Longitudinal plasma proteomics: relation to incident Alzheimer's disease dementia and biomarkers

Abstract

INTRODUCTION

We investigated whether longitudinal changes in plasma proteins were associated with baseline cognitive stages related to Alzheimer's disease (AD), their progression, and AD biomarkers.

METHODS

We analyzed longitudinal proteomics (SomaScan 7K) data (N = 347) from the Indiana AD Research Center using linear mixed‐effects models for associations with baseline cognitive stages, AD dementia (ADD) conversion, and AD imaging/plasma biomarkers, followed by machine learning analysis to evaluate predictive performance for incident ADD.

RESULTS

Our analysis identified two proteins (ACES and IGFALS) associated with baseline diagnosis stages and six proteins (ACES, C7, ZCD1, IL‐17C, CC055, and SO5A1) associated with incident ADD. Longitudinal changes of the identified proteins were also associated with AD imaging/plasma biomarkers. The inclusion of longitudinal protein changes yielded an AUC of 84.8% for predicting incident ADD.

CONCLUSION

Our findings showed molecular signatures for AD progression and the potential of dynamic changes in plasma proteins as biomarkers for predicting incident ADD.

Highlights

• Changes in plasma ACES and IGFALS linked to baseline AD cognitive stages

• Changes in ACES, C7, ZCD1, IL‐17C, CC055, and SO5A1 associated with incident ADD

• Changes in those proteins correlated with baseline AD imaging and plasma biomarkers

• Proteomics model achieved 84.8% AUC‐ROC in predicting incident ADD

1. BACKGROUND

Alzheimer's disease (AD), the most common cause of dementia, is characterized by the accumulation of amyloid beta (Aβ) plaques, neurofibrillary tangles (NFTs) composed of tau protein, and subsequent neurodegeneration. ¹ Accordingly, a disease stage classification based on biomarkers of amyloid, tau, and neurodegeneration has been proposed. ² , ³ Moreover, efforts are underway to expand this classification by incorporating biomarkers related to inflammation, vascular factors, and other co‐pathologies; however, the corresponding biomarkers for many of these co‐pathologies remain largely unexplored. ² To address this, proteomics studies have been conducted to identify additional biomarkers reflecting a broader range of AD‐related pathologic processes beyond the typical amyloid and tau pathologies. ⁴ , ⁵ , ⁶

Despite its clear advantages, there remains a relative scarcity of blood‐based proteomic research in AD compared to studies using cerebrospinal fluid (CSF) or brain tissue. To date, exploratory studies for novel biomarkers have primarily utilized CSF and brain tissue. A meta‐analysis of six CSF datasets identified 311 upregulated and 165 downregulated proteins in AD, and integration with brain proteome data improved the selection of specific proteins among these. ⁴ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² However, both brain and CSF proteome analyses are challenging to implement in routine clinical settings or for large‐scale longitudinal follow‐up due to their invasiveness and complexity. ⁴ Therefore, despite the potential for molecular changes as proteins cross the blood–brain barrier, ongoing research is focused on identifying biomarkers in plasma, which can be obtained through a simple blood sampling. ¹³ In fact, 46 proteins related to biological functions, including immune/inflammatory response, ion transport, and insulin suppression, were replicated across two or more of 17 independent cohorts. ¹⁴ These findings highlight the potential of plasma as a viable source of biomarkers, supporting continued research in plasma‐based proteomics to identify reliable and accessible biomarkers for clinical application.

Most proteomics studies in clinical cohorts have relied on cross‐sectional data. However, given that AD is a progressive neurodegenerative disorder characterized by gradual cognitive decline over an extended period, understanding the longitudinal dynamics of disease‐related molecular changes is vital. Some studies have attempted to infer temporal protein changes from cross‐sectional data. For instance, they stratified individuals by AD biomarker stage and inferred protein trajectories by comparing protein levels across stages. ¹⁵ , ¹⁶ Another approach linked protein levels to the time interval between measurement and AD dementia (ADD) onset to approximate preclinical alterations. ¹⁷ While informative, these findings remain largely inferential. One study examined longitudinal changes in NPTX2 as prognostic biomarkers, but in CSF rather than plasma. ¹⁸ Recognizing the advantages of plasma, there is a critical need for more longitudinal studies that leverage follow‐up plasma proteomic data to better characterize disease progression and identify dynamic biomarkers of AD.

In this study, we aimed to investigate whether temporal changes in individual protein levels were associated with baseline cognitive stage and ADD conversion status, using longitudinal proteomics data from the Indiana Alzheimer's Disease Research Center (IADRC). In addition, we investigated whether the change ratios of plasma proteins were associated with baseline AD imaging and plasma biomarkers. Finally, we evaluated whether incorporating the changes in these proteins improved the performance of machine learning models for predicting incident ADD. This study highlights that, beyond static protein levels, dynamic changes in plasma proteins may serve as valuable biomarkers of cognitive status and AD prognosis and may provide a foundation for understanding the temporal dynamics of AD pathophysiology.

2. METHODS

2.1. Study participants

RESEARCH IN CONTEXT

• Systematic review: We reviewed the literature on proteomics studies in Alzheimer's disease (AD), with a focus on longitudinal plasma protein data. While many prior studies have used cerebrospinal fluid or brain tissue, few have examined longitudinal changes in plasma proteins, particularly in relation to AD progression.

• Interpretation: We identified several plasma proteins whose temporal changes were associated with baseline cognitive status, incident ADD, and established imaging/plasma biomarkers. Incorporating dynamic protein changes improved the predictive performance for ADD conversion. These findings highlight the potential of longitudinal plasma proteomics to identify accessible biomarkers that reflect disease progression.

• Future directions: This study supports the importance of tracking longitudinal protein changes to understand AD pathophysiology. Future research is warranted to replicate our findings in independent larger datasets and explore the mechanisms linking identified proteins to neurodegeneration and AD risk.

Figure 1 illustrates an overview of the study design and analysis framework. Participants were recruited from IADRC as part of the Indiana Memory and Aging Study (IMAS) conducted at the Indiana University School of Medicine. We enrolled 346 participants whose plasma proteomic data were measured more than twice. The cohort included cognitively normal (CN) individuals, as well as participants with mild cognitive impairment (MCI) or ADD. CN participants were older adults without a significant performance deficit on cognitive testing. MCI was diagnosed as either amnestic or non‐amnestic MCI based on Petersen's criteria. ¹⁹ The diagnosis of ADD followed the 2011 National Institute on Aging and Alzheimer's Association criteria for probable ADD. ²⁰ AD‐specific biomarker status was not considered in these clinical diagnoses. To define ADD conversion status, individuals diagnosed with ADD at baseline were excluded from the analysis. The incident ADD group included participants whose cognitive status progressed from CN or MCI to ADD at any point during the follow‐up period. All other participants, including individuals who converted from CN to MCI, were categorized as the non‐incident ADD group. Participants underwent a comprehensive clinical visit that included clinical and neurologic assessments, neuropsychological testing with the Uniform Data Set 3 battery, and collection of blood samples. Longitudinal Clinical Dementia Rating Sum of Boxes (CDR‐SB) data were collected for all participants, with CDR‐SB assessments conducted consistently at each follow‐up visit.

FIGURE 1.

Overview of study design and analysis framework. Participants (N = 346) with plasma proteomic measurements at two or more timepoints were recruited from the IADRC. The analysis included (A) identification of significant proteins whose longitudinal changes were related to baseline cognitive stage (CN, MCI, AD) or ADD conversion status (non‐incident ADD and incident ADD groups) using linear mixed‐effects models; (B) among the proteins that were significant in Step 1, assessment of associations between longitudinal changes in plasma proteins and baseline AD biomarkers, including imaging and plasma biomarkers; and (C) machine‐learning‐based prediction of incident ADD using demographic variables, APOE ε4 carrier status, and estimated slopes of changes over time derived from significant proteins. AD, Alzheimer's disease; ADD, Alzheimer's disease dementia; CN, cognitively normal; FDR, False discovery rate; GFAP, glial fibrillary acidic protein; IADRC, Indiana Alzheimer's Disease Research Center; LMM, linear mixed‐effects model; MCI, mild cognitive impairment; MRI, magnetic resonance imaging; MTL, medial temporal lobe; NFL, neurofilament light chain; PET, positron emission tomography; p‐tau181, phosphorylated tau 181; ROI, region of interest; SUVR, standardized uptake value ratio.

2.2. Plasma sample collection

Blood samples were collected from participants by a standard venipuncture procedure into 10 mL EDTA‐treated tubes. Immediately after collection, the tubes were inverted 10 times and placed on wet ice. The samples were then centrifuged at 2000 × g for 15 min at 4°C. Following centrifugation, plasma was divided into 0.5 mL aliquots using 1.5 mL non‐sterile/skirted freezing tubes (Thermo Fisher Scientific, Catalog No.: 02‐681‐338). The blood samples were processed within 30 min of collection to ensure biomarker stability, and the aliquots were frozen and stored at −80°C within 2 h after collection.

2.3. Quality control procedure on proteomics data

Plasma protein measurements were conducted using the SomaScan 7K version 4.1 assay. ²¹ Initial normalization by SomaLogic addressed intra‐plate (via hybridization controls) and inter‐plate (via median signal) variability. Additional normalization used an external reference to correct biological variability. Aptamer‐level quality control excluded probes with scale factor deviations ≥0.5 or cross‐plate coefficient of variation ≥0.15. After log‐transformation, values beyond 1.5 × interquartile range (IQR) were set to NA (for “not applicable”). Aptamers or samples with detection in <65% were excluded, and a stricter 85% call rate was later applied. Non‐human protein targets were also removed. The final dataset comprised 7181 aptamers mapping to 6301 unique proteins. For further analysis, aptamer levels were log₂‐transformed. The detailed procedures, including normalization procedures and quality control criteria, are provided in the Supplementary Methods.

2.4. Imaging biomarkers

2.4.1. Positron emission tomography

A subset of participants underwent Aβ and tau positron emission tomography (PET) imaging using [18F]florbetapir or [18F]florbetaben for Aβ and [18F]flortaucipir for tau. PET data were reconstructed with an ordered subset expectation maximization algorithm and corrected for motion using SPM12. All scans were aligned to each participant's T1‐weighted magnetic resonance imaging (MRI) and spatially normalized to the MNI152 template. Aβ PET static images were generated from tracer‐specific time windows and normalized to the whole cerebellum using Centiloid (CL)‐defined regions of interest (ROIs). Resulting standardized uptake value ratio (SUVR) images were converted to CL units and smoothed with an 8‐mm Gaussian kernel; global cortical CL values were extracted. Tau PET images were created using 80‐ to 100‐min post‐injection data and normalized to the cerebellar crus. SUVRs were extracted from AD‐relevant brain regions based on FreeSurfer‐derived individual parcellations. CL and SUVR values were quantile‐normalized prior to further analysis. The detailed methods are available in the Supplementary Methods.

2.4.2. Magnetic resonance imaging

T1‐weighted MRI was acquired using a Siemens Prisma 3T scanner with a 64‐channel head coil, following the ADNI 2 protocol. The imaging sequence employed a three‐dimensional magnetization‐prepared rapid acquisition gradient echo (MPRAGE). Structural images were processed using FreeSurfer version 6.0 to obtain regional volumetric measures, particularly focusing on the hippocampus and total brain volume. Further details on MRI acquisition parameters and processing procedures are described in the Supplementary Methods.

2.4.3. A/T/N biomarker definition

The A biomarker was evaluated based on global cortical CL values obtained from Aβ PET scans, following quantile normalization procedures. ²² , ²³ The T biomarker was defined using the quantile‐normalized mean SUVR values from bilateral medial temporal lobe (MTL) ROI derived from tau PET imaging. The N biomarker was determined by calculating the total volume of the bilateral hippocampi from structural MRI data.

2.5. Plasma AD biomarkers

Plasma biomarkers, including Aβ42, Aβ40, phosphorylated tau 181 (p‐tau181), neurofilament light chain (NFL), and glial fibrillary acidic protein (GFAP), were quantified at two independent laboratories, and batch shipments of processed plasma were sent to the National Centralized Repository for Alzheimer's Disease and Related Dementias (NCRAD) Biomarker Assay Laboratory or the University of Gothenburg on dry ice. For both labs, the Aβ42/40 (“A”), NFL (“N”), and GFAP (“I”) biomarkers were analyzed using the multiplex N4PE assay on the Quanterix Simoa HD‐X platform. For p‐tau181 (“T”), two assays were utilized: the p‐tau181 home brew assay by the University of Gothenburg and the p‐tau181 v2 Advantage by NCRAD. Standardized laboratory protocols and quality control processes are described in the Supplementary Methods.

2.6. Statistical analysis

We compared the longitudinal changes in the proteome by baseline cognitive stage using linear mixed‐effects models. The models were fitted with time from the first visit (in months) and the level of each plasma protein as the dependent variable. To investigate whether baseline cognitive stages were associated with a distinct temporal pattern, an interaction term between time and baseline cognitive stage was included in the model. Age at blood collection, sex, apolipoprotein E (APOE) ε4 carrier status (defined as APOE ε2/ε4, ε3/ε4, ε4/ε4), and subarray (for protein analysis) were adjusted for. The regression model used is as follows:

Plasma protein level ∼ time from the first visit * baseline cognitive stage + age + sex + APOE ε4 carrier status + subarray + (1|individual)

Since the baseline cognitive stage consisted of three categories (CN, MCI, and ADD), post hoc analyses were performed to compare each pair (CN vs MCI, MCI vs ADD, and CN vs ADD). This analysis was performed for 7181 aptamers.

To compare the longitudinal changes in the proteome by ADD conversion status (incident ADD group vs non‐incident ADD group), we performed a linear mixed‐effects model using the same dependent variable and covariates. This analysis was also performed across 7181 aptamers. The model included an interaction term between time and ADD conversion status:

Plasma protein level ∼ time from the first visit * ADD conversion status + age + sex + APOE ε4 carrier status + subarray + (1|individual)

Additionally, in order to show the association between changes in CDR‐SB and the seven proteins that were significant in the prior two analyses, a linear mixed‐effects model with the interaction term between time and annualized changes in CDR‐SB was performed. Age at blood collection, sex, APOE ε4 carrier status, and subarray were adjusted for:

Plasma protein level ∼ time from the first visit (months) * annualized change in CDR‐SB + age + sex + APOE ε4 carrier status + subarray + (1|individual)

The annualized change in CDR‐SB was estimated with an independent linear mixed‐effects model with time (years since baseline visit) as the independent variable and CDR‐SB score as the dependent variable, allowing us to estimate the slope of CDR‐SB change over time for each individual. We then adjusted the estimated slopes for baseline CDR‐SB using linear regression and used the residuals as representative values of annualized CDR‐SB change.

We then investigated the association between baseline AD biomarkers and longitudinal changes in the proteome. For this analysis, only proteins that showed significant interactions in the prior models, related to baseline cognitive stage or ADD conversion status, were included as dependent variables. We constructed three separate models using global cortical CL values from Aβ PET, medial temporal lobe SUVR from tau PET, and total hippocampal volume from brain MRI as independent variables in interaction terms with time. For the Aβ PET model, covariates included age, sex, APOE ε4 carrier status, subarray, and Aβ PET type. The tau PET model included age, sex, APOE ε4 carrier status, and subarray as covariates. The MRI‐based hippocampal volume model included age, sex, APOE ε4 carrier status, subarray, and intracranial volume as covariates.

Results were reported as p values after adjusting for false discovery rate (FDR) using the Benjamini–Hochberg method. ²⁴ Statistical significance was set at p < 0.05 after FDR correction. Missing data were excluded from each analysis. All analyses were performed using R version 4.3.2 (R Foundation).

2.7. Machine learning prediction for incident ADD

To investigate whether changes in the identified proteins influenced the prediction of incident ADD, we conducted machine learning analyses to compare the performance of three models: a base model, a plasma biomarker model, and a proteomics model. The base model included baseline age, sex, and APOE ε4 carrier status. In the plasma biomarker model, baseline plasma AD biomarkers (p‐tau181 level and Aβ42/40 ratio) were added to the base model. In the proteomics model, we added the estimated slopes of proteins with significant interactions in the linear mixed‐effects models, specifically those related to baseline cognitive stage or ADD conversion status, to the plasma biomarker model.

Changes in each protein for individuals were estimated using a linear mixed‐effects model, with time from the first visit as the independent variable and subarray as a covariate. Baseline cognitive stage, ADD conversion status, and other demographic covariates were not included in this estimation. The random effect term (1 + time from the first visit | individual) accounts for individual‐specific intercepts and slopes, allowing both baseline levels and rates of protein change over time to vary across participants.

Plasma protein level ∼ time from the first visit + subarray + (1+time from the first visit|individual)

As protein levels at baseline may substantially influence the estimation of longitudinal trajectories, protein slopes were adjusted for baseline values using linear regression, and the resulting residuals were used as final estimated slopes in the model.

To address the class imbalance in the dataset (incident ADD group: N = 39; non‐incident ADD group: N = 255), we applied random downsampling to reduce the size of the non‐incident ADD group to 40 individuals. This process was conducted five times to mitigate the potential influence of sampling variability on model performance.

For the machine learning analysis, we employed STREAMLINE (Beta version 0.3.4), ²⁵ , ²⁶ a fully automated and transparent machine learning pipeline that supports end‐to‐end processing. We tested seven different machine learning algorithms (artificial neural network, decision tree, elastic net, logistic regression, naive Bayes, random forest, and support vector machine) to assess the robustness of the predictive contribution of protein changes across various model algorithms. All procedures were carried out using default configurations, including three‐fold cross‐validation with stratified sampling. Details of the hyperparameter settings are described in the Supplementary Methods. Model performance was assessed using the average values across five downsampled datasets for the following metrics: balanced accuracy, overall accuracy, F1 score, sensitivity, specificity, and area under the receiver operating characteristic curve (ROC AUC). Feature importance scores from each model were obtained from the STREAMLINE pipelines. ²⁵ , ²⁶

3. RESULTS

3.1. Demographic characteristics of study participants

Table 1 presents the demographic characteristics of the study participants at baseline and at each follow‐up visit. A total of 346 participants were included in this study (mean ± standard deviation [SD] age, 69.3 ± 9.0 years; 211 females [61.0%] and 135 males [39.0%]). Of the participants, 278 (80.3%) were non‐Hispanic White, while the remainder included 61 (17.6%) African American, five (1.4%) Hispanic, and two (0.6%) Asian. The median (IQR) follow‐up period was 24.4 (13.3 to 35.8) months; the median (IQR) number of visits was 3 (2 to 4); and the median (IQR) interval between visits was 13.0 (11.7 to 15.9) months. Figure S1 presents the distribution of visits and their intervals. Among participants with baseline cognitive stage of CN or MCI, 39 individuals converted to ADD during the follow‐up period (incident ADD group). The number of participants who underwent Aβ PET, tau PET, and brain MRI scans at baseline were 103, 72, and 201, respectively. After harmonization, baseline plasma biomarkers, including Aβ42/40 ratio, p‐tau181, NFL, and GFAP, were available for 241, 239, 250, and 241 participants, respectively.

TABLE 1.

Demographics characteristics of participants by follow‐up after baseline visit.

	Baseline (N = 346)	Year 1 follow‐up (N = 277)	Year 2 follow‐up (N = 198)	Year 3 follow‐up (N = 157)	Follow‐up after year 4 (N = 70)
Age at visit	69.3 ± 9.0	70.9 ± 8.9	71.6 ± 8.6	71.3 ± 8.6	74.3 ± 8.6
Sex
Female	211 (61.0)	173 (62.5)	128 (64.6)	99 (63.1)	41 (58.6)
Male	135 (39.0)	104 (37.5)	70 (35.4)	58 (36.9)	29 (41.4)
Race
Non‐Hispanic White	278 (80.3)	221 (79.8)	156 (78.8)	122 (77.7)	61 (87.2)
African American	61 (17.6)	51 (18.4)	39 (19.7)	30 (19.1)	8 (11.4)
Hispanic	5 (1.4)	4 (1.4)	2 (1.0)	5 (3.2)	1 (1.4)
Asian	2 (0.6)	1 (0.4)	1 (0.5)	0 (0.0)	0 (0.0)
APOE genotype a
ε4 non‐carrier	175 (50.6)	142 (51.3)	103 (52.0)	90 (57.3)	32 (45.7)
ε4 carrier	171 (49.4)	135 (48.7)	95 (48.0)	67 (42.7)	38 (54.3)
Diagnosis at visit
CN	210 (60.7)	162 (58.5)	133 (67.2)	113 (72.0)	39 (55.7)
MCI	84 (24.3)	64 (23.1)	28 (14.1)	22 (14.0)	20 (28.6)
ADD	52 (15.0)	51 (18.4)	37 (18.7)	22 (14.0)	11 (15.7)
Amyloid PET
Amyloid PET, N b	103	84	54	51	13
Global cortical centiloid	43.7 ± 47.9	43.9 ± 48.3	43.7 ± 48.8	19.4 ± 35.5	30.2 ± 49.9
Tau PET
Tau PET, N b	72	59	40	39	7
Metatemporal ROI SUVR	1.4 ± 0.4	1.3 ± 0.4	1.3 ± 0.4	1.2 ± 0.2	1.2 ± 0.4
Brain MRI
MRI, N b	201	165	114	97	43
Total hippocampal volume (mm3)	7294.1 ± 1025.3	7239.6 ± 1037.5	7330.5 ± 945.7	7631.5 ± 1049.4	7393.8 ± 989.4
Plasma biomarkers
Aβ42/40 ratio, N b	241	192	138	121	55
Aβ42/40 ratio	0.05 ± 0.01	0.05 ± 0.01	0.05 ± 0.01	0.06 ± 0.01	0.05 ± 0.01
p‐Tau181, N b	239	191	137	119	50
p‐Tau181 (pg/mL)	4.1 ± 3.2	4.1 ± 3.3	4.0 ± 3.7	3.3 ± 1.5	3.8 ± 5.0
NFL, N b	250	199	144	125	56
NFL (pg/mL)	21.5 ± 15.0	21.5 ± 13.6	22.2 ± 17.0	20.1 ± 13.0	22.2 ± 14.1
GFAP, N b	241	192	138	121	55
GFAP (pg/mL)	153.7 ± 98.7	158.2 ± 101.5	152.8 ± 99.9	135.8 ± 77.0	139.8 ± 83.9

Note: Unless otherwise noted, values are expressed as N (%) or mean ± standard deviation.

Abbreviations: Aβ, amyloid beta; ADD, Alzheimer's disease dementia; CN, cognitively normal individuals; MCI, mild cognitive impairment; GFAP, glial fibrillary acidic protein; MRI, magnetic resonance imaging; NFL, neurofilament light chain; PET, positron emission tomography; p‐tau181, phosphorylated tau 181; ROI, region of interest; SUVR, standardized uptake value ratio.

^a APOE genotype ε4 carriers were defined as individuals with the ε2/ε4, ε3/ε4, or ε4/ε4 genotype.

^b Number of participants with baseline data available for analysis.

The distribution of baseline cognitive stages across different cognitive stages at each follow‐up visit is provided in Table S1. The mean ± SD of baseline CDR‐SB scores were 0.129 ± 0.306 for the CN group, 1.304 ± 0.902 for the MCI group, and 6.135 ± 3.416 for the ADD group (Table S2). The corresponding annualized changes in CDR‐SB were 0.037 ± 0.083 for CN, 0.553 ± 0.662 for MCI, and 1.764 ± 0.949 for ADD.

3.2. Longitudinal plasma protein changes by baseline cognitive stage

Figure 2 presents two proteins, ACES and IGFALS, that showed significantly different changes over time across the cognitive stages (FDR‐adjusted p for time × cognitive stage interaction: ACES, p = 0.013; IGFALS, p = 0.013). The estimated slopes of longitudinal plasma protein changes by cognitive stage are presented in Table S3. ACES level showed a more rapid increase in MCI (p for time × cognitive stage interaction: p < 0.001) or ADD (p = 0.049) compared to CN, while IGFALS level exhibited a more rapid decrease in the MCI group compared to CN (p < 0.001) or ADD (p = 0.003).

FIGURE 2.

Longitudinal changes in plasma protein levels by cognitive status. (A) Among a total of 7181 aptamers, seq10980.11 (ACES) and seq6605.17 (IGFALS) showed significantly different change rates according to cognitive status. Post hoc analysis revealed that ACES showed significant differences between CN and MCI and between CN and ADD, while IGFALS showed significant differences between CN and MCI and between MCI and ADD. (B) The plots display observed data points and estimated trajectories of plasma protein levels. Estimations were conducted for each cognitive status using a linear mixed‐effects model, assuming average age, female sex, APOE ε4 non‐carrier genotype, and subarray in the proteomics assay. ADD, Alzheimer's disease dementia; CN, cognitively normal; MCI, mild cognitive impairment.

To assess the effect of acetylcholine esterase inhibitor (AChEI) use on ACES levels, we performed a linear mixed‐effects model including AChEI user status. Of the 346 baseline participants, medication information was available for 221 individuals. Among these, the number of AChEI users in the CN (N = 135), MCI (N = 51), ADD (N = 35) groups was one, 31, and 27, respectively. The association between cognitive stage and AChEI user status was statistically significant (Pearson chi‐squared test, p < 0.001). In this subgroup, the median (IQR) follow‐up period was 36.8 (25.4 to 42.4) months, and the median (IQR) number of visits was 3 (2 to 4). We then performed a linear mixed‐effects model in the MCI group (non‐AChEI users vs AChEI users), including time × AChEI user status interaction, with age, sex, and APOE ε4 carrier status as covariates. The interaction was not significant (p = 0.866), while AChEI user status showed a significant effect (p < 0.001) on ACES levels at baseline, indicating that AChEI use has a significant effect on baseline ACES levels, but not on their longitudinal change.

3.3. Longitudinal plasma protein changes in incident ADD group

Six different proteins showed significantly different changes in the incident ADD group compared to the non‐incident ADD group (Figure 3). The estimated slopes of longitudinal plasma protein changes by ADD conversion status are presented in Table S4. All of the proteins (ACES, C7, ZCD1, IL‐17C, CC055, SO5A1) showed a significantly increasing trajectory over time in the incident ADD group compared to the non‐incident ADD group (FDR‐adjusted p for time × ADD conversion status interaction: ACES, FDR‐adjusted p = 0.002; C7, FDR‐adjusted p = 0.005; ZCD1, FDR‐adjusted p = 0.007; IL‐17C, FDR‐adjusted p = 0.012; CC055, FDR‐adjusted p = 0.027; SO5A1, FDR‐adjusted p = 0.049).

FIGURE 3.

Longitudinal changes in plasma protein levels in incident ADD group compared to non‐incident ADD group. The incident ADD case was classified by the individual who changed the cognitive status from CN or MCI to ADD at any time point during follow‐up. (A) 6 out of 7181 aptamers showed significantly different change rates between the non‐incident ADD group and the incident ADD dementia group. (B) The plots display observed data points and estimated trajectories of plasma protein levels. Estimations were conducted for each cognitive status using a linear mixed‐effects model, assuming average age, female sex, APOE ε4 non‐carrier genotype, and subarray in the proteomics assay. ADD, Alzheimer's disease dementia; CN, cognitively normal; MCI, mild cognitive impairment.

In addition, the association between annualized CDR‐SB changes and longitudinal changes in protein were examined for the seven proteins (ACES, C7, ZCD1, IL‐17C, CC055, SO5A1, IGFALS) that were identified as significant in previous analyses related to baseline cognitive stage and ADD conversion status. Figure S2 presents the associations between annualized change in CDR‐SB and longitudinal changes in plasma proteins. Significant associations were observed for ACES (time × annualized change in CDR‐SB, β = 0.005, FDR‐adjusted p = 0.008), IL‐17C (β = 0.001¸ FDR‐adjusted p = 0.014), and SO5A1 (β = 0.001¸ FDR‐adjusted p = 0.008), indicating that increasing CDR‐SB was associated with more pronounced increases in these plasma proteins over time.

3.4. Association between longitudinal plasma protein and AD imaging biomarkers

For each AD biomarker, individual linear mixed‐effects models were constructed using each of the seven proteins (ACES, C7, ZCD1, IL‐17C, CC055, SO5A1, IGFALS) that were identified as significant in previous analyses related to baseline cognitive stage and ADD conversion status as the dependent variable.

Baseline Aβ PET uptake was significantly associated with changes in plasma ACES levels (time × global cortical CL, β = 0.004, FDR‐adjusted p = 0.028; Table 2, Figure 4A), indicating that higher amyloid burden was related to greater increases in ACES over time. Tau PET uptake showed a significant association with longitudinal changes in CC055 (time × MTL ROI SUVR, β = 0.002, FDR‐adjusted p = 0.047), demonstrating that individuals with higher baseline tau burden experienced more rapid increases in CC055 levels. In addition, baseline total hippocampal volume was significantly associated with longitudinal changes in IL‐17C (time × total hippocampal volume, β = −4.28E‐07, FDR‐adjusted p = 0.028), S05A1 (β = −4.40E‐07, FDR‐adjusted p = 0.044), indicating that greater hippocampal atrophy was linked to more pronounced increases in these plasma proteins over time. In comparison, IGFALS (β = 6.74E‐07, FDR‐adjusted p = 0.028) showed an opposite trend, with greater hippocampal atrophy associated with a more pronounced decrease in IGFALS levels.

TABLE 2.

Association between baseline AD imaging biomarkers and longitudinal changes in plasma protein level.

Aptamer	Protein name	A (amyloid PET)			T (tau PET)			N (MRI)
		β	p	FDR‐adjusted p	β	p	FDR‐adjusted p	β	p	FDR‐adjusted p
seq.10980.11	ACES	0.004	0.004	0.028	0.001	0.631	0.669	−1.40E–06	0.061	0.107
seq.13731.14	C7	0.001	0.021	0.073	0.001	0.039	0.091	−3.64E–07	0.138	0.187
seq.7745.3	ZCD1	0.001	0.181	0.246	0.002	0.061	0.107	−4.91E–07	0.160	0.187
seq.9255.5	IL‐17C	3.57E–04	0.176	0.246	0.001	0.024	0.083	−4.28E–07	0.004	0.028
seq.7939.1	CC055	0.001	0.285	0.285	0.002	0.007	0.047	−1.69E–07	0.510	0.510
seq.11669.39	SO5A1	4.23E–04	0.210	0.246	2.44E–04	0.622	0.669	−4.40E–07	0.019	0.044
seq.6605.17	IGFALS	−0.001	0.033	0.077	2.84E–04	0.669	0.669	6.74E–07	0.008	0.028

Note: Imaging biomarkers for A, T, and N defined as follows: global cortical Centiloid values for amyloid PET; SUVR of MTL ROI on tau PET; and total hippocampal volume on brain MRI.

Abbreviations: AD, Alzheimer's disease; FDR, False discovery rate; MRI, magnetic resonance imaging; MTL, medial temporal lobe; PET, positron emission tomography; ROI, region of interest; SUVR, standardized uptake value ratio.

FIGURE 4.

Association between baseline AD imaging and plasma biomarkers and longitudinal changes in plasma protein level. We investigated the association between AD imaging and plasma biomarkers and seven proteins identified as significant in previous analyses related to baseline diagnosis or incident ADD case. (A) Heatmap for imaging biomarkers. Imaging biomarkers for A, T, and N were defined as follows: global cortical Centiloid values for amyloid PET; SUVR of MTL ROI on tau PET; and total hippocampal volume on brain MRI. (B) Heatmap for plasma biomarkers. Color intensity represents sign (β) × log₁₀ (FDR‐adjusted p value), and asterisks (*) indicate FDR‐adjusted p < 0.05. AD, Alzheimer's disease; ADD, Alzheimer's disease dementia; FDR, false discovery rate; GFAP, glial fibrillary acidic protein; MRI, magnetic resonance imaging; MTL, medial temporal lobe; NFL, neurofilament light chain; PET, positron emission tomography; p‐tau181, phosphorylated tau 181; ROI, region of interest; SUVR, standardized uptake value ratio.

3.5. Association between longitudinal plasma protein and AD plasma biomarkers

Table 3 and Figure 4B show that NFL and GFAP were significantly associated with longitudinal changes in plasma protein levels, whereas the associations with Aβ42/40 ratio and p‐tau181 were statistically insignificant. Specifically, baseline NFL levels were significantly associated with longitudinal increases in C7 (time × NFL, β = 5.50E–05, FDR‐adjusted p = 0.002) and decreases in IGFALS (β = −4.46E‐05, FDR‐adjusted p = 0.026), suggesting that higher NFL levels predict rapid increases in C7 and rapid decreases in IGFALS over time. In addition, baseline GFAP levels were significantly associated with longitudinal changes in C7 (time × GFAP, β = 7.80E–06, FDR‐adjusted p = 0.014) and ZCD1 (β = 1.02E–05, FDR‐adjusted p = 0.014), indicating that higher GFAP levels were related to greater decreases in both proteins over time.

TABLE 3.

Association between baseline AD plasma biomarkers and longitudinal changes in plasma protein level.

Aptamer	Protein name	A (Aβ42/40 ratio)			T (p‐tau181)			N (NFL)			I (GFAP)
		β	p	FDR‐adjusted p	β	p	FDR‐adjusted p	β	p	FDR‐adjusted p	β	p	FDR‐adjusted p
seq.10980.11	ACES	0.035	0.587	0.587	9.17E–05	0.590	0.813	−1.40E–05	0.764	0.891	7.98E–06	0.294	0.411
seq.13731.14	C7	−0.029	0.181	0.423	1.22E–04	0.033	0.116	5.50E–05	3.03E–04	0.002	7.80E–06	0.002	0.014
seq.7745.3	ZCD1	−0.029	0.355	0.587	2.98E–04	0.093	0.217	3.89E–05	0.104	0.182	1.02E–05	0.004	0.014
seq.9255.5	IL–17C	−0.024	0.090	0.423	8.63E–05	0.015	0.104	1.89E–05	0.051	0.120	9.64E–08	0.952	0.952
seq.7939.1	CC055	−0.016	0.509	0.587	−5.07E–05	0.696	0.813	−9.58E–06	0.588	0.824	1.09E–06	0.688	0.802
seq.11669.39	SO5A1	0.010	0.574	0.587	−8.96E–07	0.984	0.984	1.13E–07	0.993	0.993	2.59E–06	0.212	0.370
seq.6605.17	IGFALS	0.032	0.170	0.423	−7.18E–05	0.230	0.402	−4.46E–05	0.008	0.026	−4.85E–06	0.0	0.182

Abbreviations: AD, Alzheimer's disease; FDR, False discovery rate; GFAP, glial fibrillary acidic protein; NFL, neurofilament light chain; p‐tau181, phosphorylated tau 181.

3.6. Machine learning prediction of models for incident ADD

To investigate whether changes in the identified proteins influenced the performance of machine learning prediction of incident ADD, seven different machine learning algorithms were run using the STREAMLINE platform. In one of the downsampled datasets with the largest variance, the median (IQR) follow‐up period was 35.1 (20.8 to 40.4) months, and the median (IQR) number of visits was 3 (2 to 4). Both the plasma biomarker model and the proteomics model yielded higher mean ROC AUC values than the base model across all machine learning algorithms (Table S5, Figure 5). In the machine learning analysis using the random forest algorithm, the base model yielded an AUC value of 0.613, the plasma biomarker model an AUC value of 0.820, and the proteomics model an AUC value of 0.848. Feature importance scores from the proteomics model using elastic net are presented in Figure S2.

FIGURE 5.

Prediction performance of incident ADD case using longitudinal change of significant plasma proteins. (A) Prediction performance for incident ADD using random forest algorithm. The base model included baseline age, sex, and APOE ε4 carrier status as predictors. The plasma biomarker model included the features from the base model and baseline plasma AD biomarkers (plasma p‐tau181 level and Aβ42/40 ratio). The proteomics model included the same predictors as the plasma biomarker model, with the addition of individual estimated slopes for plasma proteins obtained from a linear mixed‐effects model. All seven proteins identified as significant in prior analyses related to baseline diagnosis or when ADD conversion status was included. Values represent the average performance across five downsampled datasets, each evaluated using three‐fold cross‐validation. (B) ROC curve of base model. (C) ROC curve of plasma biomarker model. (D) ROC curve of proteomics model. Each plot corresponds to a single downsampled dataset whose performance metrics were closest to the overall mean across the five datasets. Detailed mean values and standard deviations for each performance metric are provided in Table S5. ADD, Alzheimer's disease dementia; APC, average precision score; AUC, area under the curve; NPV, negative predictive value; PPV, positive predictive value; PRC, precision‐recall curve; ROC, receiver operating characteristic.

4. DISCUSSION

In this study, we investigated the association between longitudinal changes in plasma proteins and baseline cognitive status or ADD conversion status, utilizing longitudinal proteomics data from the IADRC. Longitudinal changes in two different proteins were associated with baseline cognitive status, while changes in six proteins were associated with ADD conversion status. For each of these identified proteins, longitudinal changes were associated with a distinct AD imaging or plasma biomarker. Finally, we demonstrated that incorporating longitudinal changes in these plasma proteins improved the performance of machine learning models for predicting incident ADD. Taken together, our findings highlight the potential of longitudinal changes in plasma proteins as biomarkers for cognitive status and as predictive features for AD progression, while also providing a basis for investigating the longitudinal pathophysiological dynamics in AD.

We found that longitudinal changes in ACES (acetylcholine esterase) and IGFALS (insulin‐like growth factor [IGF]‐binding protein complex acid labile subunit) differed significantly depending on baseline cognitive status. Plasma ACES levels increased more rapidly in MCI and ADD compared to CN, with no significant difference between MCI and ADD. ACES, acetylcholine esterase, a hydrolytic enzyme of acetylcholine in synapses, is known to be increased around amyloid plaques and NFT and plays various roles in AD pathophysiology. ²⁷ , ²⁸ , ²⁹ In line with this, our results suggest that plasma ACES levels rise rapidly following the transition from CN to MCI and continue increasing at a steady rate through to ADD. Long‐term use of AChEI in AD patients may lead to upregulation of AChE as a compensatory response to chronic inhibition, as supported by several studies showing increased AChE levels in CSF following AChEI treatment. ²⁷ , ³⁰ , ³¹ Therefore, the effect of AChEI user status should be taken into account when evaluating changes in ACES levels. We addressed this effect by showing that there was no significant impact of AChEI use on ACES changes over time, based on a linear mixed‐effects model including AChEI user status in the MCI group. In contrast, IGFALS showed a significant decrease specifically in the MCI group. IGFALS forms a ternary complex with the IGF‐binding proteins (IGFBPs) and extends IGF half‐life. ³² , ³³ , ³⁴ , ³⁵ While the roles of IGFs and IGFBPs in AD are well established, studies on the involvement of IGFALS are lacking. It may influence AD risk by modulating IGF stability, as IGFs are neuroprotective and their reduction has been linked to increased AD risk. ³⁶ , ³⁷ , ³⁸ , ³⁹ Interestingly, elevated IGFBP levels, despite their IGF stabilizing function, have also been associated with higher AD risk, ³⁶ , ⁴⁰ , ⁴¹ often interpreted in the context of altered IGF bioavailability and insulin resistance. ³⁶ , ⁴⁰ Our study adds new insight on this complex interplay, that is, that IGFALS, a previously unrecognized factor, may play an additional role. This finding underscores the importance of longitudinal proteomics study in clinical cohorts to reveal novel aspects of AD pathophysiology.

Our second major finding was that plasma levels of ACES, C7 (complement component C7), ZCD1 (CDGSH iron‐sulfur domain‐containing protein 2), IL‐17C (interleukin‐17C), CC055 (putative uncharacterized protein PQLC2L), and SO5A1 (solute carrier organic anion transporter family member 5A1) increased over time in the incident ADD group compared to the non‐incident ADD group. Given that ACES is elevated around AD pathological structures and acetylcholine plays a critical role in cognition, ²⁷ , ⁴² , ⁴³ our finding supports this biological link. In addition, changes in immune‐related proteins, C7 and IL‐17C, appear to be associated with incident ADD. The C7 gene, a known AD risk gene, ⁴⁴ , ⁴⁵ encodes a complement component of the membrane attack complex, and its protein product has been implicated in AD pathogenesis. ⁴⁶ Interleukin‐17 family cytokines produced by Th17 cells mediate inflammatory responses and are associated with several autoimmune diseases and AD. ⁴⁷ Previous cross‐sectional studies identified immune‐related proteins such as IL‐6 and various chemokine ligands; ¹⁵ , ¹⁶ however, the proteins identified in our study differ, underscoring the potential of longitudinal proteomics to reveal novel AD targets, although these differences should be interpreted with caution. Although ZCD1, also known as CISD2, overexpression was shown to attenuate neuronal loss in the hippocampus of AD model mice, ⁴⁸ , ⁴⁹ our study found that its plasma levels increased over time in the incident ADD group. Despite this discrepancy, the consistent involvement of ZCD1 across species suggests it may still hold potential as an AD biomarker. The remaining two proteins, CC055 and SO5A1, are membrane transporters. CC055, also referred to as the putative uncharacterized protein PQLC2L, is a lysosomal cationic amino acid exporter. ⁵⁰ , ⁵¹ It recruits C9orf72 proteins, linked to amylotrophic lateral sclerosis and frontotemporal dementia, to the surface of lysosomes, ⁵⁰ , ⁵¹ suggesting relevance to another neurodegenerative disease, AD. SO5A1, a sodium‐independent organic anion transmembrane transporter, belongs to the solute carrier (SLC) family. Some studies suggest that SLC transporters are associated with AD in mouse models, ⁵² and genes from the SLC and solute carrier organicanion transporter (SLCO) families have been implicated in AD. ⁵³ , ⁵⁴ Beyond their association with incident ADD based on clinical diagnosis, the longitudinal changes in ACES, IL‐17C, and SO5A1 were also significantly associated with the annualized change in CDR‐SB. Specifically, participants with faster cognitive decline, as indicated by a steeper increase in CDR‐SB scores, showed more pronounced increases in these proteins over time. These findings are consistent with the associations observed with ADD conversion status. Taken together, our findings highlight dynamic changes in understudied proteins in AD and support their potential as novel biomarkers beyond the existing A/T/N framework.

We further investigated whether longitudinal changes in the seven proteins associated with baseline cognitive stage and its progression were also related to baseline AD imaging or plasma biomarkers. ACES was associated with Aβ PET uptake, while CC055 was associated with tau PET uptake. However, plasma biomarkers such as the Aβ42/40 ratio and p‐tau181 showed no significant associations. This may reflect that Aβ PET captures amyloid plaques, while the plasma Aβ42/40 ratio represents more soluble forms of amyloid, indicating they reflect different biological aspects of amyloid pathology. ⁵⁵ , ⁵⁶ Similarly, tau PET targets NFT, whereas p‐tau181 is not NFT‐specific and reflects both amyloid and tau burden. ⁵⁷ These PET‐plasma biomarker differences may explain the distinct protein associations. Total hippocampal volume was associated with IL‐17C, SO5A1, and IGFALS, suggesting neurodegeneration involves complex biochemical mechanisms. A previous study showed higher IGF‐1 related to larger hippocampal volume. ⁵⁸ Similar to total hippocampal volume, NFL, a neurodegeneration marker, was found to be associated with changes in IGFALS in our study, and this may be related to prior findings in progressive supranuclear palsy, where elevated NFL was associated with reduced IGF‐1. ⁵⁹ Our findings, which show that NFL and GFAP were both associated with C7, supports the complement system's role in neurodegeneration and neuroinflammation. ⁴⁶ GFAP was also related to changes in ZCD1, which align with previous findings showing that its overexpression is associated with reduced neuronal loss, less Aβ‐induced mitochondrial dysfunction, and lower GFAP levels in the hippocampus of AD mice. ⁴⁹ However, our results are based on comparisons between baseline biomarker status and longitudinal proteomic data. Therefore, a future study investigating the interaction between longitudinal dynamics of AD biomarkers and proteomic changes is required.

Our final major finding is that incorporating the longitudinal changes in these seven proteins improved the performance of machine learning‐based prediction of incident ADD compared to the base and the plasma biomarker models. Although the estimated slopes of individual protein changes were numerically small, reflecting subtle biological shifts over time, the combined use of these dynamic features noticeably enhanced predictive performance. Furthermore, the improved performance of the proteomics model compared to the plasma biomarker model suggests that incorporating markers beyond amyloid and tau contributes to more accurate prediction of AD progression. This suggests that even minimal longitudinal alterations, when integrated across multiple biological functions, may capture critical aspects of disease progression.

Our study is the first to investigate associations between AD phenotypes and longitudinal protein changes using longitudinal plasma proteomics data. Despite this strength, several limitations exist. First, the sample size was relatively modest, potentially limiting the generalizability of our findings. It will be important to assess whether the proteins identified as significant in this cohort can be replicated in independent, larger studies. Second, we modeled linear protein changes due to variability in follow‐up intervals and limited repeated measurements per participant, which constrain the use of non‐linear or spline‐based models. With additional data, exploring various temporal patterns may provide deeper insights into AD pathophysiology. Third, cognitive stage classification in this study was based on clinical diagnosis without consideration of AD biomarker status. As a result, groups defined by baseline cognitive stage or ADD conversion status may be heterogeneous in biomarker‐based frameworks. To address this limitation, we performed an association analysis between baseline AD biomarkers and longitudinal changes in proteins. For example, as our MCI group might include non‐AD MCI cases, the significant decline observed in IGFALS within the MCI group may reflect shared dynamics between AD and non‐AD MCI, rather than AD‐specific pathology. IGFALS was not associated with baseline amyloid or tau biomarkers but was significantly associated with baseline total hippocampal volume and NFL. Fourth, in the machine learning prediction, protein measurements and ADD conversion status were assessed during overlapping follow‐up periods. As a result, our proteomics model may be classifying current disease states rather than predicting future conversion. Consequently, the model has limitations for direct application in practical settings. To determine whether changes in protein levels can predict conversion to ADD beyond the follow‐up period, further validation using datasets with longer‐term cognitive follow‐up after protein measurement is needed.

In conclusion, this study demonstrates the potential of longitudinal plasma protein changes as meaningful indicators of cognitive status and predictors of AD progression. Building upon longitudinal proteomics data, we identified specific proteins whose dynamic changes reflected distinct stages of AD and biological processes, including neuroinflammation, amyloid and tau pathology, and neurodegeneration. Notably, even subtle longitudinal shifts in protein levels, particularly when integrated across multiple proteins, significantly enhanced the predictive accuracy of machine learning models for incident ADD. These findings suggest that dynamic molecular changes in the periphery may capture early and individualized disease trajectories more effectively than cross‐sectional or static biomarkers alone. As longitudinal proteomic technologies become more accessible and data availability grows, the incorporation of multiprotein plasma signatures into clinical workflows holds promise for advancing precision diagnostics and personalized prognostication in AD.

AUTHOR CONTRIBUTIONS

Eun Hye Lee and Kwangsik Nho had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design: Eun Hye Lee, Kwangsik Nho. Acquisition of data: Jeffrey L. Dage, Jared R. Brosch, Sujuan Gao, Liana G. Apostolova, Donna M. Wilcock, Shannon L. Risacher. Statistical analysis: Eun Hye Lee, Yen‐Ning Huang. Interpretation of data: Eun Hye Lee, Yen‐Ning Huang, Tamina Park, Shiwei Liu, Nicholas Adzibolosu, Soumilee Chaudhuri, Taeho Jo, Kwangsik Nho. Drafting of the manuscript: Eun Hye Lee, Yen‐Ning Huang, Tamina Park, Soumilee Chaudhuri. Critical revision of the manuscript for important intellectual content: Eun Hye Lee, Yen‐Ning Huang, Tamina Park, Shiwei Liu, Nicholas Adzibolosu, Soumilee Chaudhuri, Paula J. Bice, Jeffrey L. Dage, Jared R. Brosch, Sujuan Gao, Liana G. Apostolova, Donna M. Wilcock, Shannon L. Risacher, Andrew J. Saykin, Taeho Jo, Kwangsik Nho. Supervision: Taeho Jo, Andrew J. Saykin, Kwangsik Nho.

CONFLICT OF INTEREST STATEMENT

A.J.S. has received support from Avid Radiopharmaceuticals, a subsidiary of Eli Lilly (in‐kind contribution of PET tracer precursor) and participated on scientific advisory boards (Bayer Oncology, Eisai, Novo Nordisk, and Siemens Medical Solutions USA, Inc.) and an observational study monitoring board (MESA, NIH NHLBI), as well as several other NIA external advisory committees. He also serves as editor‐in‐chief of Brain Imaging and Behavior, a Springer‐Nature Journal. J.L.D. is an inventor on patents or patent applications assigned to Eli Lilly and Company relating to the assays, methods, reagents, and/or compositions of matter for p‐tau assays and Aβ‐targeting therapeutics. J.L.D. has/is served/serving as a consultant or on advisory boards for Eisai, AbbVie, Genotix Biotechnologies Inc., Gates Ventures, Gate Neurosciences, Dolby Family Ventures, Karuna Therapeutics, Alzheimer's Disease Drug Discovery Foundation, AlzPath Inc., Cognito Therapeutics, Inc., Eli Lilly and Company, Prevail Therapeutics, Neurogen Biomarking, Spear Bio, Rush University, University of Kentucky, Tymora Analytical Operations, MindImmune Therapeutics, Inc., Early is Good, and Quanterix. J.L.D. has received research support from ADx Neurosciences, Fujirebio, Roche Diagnostics, and Eli Lilly in the past 2 years. J.L.D. has received speaker fees from Eli Lilly and Company and LabCorp. J.L.D. is a founder and advisor for Monument Biosciences and Dage Scientific LLC. J.L.D. has stock or stock options in Eli Lilly, Genotix Biotechnologies, MindImmune Therapeutics Inc., AlzPath Inc., Neurogen Biomarking, and Monument Biosciences. Other authors have no conflicts of interest to disclose. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

The relevant Institutional Review Board approved this study. All participants provided informed consent to participate in the study in accordance with the Declaration of Helsinki.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Lee K, Huang YN, Park T, et al. Longitudinal plasma proteomics: relation to incident Alzheimer's disease dementia and biomarkers. Alzheimer's Dement. 2025;21:e70900. 10.1002/alz.70900

DATA AVAILABILITY STATEMENT

The IADRC dataset used and analyzed in this study will be provided by the corresponding author upon approval request.