Microglia protein profiles in CSF across Alzheimer’s disease clinical stages

Abstract

Microglia are implicated in the progression of Alzheimer’s disease (AD) pathology from its earliest stages, suggesting that cerebrospinal fluid (CSF) microglia profiling across clinical AD stages can aid in treatment development and monitoring. We analyzed two CSF cohorts (n = 834) that span from unimpaired controls to preclinical and dementia AD stages, identifying 109 dysregulated microglia-related proteins. Enrichment analyses revealed innate immune processes and cellular recruitment in preclinical AD, whereas AD dementia revealed adaptive immunity and macrophage responses. Next, we aligned the in vivo microglia protein profiles with ex vivo-derived microglial transcriptomic signatures, such as disease-associated microglia phenotypes. Transcriptomic signatures were not specific to either clinical stage but spanned both. We classified an 18-protein panel highlighting distinct changes between the preclinical and dementia stages. Our findings underscore the potential of microglia-based biomarker research for AD staging, offering insights into microglia dynamics in clinical AD stages and how transcriptomic signatures translate to proteomic profiles.

By integrating microglial transcriptomics with CSF proteomics, this study reveals protein markers that distinguish early and late Alzheimer’s disease and that have the potential to improve disease tracking and prediction.

Main

AD is a multifactorial disease resulting from the complex interaction of various risk factors involving age, genetics and environment^1,2. The interaction of these factors triggers various pathological processes that advance the disease. Herein, neuroinflammation and microglia are increasingly recognized as important drivers of neurodegeneration due to AD^3–11. Although microglia account for only about 10% of the total brain cells, their increasingly evidenced phenotypic diversity suggests a vital and complex involvement in the brain’s functioning^3,4,6,12. It is hypothesized that microglia are primed¹³ to remove harmful hallmark AD plaques, but, as this pathology is progressively amassed, microglia remain chronically activated and become dysfunctional^8,12,14. These processes are worsened with aging; aged microglia lose their normal functions, and they may become immunogenic¹⁴. Capturing microglial states in vivo across the clinical stages of AD is essential for the development of novel therapies aimed at modifying microglia functioning to halt AD, a challenge that has yet to be met.

Considerable progress has recently been made in CSF proteomics and single-cell transcriptomics^15–19, the latter establishing the foundations for developing a single-cell microglia atlas²⁰. Transcriptomics of ex vivo microglia reveal disease-associated heterogeneous phenotypes^15,16,18, including disease-associated microglia (DAM) phenotypes in AD brain biopsy¹⁸, postmortem tissue⁴ and experimental models^3,4,21. DAM phenotypes have been consistently shown to exhibit temporal heterogeneity^4,21–23, and this may be reflected by abnormal changes in the levels of secreted proteins in the CSF^3,12,21, a valuable in vivo proxy for assessing AD pathology¹⁰. Additionally, microglial activation in AD is hypothesized to present with two peaks: an early protective one and a later proinflammatory one^24,25. Human CSF studies show this duality in several microglia proteins, including sTREM2 (refs. ^26–30), MIF²⁷ and sAXL¹⁰, that are elevated in earlier stages of AD, possibly as a protective response to the accumulation of amyloid pathology^{10,25,28,31–33}. Employing data from both single-cell transcriptomics and proteomics may reveal the CSF profiles of microglia signatures across stages of the AD continuum^3,18,20,34.

In the present study, we combined available transcriptomic datasets with our high-throughput proximity extension assay (PEA) dataset of CSF from individuals presenting with preclinical AD and AD dementia^35–37. PEA proteomics is an antibody-based multiplexing proteomic established method for the identification of biomarker candidates that are translatable to single immunoassays and panels^35–40, critical for clinical and trial implementation. We identified and independently validated microglia-related protein profiles over the clinical stages of AD, alongside assessing their translation to transcriptomic microglia phenotypes in CSF. We applied a classification model to delineate a panel of proteins that can distinguish between preclinical asymptomatic AD and late symptomatic AD, with the aim to elucidate CSF-derived microglia-associated differentially expressed proteins (DEPs) between the two stages. For all AD microglia-associated CSF proteins of interest, we examined the effects of age, the specificity for AD compared to other dementia types and their replicability using an orthogonal proteomics technique. Capturing these proteomic profiles can provide insights into microglia activity reflected in the CSF during the transition to preclinical AD and to the AD dementia stage, which is important for the development of targeted biomarkers and treatments.

Results

The study design is summarized in Fig. 1. The discovery cohort consisted of 553 individuals spanning the AD clinical stages: a control group confirmed cognitively unimpaired with a normal CSF AD biomarker profile (controls, n = 271); a preclinical AD group with a confirmed AD CSF biomarker profile (preclin-AD, n = 48); and an AD dementia group with a confirmed AD CSF biomarker profile (AD-dem, n = 234) with CSF PEA proteomics data available (Supplementary Table 1). Validation of these proteins was conducted in the independent clinical cohort with PEA proteomics available with the same grouping (controls, n = 198; preclin-AD, n = 40; and AD-dem, n = 43; Supplementary Table 1). From these CSF PEA proteomic datasets, we selected 155 microglia-associated proteins by comparison to single-cell and bulk transcriptomic datasets of microglia isolated from fresh surgery and frozen autopsy brain tissue^15,16,18 (Supplementary Table 2). We analyzed the normalized protein expression (NPX) levels of these microglial markers to identify early and/or late dysregulation patterns through an ANCOVA with Tukey’s post hoc test and false discovery rate (FDR) correction (q < 0.05).

Fig. 1. Overview of the study.

a, From the PEA proteomics in both discovery and replication cohorts, 155 proteins were identified as microglia related based on alignment with single-cell and bulk transcriptomic datasets^15,16,18, as described further in Methods, and were measured in both the discovery and clinical replication cohorts (Supplementary Table 2). b, Cohort overview of the discovery and replication cohorts included in the main analysis. c, Proteomic profiling was assessed on a pseudo timeline of cross-sectional clinical profiles: cognitively unimpaired controls (Con), preclin-AD and AD dem. d, On the previously determined patterns of dysregulated microglial proteins, we assessed Gene Ontology (GO) analysis, matching to a microglial transcriptomic atlas for signatures such as disease-associated, age correlations and dementia specificity. e, Cross-technology replication of the patterns in a TMT-MS cohort measuring 350 CSF samples with a total protein number of 86 overlapping with the proteins in the discovery cohort. Icons in a and d created in BioRender; Hok-a-hin, Y. https://biorender.com/2h4y3uf (2026).

First, to explore the biological significance of microglia-associated CSF proteins dysregulated only in the preclinal AD group (that is, ‘Early profile’), dysregulated only in the AD dementia stage (that is, ‘Late profile’) or with sustained dysregulation in both stages (that is, ‘Sustained profile’), we conducted functional enrichment analyses using Gene Ontology. Second, to understand whether the CSF protein profiles can translate into distinct transcriptomic microglial phenotypes, we employed the recently published single-cell transcriptomic microglia atlas²⁰. In brief, this atlas is based on transcriptomic data from live microglia cells isolated from a diverse set of patients (detailed further in the Methods). These data were factorized to define biologically relevant transcriptomic signatures capturing the main axes of heterogeneity in human microglia²⁰. Our proteins of interest were matched to the top 100 markers for each transcriptomic signature in this transcriptomic atlas, to determine whether specific microglial signatures could be associated with distinct clinical AD groups from CSF.

CSF proteins uniquely upregulated in the preclinical AD stage (‘Early profile’)

The proteomic profiling of our discovery and clinical replication cohorts identified a first profile with seven proteins elevated exclusively in the preclinical AD group compared to both the control and dementia stages (Fig. 2a and Supplementary Table 3). Of these, three proteins significantly replicated in both cohorts, and four more proteins that were found in our discovery cohort followed the same trend in the clinical replication cohort, although insignificantly; so all seven were included in the biological enrichment analyses. The preclinical AD profile had significant associations with varied inflammatory and immune processes, defense and wounding responses, chemotaxis and signaling (Fig. 2d). From the seven proteins, APP, ADGRE5 (CD97), LDLR and IFNLR1 are transmembrane receptors; VEGFA and PLAT are secreted proteins; and KYNU is a cytoplasmic protein (Fig. 3 and Supplementary Table 6). All are related to microglial functioning and may activate microglia^41–43. Moreover, processes linked to endocytosis and exocytosis were enriched, potentially highlighting the glial function and surveillance. Three proteins—VEGFA, KYNU and PLAT—overlapped with microglial signatures in the transcriptomic atlas^18,20,44 (Fig. 2d and Supplementary Table 2), specifically linked to stress, immunoregulatory and motility/adhesion, respectively. Although functionally of interest, no defined signature could be attributed exclusively to the ‘Early profile’. Collectively, the Gene Ontology analysis and the overlap with microglial signatures suggest that proteins elevated exclusively during the preclinical AD phase may overall reflect increased cellular recruitment, activation of broad immune processes and enhanced motility for surveillance.

Fig. 2. Differential protein abundance profiles and their functional context.

a−c, Plotted mean log₂fold changes of each protein from the clinical groups to the control baseline (CSF AD biomarker negative (neg.)). Each line represents a protein significantly dysregulated for the stage assessed: a, in dark green, shows the seven proteins significantly upregulated in the early stage (preclinical AD stage, CSF AD biomarker positive (pos.)); b, in magenta, shows the nine proteins dysregulated in the late stage (AD dementia, CSF AD biomarker pos.); and c, in dark blue, shows the 93 proteins significantly dysregulated early and sustained to the late stage. d, Proteins dysregulated in the three profiles divided into functional categories. Bold lettering, proteins that significantly replicated the trend in both the discovery and replication cohorts; italics, proteins that follow the trends insignificantly in the replication cohort. Proteins are categorized into groups with shared functionality based on Gene Ontology enrichment analyses (magenta) and the microglial atlas (light blue). cont, continued; TLR, Toll-like receptor response.

Fig. 3. Subcellular location of the 109 DEPs from each of the three patterns.

Schematic representation of the 109 DEPs from each pattern and their subcellular localizations from three sources (GO:CC, Protein Atlas and UniProt; Supplementary Table 6). Icons created in BioRender; Hok-a-hin, Y. https://biorender.com/2h4y3uf (2026). GO:CC, Gene Ontology Cellular Component.

CSF proteins uniquely upregulated in the AD dementia stage (‘Late profile’)

The second proteomic profile that arose in our discovery and clinical replication cohorts revealed nine proteins elevated exclusively in the AD dementia group compared to the control and the preclinical AD stage (Fig. 2b and Supplementary Table 4). Out of the nine proteins, five significantly replicated the ‘Late profile’, and four more proteins that were found in our discovery cohort followed the same trend in the clinical replication cohort, although insignificantly; so all nine proteins were included in the biological enrichment analyses. Functional enrichment analysis revealed similar processes as in our ‘Early profile’, although with more proteins enriched in innate and humoral immunity and cytokine signaling; from these nine proteins, CCL2 (MCP-1), CTSH, CXCL8 (MCP-2), IL-18 and CCL8 (MIF-1α) are secreted chemokines and cytokines, whereas CD300LF (CLM-1), CEACAM1, MARCO and IL-1RN are found either as transmembrane proteins or in the cytoplasm (Figs. 2d and 3 and Supplementary Table 6). When mapping to the transcriptomic microglia atlas, only two proteins—CCL2 and CTSH—overlapped with microglial signatures (Fig. 2d and Supplementary Table 2): CCL2 with interferon 1 (IFN-1) response and chemokine signatures, whereas CTSH mapped to the antigen-presenting cell (APC) signatures (HLA^high). This means that no defined microglial signature could be assigned to only the ‘Late profile’. Taken together, the findings suggest an increase in immune response in the AD dementia stage compared to the early stage of the disease, alongside increased proinflammatory activity and cellular death proteins.

CSF proteins dysregulated in the preclinical AD stage and sustained into the AD dementia stage (‘Sustained profile’)

In the third profile of our discovery and clinical replication cohorts, 75 proteins were significantly increased in the preclinical AD stage compared to controls, with this change remaining significant into the AD dementia stage (Fig. 2c and Supplementary Table 5). In addition to these 75 proteins, 18 other proteins in this pattern in our discovery cohort were also significantly dysregulated (16 increased and two decreased) in at least one of the stages in the clinical replication cohort, although insignificantly. Most of these proteins (67) are transmembrane proteins, 24 are secreted, eight are cytoplasmic, two are Golgi and one is mitochondrial (Fig. 3 and Supplementary Table 6). Functional enrichment analysis pointed to an association with several immune and transmembrane proteins particularly enriched in apoptotic processes and adaptive, humoral and innate immune system processes, along with cytokine signaling (Fig. 2d), which fits with a combination of the above-described profiles unique for the disease stages. Notably, 44 of the 93 pattern-replicating proteins (both significantly and insignificantly) overlapped with assigned signatures in the microglia atlas, highlighting the involvement of 14 proteins from the total of 20 proteins (70%) within the disease-associated signature (GPNMB^high and APOE^high), five of the eight (62.5%) with a homeostatic signature (CX3CR1^high), 17 of the 26 (65%) with the inflammation signature (immunoregulatory, chemokine and IFN-1 response), six of the nine (67%) with the antigen presentation signature (HLA^high/APC and C1Q^high/phagocytic), nine of the 16 (56%) with the innate immunity signature (IL/IFNγ and S100/TLR signaling) and 10 of the 14 (71%) with the stress and senescence signature, among others (Fig. 2d and Supplementary Table 2). Together, the notion that the far majority of altered proteins fall in this profile suggests a profound upregulation of microglia activity throughout the two stages and that this is not limited to a specific microglia subtype reflected in the CSF.

Classification of CSF proteins dysregulated specifically between the preclinical AD and AD dementia stages

Following the above identification of the three protein profiles over the AD disease stages, we narrowed the focus onto refining a list of proteins that could be suitable for capturing the changes occurring within the progression from preclinical AD to AD dementia. Classification analysis of the microglia proteins of interest identified key proteins whose abundance patterns best distinguish between the early and late clinical AD stages and may best reflect the changing biology. We focused on CSF proteins in our discovery cohort that were most significantly differentially expressed (standardized difference > ±0.2, q < 0.05) between the preclinical AD and AD dementia stages. The 18 identified proteins are represented in Supplementary Table 7. Of these, 15 decreased across the early and late stages—ADAM23, ADGRE5, APP, AXL, CHL1, CXADR, IFNGR1, IL17RA, LAIR1, RELT, ROBO2, SCARF1, SPRY2, TNFRSF14 and VEGFA—whereas three increased across the early and late stages: CCL2, CCL3 and CTSC. These proteins originate from the three profiles: ‘Early profile’ (ADGRE5, APP and VEGFA) and ‘Late profile’ (CCL2) and the remainder in the ‘Sustained profile’ (ADAM23, AXL, CHL1, CXADR, IFNGR1, IL17RA, LAIR1, RELT, ROBO2, SCARF1, SPRY2, TNFRSF14, CCL3 and CTSC).

To begin understanding if these proteins work in an orchestrated manner to affect microglia processes, we performed STRING analysis for protein−protein interactions⁴⁵. This highlighted links between the proteins with CCL2, CCL3 and VEGFA as hubs of connectivity between the classified proteins (Fig. 4a). Functional analyses of these proteins showed broad functions in neuronal, immune response, cell signaling and apoptotic processes (Fig. 4b). The transcriptomic microglial signatures in the microglia atlas overlapped with only four proteins: CCL2, CCL3 and SPRY2 with the chemokine signature and IFNGR1 with the motility/adhesion signature (Fig. 2d and Supplementary Table 7). The later stage increase includes two proinflammatory chemokines and a protease involved in immune regulation, suggesting a shift toward chronic immune activation and potential protein dysfunction. By contrast, the 15 proteins that decrease are primarily immune receptors or receptor signaling pathways, which may indicate less cellular communication in later stages.

Fig. 4. Biological contextualization of the identified dysregulated proteins between the symptomatic AD stages.

a, STRING-generated protein−protein interaction network. Query proteins (colored nodes) with a combined score threshold of 150 are shown. Edges represent functional associations: solid lines indicate high confidence scores (>400); dashed lines denote lower confidence scores (<400). b, Functional enrichment analysis circos plot. Broadly grouped significantly enriched GO:BP terms (P_FDR < 0.05, two-tailed) are displayed on the upper arc, with associated proteins on the lower arc. c, Lollipop plot representing the two-tailed Spearmanʼs ρ correlation values of age in the cognitively unimpaired group with negative CSF AD biomarker profiles of the 18 proteins of interest. Larger sized dot represents P < 0.05; smaller dot represents non-significance. GO:BP, Gene Ontology Biological Processs; NS, not significant.

To expand the possible utility of this panel, we explored the discriminative power of these proteins between the two stages of symptomatic AD. We performed an internal cross-validation with all 18 proteins together yielding an area under the curve (AUC) of 0.94 (95% confidence interval: 0.83−1.00) and with the proteinsʼ individually tested AUCs ranging between 0.68 and 0.81 (Supplementary Table 7), providing an initial indication of their potential for staging. Although the strength of differential expression was supported in our replication cohort (standardized mean difference > ±0.2), only four proteins in this cohort replicated the significance (APP, CHL1, CXADR and IL17RA; P < 0.05). In line with this, the cross-validation in the clinical replication cohort on the 18 proteins together yielded a high overall AUC of 0.88 (0.80−0.96) and with the individually tested proteins resulting in low to medium AUCs ranging between 0.47 and 0.69 (Supplementary Table 7).

Age effects

As a major risk factor for AD, age may trigger pathogenic processes at different stages of AD progression and, as such, may modulate the protein abundances in CSF. To assess the age effects, we calculated the correlation with age on the levels of our CSF proteins of interest in our control (cognitively unimpaired) group. For the 109 replicated proteins in the previous early, late and sustained profiles, EDA2R and HSPB1 are the only proteins with a moderate correlation with age (ρ > 0.4, P < 0.05), whereas most proteins showed a weak correlation (ρ: 0.2−0.4; 61/109), a very weak correlation (ρ: 0−0.2; 35/109) or no correlation (P > 0.05; 11/109) with age (Extended Data Fig. 1). From our panel of 18 proteins discriminating the preclinical versus dementia stages (Fig. 4c), the levels of CCL2 and CHL1 did not correlate with age, whereas the remainder showed a weak or very weak correlation. We performed our receiver operating characteristic (ROC) analysis of this panel correcting for age, which led to minimal AUC changes of 0−0.05 (Supplementary Table 9). These subtle or inexistent effects of age support that dysregulations in our profiles may reflect disease effects.

Extended Data Fig. 1. Age correlations.

lollipop plot representing the Spearman rho correlation values of age in the cognitively normal group with negative CSF AD biomarkers for all proteins in the three profiles (7 in the early, 9 in the late, and 93 in the sustained, bold lettering for significant replications and italics for insignificant); extension of Fig. 4c in the main text. Dashed lines delineate the strengths of age-protein R: very weak or none < 0.2, 0.2 > weak < 0.4, 0.4 > moderate < 0.6, 0.6 > strong < 0.8, and 0.8 > very strong < 1. Larger dot represents p_FDR < 0.05, two-tailed, that is significant, smaller represents insignificant.

Specificity for AD

As an indication of specificity of the findings for AD compared to two non-AD dementias, we expanded the discovery analysis within the entire discovery cohort that included CSF from patients with a diagnosis of dementia with Lewy bodies³⁶ (DLB; n = 110) or frontotemporal dementia³⁷ (FTD; n = 143) (full demographics in Supplementary Table 8a). Six out of seven proteins showed an increase in only the preclinical AD stage, aside from KYNU, which showed an increase in DLB (Extended Data Fig. 2). From the ‘Late profile’, three out of nine proteins were specific for AD. In the ‘Sustained profile’, 59 out of 93 proteins showed specificity for AD (57 increased in both AD stages and two decreased in both AD stages). This supports that most of the dysregulation is specific to AD, whereas the reminder may be related to more generic aspects of neurodegenerative diseases.

Extended Data Fig. 2. Assessment across non-AD dementias.

lollipop plots showing the log₂(fold change) from the CN baseline of the replicated proteins (7 in the early, 9 in the late, 93 in the sustained, in bold the significantly replicating ones, in italics the non-significant) all within the discovery cohort that includes measures for DLB and FTD patients.

Technique specificity

Lastly, we performed an orthogonal technique replication—namely, tandem mass tag mass spectrometry (TMT-MS). We compared the results in our discovery cohort, PEA proteomics, to a group of patients measured with TMT-MS (n = 137; demographics in Supplementary Table 1). This analysis, with 86 proteins measured in both techniques, showed that 81% of the proteins exhibited similar upregulation or downregulation patterns across all profiles, although many lacked statistical significance (Supplementary Tables 3−5 and Extended Data Fig. 3). Profile-specific replication rates were 64% for the ‘Early profile’, 71% for the ‘Late profile’ and 84% for the ‘Sustained profile’. Eleven proteins demonstrated significantly consistent patterns in both the technical replication cohort and the discovery cohort. Of those, APP replicated in the ‘Early profile’, and 11 proteins (DSC2, GLO1, GRN, HAVCR2, IL18BP, ITGB2, SOD2, SPON1, TNFRSF12A, TNFRSF14 and VSIG4) replicated in the ‘Sustained profile’. These findings underscore the robustness of our proteomic results across different measurement techniques.

Extended Data Fig. 3. Replication heatmaps across the three cohorts.

a. Replication heatmap of the 72 proteins (only from the 109 DEPs, not the total proteins in the cohorts) that are measured in all 3 cohorts over the two proteomic techniques in the analysis (5 out of 7 in the early profile, 3 out of 9 in the late profile, and 64 out of 93 each cohort plotted with the mean z-scored fold change of each protein from the clinical groups stated to the CN baseline. The proteins in bold are the proteins that replicate the patterns. Significant associations are depicted by asterisks. **q < 0.05, ***q < 0.01, ‘ ‘ = N.S. b. Venn diagrams of the protein numbers measured in each technique over the three cohorts.

Sensitivity analyses

To evaluate the potential influence of tau pathology on these markers, we stratified groups according to AT (amyloid−tau) status. Although the A+T− group was relatively small (n = 25; Supplementary Table 8b), resulting in reduced statistical power, our findings suggest that tau may be the principal driver of differential protein expression, with 91% of proteins reaching significance only after T+ status (Supplementary Table 9). We also examined the impact of APOE ε4 carriership by repeating our analyses stratified by APOE genotype. Here, the preclinical AD groups were limited in size (APOE ε4 n = 27; APOE ε2/ε3 n = 19; Supplementary Table 8c), which may have affected statistical power. Nevertheless, our results show that 53% of the 109 proteins remain unchanged, whereas 20% are specifically associated with ε4 carriership, particularly in the late and sustained groups (Supplementary Table 9).

Discussion

In the present study, we aligned single-cell transcriptomic microglia states with CSF proteomics to describe the presence and dynamics of microglia-associated CSF proteins and signatures over the AD clinical stages. Proteomic analysis highlighted specific proteins upregulated in the preclinical AD stage (including APP, VEGFA and ADGRE5) and in the AD dementia stage (including CCL2, CTSH and CXCL8) as well as proteins that are dysregulated across both stages (including ITGB2, VSIG4 and DSC2). Functional enrichment analysis and alignment to microglial transcriptomic signatures suggested that preclinical proteins reflect increased cellular recruitment and an initial immune response, possibly representing a microglial-associated innate response to AD pathology. In the late stage, there were adaptive and humoral immune processes and apoptosis, and, throughout both stages, there was a complex pattern of immune activation and cellular death. Surprisingly, when we aligned the in vivo protein profiles with ex vivo single-cell microglial transcriptomic signatures, such as the previously described DAM, these were not specific to either clinical stage but spanned both. Our findings offer insights into microglia-associated signatures in CSF, underscoring the potential of microglial biomarker research for staging of AD and how single-cell transcriptomic signatures translate to CSF proteomic profiles. This provides a conceptual framework and foundation for tracking the microglia-associated contribution over the AD disease course for treatment development and has a potential application for predicting clinical onset of AD. Previous studies explored AD staging through CSF proteomics^40,46–49 with some even identifying glial proteins that emerge during asymptomatic phases—for example, SMOC1 and ITGAM associated with early changes and amyloid pathology⁴⁰. However, these studies did not specifically examine microglia-associated proteins across AD stages, potentially overlooking more subtle alterations arising from this minority cell population relative to neurons.

Using the single-cell transcriptomic signatures of phenotypes from the microglia atlas^18,20,44, we aimed to determine whether these phenotypes can be reflected in secreted proteomics, in relation to disease stages. Overall, we recognized a few microglia signature markers indicative of inflammation, stress and motility in the early AD stage and, in the late stage, few additional markers of inflammation and adaptive immunity with antigen presentation. As may be expected, the sustained profile containing 75 proteins significantly replicating in both cohorts (Fig. 2c,d) allowed for better mapping of transcriptomic signatures to the CSF proteome, containing markers of the following transcriptomic signature groups in order of percentage of proteins that could be mapped: homeostatic (83%), stress and senescence (80%), inflammation (76%), antigen presentation (71%), innate immunity (64%) and DAM (63%). The homeostatic signature in the transcriptomic atlas²⁰, CX3CR1^high, consists of the chemokine receptor CX3CR1, regulating and stabilizing microglia⁵⁰. The DAM signature consisted of two subdivided groups—GPNMB^high and APOE^high—representing temporal disease-associated phenotypes from experimental and pathology studies^3,4. APOE^high and GPNMB^high signatures both include high loadings for TREM2 (refs. ^20,51) with the GPNMB^high signature being more akin DAM2 signature in murine models^4,20. Interestingly, in the sustained profile, the multitude of signatures that we could assign proteins to are possibly reflecting that different microglia phenotypes can be detected in CSF at the same time. It may reflect a fundamental difference between tissue and bodily fluids (that is, tissue has specific spatial and complex cellular organizations while CSF is a cell-sparse medium), as a DAM signal arises in CSF without distinct phenotypic switching. Crucially, a recent spatial proteomics study⁵² explored the possibility of a microglial state continuum in postmortem brain tissue, highlighting a continuous gradient of microglial marker activation throughout healthy and AD tissue rather than following discrete stages. Alternatively, the lack of stage specificity could be due to individual heterogeneity of the disease progression in AD⁵³ or a mismatch between clinical and biological progression⁵⁴. These findings have implications for clinical trials, where it may, thus, be possible to detect target engagement toward each of these microglia phenotypes.

In the preclinical stage of AD—that is, asymptomatic individuals with positive AD CSF biomarker profiles—three proteins show unique upregulations: APP, VEGFA and ADGRE5. Microglia have been shown to respond early to APP, the amyloid precursor protein cleaved into amyloid-β making up AD plaques^4,55,56. Although APP is a neuronal marker, APP transcripts were detected in the available transcriptomic datasets of healthy and diseased microglia, leading to the inclusion in our data. The presence of APP here may reflect proximity to APP-expressing neighboring cells or context-dependent microglial expression under pathology²⁵, which may decrease in latter stages due to cell loss. Vascular endothelial growth factor A (VEGFA), typically protective for cognitive function and vascular integrity, when increased in AD, it increases oxidative stress and accelerates disease progression⁵⁷. ADGRE5 (CD97 antigen subunit α) is part of the complement activation and leukocyte migration pathways, often associated with early stages of AD^58,59. Early changes in protein levels may be associated with tau pathology, as demonstrated by the AT sensitivity analysis in which most proteins lost significance at the A+T− stage. This observation aligns with previous studies indicating that immune activation is linked to increasing tau levels⁶⁰. As leukocyte infiltration occurs, proinflammatory macrophages are attracted to the site of injury by chemokines such as CCL2 (also known as monocyte chemoattractant protein 1 (MCP-1)), CCL3 (MIP-1α) and CXCL8 (MCP-2)⁶¹, which increase in the dementia stage. Amyloid plaques may further lead to elevated CCL2 expression in monocytes⁶², possibly explaining the observed increase in CCL2 during late-stage AD. Our sensitivity analysis suggests that APOE ε4 carriership may slightly influence progression into the dementia stage, with a subset of proteins showing upregulation mainly driven by ε4 into the late stage⁶³. With the advancement into the dementia stage, regulators of neuroinflammation decrease in the CSF, as captured by our panel of 18 proteins in our preclinical AD versus AD dementia profile. This potential decline in immune regulation, coupled with ongoing immune cell infiltration at the site of injury, suggests that initial neuroinflammation increases may confer protection, but, as this inflammation continues to be exacerbated over time, it becomes neurotoxic^59,64,65. Interestingly, these effects were not correlated with chronological age in our study and may, therefore, be independent of processes triggered by aging. Nearly one-third of the proteins identified in our study (31/109) have been functionally characterized in genetic, animal and cell models, particularly in relation to inflammation and AD (Supplementary Table 6). This overlap demonstrates that many of the proteins identified from the CSF profiles have been established within experimental inflammatory processes. Several other proteins within the early and sustained expression patterns remain unexplored and warrant further investigation, as they may provide insights into aging and neurodegeneration and offer potential targets for future therapeutic development.

Our study has notable limitations and several strengths to consider. A key limitation lies in the challenge of the translation from ex vivo transcriptomics to clinically applicable measurements, due to biological and technical complexity⁶⁶. Although CSF offers a direct read-out of ongoing biological processes in the brain, its proteomic coverage remains limited compared to transcriptomics, and transcript levels may not accurately reflect secreted protein dynamics. There are also differences in detected transcripts per cell type per study. Here, we used single-cell transcriptomic datasets for our protein selection and signature alignment instead of the single-nucleus transcriptomics datasets, due to the lower amounts of transcripts in the nucleus as compared to the whole cell⁶⁷. Due to the scarcity of cell-type-specific proteins^38,68, contributions from astrocytes, neurons or other cell types cannot be excluded in our analysis. Additionally, with increasing age and disease, disruption of the blood−brain barrier⁶⁹ and impairment of the glymphatic system⁷⁰ can increase the influx of peripheral proteins into the CSF, further limiting the precise determination of the source of the proteins. To address the translational gap between proteomics and transcriptomics, we integrated the largest single-cell microglia transcriptomics datasets with CSF proteomics. Future strategies focusing on bridging the brain-to-CSF translation gap could include single-cell brain transcriptomics, postmortem tissue and paired proteomics, refining methods to unambiguously assign cell type origin in the CSF proteome. Second, our study is on cross-sectional groups within the cohorts, which required us to estimate temporal patterns. Long-term longitudinal studies with repeated CSF samples are required to test dynamics of CSF proteins more directly. Third, our final 18 candidates of interest were identified using targeted antibody-based proteomics. Although this approach may limit the breadth of candidate proteins, it enhances translatability to immunoassays, as previously demonstrated^71,72. Technical limitations of our study include (1) the absence of TREM2 in our discovery cohort and (2) the small sample sizes for our sensitivity analyses. Additionally, the cohorts were predominantly white, and, therefore, it can be valuable to study the generalizability in other populations. Finally, proteomic studies are inherently variable due to the multiple technologies employed for biomarker discovery; therefore, another strength of our study is that we solidified our candidate proteins through external validations. We integrated a large cohort dataset of transcriptomic-based human microglial signatures with data from two independent human CSF clinical cohorts and from two different proteomic technologies. Improving the biofluid biomarker staging addresses an important need in the AD field⁵⁴. Although the prioritized proteins in our study were not significantly replicated in the validation cohort, they demonstrated high classification performance in both cohorts (AUC > 0.88 and AUC > 0.96), suggesting strong combined discriminative power. The lack of replication may be attributable to differences in dementia group sample sizes and age distributions between cohorts. Providing the framework to bridge transcriptomic and proteomic data represents a crucial step toward clinical translation of the substantial evidence for microglia involvement in AD.

In conclusion, we show potential for the assessment of microglial phenotypes throughout the AD continuum from human CSF, showing that secreted proteins align to specific microglial phenotypes as the DAM signature. We highlight several promising protein biomarkers that can aid in inflammatory staging of patients as well as increase understanding of the contribution of microglia spanning the clinical AD stages. Furthermore, our widespread upregulation of microglial proteins throughout the AD continuum adds evidence to the increasing body of evidence of sustained microgliosis during AD progression, with relatively subtle changes over the disease course.

Methods

Ethics statement

Ethical approval was given by the institutional ethical review boards of each center, and all participants gave written informed consent. For the discovery cohort and the technical replication cohort: VU University Medical Center (VUmc): AD CSF biobank METC number 00–211; discovery cohort: additional samples from the University of Pennsylvania: language and cognitive impairment in Parkinson’s disease and Parkinson’s disease with dementia or DLB IRB069801; clinical replication cohort: DELCODE is an observational longitudinal memory clinic-based multicenter study whose protocol was approved by the ethics committees of the medical faculties of all participating sites: the ethics committees of Berlin (Charité, University Medicine), Bonn, Cologne, Göttingen, Magdeburg, Munich (Ludwig-Maximilians-University), Rostock and Tübingen. The process was led and coordinated by the ethics committee of the medical faculty of the University of Bonn. The registration number of the trial at the ethics committee in Bonn is 117/13.

Cohort description

The study was performed as a retrospective investigation in CSF samples from a cohort of individuals from the Amsterdam Dementia Cohort (ADC)⁷³ and the Center for Neurodegenerative Disease Research at the University of Pennsylvania. All participants underwent standard neurological and cognitive assessments, and diagnosis was assigned according to international consensus criteria for mild cognitive impairment⁷⁴ and AD⁷⁵. CSF Aβ42 and total-Tau (tTau), referred to as AD CSF biomarkers, were analyzed locally using commercially available kits (ADC: ELISA INNOTEST Aβ₁₋₄₂, hTAUAg, Fujirebio; University of Pennsylvania: Luminex xMAP INNO-BIA AlzBio3, Luminex). A positive CSF AD biomarker profile was defined locally as increased tTau/Aβ₄₂ in the cohorts from the ADC (>0.46 (ref. ⁷⁶)) and the University of Pennsylvania (>0.30 (ref. ⁷⁷)). We excluded one patient who progressed to FTD, two to progressive supranuclear palsy and two to vascular dementia. To capture the AD continuum, we selected a total of 553 individuals with CSF measurements for Aβ₄₂, tTau and proteomics data available, comprising three groups: cognitively unimpaired control group with negative AD CSF biomarker profiles and in whom objective cognitive investigations were normal (that is, criteria for mild cognitive impairment, dementia or any other neurological or psychiatric disorder not fulfilled⁷³) (controls; n = 277, mean age 58 years, 36.5% females); preclinical AD group with positive AD CSF biomarker profiles with no objective cognitive symptoms (preclin-AD; n = 48, mean age 68 years, 58.3% females); and AD dementia group with positive AD CSF biomarker profiles and fulfilled the dementia criteria⁷³ (n = 234, mean age 66 years, 40.6% females) for the main analysis (Supplementary Table 1). The discovery cohort was split based on AT status for sensitivity analyses, defining A+ as an Aβ₁₋₄₂ concentration of <1,092 pg ml⁻¹ and T+ as a pTau181 concentration of >24 pg ml⁻¹ or a tTau concentration of >235 pg ml⁻¹ for a small subset of patients who did not have pTau181 measurements⁷⁸ (demographics in Supplementary Table 8b).

Protein-level measurements and selection

For our discovery cohort, samples were randomly distributed across plates within each batch, including bridging samples, for proteomics measurement at the Olink facility, which was blinded to the clinical details of the samples. A total of 1,196 CSF protein levels were quantified using 11 Olink Target 96 multiplex antibody-based protein panels based on PEA technology (Cardiometabolic, Cardiovascular II and III, Cell Regulation, Development, Immune Response, Inflammation, Metabolism, Neurology, Oncology II and Organ Damage; Olink Proteomics) as in our previous studies^27,35,36,79. After quality control and normalization, the data were provided in the relative protein quantification unit, NPX, which is in a log₂ scale. All characteristics and validation data for each assay are available at the manufacturer’s website (https://olink.com/).

Based on overlap with human microglia transcriptomic datasets in both healthy and AD samples from the MGEnrichment tool⁸⁰, our protein list overlapped with two RNA sequencing datasets for homeostatic proteins^15,16 along with a single-cell RNA sequencing dataset¹⁸ for AD states of microglia, resulting in a total of 172 unique proteins, which were refined based on the independent clinical cohort availably (described below) to 155 total proteins included for the statistical analysis (transcriptomic sources for these proteins are presented in Supplementary Table 2). In addition, we obtained an indication of the cell type specificity with the EWCE (expression weighted cell type enrichment) R package⁸¹ in a whole brain cell single-nucleus dataset from ref. ⁸². In this set, 133 of our protein transcripts were detected, of which 98 in microglia, with 44 of these 98 transcripts showing to be enriched for microglia compared to the other cell types (Supplementary Table 2).

Validation cohorts and disease specificity analysis

For the clinical replication, we focused on replication in an independent cohort, using the extensive DZNE-Longitudinal Cognitive Impairment and Dementia Study (DELCODE) cohort, an observational longitudinal memory clinic-based multicenter study in Germany⁸³. The CSF AD profile was measured on a V-PLEX Aβ Peptide Panel 1 (6E10) Kit (K15200E) and a V-PLEX Human Total Tau Kit (K151LAE) (Meso Scale Diagnostics). Cutoff values for abnormal concentrations of Aβ₄₂ (<496 pg ml⁻¹)⁸⁴ and cutoff values established in Bonn based on clinical non-impaired control samples for tau (>470 pg ml⁻¹) were used to split the groups. This independent clinical cohort was measured using the Olink 3072 Explore panel, which includes 3,072 assays, of which 156 (90.7%) of the microglial proteins of interest overlapped. The cohort was split into the same groups as the discovery cohort: cognitively unimpaired (n = 198, mean age 69 years, 52% females), preclinical AD (n = 40, mean age 73 years, 27.5% females) and AD dementia (n = 43, mean age 75 years, 67.4% females) (Supplementary Table 1). Statistical analyses were performed individually in this cohort and compared to the results from the discovery cohort without the need for batch or platform corrections.

For the technical replication, we focused on mass spectrometry cohort replication. A subset of the ADC cohort was measured using TMT-MS⁸⁵; CSF biomarkers and cutoffs are the same as in the discovery cohort. The overlap with the discovery protein amounted to 109 total microglial proteins (Fig. 1a), of which 88 (51.2%) were measured in at least 50% of samples. The technical replication cohort was labeled according to the cognitive state and biomarker profiles using the same definitions as our discovery cohort: cognitively unimpaired controls (n = 126, mean age 59 years, 34.1% females), preclinical AD (n = 38, mean age 65 years, 50% females) and AD dementia (n = 186, mean age 65 years, 49.5% females) (Supplementary Table 1).

We additionally assessed, broadly, the specificity for the AD continuum by an ANCOVA comparison, the same as in our previous analysis, within our complete discovery cohort³⁵ that includes CSF from patients with diagnoses of DLB^36,86 (n = 110, mean age 69 years, 16.4% females) and FTD^37,87 (n = 143, mean age 62 years, 46.9% females) (additional demographics in Supplementary Table 8a).

Statistics and reproducibility

Data processing was performed in RStudio (version 2023-10-31) using R (version 4.3.2). No statistical methods were used to predetermine sample sizes, but our sample sizes are similar to those reported in previous publications^35,36,40. Data distribution was assumed to be normal, but this was not formally tested. Differential protein expression across the groups in the discovery cohort was performed using a one-way ANCOVA with age and sex as covariates, followed by Tukey’s post hoc test, and for multiple testing using FDR correction for the number of proteins. Differential protein expression across the groups in the clinical replication cohort was performed using a single-stratum one-way ANCOVA with age and sex as covariates comparing each group to the control group and corrected using FDR. We performed classification modeling on the selected proteins of interest that showed a standardized difference between the preclinical AD versus the AD dementia levels of more than ±0.2 and that had a significant FDR-correct P value, q value, of less than 0.05 in the profiling. Cross-validation in both the discovery and replication cohorts was performed, and the predictive performance of all models was assessed by comparing ROC curves and their AUCs. TMT-MS results are on a z-score scale, and, therefore, to ease comparison, the NPX values of the discovery and clinical replication cohorts were z-scored, and the same analyses were applied.

Differential protein expression across the groups in the technical replication cohort was performed using one-way ANCOVA with age, sex and multiple comparison testing correction and FDR, as in the discovery cohort. The replication between the two proteomic techniques was calculated by comparing the patterns of dysregulation for each profile (that is, proteins dysregulated in the same direction in both divided by the total proteins measured in both, as a percentage). The age correlation analyses were performed using the cor.test function from ‘stats’ (version 3.6.2) in R with the method set on ‘spearman’ for non-parametric tests on all proteins of interest. All plots were plotted using ggplot2 (version 3.5.1) and circlize for circos plot (version 0.4.16) in R. The figures were compiled in Adobe Illustrator, and icons in Figs. 1 and 3 were extracted from BioRender.

Biological pathway analyses and phenotypic assessment (transcriptomic atlas)

To understand the biological context of the proteins, we analyzed the dysregulated proteins in the three profiles together using clusterProfiler⁸⁸ (version 4.10.1) for Gene Ontology enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment using, as background, the entire list of proteins measured in the PEA panels and analyzing the significantly dysregulated proteins per group and the proteins that followed the same trend of dysregulation although insignificantly. Significant Gene Ontology functional terms (P < 0.05) were grouped into general categories based on their main function to broadly understand their functions.

Second, the proteins were assessed based on their overlap with transcriptomic-based microglial phenotypes^18,89,90. The assessment was done by overlapping results with the microglial transcriptomic atlas derived from single-cell hierarchical Poisson factorization^89,90 clustering on single-cell RNA sequencing of microglial cells of various diseased samples^18,20,44. The top 100 markers associated with each transcriptomic signature were assigned to 26 groups and subgroups (innate immunity: IL/IFNγ signaling, TLR/MAPK signaling and S100/TLR signaling; disease-associated: APOE^high and GPNMB^high; antigen presentation: HLA^high/APC and C1Q^high/phagocytic; metabolic: glycolysis, OxPhos-1, OxPhos-2 and OxPhos-3; inflammation: chemokine, IFN-1 response and immunoregulatory; cellular stress: stress and senescence; homeostatic: CX3CR1^high; other: NPY1R^high, motility/adhesion, GRID2^high and actin folding; and unknown: CIITA^high and PLCG2^high). In Supplementary Tables 2 and 7, we highlight (in gray) the markers that had any detectable association over score 1 for each signature in the whole dataset (not only the top 100).

The panel of 18 proteins were connected using STRING version 11.5 using the STRINGdb⁹¹ (version 2.14.3) in R, using the score threshold 200 on all interaction sources.

Reporting summary

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

Author contributions

Conceptualization: E.-R.B., C.E.T. and L.V. Methodology: E.-R.B., C.E.T. and L.V. Proteomic analysis: E.-R.B., P.V.M.-A., E.M.V., B.M.T. and L.V. Transcriptomic data analysis: V.S.M. and P.L.D.J. Project administration and funding acquisition for the clinical cohorts and projects (for example, CSF collection, PEA/TMT-MS runs and clinical data): Y.S.H., W.A.B., D.J.I., A.S.C.-P., A.W.L., Y.P., W.M.v.d.F., O.P., J.H.-R., J.P., A.S., J.W., F.J., E.D., K.B., R.P., S.T., C.L., F.B., DELCODE, M.d.C., P.J-V., B.M.T. and A.R. Writing—original draft: E.-R.B., P.V.M.-A., A.R., C.E.T. and L.V. Writing—reviewing and editing: E.-R.B., P.v.B., P.V.M.-A., V.S.M., E.M.V., Y.S.H., W.A.B., D.J.I., A.S.C.-P., A.W.L., Y.P., W.M.v.d.F., O.P., J.H.-R., J.P., A.S., J.W., F.J., E.D., K.B., R.P., S.T., C.L., F.B., M.d.C., R.W., P.J.-V., B.M.T., P.L.D.J., A.R., C.E.T. and L.V. Visualization: E.R.B. Supervision: C.E.T. and L.V.

Peer review

Peer review information

Nature Aging thanks David Morgan, Bo Peng and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Data availability

The discovery PEA proteomics data that support the findings of this study are available at https://www.synapse.org/PRIDE_AD. The clinical replication PEA proteomics data that support the findings of this study are available from the DELCODE authors, but restrictions apply to the availability of these data, and so they are not publicly available. Data are, however, available from the authors upon reasonable request and with permission from the cohorts’ steering committees (contact for and information on data access: https://www.dzne.de/en/research/studies/clinical-studies/delcode/). The technical replication mass spectrometry proteomics data supporting the findings of this study are available through the ADDI workbench (10.58085/HR6S-2991).

Code availability

No custom code was developed specifically for this study.

Competing interests

A.C.P. is supported by the NIH, SPARK-NS, the Parker Family Chair and the Lipman Fund. She has received consulting fees for serving on the scientific advisory board of Novartis Neuroscience. She receives royalties as an inventor of a patent held by the Children’s Hospital of Philadelphia on therapies to treat frontotemporal dementia. As of 11 January 2025, W.M.v.d.F. is executive director at Alzheimer Nederland in Amersfoort, The Netherlands. Before 1 November 2025, research programs of W.M.v.d.F. were funded by ZonMW, NWO, EU-JPND, EU-IHI, Alzheimer Nederland, Hersenstichting CardioVascular Onderzoek Nederland, Health~Holland (Topsector Life Sciences & Health), Stichting Dioraphte, the Noaber Foundation, Pieter Houbolt Fonds, Gieskes-Strijbis Fonds, Stichting Equilibrio, Edwin Bouw Fonds, Pasman Stichting, Philips, Biogen, Novartis NL, Life-MI, AVID, Roche BV, Eli Lilly NL, Fujifilm, Eisai and Combinostics. W.M.v.d.F. is a recipient of ABOARD, which is a public−private partnership receiving funding from ZonMW (no. 73305095007) and Health~Holland (Topsector Life Sciences & Health) (PPP allowance; no. LSHM20106). Before 11 January 2025, W.M.v.d.F. was an invited speaker at Biogen, Danone, Eisai, WebMD Neurology (Medscape), Novo Nordisk, Springer Healthcare and the European Brain Council. W.M.v.d.F. has been a consultant to Oxford Health Policy Forum CIC, Roche, Biogen, Eisai, Eli Lilly, Owkin France and Nationale Nederlanden Ventures. W.M.v.d.F. has participated in advisory boards of Biogen, Roche and Eli Lilly. All funding is paid to her institution. In 2024−2025, W.M.v.d.F. was a member of the steering committee of the phase 3 EVOKE/EVOKE+ studies (Novo Nordisk). In 2025, W.M.v.d.F. was a member of the steering committee of the phase 3 trontinemab study (Roche). All funding has been paid to Amsterdam UMC. In 2020−2021, W.M.v.d.F. was an associate editor of Alzheimerʼs Research & Therapy. In 2021−2025, W.M.v.d.F. was an associate editor of Brain. W.M.v.d.F. is chair of the Scientific Leadership Group of InRAD. W.M.v.d.F. is a member of the Supervisory Board (Raad van Toezicht) of the Trimbos Instituut. S.T. served on advisory boards of Eli Lilly, Eisai and GE Healthcare. He was member of the Data and Safety Monitoring Board of the ENVISION study (Biogen). K.B. has received payment or honoraria for lectures from Eisai and was supported for attending meetings by Eli Lilly Deutschland and Novo Nordisk. J.W. has received consulting fees from Immunogenetics, Noselab and Roboscreen; has received payment or honoraria for lectures from Beeijing Yibai Science and Technology, Gloryren, Janssen-Cilag, Pfizer, Med Update GmbH, Roche Pharma and Eli Lilly; is on the advisory board of Biogen, Abbott, Boehringer Ingelheim, Eli Lilly, Merck Sharp & Dohme and Roche; and has a fiduciary role with the Working Group for Neuropsychopharmacology and Pharmacopsychiatry (AGNP), the German Society for CSF Diagnostics and Clinical Neurochemistry (DGLN), the German Association for Psychiatry, Psychotherapy and Psychosomatics (DGPPN), Deutsche Hirnliga and the CSF Society. M.C. has been an invited speaker at Eisai and Novo Nordisk and has been an invited writer for Springer Healthcare. She is an associate editor at Alzheimerʼs Research & Therapy and a scientific advisor of the Michael J. Fox Foundation. C.E.T. has research contracts with Acumen, ADx Neurosciences, AC-Immune, Alamar, Aribio, Axon Neurosciences, Beckman Coulter, BioConnect, Bioorchestra, Brainstorm Therapeutics, C2N Diagnostics, Celgene, Cognition Therapeutics, EIP Pharma, Eisai, Eli Lilly, Fujirebio, Instant Nano Biosensors, Merck, Muna, Nitrase Therapeutics, Novo Nordisk, Olink, PeopleBio, Quanterix, Roche, Sysmex, Toyama, Vaccinex and Vivoryon. She is editor-in-chief of Alzheimerʼs Research & Therapy; serves on the editorial boards of Molecular Neurodegeneration, Alzheimer’s & Dementia, Neurology: Neuroimmunology & Neuroinflammation and Medidact Neurologie/Springer; and is a committee member to define guidelines for cognitive disturbances and one for acute neurology in The Netherlands. She has consultancy/speaker contracts with Aribio, Biogen, Beckman Coulter, Cognition Therapeutics, Danaher, Eisai, Eli Lilly, Janssen, Merck, Neurogen Biomarking, Nordic Biosciences, Novo Nordisk, Novartis, Olink, Quanterix, Roche, Sanofi and Veravas. D.J.I. receives research funding from the NIH, the Michael J. Fox Foundation and the Lewy Body Dementia Association and research funding paid to the institution for clinical trials by Alector, Cervo Med, Denali, Passage Bio and Prevail. The remaining authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s43587-026-01088-0.

Supplementary information

The online version contains supplementary material available at 10.1038/s43587-026-01088-0.