Single-cell multiomics: technologies and data analysis methods. Exp. Mol. Med.52, 1428–1442 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Spitzer, M. H. & Nolan, G. P. Mass cytometry: single cells, many features. Cell165, 780–791 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Asp, M., Bergenstrahle, J. & Lundeberg, J. Spatially resolved transcriptomes-next generation tools for tissue exploration. Bioessays42, e1900221 (2020). [DOI] [PubMed] [Google Scholar]
23.Black, S. et al. CODEX multiplexed tissue imaging with DNA-conjugated antibodies. Nat. Protoc.16, 3802–3835 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Phillips, D. et al. Highly multiplexed phenotyping of immunoregulatory proteins in the tumor microenvironment by CODEX tissue imaging. Front. Immunol.12, 687673 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Schurch, C. M. et al. Coordinated cellular neighborhoods orchestrate antitumoral immunity at the colorectal cancer invasive front. Cell182, 1341–1359 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Heinrich, L. et al. Multiplex imaging of human induced pluripotent stem cell-derived neurons with CO-Detection by indEXing (CODEX) technology. J. Neurosci. Methods378, 109653 (2022). [DOI] [PubMed] [Google Scholar]
27.Goyal, V. K. Lipofuscin pigment accumulation in human brain during aging. Exp. Gerontol.17, 481–487 (1982). [DOI] [PubMed] [Google Scholar]
28.Moreno-Garcia, A., Kun, A., Calero, O., Medina, M. & Calero, M. An overview of the role of lipofuscin in age-related neurodegeneration. Front. Neurosci.12, 464 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Riga, D., Riga, S., Halalau, F. & Schneider, F. Brain lipopigment accumulation in normal and pathological aging. Ann. N. Y. Acad. Sci.1067, 158–163 (2006). [DOI] [PubMed] [Google Scholar]
30.Liu, X. et al. The superficial glia limitans of mouse and monkey brain and spinal cord. Anat. Rec.296, 995–1007 (2013). [Google Scholar]
31.Holtzman, D. M. et al. Apolipoprotein E isoform-dependent amyloid deposition and neuritic degeneration in a mouse model of Alzheimer’s disease. Proc. Natl Acad. Sci. USA97, 2892–2897 (2000). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Nguyen, A. T. et al. APOE and TREM2 regulate amyloid-responsive microglia in Alzheimer’s disease. Acta Neuropathol.140, 477–493 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Pey, P., Pearce, R. K., Kalaitzakis, M. E., Griffin, W. S. & Gentleman, S. M. Phenotypic profile of alternative activation marker CD163 is different in Alzheimer’s and Parkinson’s disease. Acta Neuropathol. Commun.2, 21 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Munoz-Castro, C. et al. Monocyte-derived cells invade brain parenchyma and amyloid plaques in human Alzheimer’s disease hippocampus. Acta Neuropathol. Commun.11, 31 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Bennett, F. C. et al. A combination of ontogeny and CNS environment establishes microglial identity. Neuron98, 1170–1183 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.DeTure, M. A. & Dickson, D. W. The neuropathological diagnosis of Alzheimer’s disease. Mol. Neurodegener.14, 32 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Sadleir, K. R. et al. Presynaptic dystrophic neurites surrounding amyloid plaques are sites of microtubule disruption, BACE1 elevation, and increased Abeta generation in Alzheimer’s disease. Acta Neuropathol.132, 235–256 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Philips, T. & Rothstein, J. D. Oligodendroglia: metabolic supporters of neurons. J. Clin. Invest.127, 3271–3280 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Bernstein, H. G., Keilhoff, G., Dobrowolny, H., Guest, P. C. & Steiner, J. Perineuronal oligodendrocytes in health and disease: the journey so far. Rev. Neurosci.31, 89–99 (2019). [DOI] [PubMed] [Google Scholar]
40.Simons, M., Gibson, E. M. & Nave, K. A. Oligodendrocytes: myelination, plasticity, and axonal support. Cold Spring Harb. Perspect. Biol.16, a041359 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Tay, T. L. et al. A new fate mapping system reveals context-dependent random or clonal expansion of microglia. Nat. Neurosci.20, 793–803 (2017). [DOI] [PubMed] [Google Scholar]
42.Verdonk, F. et al. Phenotypic clustering: a novel method for microglial morphology analysis. J. Neuroinflammation13, 153 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Fernandez-Arjona, M. D. M., Grondona, J. M., Granados-Duran, P., Fernandez-Llebrez, P. & Lopez-Avalos, M. D. Microglia morphological categorization in a rat model of neuroinflammation by hierarchical cluster and principal components analysis. Front. Cell Neurosci.11, 235 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Colombo, G. et al. A tool for mapping microglial morphology, morphOMICs, reveals brain-region and sex-dependent phenotypes. Nat. Neurosci.25, 1379–1393 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.van Weering, H. R. J., Nijboer, T. W., Brummer, M. L., Boddeke, E. & Eggen, B. J. L. Microglia morphotyping in the adult mouse CNS using hierarchical clustering on principal components reveals regional heterogeneity but no sexual dimorphism. Glia71, 2356–2371 (2023). [DOI] [PubMed] [Google Scholar]
46.Sankowski, R. et al. Mapping microglia states in the human brain through the integration of high-dimensional techniques. Nat. Neurosci.22, 2098–2110 (2019). [DOI] [PubMed] [Google Scholar]
47.Raj, D. et al. Increased white matter inflammation in aging- and Alzheimer’s disease brain. Front. Mol. Neurosci.10, 206 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Xu, Q. et al. Profile and regulation of apolipoprotein E (ApoE) expression in the CNS in mice with targeting of green fluorescent protein gene to the ApoE locus. J. Neurosci.26, 4985–4994 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Yamazaki, Y. et al. ApoE (Apolipoprotein E) in brain pericytes regulates endothelial function in an isoform-dependent manner by modulating basement membrane components. Arterioscler. Thromb. Vasc. Biol.40, 128–144 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
50.Dolan, M. J. et al. Exposure of iPSC-derived human microglia to brain substrates enables the generation and manipulation of diverse transcriptional states in vitro. Nat. Immunol.24, 1382–1390 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
51.Mancuso, R. et al. Xenografted human microglia display diverse transcriptomic states in response to Alzheimer’s disease-related amyloid-beta pathology. Nat. Neurosci.27, 886–900 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Priego, N. et al. STAT3 labels a subpopulation of reactive astrocytes required for brain metastasis. Nat. Med.24, 1024–1035 (2018). [DOI] [PubMed] [Google Scholar]
53.Sala Frigerio, C. et al. The major risk factors for Alzheimer’s disease: age, sex, and genes modulate the microglia response to Abeta plaques. Cell Rep.27, 1293–1306 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
54.Fabriek, B. O. et al. CD163-positive perivascular macrophages in the human CNS express molecules for antigen recognition and presentation. Glia51, 297–305 (2005). [DOI] [PubMed] [Google Scholar]
55.Rajan, W. D. et al. Defining molecular identity and fates of CNS-border associated macrophages after ischemic stroke in rodents and humans. Neurobiol. Dis.137, 104722 (2020). [DOI] [PubMed] [Google Scholar]
56.Dermitzakis, I. et al. CNS border-associated macrophages: ontogeny and potential implication in disease. Curr. Issues Mol. Biol.45, 4285–4300 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Wang, Y. et al. TREM2-mediated early microglial response limits diffusion and toxicity of amyloid plaques. J. Exp. Med.213, 667–675 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Masuda, T. et al. Specification of CNS macrophage subsets occurs postnatally in defined niches. Nature604, 740–748 (2022). [DOI] [PubMed] [Google Scholar]
59.van Olst, L. et al. Microglial mechanisms drive amyloid-beta clearance in immunized patients with Alzheimer’s disease. Nat. Med.31, 1604–1616 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
60.Bottcher, C. et al. Single-cell mass cytometry of microglia in major depressive disorder reveals a non-inflammatory phenotype with increased homeostatic marker expression. Transl. Psychiatry10, 310 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
61.Green, M. S., Kaye, J. A. & Ball, M. J. The Oregon brain aging study: neuropathology accompanying healthy aging in the oldest old. Neurology54, 105–113 (2000). [DOI] [PubMed] [Google Scholar]
62.Benmamar-Badel, A., Owens, T. & Wlodarczyk, A. Protective microglial subset in development, aging, and disease: lessons from transcriptomic studies. Front. Immunol.11, 430 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
63.Schroder, B. The multifaceted roles of the invariant chain CD74—more than just a chaperone. Biochim. Biophys. Acta1863, 1269–1281 (2016). [DOI] [PubMed] [Google Scholar]
64.Schetters, S. T. T., Gomez-Nicola, D., Garcia-Vallejo, J. J. & Van Kooyk, Y. Neuroinflammation: microglia and T cells get ready to Tango. Front. Immunol.8, 1905 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Cristovao, J. S. & Gomes, C. M. S100 proteins in Alzheimer’s disease. Front. Neurosci.13, 463 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
66.Du, Z. et al. Qualifying antibodies for image-based immune profiling and multiplexed tissue imaging. Nat. Protoc.14, 2900–2930 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Rajkowska, G., Selemon, L. D. & Goldman-Rakic, P. S. Neuronal and glial somal size in the prefrontal cortex: a postmortem morphometric study of schizophrenia and Huntington disease. Arch. Gen. Psychiatry55, 215–224 (1998). [DOI] [PubMed] [Google Scholar]
68.Wolf, F. A., Angerer, P. & Theis, F. J. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol.19, 15 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
69.Leek, J. T., Johnson, W. E., Parker, H. S., Jaffe, A. E. & Storey, J. D. The sva package for removing batch effects and other unwanted variation in high-throughput experiments. Bioinformatics28, 882–883 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
70.Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics8, 118–127 (2007). [DOI] [PubMed] [Google Scholar]
71.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (II). Zenodo10.5281/zenodo.18903414 (2026).
72.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain. Zenodo10.5281/zenodo.18878505 (2026).
73.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (Dataset I). Zenodo10.5281/zenodo.20725082 (2026).
74.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (Dataset II). Zenodo10.5281/zenodo.20706335 (2026).
75.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (Dataset III). Zenodo10.5281/zenodo.20726162 (2026).

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information^{(10.8MB, pdf)}

Supplementary Tables 1–4 and Figs. 1–11.

Reporting Summary^{(93.9KB, pdf)}

Supplementary Data 1^{(23.9KB, xlsx)}

Statistical source data for Supplementary figures.

Source Data^{(63.8KB, xlsx)}

Statistical source data for main and extended data figures.

Data Availability Statement

Computer code for this project can be provided upon reasonable request and found at https://github.com/BaharehAjami/CODEX-CNS.

[CR1] 1.Serrano-Pozo, A., Frosch, M. P., Masliah, E. & Hyman, B. T. Neuropathological alterations in Alzheimer disease. Cold Spring Harb Perspect Med1, a006189 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR2] 2.Mathys, H. et al. Single-cell transcriptomic analysis of Alzheimer’s disease. Nature570, 332–337 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR3] 3.Wang, Q. et al. Single cell transcriptomes and multiscale networks from persons with and without Alzheimer’s disease. Nat. Commun.15, 5815 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR4] 4.Mathys, H. et al. Single-cell multiregion dissection of Alzheimer’s disease. Nature632, 858–868 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Ajami, B. et al. Single-cell mass cytometry reveals distinct populations of brain myeloid cells in mouse neuroinflammation and neurodegeneration models. Nat. Neurosci.21, 541–551 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR6] 6.Siletti, K. et al. Transcriptomic diversity of cell types across the adult human brain. Science382, eadd7046 (2023). [DOI] [PubMed] [Google Scholar]

[CR7] 7.Batiuk, M. Y. et al. Identification of region-specific astrocyte subtypes at single cell resolution. Nat. Commun.11, 1220 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR8] 8.Bottcher, C. et al. Human microglia regional heterogeneity and phenotypes determined by multiplexed single-cell mass cytometry. Nat. Neurosci.22, 78–90 (2019). [DOI] [PubMed] [Google Scholar]

[CR9] 9.Keren-Shaul, H. et al. A unique microglia type associated with restricting development of Alzheimer’s disease. Cell169, 1276–1290 (2017). [DOI] [PubMed] [Google Scholar]

[CR10] 10.Olah, M. et al. Single cell RNA sequencing of human microglia uncovers a subset associated with Alzheimer’s disease. Nat. Commun.11, 6129 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR11] 11.Krasemann, S. et al. The TREM2-APOE pathway drives the transcriptional phenotype of dysfunctional microglia in neurodegenerative diseases. Immunity47, 566–581 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR12] 12.Chen, Y. & Colonna, M. Microglia in Alzheimer’s disease at single-cell level. Are there common patterns in humans and mice? J. Exp. Med.218, e20202717 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Srinivasan, K. et al. Alzheimer’s patient microglia exhibit enhanced aging and unique transcriptional activation. Cell Rep.31, 107843 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Friedman, B. A. et al. Diverse brain myeloid expression profiles reveal distinct microglial activation states and aspects of Alzheimer’s disease not evident in mouse models. Cell Rep.22, 832–847 (2018). [DOI] [PubMed] [Google Scholar]

[CR15] 15.Zhou, Y. et al. Human and mouse single-nucleus transcriptomics reveal TREM2-dependent and TREM2-independent cellular responses in Alzheimer’s disease. Nat. Med.26, 131–142 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Del-Aguila, J. L. et al. A single-nuclei RNA sequencing study of Mendelian and sporadic AD in the human brain. Alzheimers Res. Ther.11, 71 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR17] 17.Hasselmann, J. et al. Development of a chimeric model to study and manipulate human microglia in vivo. Neuron103, 1016–1033 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Geirsdottir, L. et al. Cross-species single-cell analysis reveals divergence of the primate microglia program. Cell179, 1609–1622 (2019). [DOI] [PubMed] [Google Scholar]

[CR19] 19.Tuddenham, J. F. et al. A cross-disease resource of living human microglia identifies disease-enriched subsets and tool compounds recapitulating microglial states. Nat. Neurosci.27, 2521–2537 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR20] 20.Lee, J., Hyeon, D. Y. & Hwang, D. Single-cell multiomics: technologies and data analysis methods. Exp. Mol. Med.52, 1428–1442 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Spitzer, M. H. & Nolan, G. P. Mass cytometry: single cells, many features. Cell165, 780–791 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR22] 22.Asp, M., Bergenstrahle, J. & Lundeberg, J. Spatially resolved transcriptomes-next generation tools for tissue exploration. Bioessays42, e1900221 (2020). [DOI] [PubMed] [Google Scholar]

[CR23] 23.Black, S. et al. CODEX multiplexed tissue imaging with DNA-conjugated antibodies. Nat. Protoc.16, 3802–3835 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Phillips, D. et al. Highly multiplexed phenotyping of immunoregulatory proteins in the tumor microenvironment by CODEX tissue imaging. Front. Immunol.12, 687673 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR25] 25.Schurch, C. M. et al. Coordinated cellular neighborhoods orchestrate antitumoral immunity at the colorectal cancer invasive front. Cell182, 1341–1359 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.Heinrich, L. et al. Multiplex imaging of human induced pluripotent stem cell-derived neurons with CO-Detection by indEXing (CODEX) technology. J. Neurosci. Methods378, 109653 (2022). [DOI] [PubMed] [Google Scholar]

[CR27] 27.Goyal, V. K. Lipofuscin pigment accumulation in human brain during aging. Exp. Gerontol.17, 481–487 (1982). [DOI] [PubMed] [Google Scholar]

[CR28] 28.Moreno-Garcia, A., Kun, A., Calero, O., Medina, M. & Calero, M. An overview of the role of lipofuscin in age-related neurodegeneration. Front. Neurosci.12, 464 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR29] 29.Riga, D., Riga, S., Halalau, F. & Schneider, F. Brain lipopigment accumulation in normal and pathological aging. Ann. N. Y. Acad. Sci.1067, 158–163 (2006). [DOI] [PubMed] [Google Scholar]

[CR30] 30.Liu, X. et al. The superficial glia limitans of mouse and monkey brain and spinal cord. Anat. Rec.296, 995–1007 (2013). [Google Scholar]

[CR31] 31.Holtzman, D. M. et al. Apolipoprotein E isoform-dependent amyloid deposition and neuritic degeneration in a mouse model of Alzheimer’s disease. Proc. Natl Acad. Sci. USA97, 2892–2897 (2000). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR32] 32.Nguyen, A. T. et al. APOE and TREM2 regulate amyloid-responsive microglia in Alzheimer’s disease. Acta Neuropathol.140, 477–493 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR33] 33.Pey, P., Pearce, R. K., Kalaitzakis, M. E., Griffin, W. S. & Gentleman, S. M. Phenotypic profile of alternative activation marker CD163 is different in Alzheimer’s and Parkinson’s disease. Acta Neuropathol. Commun.2, 21 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR34] 34.Munoz-Castro, C. et al. Monocyte-derived cells invade brain parenchyma and amyloid plaques in human Alzheimer’s disease hippocampus. Acta Neuropathol. Commun.11, 31 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR35] 35.Bennett, F. C. et al. A combination of ontogeny and CNS environment establishes microglial identity. Neuron98, 1170–1183 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.DeTure, M. A. & Dickson, D. W. The neuropathological diagnosis of Alzheimer’s disease. Mol. Neurodegener.14, 32 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR37] 37.Sadleir, K. R. et al. Presynaptic dystrophic neurites surrounding amyloid plaques are sites of microtubule disruption, BACE1 elevation, and increased Abeta generation in Alzheimer’s disease. Acta Neuropathol.132, 235–256 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Philips, T. & Rothstein, J. D. Oligodendroglia: metabolic supporters of neurons. J. Clin. Invest.127, 3271–3280 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR39] 39.Bernstein, H. G., Keilhoff, G., Dobrowolny, H., Guest, P. C. & Steiner, J. Perineuronal oligodendrocytes in health and disease: the journey so far. Rev. Neurosci.31, 89–99 (2019). [DOI] [PubMed] [Google Scholar]

[CR40] 40.Simons, M., Gibson, E. M. & Nave, K. A. Oligodendrocytes: myelination, plasticity, and axonal support. Cold Spring Harb. Perspect. Biol.16, a041359 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR41] 41.Tay, T. L. et al. A new fate mapping system reveals context-dependent random or clonal expansion of microglia. Nat. Neurosci.20, 793–803 (2017). [DOI] [PubMed] [Google Scholar]

[CR42] 42.Verdonk, F. et al. Phenotypic clustering: a novel method for microglial morphology analysis. J. Neuroinflammation13, 153 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR43] 43.Fernandez-Arjona, M. D. M., Grondona, J. M., Granados-Duran, P., Fernandez-Llebrez, P. & Lopez-Avalos, M. D. Microglia morphological categorization in a rat model of neuroinflammation by hierarchical cluster and principal components analysis. Front. Cell Neurosci.11, 235 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Colombo, G. et al. A tool for mapping microglial morphology, morphOMICs, reveals brain-region and sex-dependent phenotypes. Nat. Neurosci.25, 1379–1393 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR45] 45.van Weering, H. R. J., Nijboer, T. W., Brummer, M. L., Boddeke, E. & Eggen, B. J. L. Microglia morphotyping in the adult mouse CNS using hierarchical clustering on principal components reveals regional heterogeneity but no sexual dimorphism. Glia71, 2356–2371 (2023). [DOI] [PubMed] [Google Scholar]

[CR46] 46.Sankowski, R. et al. Mapping microglia states in the human brain through the integration of high-dimensional techniques. Nat. Neurosci.22, 2098–2110 (2019). [DOI] [PubMed] [Google Scholar]

[CR47] 47.Raj, D. et al. Increased white matter inflammation in aging- and Alzheimer’s disease brain. Front. Mol. Neurosci.10, 206 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR48] 48.Xu, Q. et al. Profile and regulation of apolipoprotein E (ApoE) expression in the CNS in mice with targeting of green fluorescent protein gene to the ApoE locus. J. Neurosci.26, 4985–4994 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Yamazaki, Y. et al. ApoE (Apolipoprotein E) in brain pericytes regulates endothelial function in an isoform-dependent manner by modulating basement membrane components. Arterioscler. Thromb. Vasc. Biol.40, 128–144 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR50] 50.Dolan, M. J. et al. Exposure of iPSC-derived human microglia to brain substrates enables the generation and manipulation of diverse transcriptional states in vitro. Nat. Immunol.24, 1382–1390 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR51] 51.Mancuso, R. et al. Xenografted human microglia display diverse transcriptomic states in response to Alzheimer’s disease-related amyloid-beta pathology. Nat. Neurosci.27, 886–900 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR52] 52.Priego, N. et al. STAT3 labels a subpopulation of reactive astrocytes required for brain metastasis. Nat. Med.24, 1024–1035 (2018). [DOI] [PubMed] [Google Scholar]

[CR53] 53.Sala Frigerio, C. et al. The major risk factors for Alzheimer’s disease: age, sex, and genes modulate the microglia response to Abeta plaques. Cell Rep.27, 1293–1306 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR54] 54.Fabriek, B. O. et al. CD163-positive perivascular macrophages in the human CNS express molecules for antigen recognition and presentation. Glia51, 297–305 (2005). [DOI] [PubMed] [Google Scholar]

[CR55] 55.Rajan, W. D. et al. Defining molecular identity and fates of CNS-border associated macrophages after ischemic stroke in rodents and humans. Neurobiol. Dis.137, 104722 (2020). [DOI] [PubMed] [Google Scholar]

[CR56] 56.Dermitzakis, I. et al. CNS border-associated macrophages: ontogeny and potential implication in disease. Curr. Issues Mol. Biol.45, 4285–4300 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR57] 57.Wang, Y. et al. TREM2-mediated early microglial response limits diffusion and toxicity of amyloid plaques. J. Exp. Med.213, 667–675 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR58] 58.Masuda, T. et al. Specification of CNS macrophage subsets occurs postnatally in defined niches. Nature604, 740–748 (2022). [DOI] [PubMed] [Google Scholar]

[CR59] 59.van Olst, L. et al. Microglial mechanisms drive amyloid-beta clearance in immunized patients with Alzheimer’s disease. Nat. Med.31, 1604–1616 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR60] 60.Bottcher, C. et al. Single-cell mass cytometry of microglia in major depressive disorder reveals a non-inflammatory phenotype with increased homeostatic marker expression. Transl. Psychiatry10, 310 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR61] 61.Green, M. S., Kaye, J. A. & Ball, M. J. The Oregon brain aging study: neuropathology accompanying healthy aging in the oldest old. Neurology54, 105–113 (2000). [DOI] [PubMed] [Google Scholar]

[CR62] 62.Benmamar-Badel, A., Owens, T. & Wlodarczyk, A. Protective microglial subset in development, aging, and disease: lessons from transcriptomic studies. Front. Immunol.11, 430 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR63] 63.Schroder, B. The multifaceted roles of the invariant chain CD74—more than just a chaperone. Biochim. Biophys. Acta1863, 1269–1281 (2016). [DOI] [PubMed] [Google Scholar]

[CR64] 64.Schetters, S. T. T., Gomez-Nicola, D., Garcia-Vallejo, J. J. & Van Kooyk, Y. Neuroinflammation: microglia and T cells get ready to Tango. Front. Immunol.8, 1905 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR65] 65.Cristovao, J. S. & Gomes, C. M. S100 proteins in Alzheimer’s disease. Front. Neurosci.13, 463 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR66] 66.Du, Z. et al. Qualifying antibodies for image-based immune profiling and multiplexed tissue imaging. Nat. Protoc.14, 2900–2930 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR67] 67.Rajkowska, G., Selemon, L. D. & Goldman-Rakic, P. S. Neuronal and glial somal size in the prefrontal cortex: a postmortem morphometric study of schizophrenia and Huntington disease. Arch. Gen. Psychiatry55, 215–224 (1998). [DOI] [PubMed] [Google Scholar]

[CR68] 68.Wolf, F. A., Angerer, P. & Theis, F. J. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol.19, 15 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR69] 69.Leek, J. T., Johnson, W. E., Parker, H. S., Jaffe, A. E. & Storey, J. D. The sva package for removing batch effects and other unwanted variation in high-throughput experiments. Bioinformatics28, 882–883 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR70] 70.Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics8, 118–127 (2007). [DOI] [PubMed] [Google Scholar]

[CR71] 71.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (II). Zenodo10.5281/zenodo.18903414 (2026).

[CR72] 72.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain. Zenodo10.5281/zenodo.18878505 (2026).

[CR73] 73.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (Dataset I). Zenodo10.5281/zenodo.20725082 (2026).

[CR74] 74.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (Dataset II). Zenodo10.5281/zenodo.20706335 (2026).

[CR75] 75.Ajami, B. & Rosmus, D.-D. Spatial proteomic analysis enables identification of microenvironment-dependent myeloid cell heterogeneity in Alzheimer’s disease human brain (Dataset III). Zenodo10.5281/zenodo.20726162 (2026).

PERMALINK

Spatial proteomic analysis in human Alzheimer’s disease brains enables identification of microenvironment-dependent microglial cell states

Paula Sanchez-Molina

Dennis-Dominik Rosmus

Dillon Brownell

Mert Meral

Cavanagh Gohlich

Aditya Pratapa

Yaser Peymanfar

Alyssa Whitley

Yue Hou

Nadezhda Nikulina

Alina Bogachuk

Ellen Lara Bouchard

Aude Chiot

Heidrun Kuhrt

Peter Wieghofer

Randall Woltjer

Fabian Svara

Oliver Braubach

Bahareh Ajami

Abstract

Main

Results

Development of CODEX-CNS to study the human brain

Fig. 1. CODEX workflow for neuroscience applications.

Extended Data Fig. 1. Imaging of human FFPE brain slides using CODEX-CNS method.

Extended Data Fig. 2. Cellular composition in healthy and AD brains.

AD brains exhibit specific cell–cell interactions

Fig. 2. Multicellular characterization of the Aβ plaque microenvironment at the proteomic level.

Fig. 3. CODEX-CNS reveals specific cellular interactions in human AD brain.

Extended Data Fig. 3. Cellular interactions in the white matter of healthy and AD brains.

Distinct microglial morphologies across tissue microenvironments

Fig. 4. Morphological analysis of myeloid cells in the human frontal cerebral cortex.

Extended Data Fig. 4. Morphological analysis of human myeloid cells.

Brain myeloid cells form five protein-based clusters characterized by their spatial distribution

Fig. 5. Microglia/macrophage protein signatures correlate with specific tissue microenvironments in the human frontal cerebral cortex.

Extended Data Fig. 5. Myeloid cell clustering based on protein expression.

The BLM subpopulation forms the strongest association with dense Aβ plaques

Fig. 6. Myeloid cells in close proximity to dense Aβ plaques represent a spatial and temporal state between PVMs and microglia.

Extended Data Fig. 6. In situ validation of BLMs as a unique transitional state between microglia and PVMs.

Brain microenvironments drive the accumulation of specific myeloid phenotypes in the human AD cortex

Fig. 7. Neighborhood-based clustering of myeloid cells reveals a specific myeloid phenotype enriched in AD human brains.

Extended Data Fig. 7. Distribution of neighborhood analysis-defined cell clusters across AD samples.

Extended Data Fig. 8. Comparison of HPAMs between AD patients and healthy controls.

Extended Data Fig. 9. CODEX-CNS identifies BLM and HPAM subpopulations in myeloid cells in an independent validation cohort.

Extended Data Fig. 10. Myeloid cell features can be applied to predict Aβ plaques in their local environment.

Discussion

Methods

Postmortem brain samples

Design of CODEX-CNS antibody panel

Tissue preparation and staining

CODEX-CNS

Immunofluorescence

Tissue imaging

CODEX

Microscopy

Image analysis

Tissue annotation

Cell segmentation

Blood vessel segmentation

Aβ plaque segmentation

Cell interaction analysis

Myeloid cell clustering based on protein expression

Spatial neighborhood analysis

Morphological analysis

Myeloid cell clustering based on morphology

Distance analysis

Pseudotime analysis

Myeloid cell clustering based on neighborhood

Aβ plaque prediction based on myeloid cell features

Statistics

Reporting summary

Online content

Supplementary information

Source data

Acknowledgements

Extended data

Author contributions

Peer review