Cerebrospinal fluid protein elevations in Alzheimer's disease are dissociated from immune cell composition

Abstract

INTRODUCTION

Elevated cerebrospinal fluid (CSF) inflammatory proteins have been reported in Alzheimer's disease (AD), yet whether they reflect peripheral immune cell infiltration or central nervous system (CNS)–intrinsic pathological processes remains unclear. We tested whether AD exhibits a distinct CSF protein signature and how it relates to CSF immune cell composition.

METHODS

We conducted a cross‐sectional study of 64 participants: 22 with biomarker‐confirmed AD (bioAD), 22 non‐AD biomarker‐negative cognitive controls (NAD‐CC), and 20 other neurological disease controls (OND). CSF proteins were quantified using the Olink Target 96 Inflammation panel, and immune cell populations were characterized by multiparameter spectral flow cytometry. Cell–protein associations were assessed using hierarchical generalized linear models adjusted for age and sex.

RESULTS

BioAD showed widespread inflammatory and trophic protein elevation versus NAD‐CC, with 21 proteins significantly increased (false discovery rate [FDR] q < 0.05), whereas NAD‐CC was largely indistinguishable from OND. MMP‐10 showed the largest effect (geometric mean ratio [GMR] = 2.10, 95% confidence interval [CI]: 1.53–2.88, p < 0.001). CSF immune cell populations did not differ significantly between groups; only discrete T‐cell alterations were detected: reduced regulatory T cells in bioAD relative to NAD‐CC (odds ratio [OR] = 0.69, 95% CI: 0.54–0.88, p = 0.023), but not relative to OND and minor memory subset shifts. CD8+ effector memory T cells and TNFSF14 showed the only detected AD‐specific association: in bioAD, a 10‐fold increase in CD8+ EM2 cells corresponded to 70% higher TNFSF14 (GMR = 1.70, 95% CI: 1.39–2.09, p < 0.0001).

DISCUSSION

In this feasibility‐sized, cross‐sectional cohort, CSF inflammatory proteins in AD appeared dissociated from immune‐cell composition, potentially reflecting CNS‐intrinsic processes and/or altered CSF barrier dynamics. The identified signature, including MMP‐10, is hypothesis‐generating and may complement conventional biomarkers for diagnosis and treatment monitoring, pending longitudinal validation.

Highlights

• Alzheimer's disease (AD) cerebrospinal fluid (CSF) shows widespread elevation of 21 inflammatory and trophic proteins

• CSF protein elevation is dissociated from immune cell composition in AD

• Reduction in T regulatory cells

• AD‐specific coupling between CSF CD8 T cells and TNFSF14, and between GFAP and monocytes

1. INTRODUCTION

Alzheimer's disease (AD) is neuropathologically defined by amyloid beta (Aβ) plaques and tau neurofibrillary tangles, yet limited amyloid‐lowering efficacy implicates other contributing mechanisms, including neuroinflammation. ¹ , ² , ³

Cerebrospinal fluid (CSF) inflammatory profiling in AD has revealed widespread elevation of proteinases, chemokines, and cytokines, ⁴ , ⁵ , ⁶ whereas emerging single‐cell analyses point to clonal expansion of CD8+ effector T cells, reduced CD4+ T cell receptor diversity, and dysregulation of regulatory T cells. ⁷ , ⁸ , ⁹ However, these two lines of evidence have developed in parallel, with analyses conducted in separate cohorts.

Whether inflammatory protein and immune cell changes in the CSF co‐vary within individuals remains untested. Alternatively, CSF proteins may reflect glia, neuronal injury, or defective clearance, thus mirroring CNS‐intrinsic pathology. Furthermore, the lack of appropriate disease control groups in several studies has made it difficult to distinguish AD‐specific inflammatory changes from those accompanying cognitive impairment of other etiologies.

To address these gaps, we performed an integrated analysis combining high‐throughput profiling of inflammatory proteins with multiparameter flow cytometry–based immune cell phenotyping in the same CSF samples from biomarker‐confirmed AD (bioAD), non‐AD cognitive controls, and other neurological disease controls. Our objectives were to: (1) identify an AD‐specific CSF inflammatory protein signature, (2) characterize CSF immune cell differences across diagnostic groups, (3) evaluate whether protein elevations are coupled to immune cell changes, and (4) delineate AD‐specific cell–protein associations that may inform disease mechanisms.

2. METHODS

2.1. Study cohort

Study participants were recruited from individuals undergoing diagnostic lumbar puncture at the Department of Neurology, Medical University of Vienna.

RESEARCH IN CONTEXT

1. Systematic review: Prior studies have documented elevated proteins in Alzheimer's disease (AD) cerebrospinal fluid (CSF), including cytokines, chemokines, and matrix metalloproteinases, yet their cellular origin has remained unresolved. Whether these signals reflect peripheral immune cell infiltration or central nervous system (CNS)–intrinsic neuroinflammation is a critical unanswered question with direct implications for disease monitoring and therapeutic targeting.

2. Interpretation: In matched CSF samples, we identified a robust AD‐specific inflammatory and trophic protein signature largely uncoupled from immune cell composition, suggesting that CSF proteins reflect CNS‐intrinsic processes, rather than immune cell infiltration. Two disease‐specific axes (CD8+ effector‐memory 2 [EM2]‐ tumor necrosis factor superfamily member 14 [TNFSF14], monocyte‐glial fibrillary acidic protein [GFAP]) suggest discrete immune:glia interactions.

3. Future directions: The dissociation between protein and cellular CSF profiles reframes the interpretation of inflammatory findings and suggests that these signatures may serve as biomarkers of CNS‐intrinsic pathological processes. If confirmed in longitudinal studies, they could complement established AD biomarkers for differential diagnosis and treatment monitoring.

2.1.1. Biomarker‐confirmed Alzheimer's disease (bioAD) group

Participants recruited from the memory outpatient clinic (2023–2025) met international working group (IWG) criteria for highly probable AD and 2018 National Institute on Aging–Alzheimer's Association (NIA‐AA) AD criteria. ¹⁰ , ¹¹

Amyloid positivity was determined by CSF Amyloid Tau Index, diminished Aβ40/Aβ42 ratio, and amyloid‐PET (positron emission tomography) when available. The tau pathology was confirmed by elevated CSF phosphorylated tau‐181 (p‐tau181).

2.1.2. Non‐AD cognitive controls (NAD‐CC)

Participants from the memory outpatient clinic had subjective or objective cognitive impairment by Mini‐Mental State Examination (MMSE) without amyloid or tau biomarker positivity (Table 1). ¹² , ¹³

TABLE 1.

Baseline demographic and clinical characteristics.

Characteristic	Overall (N = 64) a	OND (N = 20) a	NAD‐CC (N = 22) a	bioAD (N = 22) a	p‐value d
Age, years	62 (48, 70)	41 (33, 55)	66 (60, 72)	66 (60, 73)	<0.001
Sex					>0.9
Female	36 (56%)	11 (55%)	12 (55%)	13 (59%)
Male	28 (44%)	9 (45%)	10 (45%)	9 (41%)
Age at onset					0.73
<65	22 (50%)	—	10 (45%)	12 (55%)
≥65	22 (50%)	—	12 (55%)	10 (45%)
CSF volume, mL	9 (6, 10)	10 (8, 12)	8 (5, 10)	9 (6, 9)	0.092
CSF cell count, cells/µL	2.5 (2, 3)	3 (2, 4)	2 (1, 3)	2 (1, 3)	0.3
CSF/serum albumin ratio	5 (4, 6)	4 (3, 6)	5 (4, 6)	4.5 (4, 8)	0.2
CSF total tau (pg/mL) b	391 (215, 753)	—	215 (175, 252)	753 (507, 919)	<0.001
CSF phospho‐tau (pg/mL) b	66 (30, 111)	—	30 (22, 34)	111 (75, 135)	<0.001
CSF Aβ42, pg/mL b	697 (471, 921)	—	921 (695, 1,032)	529 (440, 699)	<0.001
CSF Aβ40, pg/mL b	9,456 (7,902, 12,380)	—	8,797 (7,494, 10,889)	11,355 (7,980, 14,339)	0.071
CSF Aβ42/Aβ40 ratio b	0.07 (0.05, 0.1)	—	0.1 (0.09, 0.11)	0.05 (0.04, 0.06)	<0.001
APOE Genotype					<0.001
2/3	4 (8%)	3 (23%)	1 (6%)	0 (0%)
3/3	28 (54%)	9 (69%)	15 (88%)	4 (18%)
3/4	16 (31%)	1 (8%)	1 (6%)	14 (64%)
4/4	4 (8%)	0 (0%)	0 (0%)	4 (18%)
Amyloid PET status c					0.007
Normal	2 (11%)	—	2 (100%)	0 (0%)
Pathological	16 (89%)	—	0 (0%)	16 (100%)
CSF NfL, pg/mL	1,122 (565, 1,894)	531 (306, 1,034)	787 (604, 1,358)	1,710 (1,160, 2,214)	<0.001
CSF GFAP, pg/mL	8,336 (4,856, 12,099)	4,749 (3,385, 7,053)	8,086 (4,899, 9,923)	11,954 (9,807, 16,287)	<0.001
MMSE	25 (22, 27)	‐	26 (25, 28)	23 (20, 26)	<0.001

Abbreviations: Aβ, amyloid beta; APOE, apolipoprotein E; bioAD, biomarker‐confirmed Alzheimer's disease; CSF, cerebrospinal fluid; GFAP, glial fibrillary acidic protein; IQR, interquartile range; MMSE, Mini Mental State Examination; NAD‐CC, non‐Alzheimer's disease cognitive complaints; NfL, neurofilament light chain; OND, other neurological diseases; PET, positron emission tomography.

^a Data presented as median (IQR) for continuous variables, and n (%) for categorical variables.

^b Not measured in OND group as part of routine diagnostic evaluation.

^c Available for subset of patients: NAD‐CC (n = 2) and bioAD (n = 16).

^d Kruskal‐Wallis test for continuous variables; Fisher's exact test for categorical variables with expected cell counts <5; chi‐square test for sex.

2.1.3. Other neurological diseases (OND) group

Participants from the general neurology outpatient clinic had no cognitive impairment and no systemic or CNS immunological, infectious, or neurodegenerative disease. Diagnoses included headache disorders (n = 8), idiopathic intracranial hypertension (n = 5), and nonspecific neurological symptoms without radiological abnormalities (n = 7).

2.1.4. Inclusion criteria

Age ≥18 years and capacity for informed consent.

2.1.5. Exclusion criteria

CNS infection, autoimmune/inflammatory disorders, malignancy with CNS involvement, or immunosuppressive therapy

2.2. CSF and blood processing

Fresh CSF samples were transported on ice to the laboratory within 15 min after lumbar puncture, were screened using ChemStrip (20 µL), and the concentration of leukocytes was determined by counting in a Fuchs Rosenthal chamber prior to centrifugation. Absolute subset concentrations were calculated by multiplying subset frequencies by the total CSF leukocyte count (cells/µL) obtained using: cells/mL = frequency × total CSF leukocytes × 1000. Samples with ≥50 erythrocytes/µL were excluded. CSF was centrifuged (400 x g, 10 min, 4°C). For staining, samples were incubated with Fc‐Block (10 min) followed by the antibody cocktail (30 min) at 4°C in the dark (Table S1), washed, and centrifuged (400 x g, 5 min, 4°C). The final pellet was resuspended in 200 µL stain buffer and directly analyzed using a Cytek Aurora flow cytometer (16UV‐16V‐14B‐10YG‐8R).

Whole blood (WB) and peripheral blood mononuclear cells (PBMCs) were used solely for panel validation, spectral unmixing, gate setting, and fluorescence minus one (FMO) reference controls, and were not included in any patient analysis. WB samples underwent red blood cell (RBC) lysis (1:10 dilution) before undergoing the same processing steps as CSF. Data were analyzed using SpectroFlo v3.1.0 and FlowJo v10.9.0 (Gating: Figure S1). ¹⁴ CSF aliquots for protein assays were stored at −70°C in the neurological biobank.

2.3. CSF protein measurements

Neurofilament light chain (NfL) and glial fibrillary acidic protein (GFAP) were quantified by single‐molecule array (SIMOA Neurology 2 Plex B, SR‐X Analyzer, Quanterix) on biobanked samples by a blinded investigator as described previously. ¹⁴

2.4. Diagnostic biomarkers

CSF concentrations of Aβ40, Aβ42, p‐tau181, and total tau (t‐tau) were measured using the Innotest enzyme‐linked immunosorbent assays (ELISAs) (Fujirebio Europe) in a routine clinical diagnostic setting ¹⁵ and, therefore, were available only for the bioAD and NAD‐CC groups.

Amyloid‐PET was performed on 18 participants using [^{1 1}C]PiB (Pittsburgh compound B; ∼400 MBq; 40‐min uptake) or [^{1 8}F]flutemetamol (185 MBq; 90‐min uptake) on Siemens Biograph 64 TruePoint or GE Advance scanners, visually rated amyloid‐positive or amyloid‐negative by an experienced nuclear medicine physician per manufacturer guidelines. ¹⁶

2.5. Inflammation‐related biomarkers

Proteins were quantified using the Olink Target 96 Inflammation panel (Olink Proteomics, Uppsala, Sweden, at Olink Services (AIT, Vienna). Proteins detected in ≥75% of samples were retained for downstream analysis (78 of 92 proteins; Table S2). All samples passed quality control (QC; intra‐assay coefficient of variation (CV) ≤12%).

2.6. Statistics

Overall approach: Baseline characteristics were compared by Kruskal–Wallis (continuous) and Fisher's exact tests (categorical). Because the groups differed in age, all subsequent comparisons used regression models of the form outcome ∼ group + age + sex, with pairwise differences obtained as estimated marginal means (emm) (group means predicted after accounting for age and sex). Heteroscedasticity‐consistent standard errors avoided reliance on equal variances across the modestly sized groups. Assumptions were checked by diagnostic plots and Shapiro–Wilk tests.

Effects are reported with 95% confidence intervals (CIs): log‐scale outcomes as geometric mean ratios (GMR = exp(β); 2^β for log₂ normalized protein eXpression [NPX]), the fold‐difference between groups, and native‐scale outcomes as β, the additive difference.

Data transformation: Right‐skewed markers (total tau, p‐tau181, NfL, GFAP) were natural‐log transformed to reduce skew and stabilize variance. Approximately normal amyloid measures and NPX (already log₂) were analyzed untransformed.

Proteins and pathways: Each protein was modeled separately (NPX ∼ group + age + sex) and the three pairwise contrasts expressed as GMRs, with Benjamini–Hochberg false discovery rate (FDR) correction across all proteins within each contrast (q < 0.05 considered differentially expressed). Pathway‐level changes were assessed over MSigDB Hallmark and Reactome sets with ≥5 panel genes, by gene‐set enrichment analysis (GSEA; fgseaMultilevel, 10,000 permutations) of proteins ranked on age‐ and sex‐adjusted Cohen's d and by over‐representation analysis (ORA) of the q < 0.05 proteins. Enrichment was treated as hypothesis‐generating (q < 0.25) given the panel's limited proteome coverage.

Immune‐cell composition: Table S3 lists all populations with their markers, gating parents, per‐group distributions, and detection rates. Populations quantifiable (≥25 events in ≥15 participants) were modeled: frequencies, being proportions bounded by 0 and 1, by beta regression with a logit link and reported as odds ratios (ORs), and counts by linear regression on log₁₀ cells/mL. Sparser populations were not modeled but reported descriptively. Their detection frequencies were compared exploratorily (Fisher's exact test, or Firth's penalized logistic regression when a group lacked positive samples). FDR correction was applied within cell‐subset lineages, mirroring the nested gating hierarchy.

Cell–protein integration: To assess whether increases in protein levels could be explained by cell abundance, a two‐step regression analysis of differentially expressed proteins on cell counts was performed. A pooled model (NPX ∼ cell + group + age + sex) estimated the association shared by all groups, with the effect of cell abundance expressed as GMR, representing the protein fold‐change per 10‐fold increase in cells. Pairs with q < 0.10 within protein were assessed in a second model (NPX ∼ cell × group + age + sex) with a cell‐by‐group interaction, tested by Type III analysis of variance (ANOVA). Interactions at q < 0.10 were followed by group‐specific slopes and their pairwise differences were estimated by emm.

No a priori sample‐size calculation was performed. Sample size was feasibility‐driven, matching the exploratory design.

2.7. Standard protocol approvals

The study was approved by the ethics committee of the Medical University of Vienna (ethical approval number: 1715/2023).

3. RESULTS

The study included 64 participants: 20 patients with OND, 22 non‐AD cognitive controls (NAD‐CC), and 22 patients with biomarker‐confirmed AD (bioAD) (Table 1). Patients with bioAD and NAD‐CC were significantly older than the patients in the OND group. Disease onset before age 65 occurred in 55% of bioAD patients, compared to 45% in the NAD‐CC group. CSF volume for analysis and cellularity was comparable across groups. The bioAD group showed a strong enrichment of apolipoprotein E (APOE) ε4 risk allele carriers (82% compared to 6% in NAD‐CC and 8% in OND), with homozygous ε4 carriers observed exclusively in this cohort.

3.1. Fluid biomarkers of Alzheimer's pathology, neural injury, and glial reactivity

BioAD participants showed significantly reduced CSF Aβ42 and Aβ42/Aβ40 ratio compared to NAD‐CC (β = −340.0 pg/mL, 95% CI: −488.9 to −191.2, p < 0.0001; β = −0.050, 95% CI: −0.059 to −0.041, p < 0.0001) (Figure 1). In contrast, tau pathology markers were robustly elevated in the bioAD group, with total tau levels 3.28‐fold higher (GMR = 3.28, 95% CI: 2.43–4.44, p < 0.0001), and p‐tau181 concentrations 3.81‐fold higher (GMR = 3.81, 95% CI: 2.84–5.12, p < 0.0001) relative to NAD‐CC.

FIGURE 1.

Fluid biomarkers of Alzheimer's pathology, neural injury, and glial reactivity. (A–F) Boxplots showing CSF concentrations of core AD biomarkers (A: Aβ42, B: Aβ42/40 ratio, C: total tau, D: p‐tau181) and neuroglial injury markers (E: NfL, F: GFAP). Boxes represent the median and interquartile range (IQR); whiskers extend to 1.5×IQR; points represent individual samples. (A–D) Comparisons between biologically defined AD (bioAD, red) and non‐AD cognitive change (NAD‐CC, blue). (E–F) Comparisons across all groups. Statistics: p‐values derived from linear regression models adjusted for age and sex. Significance indicates differences relative to the bioAD group. Aβ42, ratio, total tau, p‐tau181, NfL, and GFAP values were analyzed on a log₁₀‐scale where appropriate for normality (bioAD n = 22, NAD‐CC n = 22, OND n = 18). AD, Alzheimer's disease; CSF, cerebrospinal fluid; OND, other neurological diseases; NfL, neurofilament light chain; GFAP, glial fibrillary acidic protein.

Markers of neural injury and glial reactivity were assessed across all three groups with age adjustment. NfL levels were ≈2‐fold higher in bioAD relative to both NAD‐CC (GMR = 1.84, 95% CI: 1.28–2.65, p = 0.0023) and OND (GMR = 2.18, 95% CI: 1.00–4.77, p = 0.062). NfL did not differ between NAD‐CC and OND. GFAP showed a similar pattern: 1.65‐fold higher in bioAD than NAD‐CC (GMR = 1.65, 95% CI: 1.17–2.34, p = 0.0078) and 2.44‐fold higher than OND (GMR = 2.44, 95% CI: 1.46–4.10, p = 0.002), with no NAD‐CC versus OND difference after age adjustment.

Descriptive statistics for all biomarkers are provided in Table S4.

3.2. Distinct protein signature characterizes biological AD from non‐AD individuals

To characterize CSF inflammatory proteins across diagnostic groups, we performed differential expression analysis using the Olink Inflammation panel, identifying 21 of 78 analyzed proteins with significantly elevated levels in bioAD compared to NAD‐CC (Figure 2A).

FIGURE 2.

AD‐specific protein alterations in CSF. (A) Volcano plot showing differential protein expression between bioAD and NAD‐CC. The x‐axis represents the log₂ fold change (estimated difference in NPX), and the y‐axis represents statistical significance (–log₁₀ FDR). Red points indicate significantly upregulated proteins (FDR < 0.05). (B) Highlighting the top 10 most strongly upregulated proteins in bioAD (ranked by effect size). Bars represent the geometric mean ratio (GMR) relative to NAD‐CC. (C) Gene Set Enrichment Analysis (GSEA) results showing the two non‐redundant enriched pathway clusters in the bioAD proteomic signature. The x‐axis (NES) indicates pathway activation strength; dot size represents the number of proteins in the leading edge; dot color indicates the FDR significance category (q < 0.01; q < 0.10). (D) Volcano plot showing differential protein expression: NAD‐CC versus OND. Statistics: Differential expression based on linear models (NPX ∼ Group + Age + Sex) with Benjamini–Hochberg FDR correction. GSEA performed on ranked differential expression lists. AD, Alzheimer's disease; bioAD, biomarker‐confirmed Alzheimer's disease; CSF, cerebrospinal fluid; FDR, false discovery rate; NAD‐CC, non‐AD biomarker‐negative cognitive controls; NES, normalized enrichment score; NPX, normalized protein expression; OND, other neurological diseases.

Matrix metalloproteinase‐10 (MMP‐10) showed the largest effect (GMR = 2.10, 95% CI: 1.53–2.88, p = 0.00066). Other elevated proteins included C‐C motif chemokine ligand 3 (CCL3) (GMR = 1.45, 95% CI: 1.25–1.70, p = 0.00066), programmed cell death ligand 1 (PD‐L1) (GMR = 1.47, 95% CI: 1.20–1.80, p = 0.0067), and hepatocyte growth factor (HGF; GMR = 1.43, 95% CI: 1.21–1.68, p = 0.0014), alongside growth factors (fibroblast growth factor 19 [FGF‐19], FGF‐5, stem cell factor [SCF], vascular endothelial growth factor A [VEGFA]) and metabolism/survival regulators (sirtuin 2 [SIRT2], colony‐stimulating factor 1 [CSF‐1], TNF‐related weak inducer of apoptosis [TWEAK]) (Figure 2A, B). The bioAD versus OND comparison corroborated these findings, with the largest effects for MMP‐10 (GMR = 2.30, 95% CI: 1.52–3.47, p = 0.011) and urokinase plasminogen activator (uPA; GMR = 1.53, 95% CI: 1.23–1.90, p = 0.011)

Conversely, interleukin‐2 receptor beta (IL‐2RB) was the sole protein decreased in bioAD relative to OND (GMR = 0.78, 95% CI: 0.67–0.91, p = 0.047).

The NAD‐CC inflammatory profile was largely indistinguishable from OND: only IL‐2RB trended lower in NAD‐CC (GMR = 0.79, 95% CI: 0.69–0.91, p = 0.095), and MMP‐10 did not differ (Figure 2D, Table S4), indicating that bioAD elevations reflect disease‐specific pathophysiology rather than general neurological dysfunction.

GSEA identified two non‐redundant enrichment clusters in bioAD versus NAD‐CC: receptor tyrosine kinase (RTK)/PI3K‐Akt cluster (representative pathway: constitutive signaling by aberrant PI3K; NES = 1.87, FDR q = 0.007; driven by HGF, FGF‐19, SCF, Flt3L, FGF‐5, TGFα), and a TNFR2 non‐canonical NF‐κB signaling cluster (NES = 1.63, FDR q = 0.060; driven by TWEAK, TNF, TNFRSF9, CD40, TNFSF14, TNFRSF11B). ORA corroborated this, with RTK signaling the top Reactome pathway. This pattern was absent in NAD‐CC and OND (Figure 2C).

3.3. Preserved cellular equilibrium contrasts marked protein dysregulation in CSF of AD

In contrast to the widespread protein alterations, CSF leukocyte immunophenotyping revealed a largely preserved cellular landscape, with isolated changes confined to the T‐cell compartment. Regulatory T cells (Tregs) were reduced in bioAD versus NAD‐CC, both as a proportion of total leukocytes (OR = 0.69, 95% CI: 0.54–0.88, p = 0.023) and within the CD4+ lineage (OR = 0.78, 95% CI: 0.64–0.96, p = 0.083) (Figure 3A and B) but did not differ from OND. NAD‐CC did not differ significantly from OND (OR = 1.14, 95% CI: 0.83–1.60, p = 0.81; OR = 1.10, 95% CI: 0.83–1.44, p = 0.77, respectively).

FIGURE 3.

Cellular landscape stability and selective cell: protein coupling in bioAD. (A) Volcano plot showing global differences in CSF immune cell frequencies (% of CD45) between bioAD and NAD‐CC. x‐axis: log₂ odds ratio (OR; effect size); y‐axis: –log₁₀ p‐value. Solid points: FDR < 0.05; Hollow points: Nominal p < 0.05. (B) Forest plot of T‐cell differentiation subset alterations (expressed as percentage of parent lineage). Points represent ORs with 95% confidence intervals (CIs) derived from beta‐regression models adjusted for age and sex. Asterisks indicate FDR < 0.05. (C) Interaction plot showing the relationship between CD8+ Effector Memory 2 (EM2) cells (log₁₀ cells/mL) and CSF TNFSF14 levels (NPX). Lines represent model‐predicted values (adjusted for age and sex) from a linear interaction model (P‐interaction = 0.0017); shaded areas indicate 95% CIs. (D) Group‐stratified relationships between classical monocytes (log₁₀ cells/mL) and glial injury marker GFAP (log₁₀ pg/mL). Lines represent age/sex‐adjusted predictions derived from separate log‐log linear models for each diagnostic group. Statistics: (A,B) Beta‐regression with Benjamini‐Hochberg FDR correction. (C) Linear interaction model stats. (D) Geometric mean ratios (GMRs) per 10‐fold increase in monocytes and p‐values calculated from group‐stratified models. bioAD, biomarker‐confirmed Alzheimer's disease; CI, confidence interval; CSF, cerebrospinal fluid; EM2, effector memory 2; FDR, false discovery rate; GFAP, glial fibrillary acidic protein; GMR, geometric mean ratio; NAD‐CC, non‐AD biomarker‐negative cognitive controls; NPX, normalized protein expression; OND, other neurological diseases; OR, odds ratio.

Central Memory CD4+ T cells were more frequent in bioAD than NAD‐CC (OR = 1.83, 95% CI: 1.22–2.76, p = 0.031), with no difference versus OND (OR = 1.19, 95% CI: 0.80–1.76, p = 0.40).

CD8+ Effector Memory 1 (EM1) T cells were reduced relative to the OND (OR = 0.46, 95% CI: 0.29–0.75, p = 0.012) but did not differ from NAD‐CC (OR = 1.00, 95% CI: 0.59–1.72, p = 0.99), suggesting an age‐related effect.

B and NK cells showed no disease‐associated changes. Monocyte subsets displayed a reciprocal trend: Classical monocytes (CD14+CD16–) trended lower in bioAD than OND (OR = 0.61, 95% CI: 0.36–1.03, p = 0.064) and NAD‐CC (OR = 0.72, 95% CI: 0.49–1.07, p = 0.10), whereas non‐classical/intermediate monocytes (CD16+) trended higher than NAD‐CC (OR = 1.22, 95% CI: 0.79–1.87, p = 0.37).

3.4. Preserved cell: protein associations despite protein dysregulation in the CSF of AD

To test whether protein elevations were explained by cellular abundance, we evaluated 572 cell–protein pairs using a two‐stage hierarchical approach adjusted for age and sex, with formal group–interaction testing restricted to the pairs showing a pooled association (see Methods).

Pooled across diagnoses, 25 of 572 pairs (4.4%) showed significant overall associations (p < 0.10), all positive and primarily involving CD8+ T cells. CD8+ Effector Memory 2 (EM2) cells correlated with CCL3 (MIP‐1α) (GMR = 1.32, 95% CI: 1.13–1.54, p = 0.010) and TNFSF14 (LIGHT) (GMR = 1.25, 95% CI: 1.09–1.43, p = 0.048). Double‐positive T cells were associated with multiple mediators including CCL3 (GMR = 1.38, 95% CI: 1.14–1.66, p = 0.010) and IL‐10RB (GMR = 1.36, 95% CI: 1.11–1.66, p = 0.085).

We next tested whether these cell–protein relationships were altered in bioAD. Twenty‐four of 25 associations showed no detectable group heterogeneity at this sample size. A single pair met our pre‐specified interaction threshold, CD8+ EM2 cells × TNFSF14 (p = 0.0017): OND showed no association (GMR = 0.97, 95% CI: 0.73–1.29, p = 0.86), whereas bioAD showed a positive association (GMR = 1.70, 95% CI: 1.39–2.09, p < 0.0001), indicating that a 10‐fold rise in EM2 cells associates with a 70% increase in TNFSF14 (Figure 3C). Because this interaction was identified from a small, non‐replicated dataset, its effect size and CI are likely to be optimistic and should be viewed as hypothesis‐generating.

We further assessed the relationship between myeloid cells and markers of neural injury. Log‐log regression revealed a disease‐specific association between classical monocytes (CD14+CD16–) and GFAP: in bioAD, a 10‐fold increase in classical monocytes corresponded to a 1.95‐fold rise in GFAP (GMR = 1.95, 95% CI: 1.14–3.34, p = 0.025), explaining 23.5% of GFAP variance (Figure 3D). The association was not significant in OND (GMR = 1.56, 95% CI: 0.92–2.63, p = 0.12) or NAD‐CC (GMR = 0.97, p = 0.95). No significant association was detected with NfL.

Overall, widespread protein elevation in bioAD showed no detectable association with CSF immune cell abundance, with the exception of two exploratory group‐specific signals, CD8⁺ EM2:TNFSF14 and monocyte:GFAP. Given the limited sample size within each group, the number of pairs tested and the absence of replication, these findings indicate a lack of detected associations rather than definitive evidence that such relationships are absent.

4. DISCUSSION

In this cross‐sectional feasibility study, we observed widespread elevation of trophic and inflammatory proteins in bioAD without a detectable change in leukocyte populations, except for discrete shifts in T cells. These data most strongly support the interpretation that CSF inflammatory proteins in bioAD are dissociated from the measured immune‐cell composition, not that they arise from CNS resident cells alone. These proteins may instead reflect CNS‐intrinsic processes, altered CSF‐barrier dynamics, and/or a contribution from peripherally derived proteins.

BioAD showed the canonical CSF profile of amyloid depletion, tau elevation, and increased NfL/GFAP, consistent with active neurodegeneration and astroglial reactivity. ¹⁷ Beyond these established markers, a distinct AD‐specific proteomic signature was observed. Because the NAD‐CC and OND profiles were comparable despite age‐related differences, the identified associations are likely specific to AD pathobiology and not reflective of age‐related or general consequences of neurodegeneration. ⁴ , ¹⁸ , ¹⁹

The observed pattern of increased protein concentration without cellular changes may be attributable to several overlapping and mutually influencing factor. These include CNS‐resident cell production, increased BBB/CSF‐brain barrier permeability, and impaired glymphatic clearance, all implicated in AD. ²⁰ Enhanced barrier permeability may enable protein exchange between the blood and CSF, consistent with the possibility that a proportion of elevated CSF proteins could be blood‐derived. Notably, APOE ε4 was markedly enriched in the bioAD group. Given its established role in BBB vulnerability, pericyte degeneration, and extracellular matrix dysregulation, APOE ε4–dependent vascular dysfunction may contribute to the CSF protein profile observed here. ²¹

MMP‐10 was the most robustly elevated protein, consistent with prior studies, yet showed no correlation with CSF leukocyte populations. As MMP‐10 is constitutively expressed by microglia and astrocytes, these cells represent a plausible source, although peripheral expression means a blood‐derived contribution remains possible. ²² , ²³ Its elevation may reflect extracellular matrix remodeling and BBB dysfunction in AD. ²⁴ , ²⁵ Similarly, HGF is produced by neurons and reactive glia surrounding amyloid plaques. ²⁶ Growth factor upregulation (VEGFA, FGF‐5, SCF) may reflect glial/neuronal compensation, consistent with CSF VEGF's neuroprotective associations in AD biomarker‐positive individuals. ²⁷

Pathway enrichment implicated PI3K‐Akt and RTK signaling, which in neurons promotes survival and inhibits tau phosphorylation and in microglia facilitates amyloid clearance via TREM2 and CSF‐1R. ²⁸

Elevated CSF1 is consistent with glial‐mediated microglial activation. ²⁹

The absence of detected shifts in CSF immune cell populations is consistent with, although not proof of, a non‐CSF cellular origin. Canonical T‐cell or monocyte‐driven inflammation would elevate interferon and antigen‐presentation proteins; instead, CCL3 and CD8A correlated with leukocyte subsets across all groups without group‐specific differences, consistent with constitutive immune surveillance. Stable overall T‐cell frequencies with only minor shifts in memory subsets may appear contrary to findings of parenchymal CD8 T‐cell increase in AD. ³⁰ , ³¹ Yet these are predominantly tissue‐resident phenotypes, which may be underrepresented in CSF. The absence of compositional shifts in CSF, therefore, need not reflect a lack of T‐cell involvement. Prior TCR studies similarly report clonal CD8⁺/CD4⁺ expansion without compositional shifts. ⁸ , ⁹ Surface‐marker phenotyping cannot capture altered per‐cell function, that would also manifest as protein–cell dissociation. In fact, we observe that CD8⁺ EM cells selectively correlated with TNFSF14 in AD, consistent with transcriptomic evidence of elevated TNFSF14 expression in clonally expanded CSF CD8⁺ T cells, implicating chronically activated CD8⁺ T cells as a possible source. ³² , ³³ , ³⁴ We emphasize that this is a single exploratory association in a feasibility‐sized cohort and should be read as a hypothesis to be tested rather than as an established axis of AD pathobiology.

Tregs did not differ between bioAD and OND. The decrease compared with NAD‐CC was driven partly by higher Treg proportions in NAD‐CC, making this finding specific to that comparison. If confirmed as a true reduction, it would be consistent with studies showing that Treg dysfunction accelerates AD pathology in experimental models, ³⁵ as well as with human studies that, despite inconsistent results regarding frequency, consistently point to reduced immunosuppressive function of Tregs. ³⁶ , ³⁷ A recent Phase‐2a trial of low‐dose IL‐2 in AD supports therapeutic relevance, with Treg expansion and stabilization of CSF NfL. ³⁸

The AD‐specific classical monocyte:GFAP correlation mirrors our prior findings, suggesting astrocyte‐to‐myeloid crosstalk amplified by amyloid/tau pathology. ¹⁴ Whether this reflects compensatory cytokine signaling/phagocytosis or maladaptive changes linked to glymphatic dysfunction may be clarified by studies with longitudinal sampling with glymphatic imaging. ³⁹

Several limitations should be noted: cross‐sectional design (precluding causal inference), unmatched OND age (residual confounding possible despite adjustment), feasibility‐driven sample size limiting power for subtle effects, and correlation analysis of surface‐marker defined phenotyping neither establishing direct cellular production nor excluding altered per‐cell function constraining mechanistic interpretation. Measurements of CSF cellularity at normal cell counts may introduce uncertainty into absolute concentrations. As this error is non‐differential, it would be expected to attenuate rather than create correlations, which makes zero values less informative. Subgroup frequencies were unaffected, and predefined event thresholds limited inaccuracy for low‐frequency events. The cohort represents symptomatic, biomarker‐confirmed AD; therefore, the observed cell–protein associations may not generalize across the full disease continuum. Earlier disease stages may show different or stronger relationships between CSF cellular phenotypes and protein concentrations, whereas in later stages, CNS‐intrinsic processes or impaired protein clearance may become more dominant. Future studies should prioritize longitudinal serial CSF sampling to clarify temporal cell–protein relationships across disease stages, and pair CSF with post‐mortem brain tissue to validate cellular sources of elevated proteins.

In conclusion, integrated CSF profiling in bioAD reveals widespread protein elevation with limited detectable associations with cell composition. These results raise the possibility that the AD‐specific signature may originate, at least in part, from CNS‐resident cells or altered CSF dynamics, but the absence of direct measurements of protein cellular sources prevents this from being established. If validated in longitudinal cohorts, this signature may serve as a differential diagnostic biomarker and as a mirror of CNS‐intrinsic responses to disease‐modifying therapies, including amyloid‐targeting antibodies.

CONFLICT OF INTEREST STATEMENT

S.Z., T.K., V.F., K.S., S.S., T.P., and S.M. declare that they have no competing interests. O.K. has received support to attend scientific conferences or meetings from Amgen, Biogen, Lilly, Novartis, Roche, Sanofi, and UCB; speaker honoraria from Biogen and Roche; and project funding from ArgenX. N.K. has participated in meetings sponsored by, and received speaker honoraria or travel funding from, Alexion, BMS/Celgene, Janssen‐Cilag, Merck, Neuraxpharm, Novartis, Roche, and Sanofi‐Genzyme. N.K. also held a grant for a Multiple Sclerosis Clinical Training Fellowship Programme from the European Committee for Treatment and Research in Multiple Sclerosis (ECTRIMS). R.W. has participated in meetings sponsored by, and received speaker honoraria or travel funding from, Roche Diagnostics, Eisai, and Eli Lilly. P.R. has received honoraria for consultancy and speaking activities from Alexion/AstraZeneca, Almirall, Amgen/Horizon, Amicus, Biogen, Merck, Novartis, Roche, Sandoz, and Sanofi. P.R. has also received research grants from Amicus, Biogen, Merck, and Roche. F.L. has participated in meetings sponsored by, and received speaker honoraria or travel funding from, Actelion, Almirall, Biogen, Celgene, Johnson & Johnson, MedDay, Merck, Novartis, Roche, Sanofi‐Genzyme, and Teva. F.L. has also received consulting honoraria from Biogen, Celgene, Merck, Novartis, Roche, Sanofi‐Genzyme, and Teva. G.B. has participated in meetings sponsored by, and received speaker honoraria or travel funding from, Biogen, Celgene/BMS, Janssen, Lilly, Medwhizz, Merck, Novartis, Roche, Sanofi‐Genzyme, and Teva. G.B. has received consulting honoraria from Adivo Associates, Biogen, Celgene/BMS, Janssen, Merck, Novartis, Roche, Sanofi‐Genzyme, and Teva. G.B. has received unrestricted research grants from Celgene/BMS and Novartis. In addition, G.B. serves as an Executive Committee Member of the European Committee for Treatment and Research in Multiple Sclerosis (ECTRIMS). E.S. has received honoraria for lectures, advisory boards, and consultations from Biogen, Roche, Novo Nordisk, Eisai, Novartis, Sanofi, and Lilly. E.S. has held leadership or fiduciary roles in scientific societies, including the Austrian Alzheimer Association and the EAN Scientific Panel Dementia. E.S.’s institution (Medical University of Vienna, Austria) has received financial support for participation in clinical trials from Lundbeck, Biogen, Novo Nordisk, and Roche Diagnostics, and research grants from Roche, Eisai, the EU Horizon 2020 programme, and the Austrian Alzheimer Association. T.B. has participated in meetings sponsored by, and received honoraria for lectures, advisory boards, and consultations from Amgen, Allergan, Bayer, Biogen, Bionorica, BMS, Genesis, GSK, GW/Jazz Pharma, Horizon, Janssen‐Cilag, MedDay, Merck, Neuraxpharm, Novartis, Octapharma, Roche, Sandoz, Sanofi, Teva, and UCB. T.B.’s institution has received unrestricted research grants from Biogen, Bayer, BMS, Merck, Novartis, Roche, Sanofi, and Teva, as well as financial support for participation in clinical trials sponsored by Alexion, Bayer, Biogen, Merck, Novartis, Octapharma, Roche, Sanofi‐Genzyme, and Teva. T.Z. has participated in meetings sponsored by, or received travel funding from, Biogen, Merck, Novartis, Roche, and Teva. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

Written informed consent was obtained from all study participants.

Supporting information

Supporting Information: trc270321‐sup‐0001‐SuppMat.pdf

Supporting Information: trc270321‐sup‐0002‐ICMJE.pdf

DATA AVAILABILITY STATEMENT

Data supporting the findings of this study are available from the corresponding author on reasonable request by a qualified researcher and on approval by the ethics committee of the Medical University Vienna.