Cerebrospinal Fluid Over Plasma Links Analytes to Cognitive Decline in Older Adults at Risk for Alzheimer's Disease Authors: Jagan A Pillai, Audrey Zhu, Chunxuan Ma, Ming Wang, Yifan Yang, Lynn M Bekris, Maria Khrestian, Frank P DiFilippo, Mikhail Nasrallah, Aaron Burberry, James B Leverenz, Yuan Hou, Feixiong Cheng, Stephen M Rao Published: 2026 ABSTRACT ABSTRACTObjectiveTo identify inflammatory analytes in cerebrospinal fluid (CSF) and plasma associated with cognitive decline in cognitively normal (CN) older adults at risk for Alzheimer's disease (AD).MethodsIn a longitudinal study of 118 CN older adults (65–80 years, 54% APOE ε4, 26% preclinical AD), 1331 CSF and 1501 plasma analytes were quantified at baseline and 2‐year follow‐up using SomaLogic, with key inflammatory findings validated on the Luminex platform. Linear models (Limma), adjusted for age, sex, APOEε4, and amyloid‐positive status, identified pathology‐associated analytes. Co‐expression network and multivariable modeling defined hub analytes and enriched pathways. To assess robustness, a targeted panel of 94 inflammation‐related analytes was analyzed using best subsets regression to derive parsimonious models based on adjusted R2 improvement (≥ 0.01) and ≥ 5 observations per predictor. Subgroup analyses by amyloid and APOE ε4 status were performed, and a two‐stage elastic‐net approach additionally validated analyte selection.ResultsCSF proteomics revealed stronger APOEε4 and amyloid‐associated analyte signatures than plasma. Forty‐four CSF analytes were co‐regulated by APOEε4 and amyloid‐positive status, forming central network hubs (e.g., EFNB2, NPTN, UNC5D) enriched in axon guidance, synaptic signaling, and extracellular matrix pathways. In contrast, inflammatory analytes including eotaxin‐1 and IL‐17 pathway‐related molecules were associated with longitudinal cognitive decline, with stronger effects observed in females. Eotaxin‐1 demonstrated the most consistent predictive performance across analytic methods, subsets, and assay platforms, whereas network hub analytes were not predictive of cognitive outcomes.InterpretationCSF reflects strong APOEε4/amyloid related proteomic network alterations linked to cognitive decline, while plasma provides weaker but complementary signals. CSF inflammatory signaling, particularly eotaxin‐1, may serve as a correlate of longitudinal cognitive decline in cognitively normal older adults at risk for AD.Keywords: alzheimer's disease, cerebrospinal fluid, inflammation, longitudinal cognitive decline, older adults, plasma, preclinical AD, proteomics, systems biology Introduction No cure for Alzheimer's disease (AD), the leading etiology of dementia worldwide, has been identified. An increasing body of research has noted additional pathophysiological processes beyond amyloid and tau pathology, including changes in neuroinflammation and immune response, lipid metabolism, vascular dysfunction, and mitochondrial dysfunction [1, 2, 3, 4, 5]. These findings suggest that, within a pathobiological AT(N) (amyloid/tau/neurodegeneration) framework of defining AD biology in human subjects, additional markers of AD‐related pathophysiology, including inflammation and vascular change, are also of increasing interest [6]. Currently, AD is considered a biological process that begins with the appearance of AD neuropathologic change in asymptomatic individuals [6]. There is consensus that effective disease‐modifying treatment may be administered before the onset of AD clinical symptoms (preclinical AD) [7, 8, 9]. However, biomarker changes that track meaningful clinical outcomes in this preclinical AD stage [10] are still being clarified [11]. Among biomarkers, changes in plasma and cerebrospinal fluid (CSF) related to inflammation and their impact on clinical outcomes have been of most interest. Early studies have noted that CSF sTREM2 [12], YKL‐40 [13], ICAM1, VCAM1, IL‐15 [14], and NPTX2 alone [15], or as a ratio with YWHAG [16], as well as plasma GFAP [17], relate to AT(N) biomarkers and cognitive outcomes in the preclinical AD and earlier mild cognitive impairment (MCI) stage of AD. Unfortunately, univariate inflammation biomarker studies fail to capture the richness of the early neuroinflammatory changes involving multiple inflammation‐related pathways activated in AD [5, 18] in impacting the subsequent rate of cognitive decline. Additionally, the nature of inflammatory changes can vary dynamically between preclinical AD and MCI stages of AD [19, 20, 21] supporting the need for more careful evaluation of these changes in cognitively normal (CN) older adults at risk for future onset of clinical symptoms from AD. To address these needs, we conducted a two‐year longitudinal study in cognitively normal older adults, including participants at high risk for AD (preclinical AD and APOEε4 carriers). Using the validated multiplex proteomic platform SOMAscan, we simultaneously assessed proteomic changes in plasma and CSF to characterize longitudinal interactions among multiple biomarkers rather than single correlates of disease progression. We aimed to identify CSF and plasma proteomic changes that best predict cognitive decline over two years. Key SOMAscan findings were validated using the MILLIPLEX platform. To ensure robustness, we also validated co‐expression network and multivariable modeling results using complementary regression approaches, including elastic‐net regression, to address multicollinearity, reduce overfitting, and improve estimate stability [22]. Methods Participants We conducted a two‐year longitudinal cohort study that included 118 CN participants between 65 and 80 years old. Participants had completed at least ten years of education and were fluent in English. Detailed inclusion/exclusion criteria are provided in Figure S1. The Cleveland Clinic Institutional Review Board approved the study, and individual patient consent was obtained before study participation. The study overview is provided in (Figure 1). A framework for comparing dynamic analyte levels between baseline and 2‐year follow‐up in cognitive decline. CSF AD Biomarkers CSF was collected by lumbar puncture in the morning after fasting to avoid preanalytical variations. Details on sample analysis have been previously reported [23]. In brief, the CSF was aliquoted into 500‐μL amber polypropylene tubes, immediately frozen, and stored at −80°C. CSF levels of Aβ40 and Aβ42 were measured using the MILLIPLEX Human Amyloid Beta Tau Magnetic Bead Panel (HNABTMAG‐68 K, MilliporeSigma) on the Luminex platform, modified by a 1:10 dilution. The lower limit of detection for the analytes in the kit is as follows: Aβ 40, 6.2 pg/mL; Aβ 42, 1.3 pg/mL; CSF Aβ 42/40 ratio ≤ 0.1399 was considered to meet the threshold of amyloid positivity (A+) or preclinical AD diagnosis [23] based on cut‐offs valid for our center. CSF And Plasma Analytes Biofluid samples were processed according to a standardized collection protocol. CSF samples were centrifuged at 2,000 × g for 10 min and frozen immediately after collection and processing (at −80°C). Plasma was processed with a mean processing time of 2 h 50 min from collection to freezing (range 1–6 h) and stored at −80°C until analysis. Frozen aliquots were shipped directly to SomaLogic for proteomic analysis without thawing at the study site. Samples underwent a single freeze–thaw cycle prior to assay. Multiplex proteomic assessment of chemokines and cytokines in both plasma and CSF collected in parallel at the same time was completed using a Slow Off‐rate Modified Aptamer (SOMAmer)–based capture array called “SOMAscan” (SomaLogic Inc., Boulder, CO). This approach uses chemically modified nucleotides to transform a protein signal into a nucleotide signal that can be quantified using relative fluorescence on microarrays [24]. SOMAscan has been used in national datasets including the Alzheimer's Disease Neuroimaging I (ADNI) study [25]. Proteomic profiling was conducted using the SomaScan platform. Raw SomaScan data underwent SomaLogic's standard quality control and normalization procedures, including hybridization normalization, intraplate median signal normalization, plate scaling, calibration, and quality control using calibrator and pooled quality control (QC) samples to assess inter‐plate reproducibility. Baseline samples were run on one SomaScan plate, while 2‐year follow‐up samples were run across two plates with pooled‐sample calibrator replicates on each plate. Plate‐specific scaling and protein‐specific calibration factors were applied to reduce inter‐plate and technical variability relative to external calibrator references. QC replicates were then evaluated for agreement with the QC reference and cross‐plate coefficient of variation (CVs) to assess assay performance and reproducibility. Plates were required to meet SomaLogic acceptance criteria, including a plate scaling factor between 0.4 and 2.5 and a QC percent in the tails of < 15%. Measurements below the estimated lower limit of detection (ELoD) were treated as missing values, and analytes with ≥ 50% of measurements below the EloD or ≥ 50% missing values after EloD censoring were excluded prior to downstream analyses. AD GWAS loci are enriched in pathways related to immune function, lipid metabolism, endosomal–lysosomal trafficking, and cytoskeletal/synaptic biology [26]. Accordingly, analytes mapping to these biological processes were identified using Gene Ontology (GO) [27] and KEGG pathway [28] annotations and selected for downstream analysis. The final dataset contained 1331 CSF analytes and 1501 plasma analytes derived from the SomaScan 7 K assay. Key plasma results from the SomaLogic platform were validated using the MILLIPLEX Human Cytokine/Chemokine/Growth Factor Panel A (HCYTA‐60 K, MilliporeSigma) on the Luminex platform. The assay utilizes a capture antibody coupled to the surface of magnetic beads and a detection antibody with a fluorescent label. In a 96‐well plate, the sample is mixed with antibody‐coupled beads. Following a series of incubations and washes with the detection antibodies and the fluorescent label, the beads are resuspended and analyzed using the Luminex 200 × PONENT System. The kit was completed according to the manufacturer's instructions with all standards, controls, and plasma samples run in duplicate. CSF amyloid and tau concentrations were measured using the MILLIPLEX MAP Human Amyloid Beta and Tau Magnetic Bead Panel according to the manufacturer's instructions. Standard curves were generated for each assay run using the kit‐provided lyophilized standards, and assay performance was assessed based on standard curve quality and sample distribution. Two kit‐provided quality control (QC) samples with manufacturer‐specified acceptable ranges were included in each run, and assays were repeated if QC values fell outside these ranges. All standards, QCs, and samples were analyzed in duplicate. Measurements with a coefficient of variation (CV) > 25% between duplicate wells or those failing to meet the minimum bead count required for valid quantification were excluded from analysis. Amyloid PET Amyloid PET data were acquired according to standardized protocols of the ADNI (Alzheimer's Disease Neuroimaging Initiative) and the SCAN Initiative (http://scan.naccdata.org). Participants received 370 MBq (10 mCi) of 18F‐florbetapir intravenously and were scanned on a Biograph mCT 4‐ring PET/CT system (Siemens Healthineers). PET data were acquired in list mode for 20 min, starting 50 min post‐injection of 18F‐florbetapir. Dynamic PET images were reconstructed as four 5‐min frames, 400 × 400 matrix, 2 mm slices, zoom 2.0, 4 iterations, 24 subsets, and no postfilter. Amyloid burden at the baseline visit was quantified according to the Centiloid scale [29] using the methods and 18F‐florbetapir cross‐calibration of Su et al. [30]. Briefly, dynamic PET images were co‐registered to T1 MRI images (parcellated using FreeSurfer). PET SUVR (standardized uptake value ratio) was calculated using the PET Unified Pipeline [31] for the parcellated cortical regions relative to the reference region (cerebellar cortex). Regional partial volume correction was applied to the SUVR calculations. Centiloid was calculated from SUVR according to the published linear calibration for this processing pipeline [30]. A centiloid value of 24.1 was considered to meet the threshold of amyloid positivity (A+) or preclinical AD diagnosis at baseline visit. Cognitive Measures Baseline and 2‐year follow‐up cognitive measures included the verbal Rey Auditory Verbal Learning Test—Delayed Recall (RAVLT‐DR) [32], nonverbal Brief Visuospatial Memory Test Revised—Delayed Recall (BVMTR‐DR) [33]. Statistics This study analyzed longitudinal CSF and plasma proteomic data from 118 participants. Individuals missing either baseline or 2‐year follow‐up samples, or lacking required metadata, were excluded separately for each biofluid. The final analytic samples included 76 participants for CSF and 82 for plasma. Analyses were performed independently for each biofluid. There were 1331 analytes in CSF and 1501 in plasma. To assess associations between longitudinal changes in analyte levels (2‐year follow‐up relative to baseline) and AD‐related pathology, Linear Models for Microarray Data (Limma) were applied to compare log2‐transformed analyte change ratios by amyloid status (amyloid positive, A+ vs. amyloid negative, A−) and APOEε4 status (E4+ vs. E4−). Models were adjusted for age, sex, and amyloid and APOEε4 status, as appropriate. The first three proteomic principal components (PC1–PC3) were included to account for technical and biological variability. Statistical analyses were performed in R (version 4.4) using two‐tailed tests with a significance level of 0.05. Analytes with a false discovery rate (FDR) < 0.05 were considered statistically significant. Analytes significantly associated with both amyloid and APOEε4 status in CSF were further evaluated using co‐expression network analysis. Louvain clustering with 100 bootstrap iterations was used to identify stable modules, followed by pathway enrichment analysis to characterize underlying biological processes. Sensitivity analysis was performed to evaluate Tau181 change (Δp‐tau181 = 2 years—baseline) and amyloid status × APOE ε4 status interaction by including those terms in the multivariable Limma model, separately. To assess proteins whose levels, remain stable over 2 years, the median ratio, first and third quartiles (Q1/Q3), and interquartile range (IQR) were calculated across samples; stability was quantified as the median absolute log2 ratio and the percentages of measurements within ±10% and ±20% of the expected ratio (1.0), with proteins ranked by increasing median absolute log2 ratio. Internal Validation of Key Inflammatory Analytes Results and Cognitive Decline Following identification of inflammatory signals in prior analysis, an internal validation using a complementary statistical framework was undertaken. A targeted panel of 94 inflammation‐related analytes based on prior literature [24–29] was curated from the original proteomic panel. Of these, 86 analytes met quality control criteria, defined as ≥ 50% complete data (i.e., above the estimated limit of detection) in both plasma and CSF, and were log10‐transformed [34] and included in subsequent analyses. A complete list of analytes and their lower limit of detection is provided in Table S1. Pearson correlation analyses were first used to identify baseline inflammatory analytes associated with 2‐year change in cognitive performance measured by RAVLT‐DR and BVMTR‐DR. Analytes showing significant associations were subsequently evaluated using multivariable regression models adjusted for age, sex, APOEε4 status, and education. Stratified analyses were additionally performed by amyloid and APOEε4 status. Best subsets regression was used to identify parsimonious predictive models containing 1–4 analytes for each continuous cognitive outcome. Model selection was based on improvement in adjusted R2 (≥ 0.01) while maintaining a minimum ratio of five observations per predictor to reduce overfitting. As a sensitivity analysis, outliers, defined as values at least 4 median absolute deviations (MADs) from the sample median, were excluded from the best subsets regression analyses to ensure that analyte selection was not unduly influenced by extreme values. To further evaluate model stability, a two‐stage elastic‐net regression approach was applied, consisting of initial variable selection followed by refitting a simplified multivariable model using the selected analytes. All analyses were conducted using a complete‐case approach, with participants excluded only from analyses requiring missing variables. Receiver operating characteristic (ROC) curve analyses were performed to assess the ability of key baseline cerebrospinal fluid and plasma analytes to predict clinically significant cognitive decline, defined as at least one standard deviation decline in cognitive scores over 2 years, via logistic regression models adjusted for the same covariates as above. Again, as a sensitivity analysis, outliers, defined as values at least 4 median absolute deviations (MADs) from the sample median, were excluded from the logistic regression analyses to ensure that the model results were not unduly influenced by extreme values. Functional Pathway Analysis on Analytes of Interest Analytes of significance identified in the analyte subgroups were entered into STRING functional protein association networks for pathway enrichment analysis [35]. The most significant analyte terms related to gene–tissue associations identified in an expression database (TISSUES 2.0) [36] (FDR p < 0.05) were summarized. The Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database resource for representation and analysis of biological systems [28] and the Reactome Pathway Knowledgebase [37] were used in enrichment analysis to identify targeted analyte pathways of interest in any secondary analysis. Results Among a total of 118 individuals, the final analysis included 76 individuals for the CSF dataset and 82 individuals for the plasma dataset with both baseline and 2‐year follow‐up data. (Table 1) and Table 2. TABLE 1Participant demographics.CharacteristicCSFPlasma N = 76 N = 82DemographicsAge, Mean (SD)69.88 (3.81)69.90 (3.83)Sex, n (%)Female35 (46.1)38 (46.3)Male41 (53.9)44 (53.7)Race, n (%)White68 (89.5)75 (91.5)Black or African American8 (10.5)7 (8.5)Ethnicity, n (%)Not Hispanic or Latino75 (98.7)81 (98.8)Hispanic or Latino1 (1.3)1 (1.2)Education (yrs), mean (SD)17.08 (2.28)17.01 (2.35)Pathology indicatorsAmyloid status, n (%)A—54 (71.1)61 (74.4)A+22 (28.9)21 (25.6)APOE ε4 status, n (%)E4+41 (53.9)44 (53.7)E4—35 (46.1)38 (46.3)CSF pTau 181 (baseline), mean (SD)78.31 (40.29) N/ACSF pTau 181 (2‐year), mean (SD)82.60 (46.72) N/ACognitive assessmentsRAVLT‐DR (baseline), mean (SD)10.30 (2.70)10.21 (2.72)RAVLT‐DR (2‐year), mean (SD)9.63 (2.91)9.68 (2.89)RAVLT‐DR (change), mean (SD)−0.73 (2.72)−0.58 (2.73)BVMTR‐DR (baseline), mean (SD)7.64 (2.73)7.67 (2.69)BVMTR‐DR (2‐year), mean (SD)6.91 (2.90)6.94 (2.83)BVMTR‐DR (change), mean (SD)−0.72 (2.52)−0.72 (2.54) Participant demographics. TABLE 2Inflammatory analytes significantly correlated with longitudinal cognitive change.CSFPlasmaNameGeneNameGeneα2‐macroglobulinA2MC‐C motif chemokine 5 (RANTES)CCL5α‐enolaseENO1C‐C motif chemokine 8 (MCP‐2)CCL8C‐C motif chemokine 2 (MCP‐1)CCL2C‐C motif chemokine 16 (HCC‐4)CCL16C‐C motif chemokine 14 (HCC‐1)CCL14C‐C motif chemokine 25 (TECK)CCL25C‐C motif chemokine 16 (HCC‐4)CCL16C‐C motif chemokine 20 (MIP‐3a)CCL20C‐C motif chemokine 18 (PARC)CCL18C‐C motif chemokine 23 (MPIF‐1)CCL23C‐C motif chemokine 25 (TECK)CCL25C‐reactive proteinCRPC‐reactive proteinCRPC‐X‐C motif chemokine 10 (IP‐10)CXCL10C‐X‐C motif chemokine 16CXCL16C‐X‐C motif chemokine 13 (BLC)CXCL13Eotaxin‐1CCL11C‐X‐C motif chemokine 14 (BRAK)CXCL14MatrilysinMMP7Eotaxin‐1CCL11Plasminogen activator inhibitor 1SERPINE1FerritinFTL|FTH1Glial fibrillary acidic proteinGFAPInterleukin‐2IL2Interleukin‐5IL5Interleukin‐9IL9Interleukin‐12 subunit beta (IL‐12 p40)IL12BInterleukin‐16IL16Interleukin‐17AIL17AInterleukin‐17BIL17BInterleukin‐17FIL17FInterleukin‐22IL22Matrix metalloproteinase‐9MMP9Intercellular adhesion molecule 1ICAM1Stromelysin‐2MMP10 Note: |r| > 0.2 = only correlations with an absolute correlation coefficient greater than 0.2 were included. Inflammatory analytes significantly correlated with longitudinal cognitive change. Note: |r| > 0.2 = only correlations with an absolute correlation coefficient greater than 0.2 were included. Amyloid‐ and APOEε4 ‐Related Analyte Comparisons Amyloid Positive (A+) Vs. Amyloid Negative (A‐) Group Differences We first sought to determine whether analytes measured in plasma or CSF were consistently altered at baseline when comparing A+ and A‐ individuals. Among 1331 analytes from CSF analyzed by Limma regression, 48 (3.61%) follow‐up‐to‐baseline analyte ratios were significantly associated with amyloid status after FDR correction (Figure 2A). The sensitivity analysis showed consistent findings, with 44 of 48 analytes remaining significant after adjustment for Δp‐tau181 (Supporting Information 2). Among 1501 analytes from plasma analyzed by Limma regression, one analyte ratio, FABP6 (β = −0.11, FDR = 0.02), was significantly decreased compared to the ratios in the A‐ group after FDR correction (Figure 2B). Volcano plots of analytes with significantly increased and decreased abundance associated with amyloid‐positive status in CSF (A), plasma (B), and with APOEε4 status in CSF (C) and plasma (D). APOEε4 Positive (E4+) Vs. APOEε4 Negative (E4‐) Group Differences Next, we considered whether changes in analyte abundance in CSF and plasma were related to the presence of an APOE ε4 risk allele in study participants. Among the whole cohort, analytes in CSF analyzed by Limma regression, 350 (26.30%) follow‐up‐to‐baseline analyte ratios were significantly associated with APOE status after FDR correction (Figure 2C). Most associations reflected higher analyte ratios in the E4+ group compared with the E4‐ group. The sensitivity analysis showed consistent findings, with 322 of 350 analytes remaining significant after adjustment for Δp‐tau181 (Supporting Information 2). Among whole cohort analytes in plasma analyzed by Limma regression, 91(6.06%) follow‐up‐to‐baseline ratios were significantly associated with APOE status after FDR correction (Figure 2D). No statistically significant Amyloid status × APOE ε4 interaction effects were identified for any CSF or plasma analytes after false discovery rate (FDR) correction (Supporting Information 3 and Supporting Information 4). Analytes Associated With Both Amyloid and APOEε4 Status at Baseline in CSF To identify CSF analytes jointly regulated by both amyloid and APOEε4 status, we identified the significant analytes from both analyses. Of the 350 differentially abundant analytes, 44 analytes were shared between the two conditions (Figure 3A). All 44 overlapping analytes exhibited a concordant upregulation pattern (Figure 3B). Analytes significantly related to both amyloid status and APOE status in CSF. (A) Venn plot shows a total of 44 analytes significantly related to both amyloid‐positive status and APOEε4 status. (B) All 44 analytes show positive effects. (C) Louvain clustering stability heatmap with Eigen centrality modules classification after 100 times runs. These findings indicate that amyloid pathology and APOE genotype converge on a shared longitudinal CSF proteomic program, rather than acting through entirely independent molecular pathways. To detect the function of these 44 co‐regulated analytes, we constructed a co‐expression network and applied Louvain community detection with 100 bootstrap replicates. Modules exhibiting greater than 80% membership stability were retained, ensuring robust and reproducible module assignments (Figure 3C). The analysis classed six distinct modules and seven models with only one analyte. Among those, clusters 2 and 4 are the two largest modules, including EFNB2, NPTN, UNC5D, CNTN4, and MDH2. To understand the relationship between analytes that changed in CSF abundance over the two‐year period, we constructed a co‐expression network and quantified nodal importance using eigenvector centrality (Figure 4A). Hub nodes, defined as the top 20% of analytes by eigenvector centrality score, were clustered within the two largest modules (Core modules) and were distinguished as diamonds in the network. Core modules exhibited significantly higher eigenvector centrality scores compared to other modules (Wilcoxon rank‐sum test, p < 0.001), indicating that the overlapping analytes in CSF are organized around a central regulatory backbone rather than uniformly distributed interactions (Figure 4B). (A) Co‐expression network among CSF co‐regulated analytes mapping eigen centrality modules. (B) Comparison of eigen centrality between core and peripheral modules using Wilcoxon rank‐sum test. (C) Pathways centralized across core modules (model 2 and model 4). (D) Baseline and 2‐year levels of EFNB2, NPTN, and UNC5D between amyloid and APOEε4 status. In the network, the identified hub analytes, EFNB2, NPTN, and UNC5D, occupied central positions within modules 2 and 4, consistent with their broad connectivity to co‐regulated neighbors. Given that all analytes were quantified as longitudinal ratios (2‐year follow‐up/baseline), the hub analytes reflect their rate of change over time, which suggests that these molecules undergo coordinated dynamic remodeling in response to amyloid‐ and APOE‐driven risk. Pathway enrichment analysis of the 44 analytes shows that each core module independently revealed convergent enrichment across axon guidance, nervous system development, and netrin‐mediated repulsion signaling. These pathways converged on a set of core analytes within the central modules, among which EFNB2, NPTN, and UNC5D emerged as representative molecular effectors linking pathway enrichment to network topology. Pathways revealed a structured division of biological functions across modules. Module 2 contributed predominantly to pathways involving cell–cell signaling and synaptic interface organization, including LGI–ADAM interactions and GPI‐anchored protein synthesis, which are critical for membrane protein localization and synaptic adhesion. In contrast, module 4 showed stronger enrichment in NCAM1 interactions and heparan sulfate (HS‐GAG) metabolism, pathways that regulate extracellular matrix composition, neurite outgrowth, and synaptic plasticity (Figure 4C). EFNB2, NPTN, and UNC5D map to complementary aspects of these enriched pathways. EFNB2 is involved in contact‐dependent signaling and synaptic interface remodeling. NPTN is positioned within pathways governing synaptic stabilization and membrane protein organization, reflecting its role in maintaining functional synaptic architecture. UNC5D corresponds to module 4, enriched in axon‐guidance and extracellular matrix‐related pathways, consistent with its role in repulsive signaling and structural reorganization of neural circuits. These pathway‐level associations are supported by direct longitudinal evidence (Figure 4D). Across amyloid and APOEε4 status, EFNB2, NPTN, and UNC5D exhibited consistently elevated analyte levels at both baseline and 2‐year follow‐up, with preserved group differences over time. Thus, these analytes demonstrated both relatively high baseline abundance and sustained longitudinal elevation, suggesting that certain components of the A+/E4+ molecular program may reflect stable disease‐associated set points rather than purely dynamic changes. We also examined whether the overlapping analyte network was modulated by sex. No significant interaction effects between analyte levels and sex were observed across the 44 co‐regulated analytes, indicating that the network structure is consistent across sexes. We next evaluated whether longitudinal changes in analyte levels were associated with cognitive decline in a sex‐specific manner. No substantial differences in analyte ratio distributions were observed between females and males (Figure 5A). However, sex‐stratified analyses revealed consistent directional differences in RAVLT‐DR outcomes (Figure 5B). In females, increased analyte ratios were positively associated with cognitive change across EFNB2, NPTN, and UNC5D (β = 3.96, 3.53, and 3.07, respectively), whereas in males, the associations were consistently negative (β = −2.59, −2.05, and −1.73). Notably, NPTN was detected as a statistically significant association in females (p = 0.02). No significant associations were observed for BVMTR‐DR (Figure 5C), suggesting that these effects may be specific to verbal memory domains. Sex‐dependent associations between analyte dynamics and cognitive change. (A) Distribution of longitudinal cytokine ratios for EFNB2, NPTN, and UNC5D stratified by sex. (B) Associations between analyte ratios and RAVLT‐DR change stratified by sex. (C) Associations between analyte ratios and BVMTR‐DR change stratified by sex. Both CSF and plasma show most proteins clustered at low median absolute log2 ratios (indicating good stability over two years), but plasma had a broader distribution and more proteins with larger deviations, suggesting overall lower protein stability over two years than CSF Figure S1 and Figure S2. Rank list of CSF and plasma proteins based on their stability over two years is provided in Supporting Information 5 and Supporting Information 6. Internal Validation of Key Inflammatory Analytes and Cognitive Decline Table 2 provides inflammatory analytes significantly correlated with longitudinal cognitive change. Best subsets regression identifies optimal predictor combinations, while elastic‐net regression stabilizes selection under correlation; comparing both results distinguishes robust predictors of cognitive decline from model‐sensitive ones and tests whether static or dynamic measures best predict outcomes. Of the eighty‐six analytes that were used in the internal validation analysis, twelve CSF inflammatory analytes and 25 plasma inflammatory analytes were identified as being significantly correlated with either change in RAVLT‐DR score or change in BVMTR‐DR score over two years. Linear Regression Estimates of Associations Between Baseline Inflammatory Analytes and Longitudinal Cognitive Changes Of the CSF analytes, α2‐macroglobulin, MCP‐1/CCL2, eotaxin‐1/CCL11, HCC‐1/CCL14, HCC‐4/CCL16, PARC/CCL18, TECK/CCL25, and MMP‐7 were significantly associated with cognitive change in at least one of the statistical models selected by best subsets regression analysis. Some of the analytes were significant in more than one model. Eotaxin‐1 was significantly and positively associated with change in BVMTR‐DR score in both the whole cohort and in every subgroup. In addition, α2‐macroglobulin was significantly and positively associated with change in RAVLT‐DR in the whole cohort, the A‐ group, and the E4‐ group (Table 3, Table S2 and Table S3). TABLE 3Summary of significant best subsets regression and elastic‐net regression results.BiofluidAnalyteBest subsets regressionElastic‐net regressionCognitive outcomeDirectionCSFα2‐macroglobulinWhole cohort, A‐, E4—Whole cohort, E4—RAVLT‐DRPositiveEotaxin‐1/CCL11Whole cohort, all subgroupsWhole cohort, A—BVMTR‐DRPositiveMCP‐1/CCL2Whole cohort, A—Whole cohort, A—RAVLT‐DRNegativeMMP‐7E4‐E4—RAVLT‐DRPositivePARC/CCL18A—A—BVMTR‐DRNegativeTECK/CCL25A‐, E4+Whole cohort, A‐, E4+BothNegativePlasmaHCC‐4/CCL16Whole cohort, E4+Whole cohort, E4+RAVLT‐DRNegativeIL‐17AWhole cohort, A‐, E4‐, E4+E4+BothPositiveIL‐2E4‐, E4+E4+BothPositiveIL‐5Whole cohort, A+, E4—Whole cohortRAVLTNegative Summary of significant best subsets regression and elastic‐net regression results. Nineteen of the plasma analytes (RANTES/CCL5, MCP‐2/CCL8, eotaxin‐1/CCL11, HCC‐4/CCL16, MIP‐3A/CCL20, MPIF‐1/CCL23, TECK/CCL25, CRP, IP‐10/CXCL10, BLC/CXCL13, ferritin, GFAP, IL‐2, IL‐5, IL‐17A, IL‐17B, IL‐17F, IL‐22, and sICAM‐1) were significantly associated with cognitive change in at least one of the statistical models selected by best subsets regression analysis. IL‐17B was significantly and negatively associated with BVMTR‐DR change in the whole cohort and all subgroups except E4‐. IL‐17A was significantly and positively associated with BVMTR‐DR change in the whole cohort, the A‐ group, and the E4‐ group. IL‐5 was significantly and negatively associated with RAVLT‐DR change in the whole cohort, the A+ group, and the E4‐ group (Table S4 and Table S5). Validation by Elastic‐Net Regression of Baseline Inflammatory Analytes and Bioinformatics Analysis Six of the 8 significant CSF analytes identified by best subsets regression (α2‐macroglobulin, MCP‐1/CCL2, TECK/CCL25, eotaxin‐1/CCL11, PARC/CCL18, and MMP‐7) were also identified by elastic‐net regression as significantly associated with cognitive change in selected models (Table S6 and Table S7, Figure 6a and Figure 7a). The IL‐17 signaling pathway was a key pathway identified by KEGG (FDR 0.0274) for the analytes identified above. CSF eotaxin‐1 was an especially robust analyte given its association with BVMTR‐DR change across all subgroups. Coefficients for log10‐transformed (A) CSF analytes and (B) plasma analytes estimated by models selected by best subsets regression analysis performed on the whole cohort, subset by amyloid‐positive status (A+, A‐), and subset by APOEε4 (E4+, E4‐) status. Significant coefficients are highlighted in red. Coefficients for log10‐transformed (A) CSF analytes and (B) plasma analytes selected by elastic‐net regression models performed on the whole cohort, subset by amyloid status (A+, A‐), and subset by APOEε4 (E4+, E4‐) status. Significant coefficients are highlighted in red. No analytes were selected for models predicting change in BVMTR‐DR. Among the plasma analytes, HCC‐4/CCL16, IL‐2, IL‐5, and IL‐17A were also identified as significantly associated with change in RAVLT‐DR in models selected via elastic‐net regression. Elastic‐net regression did not select any plasma analytes for the models predicting change in BVMTR‐DR. These analytes were again enriched for the IL‐17 signaling pathway within KEGG (FDR 0.0064). (Table S8, Figure 6b and Figure 7b). Validation Between SomaLogic and MILLIPLEX Multianalyte Platforms Eotaxin‐1's area under the curve (AUC) was calculated for logistic regression models predicting at least one standard deviation of decline in RAVLT‐DR over 2 years using baseline CSF and plasma eotaxin‐1 in the whole cohort, A+, and E4+ subgroups using the SomaLogic platform. Because multiplexed platforms can exhibit differences in limit of detection and specificity due to alternative methods of analyte capture and visualization, these SomaLogic results were validated for plasma eotaxin‐1 measured using the MILLIPLEX platform. When accounting for age, sex, E4 status and CSF ptau181, the results were consistent for plasma eotaxin‐1 measured across the two platforms (SomaLogic, AUC 0.79 [95% CI 0.67 to 0.91]; MILLPLEX, AUC 0.78 [95% CI 0.65 to 0.92]) and similar to CSF eotaxin‐1 measured with the SomaLogic platform (AUC 0.81 [95% CI, 0.70 to 0.92]). Additional sensitivity analysis completed by removing outliers supports the robustness of the CSF and plasma SomaLogic and plasma/Milliplex results. See (Table S9a,b). Summary of Validation Results Eotaxin‐1 (CCL11) was the strongest predictor of longitudinal cognitive decline, including clinically significant (> 1 SD) decline in RAVLT‐DR over 2 years, consistent across regression models and assay platforms. CSF analytes, including α2‐macroglobulin, MCP‐1, PARC, TECK, and MMP‐7, showed stable associations between baseline levels and later cognitive decline in both best subsets and elastic‐net models, indicating predictive value of baseline inflammatory burden independent of longitudinal change. Plasma analytes (HCC‐4, IL‐2, IL‐5, IL‐17A) showed weaker but reproducible associations. In contrast, EFNB2, NPTN, and UNC5D were amyloid‐ and APOE‐related network hubs but were not predictive of cognitive decline. Discussion In this longitudinal study of cognitively normal older adults enriched for risk of AD, we evaluated multiplex analytes in CSF and plasma in relation to amyloid status, APOEε4 genotype, and cognitive change over two years. Several key observations were striking. First, we note that both amyloid positivity and APOEε4 carrier status were associated with distinct longitudinal proteomic signatures, particularly in CSF. Notably, a subset of 44 analytes demonstrated consistent upregulation across both conditions, suggesting convergence on a shared molecular program. Network and pathway analyses further indicated that these analytes are organized within coordinated modules enriched for processes related to axon guidance, synaptic organization, and extracellular matrix signaling. The identification of central hub analytes, including EFNB2, NPTN, and UNC5D, supports the presence of a potentially coordinated biological response rather than isolated biomarker changes. However, these findings should be interpreted as evidence of association, and the functional roles of these analytes in AD pathophysiology require further validation. The modular organization and identification of hub analytes suggest that inflammatory changes may occur through coordinated network‐level processes rather than isolated biomarker effects. The substantially larger number of CSF analytes associated with APOEε4 status relative to amyloid status suggests that APOE‐related processes may exert broader effects on proteomic regulation in the preclinical stage. Second, inflammatory analytes associated with longitudinal cognitive change were identified in both CSF and plasma. Across the different statistical approaches used, a subset of CSF analytes (including α2‐macroglobulin, MCP‐1/CCL2, TECK/CCL25, eotaxin‐1/CCL11, PARC/CCL18, and MMP‐7) demonstrated consistent associations with cognitive outcomes, suggesting greater robustness relative to other markers. In plasma, fewer analytes showed consistent selection across modeling approaches, and no plasma markers were retained in elastic‐net models predicting visuospatial memory change, indicating weaker or less stable effects in peripheral measures. Plasma MCP1/CCL2, which has been reported previously to be related to longitudinal verbal delayed recall decline among older adults [38], and CSF MCP1/CCL2, associated with a faster rate of progression in the mild cognitive impairment stage of AD [39], were again notable predictors of verbal memory delayed recall decline in the whole cohort in the CSF. Importantly, several of the identified analytes converged on inflammatory signaling pathways, including IL‐17‐related processes. It is noted that IL‐17 pathway activation is related to downstream upregulation of pro‐inflammatory cytokines including IL‐6 and TNF‐α, as well as chemokines CXCL1, CXCL2, and CXCL8 that contribute to the recruitment of neutrophils and enhanced tissue inflammation [40]. However, the variability across models and outcomes suggests that these pathway‐level findings should be considered exploratory. Moreover, given the limited two‐year follow‐up, associations between inflammatory analytes and cognitive change should be interpreted cautiously and require validation in larger cohorts with longer longitudinal follow‐up. Among individual analytes, eotaxin‐1/CCL11, a chemokine that binds to the CCR3 receptor and attracts eosinophils, emerged as a relatively consistent correlate of cognitive change across subgroups and analytic methods, and showed concordant performance across assay platforms. This cross‐platform consistency supports its potential utility as a candidate biomarker; however, given the pleiotropic nature of chemokines and their sensitivity to systemic inflammatory states, its AD‐specific role cannot be commented on based on this analysis alone. However, a role for plasma eotaxin‐1 in memory dysfunction in the context of varied MCI and AD dementia along with MCP1 has been previously reported [41]. Plasma eotaxin‐1 levels were also correlated with disease age of onset of symptoms [42], and the gene locus related to it was associated with a protective effect in presenilin‐1 gene carriers [42]. Eotaxin‐1 is able to pass the blood–brain barrier [43], and in that context it is notable that AUCs for eotaxin‐1 relating to a clinically meaningful > 1 standard deviation of decline in RAVLT‐DR over two years in the current cohort of older individuals at high risk for AD were similar for both CSF and plasma. The role of eotaxin‐1 as a potential biomarker of cognitive decline to complement classical AD biomarkers needs validation in other cohorts. Sex‐stratified analyses revealed directional differences in associations between selected analytes and verbal memory outcomes, with positive associations observed in females and negative associations in males. The absence of significant interaction effects in the full cohort, alongside directional differences in stratified analyses, suggests that these findings may reflect limited power or model instability rather than true sex‐specific biological effects. The study results also suggest that the availability of data simultaneously from both CSF and peripheral blood may be a unique opportunity to distinguish incidental peripheral processes that change inflammatory response (e.g., medication use) from inflammatory changes of significance to AD‐related outcomes. The markedly stronger associations observed in CSF relative to plasma, particularly for amyloid status, likely reflect the greater proximity of CSF to central nervous system processes, whereas plasma measures may be diluted by systemic influence. Taken together, our results suggest that accumulating AD pathology may play a more direct role in shaping inflammatory responses in both CSF and plasma, but they are in contrast to results from evaluating pre‐diagnostic biomarkers of other neurodegenerative diseases such as Parkinson's disease and amyotrophic lateral sclerosis, in which extensive biofluidic profiling prior to disease onset has not uncovered reproducible predictors of phenoconversion [44, 45, 46, 47]. Several important limitations should be considered. Given the challenges in recruiting a deeply characterized normal aging at‐risk for AD cohort, the sample size was modest. This limits statistical power and increases the risk of model instability given the high dimensionality of the proteomic data. To address these limitations, we applied FDR correction and complementary modeling approaches, including elastic‐net regression, with key findings further validated across independent assay platforms. While some inconsistency between modeling approaches was observed, multiple convergent findings across methods support the robustness of core results, though validation in independent cohorts remains an important next step. The two‐year follow‐up period may also have limited detection of meaningful cognitive decline in cognitively normal individuals, particularly during preclinical AD. Use of domain‐specific cognitive measures may have contributed to variability across outcomes. The cohort had limited racial and ethnic diversity. Because inflammatory biomarkers are influenced by systemic factors, participants were screened for inflammatory and autoimmune diseases, uncontrolled medical conditions, infections, recent hospitalizations, and neurological or psychiatric disorders, though lifestyle factors may still represent residual confounding. These limitations may affect generalizability and the specificity of observed associations to AD‐related processes, requiring validation in real‐world cohorts. Finally, while pathway enrichment and network analyses provide useful frameworks for interpreting high‐dimensional data, they are inherently sensitive to input selection and may be viewed as hypothesis‐generating, requiring replication in independent cohorts. Despite these limitations, this study has several strengths, including the simultaneous assessment of CSF and plasma analytes, longitudinal design, integration of proteomic, genetic, and imaging data, and the use of multiple analytic approaches to assess robustness. The convergence of a subset of analytes across methods and their organization within coherent biological networks provides a foundation for future investigation. Conclusions Our longitudinal proteomic analysis shows that CSF, more than plasma in cognitively normal older adults, captures APOEε4 and amyloid‐driven proteomic changes linked to cognitive function in individuals at risk for AD. Core hub analytes, including EFNB2, NPTN, and UNC5D, show coordinated dynamic changes and central network roles, reflecting biologically meaningful remodeling tied to disease risk. Inflammatory proteins in CSF and plasma are associated with cognitive decline, suggesting potential early biomarkers requiring validation. These findings highlight the value of CSF proteomics for elucidating mechanisms of cognitive impairment and identifying targets for monitoring and intervention. Author Contributions J.A.P. obtained funding, design and conceptualized study, interpreted the data; drafted the manuscript for intellectual content. A.Z., C.M., M.W., Y.Y., Y.H. and F.C. analyzed the data; revised the manuscript for intellectual content. L.M.B. and M.K. helped perform CSF AD biomarker and Luminex biomarker evaluations; revised the manuscript for intellectual content. F.P.D. helped perform Amyloid. PET; revised the manuscript for intellectual content. M.N., A.B. and J.B.L. interpreted the data; revised the manuscript for intellectual content. S.M.R. obtained funding, study design and interpreted the data; drafted the manuscript for intellectual content. Funding Funding from K23AG055685, P30AG072959, R01AG078763, R01AG022304, Annabelle Foundation, and Keep Memory Alive Foundation. National Institute on Aging (NIA). Conflicts of Interest Ming Wang: Research funding from the National Institutes of Health. Lynn M. Bekris: Research funding from the National Institutes of Health. Frank P. DiFilippo: Research funding from the National Institutes of Health, Michael J Fox Foundation. Aaron Burberry: Research funding from the National Institutes of Health, Target ALS, Packard Center for ALS. James B. Leverenz: Research funding from the National Institutes of Health, Aging Mind. Foundation, Amprion, GE Healthcare, Lewy Body Dementia Association, Michael J Fox Foundation. Jagan A. Pillai: Research funding from the National Institutes of Health, Department of Defense, Annabelle Foundation, and Keep Memory Alive Foundation. Stephen Rao: Research funding from the National Institutes of Health. The other authors listed have none. Supporting information Figure S1: Plot of number of CSF proteins (blue) vs. median absolute log2 ratio over two years to assess the overall extent and distribution of protein‐level changes. Figure S2: Plot of number of Plasma proteins (red) vs. median absolute log2 ratio over two years to assess the overall extent and distribution of protein‐level changes. Supporting Information: 1 Inclusion/exclusion criteria of the cohort. Supporting Information: 2 Sensitivity analysis of the Limma model adjusted for longitudinal change in CSF p‐tau181. Supporting Information: 3 Sensitivity analysis of the Limma model including the amyloid status × APOE ε4 status interaction term in CSF. Supporting Information: 4 Sensitivity analysis of the Limma model including the amyloid status × APOE ε4 status interaction term in Limma. Supporting Information: 5 Rank list of CSF proteins based on stability over two years. Supporting Information: 6 Rank list of plasma proteins based on stability over two years. Table S1: List of CSF and plasma inflammatory analytes evaluated and estimated limits of detection (LoD). Table S2: Best subsets of regression‐selected CSF biomarkers predicting RAVLT Delayed Recall change after adjustment for age, sex, APOE ε4 status, and education. Table S3: Best subsets of regression‐selected CSF biomarkers predicting BVMT Delayed Response change after adjustment for age, sex, APOE ε4 status, and education. Table S4: Best subsets of regression‐selected plasma biomarkers predicting RAVLT Delayed Recall change after adjustment for age, sex, APOE ε4 status, and education. Table S5: Best subsets of regression‐selected plasma biomarkers predicting BVMT Delayed Response change after adjustment for age, sex, APOE ε4 status, and education. Table S6: Elastic net regression‐selected CSF biomarkers predicting RAVLT Delayed Recall change over 2 years after adjustment for age, sex, APOE ε4 status, and education. Table S7: Elastic net regression‐selected CSF biomarkers predicting BVMT Delayed Response change over 2 years after adjustment for age, sex, APOE ε4 status, and education. Table S8: Elastic net regression‐selected plasma biomarkers predicting RAVLT Delayed Recall change over 2 years after adjustment for age, sex, APOE ε4 status, and education. Table S9: (a) Area under the curve and 95% Confidence interval for logistic regression models predicting > 1SD decline in RAVLT Delayed Recall over 2 years using baseline CSF and plasma eotaxin in various subgroups using two platforms, Somalogic and MILLIPLEX. (b) Sensitivity analysis after removing outliers notes the robustness of the CSF and plasma Somalogic and plasma MILLIPLEX results in a. Acknowledgments We thank the participants and families who took part in the study at Cleveland Clinic Center for Brain Health. Funding Statement This work was funded by National Institute on Aging (NIA), National Institutes of Health grants K23AG055685, P30AG072959, R01AG022304, and R01AG078763. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Associated Data Supplementary Materials Figure S1: Plot of number of CSF proteins (blue) vs. median absolute log2 ratio over two years to assess the overall extent and distribution of protein‐level changes. Figure S2: Plot of number of Plasma proteins (red) vs. median absolute log2 ratio over two years to assess the overall extent and distribution of protein‐level changes. Supporting Information: 1 Inclusion/exclusion criteria of the cohort. Supporting Information: 2 Sensitivity analysis of the Limma model adjusted for longitudinal change in CSF p‐tau181. Supporting Information: 3 Sensitivity analysis of the Limma model including the amyloid status × APOE ε4 status interaction term in CSF. Supporting Information: 4 Sensitivity analysis of the Limma model including the amyloid status × APOE ε4 status interaction term in Limma. Supporting Information: 5 Rank list of CSF proteins based on stability over two years. Supporting Information: 6 Rank list of plasma proteins based on stability over two years. Table S1: List of CSF and plasma inflammatory analytes evaluated and estimated limits of detection (LoD). Table S2: Best subsets of regression‐selected CSF biomarkers predicting RAVLT Delayed Recall change after adjustment for age, sex, APOE ε4 status, and education. Table S3: Best subsets of regression‐selected CSF biomarkers predicting BVMT Delayed Response change after adjustment for age, sex, APOE ε4 status, and education. Table S4: Best subsets of regression‐selected plasma biomarkers predicting RAVLT Delayed Recall change after adjustment for age, sex, APOE ε4 status, and education. Table S5: Best subsets of regression‐selected plasma biomarkers predicting BVMT Delayed Response change after adjustment for age, sex, APOE ε4 status, and education. Table S6: Elastic net regression‐selected CSF biomarkers predicting RAVLT Delayed Recall change over 2 years after adjustment for age, sex, APOE ε4 status, and education. Table S7: Elastic net regression‐selected CSF biomarkers predicting BVMT Delayed Response change over 2 years after adjustment for age, sex, APOE ε4 status, and education. Table S8: Elastic net regression‐selected plasma biomarkers predicting RAVLT Delayed Recall change over 2 years after adjustment for age, sex, APOE ε4 status, and education. Table S9: (a) Area under the curve and 95% Confidence interval for logistic regression models predicting > 1SD decline in RAVLT Delayed Recall over 2 years using baseline CSF and plasma eotaxin in various subgroups using two platforms, Somalogic and MILLIPLEX. (b) Sensitivity analysis after removing outliers notes the robustness of the CSF and plasma Somalogic and plasma MILLIPLEX results in a. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.