Associations of potential ADRD plasma biomarkers in cognitively normal volunteers

1 -. Introduction

Dementia is characterized by cognitive impairment that affects memory and other cognitive functions (e.g., the ability to reason, plan, or effectively communicate)¹ and is most prevalent after age 65². By 2030, approximately 21% of the US population³ will reach age 65 years⁴. By 2020, more than five million Americans were living with Alzheimer’s disease (AD), and one in three deaths among adults aged 65 and older were associated with dementia⁵.

Decades of research have characterized the pathophysiology and natural history of diseases that cause dementia, with a focus on AD⁶. However, both dementia and AD remain incurable and without proven prevention measures⁷. Moreover, diseases that cause dementia are complicated⁸, and the gold-standard diagnosis for AD requires brain autopsy. Clinical diagnosis of AD is based on symptoms, patient history, and possibly cerebrospinal fluid (CSF)⁹ or neuroimaging biomarkers¹⁰. Neuroimaging can be expensive, and attaining CSF is invasive¹¹. These, among other reasons, reduce feasibility of CSF and neuroimaging biomarkers for widespread use¹². Given such limitations, there is a need for valid, accurate, and easily measurable biomarkers of dementia-causing diseases.

Blood-based biomarkers for AD and related dementias (ADRD) have not yet been optimized. However, there is hope that blood-based biomarkers can help identify disease states and predict cognitive trajectories, while mitigating difficulties associated with more invasive and expensive testing^11,12. Quanterix Single Molecule Array (SiMOA) technology, more sensitive than previous blood assays, allows measurement of AD-relevant plasma biomarker concentrations, which are much lower than concentrations in CSF¹³.

The current study investigated the relationships between three types of plasma biomarkers (inflammatory, vascular, and neurodegenerative; defined below) and participant demographics, health related factors, and technological factors. Our objective was to better understand which, if any, features outside of neurodegenerative or cerebrovascular disease may influence biomarker values¹⁴. In participants with initially normal cognition, cross-sectional associations were estimated at two time points, five years apart, to assess robustness of these associations. We also quantified the proportion of biomarker variance explained by covariates.

2 -. Methods

2.1 -. Setting

Data for the current study were drawn from the community-based longitudinal cohort of brain aging and cognition at the University of Kentucky Alzheimer’s Disease Research Center (UKADRC). Since cohort recruitment began in 1989, over 1000 participants have agreed to be followed approximately annually until death, most consenting to brain donation¹⁵. In 2019, the UKADRC’s Biomarker Core began using SiMOA to measure biomarkers in plasma donated by UKADRC participants at annual visits¹⁵. Sampling banked plasma for analysis began with participants with two visits five years apart, initially with visits in 2012 and in 2017, as plasma collection at UKADRC was transitioned to ethylenediaminetetraacetic acid (EDTA) from heparinized vacutainer tubes in 2012. The UK Institutional Review Board (IRB) approved all study procedures, and all participants provided written informed consent.

2.2 -. Study Design

We conducted retrospective cross-sectional studies on a subset of the UKADRC cohort with available biomarker and clinical data¹⁵. Inclusion criteria for the current study were available blood-based biomarker data and normal cognition. The primary analysis focused on two pairs of visits: 2012 and 2017 (12/17), and 2013 and 2018 (13/18). If participants were included in both pairs, the 12/17 data were used. Following suggestions from peer reviewers, we also performed a second sensitivity analysis for batch effects and robustness of results. Here, we included new data, which were collected from UKADRC particpants in 2014 and 2019. Combined with the data from 2013, these additional data allowed us to evaluate batch effects on data from the same machine. If individuals were present in both 13/18 and 14/19, the 13/18 data were used. The details for all analyses are below.

2.3 -. Plasma Biomarkers

Commercial Quanterix SiMOA kits were used for all biomarker measurements, and the kits used were consistent across the study years. Multiplex kits were used for amyloid-beta_1–40 (Aβ40), amyloid-beta_1–42 (Aβ42), and total tau, while uniplex kits were used for neurofilament light chain (NfLight), tumor-necrosis factor-alpha (TNFα), interleukin 6 (IL6), interleukin 8 (IL8), interleukin 10 (IL10), interleukin 1Beta (IL1B), matrix metallopeptidase 9 (MMP9), and placental growth factor (PlGF). Phosphorylated tau (p-tau) was not included in this analysis because the assay used in 12/17 (p-tau 231) demonstrated poor reliability in validation testing (data not shown), and a different assay was used in 13/18 and 14/19 (p-tau 181).

Biomarker data were collected at a single facility using standard NIA/NACC biospecimen best practice protocols, and assays were run in the single UKADRC biomarker core biosample laboratory [Table 1]. Between processing of the 12/17 and 13/18 samples, the machine used to run the assays was updated from the Quanterix HD1 Analyzer to the HDX Analyzer, which incorporated sophisticated control systems to enhance reproducibility and included essential temperature control that the HD1 lacked¹⁶. It was unclear a priori whether the data produced on the two instruments are interchangeable, though the same kits were used on both machines. Quanterix no longer sells or supports the HD1 Analyzer¹⁷, and we are unaware of any publications comparing the performance of the HD1 and HDX machines. In addition to the covariates described below, all analyses in the primary study included a batch indicator (12/17 vs 13/18). To assess the potential for batch effects when using HDX only (rather than changing machines), we also included a batch indicator in the post hoc sensitivity analysis (13/18 vs 14/19).

Table 1:

Distribution of plasma biomarkers at baseline and five years later among cognitively normal older adult research volunteers

Biomarker Distribution Statistics
	Baseline			5 years later
Biomarker	Mean (SD)	Median (IQR)	Effective N	Mean (SD)	Median (IQR)	Effective N
Aβ40	193.59 (104.63)	172 (122.88)	267	237.29 (122.7)	215.12 (130)	266
Aβ42	11.39 (6.97)	10.1 (9.3)	273	14.07 (8.15)	13.25 (10.46)	275
Aβ42/40	0.07 (0.06)	0.05 (0.03)	264	0.06 (0.05)	0.05 (0.03)	264
Tau	6.95 (6.75)	4.94 (3.97)	276	8.46 (47.34)	4.22 (3.38)	277
Tau/Aβ42	0.88 (1.25)	0.52 (0.5)	273	0.64 (3.15)	0.34 (0.22)	275
NfLight	20.61 (23.47)	16.48 (12.89)	276	25.4 (14.49)	21.55 (17.77)	275
PlGF	27.02 (70.19)	4.03 (5.58)	267	26.15 (67.58)	4.29 (5.55)	273
MMP9*	50.72 (66.81)	27.45 (40.45)	273	78.80 (12.59)	38.90 (64.15)	259
IL6	1.61 (4.36)	0.8 (0.91)	275	2.46 (8.14)	0.99 (1.27)	274
IL8	0.41 (1.34)	0.19 (0.33)	250	0.29 (1.3)	0.14 (0.18)	242
IL10	0.7 (1.42)	0.5 (0.35)	274	0.85 (1.69)	0.57 (0.46)	277
IL1b	0.44 (2.63)	0.07 (0.09)	244	0.33 (1.54)	0.02 (0.04)	245
TNFα	1.89 (2.7)	1.21 (0.82)	271	2.69 (9.26)	1.32 (1.07)	266

NOTE: SD: Standard Deviation; IQR: Interquartile Range; Effective N: number of observations where biomarker is not missing out of possible 277 total individuals (2012/13 data). Aβ: Amyloid Beta; NfLight: Neurofilament Light Chain; PlGF: Placental Growth Factor; MMP9: Matrix Metallopeptidase 9; IL: Interleukin; TNFα: Tumor Necrosis Factor alpha.

^*MMP9 values are presented in thousands.

2.4 -. Covariate Selection

Because our interest in these biomarkers relates to their potential association with ADRD and vascular cognitive impairment and dementia (VCID), we identified covariates that would mitigate confounding, defined as distortion in the association between biomarkers and cognition arising from shared causes. To facilitate covariate selection, we classified biomarkers into clinically relevant subgroups: (1) nonspecific neurodegenerative and AD markers (NfLight, tau, Aβ40, Aβ42, and the ratios Aβ42/Aβ40 and tau/Aβ42); (2) vascular markers (PlGF and MMP9); and (3) inflammatory markers (i.e., cytokines: TNFα, IL6, IL8, IL10, and IL1β). We created directed acyclic graphs (DAGs) to encode our theoretical model¹⁸ [Supplementary Figures 1–3]. We set the biomarker types (vascular, inflammatory, and neurodegenerative/AD) as the exposure, assuming that biomarkers of the same type would have similar causes and effects. Covariate selection was guided via sufficient adjustment sets in each DAG [Table 2], which theoretically eliminate confounding and bias. Following suggestions by peer reviewers, we also performed post hoc analyses including all potential confounders as covariates.

Table 2:

Selected covariates by biomarker group

Covariates	Neurodegenerative	Vascular	Inflammatory
Age	X	X	X
APOE	X
B12			X
Body mass index			X
Cancer			X
Cerebrovascular disease		X	X
COPD			X
Cardiovascular disease			X
Depression			X
Diabetes		X	X
Hypercholesterolemia		X	X
Hypertension	X	X	X
Gender	X	X	X
Smoker		X	X

NOTE: APOE: Apolipoprotein E; COPD: Chronic Obstructive Pulmonary Disease.

Participant age (in years), gender, BMI (calculated via measured height and weight), lifetime smoking status (ever vs. never), and Apolipoprotein E (APOE) were included as covariates. APOE, the strongest genetic risk factor for late onset AD¹⁹, is included as no e4 alleles vs any e4 alleles, as our sample size did not warrant finer categorization.

Self-reported medical conditions (coded ever vs. never, unless otherwise specified) were cancer, cardiovascular conditions (CV; any vs none), cerebrovascular conditions (CB; any vs none), COPD, depression, diabetes, hypercholesterolemia, hypertension, and vitamin B12 deficiency. CV was operationalized as a single variable to indicate whether a participant had at least one cardiovascular condition, based on self-reported atrial fibrillation, angina, angioplasty, coronary bypass, congestive heart failure, or heart attack. CB was similarly operationalized based on self-reported ischemic stroke and transient ischemic attack. Medical conditions were updated at annual visits.

2.5 -. Cognition

Participants undergo cognitive testing at each study visit. These data are used, with results of clinical examinations, to ascertain syndromic cognitive diagnosis: normal cognition, MCI, or dementia²⁰. Explicit guidelines for clinical diagnosis were followed to reduce biases from subjective clinical diagnoses¹⁵.

2.6 -. Statistical Analysis

Biomarkers were investigated individually to estimate their associations with covariates. Adjusted analyses were implemented as 13 linear regression models, with individual biomarkers or their ratios as dependent variables. Biomarker values were log-transformed^21,22 to improve plausibility of linear model assumptions²³. Log-transformations also present an advantage for analyzing the biomarker ratios, in that log-transforming makes the results invariant to choice of numerator and denominator²⁴. Although some values appeared as possible outliers [Supplementary Figure 4], they were checked and confirmed.

In the 13 linear models, independent variables were based on DAG sufficient adjustment sets, along with the batch indicator. Missingness in the data was minimal; there were at most 5 missing observations per covariate, and at most 13% missing observations for any biomarker, though most biomarkers had fewer than 5% missing values. Thus we used only complete cases. Goodness-of-fit was assessed via Normal Q-Q plots and plots of residuals vs fitted values. R-squared was used to estimate the proportion of variance explained by covariates.

As described above, the first set of analyses focused on baseline levels of biomarkers (i.e., first year in the pair 12/17 or 13/18). After analyzing these data, we performed our pre-planned sensitivity analysis using information from the second visit (5 years later) among participants who remained cognitively normal. The same sufficient adjustment covariate sets were used.

In the post hoc sensitivity analyses, all biomarker models were refit: (a) to the main data (12/17 and 13/18) with all potential confounders listed in Table 2 (rather than just the DAG-identified sufficient adjustment sets); (b) to the combined 2013/2014 data with DAG-derived variables; (c) and, to the 2013/14 data with all potential confounders.

A final post hoc analysis arose upon reviewing results of the 13 models from the main analysis of the primary study. The model with the highest R-squared involved a ratio of two biomarkers (Aβ42/Aβ40). These two biomarkers had a pairwise correlation of 0.629 [Supplementary Table 1], which was much stronger than the correlations between all other pairs of biomarkers, except for TNFα and PlGF, which had a correlation of 0.694. We pursued a post hoc analysis using log(TNFα/PlGF) as an outcome to see if a linear model for this ratio would produce a similarly large R-squared. We used all variables from each relevant sufficient adjustment set (inflammatory and vascular) as covariates in this post hoc analysis.

We used a 5% significance level when interpreting results. Analyses were performed with R version 3.6.2 in RStudio, using packages readxl, haven, tidyverse, plyr, ggplot2, gridExtra, and r2symbols^25–32.

3 -. Results

A total of 237 UKADRC participants with study visits in 2012/17 and 2013/18 met inclusion criteria. The average baseline age was 82.6 years, 63% were female, and 19% had at least one APOE e4 allele [Table 3]. Characteristics of the 239 participants included in the post hoc sensitivity analyses were comparable to the original study cohort [Table 3]. Overall, the selected demographic and clinical features did not explain much variance in biomarker values. The mean and median R-squared values were 0.117 and 0.088 [Table 4]. The highest R-squared was 0.363 for log(Aβ42/Aβ40). In sections 3.1 – 3.3, we describe the results of the primary study, and in sections 3.4 and 3.5 we describe the results of the post hoc analyses.

Table 3:

Included University of Kentucky Alzheimer’s Disease Research Center baseline participant characteristics for new and old data sets.

Variable	12/13 Data	13/14 Data
Age (mean ± sd)	82.57 ± 7.08	76.46 ± 6.10
BMI (mean ± sd)	26.31 ± 4.73	25.93 ± 4.42
Batch (2013/2018)	51 (22)	172 (72)
Gender (F)	149 (63)	156 (65)
APOE (any e4 allele)	68 (29)	74 (31)
Cerebrovascular disease	13 (6)	15 (6)
Cardiovascular disease	49 (21)	65 (27)
Hypertension	145 (61)	137 (58)
Diabetes	32 (14)	30 (13)
Hypercholesterol	150 (63)	149 (63)
Smoker	110 (46)	118 (49)
B12 Deficiency	20 (8)	16 (7)
COPD	18 (8)	12 (5)
Cancer	55 (23)	51 (21)
Depression	41 (18)	52 (22)
Total N	237	239

NOTE: Unless otherwise stated in Variable column, statistics reported are number (%) and variables are coded as 0=never having condition and 1=ever having condition, unless otherwise stated. BMI: Body mass index; APOE: Apolipoprotein E; COPD: Chronic obstructive pulmonary disease, N: sample size.

Table 4:

Main Analysis Model Results by Biomarker Group

Neurodegenerative/AD Biomarker Model Results
Outcome:	log(AB40)			log(AB42)			log(AB42/AB40)			log(Tau)			log(Tau/AB42)			log(NfLight)
Predictor	Est.	SE	P-val	Est.	SE	P-val	Est.	SE	P-val	Est.	SE	P-val	Est.	SE	P-val	Est.	SE	P-val
Age	0.013	0.006	0.033	−0.006	0.008	0.424	−0.019	0.006	0.001	−0.017	0.008	0.030	−0.011	0.009	0.221	0.039	0.006	<0.001
APOE	0.042	0.084	0.615	−0.141	0.105	0.181	−0.194	0.079	0.015	−0.073	0.103	0.478	0.056	0.119	0.639	0.091	0.085	0.282
Batch	−0.079	0.103	0.447	0.534	0.132	<0.001	0.671	0.097	<0.001	0.043	0.129	0.740	−0.509	0.149	0.001	0.662	0.106	<0.001
Gender	0.051	0.077	0.510	0.058	0.097	0.552	0.018	0.072	0.800	−0.050	0.095	0.597	−0.107	0.109	0.331	0.129	0.078	0.099
Hypertension	0.191	0.076	0.013	0.076	0.096	0.429	−0.136	0.072	0.059	0.202	0.095	0.034	0.113	0.109	0.300	0.057	0.077	0.464
R-squared	0.075			0.117			0.363			0.046			0.063			0.189
Effective N	227			233			225			235			233			235

NOTE: Model results for Neuro/AD biomarkers. Est: coefficient estimate; SE: Standard Error; Aβ: Amyloid Beta; NfLight: Neurofilament Light Chain; APOE: Apolipoprotein E.

3.1 -. Demographics and Genetics

Of the 13 biomarkers, only Aβ40, NfLight, MMP9, and IL10 had higher concentrations significantly associated with increasing age. For a 5-year increase in age, the models predicted 6.2% (95% CI: 0.5, 13.6), 21.5% (95% CI: 14.3, 29.4), 16.2% (95% CI: 3.0, 31.4), and 11.1% (95% CI: 3.2, 20.1) increases in these biomarkers, respectively. Increasing age was also significantly associated with lower concentrations of Aβ42/40 and tau. For a 5-year increase in age, the models predicted 9.1% (95% CI: 3.5, 13.9) and 8.1% (95% CI: 0.8, 14.8) decreases in Aβ42/40 and tau, respectively.

Gender was not significantly associated with any biomarker, and APOE, included in the six neurodegenerative biomarker models, was only significantly associated with log(Aβ42/Aβ40). This biomarker ratio was predicted to be 18% lower among participants who had at least one e4 allele (95% CI: 4, 30).

3.2 -. Medical Conditions

Hypertension was significantly positively associated with Aβ40 and tau. Cancer history, included only in the five inflammatory biomarker models, was significantly positively associated with IL6. No other medical conditions had significant associations with biomarkers.

3.3 –. Possible machine effect (HD1 vs HDX)

Data produced by the HD-1 and HD-X machines were assessed in multiple ways. The coefficients of variation (CV) for each biomarker on each machine were calculated [Supplementary Table 2]. Except for IL1β, the CVs from the HD-X were below 12% (thus, also below the commonly used 20% threshold³³). The HD-X average CVs with and without IL1β were 12.25% and 8.49%, respectively. The HD-1 CVs were below 19%, except for IL1β, with average CVs of 15.88% and 12.69%, respectively.

Eight of 13 models produced a significant batch coefficient. Effect size, defined as the estimated absolute number of standard deviations by which the mean log biomarker concentration changed when the batch changed (i.e., absolute value of batch coefficient, divided by outcome marginal standard deviation), was calculated. Effect sizes ranged from 0.06 to 1.17, averaging 0.56 [Supplementary Table 3]. Ten of 13 effect sizes exceeded one third, and three exceeded one, meaning that in the latter cases predicted biomarker measurements from the HDX were more than one standard deviation different from those of the HD1.

3.4 –. Pre-planned Sensitivity Analysis

After 5 years, most participants remained cognitively normal, while ~20% transitioned to MCI or dementia [Supplementary Table 4]. We repeated analyses on the 190 individuals remaining cognitively normal after 5 years [Supplementary Table 5]. APOE was again significantly associated with one outcome, though here it was tau/Aβ42. Six models produced a significant batch effect (possibly machine related). Otherwise, we saw no discernable patterns.

3.5 -. Post Hoc Analyses

Direction, magnitude, and significance of beta coefficients varied across models and data [Figure 1, Supplementary Figure 5]. Throughout the original and post hoc sensitivity analyses, Neuro/AD and Inflammatory biomarkers tended to acquire more significant coefficients than Vascular biomarkers. Within datasets, coefficient estimates corresponding to the same biomarker and the same covariate tended to be similar, regardless of whether the covariates included just the DAG-identified variables (the reduced set) or all potential confounders [Supplementary Tables 6–8]. The differences across analyses (mentioned below) may reflect differing quality of data based on biomarker collection and highlight the importance of considering issues of data acquisition and reproducibility.

Figure 1:

Forest Plots for Main and Post Hoc Senstivitity Analyses

NOTE: Forest plots comparing estimated coefficient confidence intervals for models with Aβ42/40, MMP9, and TNFα as outcomes. Est: coefficient estimate; SE: Standard Error; Aβ: Amyloid Beta; NfLight: Neurofilament Light Chain; APOE; Apolipoprotein E.; BMI: Body mass index; CB: Cerebrovascular conditions; CV: Cardiovascular conditions; COPD: Chronic obstructive pulmonary disease.

Age, which was significant in six of 13 original models (Aβ40, Aβ42/40, Tau, NfLight, MMP9, and IL10), is significant in four when including all variables using 12/13 data (Aβ42/40, NfLight, MMP9, and IL10), and two when using 13/14 data and either variable set (NfLight and IL10). Batch, which was significant for nine of 13 original models (Aβ42, Aβ42/40, Tau/Aβ42, NfLight, PlGF, MMP9, TNFα, IL8, and IL1β), is significant in eight when using all variables with 12/13 data (Aβ42, Aβ42/40, Tau/aβ42, NfLight, PlGF, MMP9, IL6, and IL8), and nine when using 13/14 data and either variable set (Aβ42, Aβ42/40, Tau, Tau/Aβ42, NfLight, MMP9, TNFα, IL8, and IL10). APOE was significant for log(Aβ42/40) when using 12/13 data and either variable set. BMI which was not significant in any original models but has elsewhere been significantly associated with numerous biomarkers in serum and plasma^34–36, was significant for NfLight when using 12/13 data and all variables, IL6 and IL10 when using 13/14 data and the reduced variable set, and for Tau, Tau/Aβ42, NfLight, IL6, and IL10 when using 13/14 data and all variables.. Depression was significant for IL8 and IL10 when using 13/14 data and either variable set.

In the final post hoc analysis, the TNFα/PlGF ratio produced an R-squared of 0.098, similar to the average R-squared for the original 13 models [Supplementary Table 9]. The only variable with a significant coefficient was batch (effect size 0.68 [Supplementary Table 3]).

4 -. Discussion

We evaluated participant and technical characteristics as predictors of concentrations of plasma-based biomarkers in cognitively normal research volunteers. Biomarker concentrations were not strongly associated with age, gender, nor medical conditions, suggesting that changes in biomarkers, when observed, may be neuropathological. This study adds to the literature on plasma biomarkers in cognitively normal individuals given that, of existing studies reporting on these SiMOA-measured biomarkers, all but one focused on cognitively impaired participants³⁷.

Arguably, our most robust results were for Aβ42/Aβ40, which was associated with age, APOE, and batch in the main analysis; age, APOE, batch and Hypertension in the post hoc sensitivity analysis using 12/13 data and all variables; and batch in both post hoc sensitivity analyses using 13/14 data. In all 5 analyses (including primary and both post hoc analyses) this outcome had either the first or second largest R-squared value, making it the overall most predictable outcome.

Relationships with age were not consistent across biomarkers. Other studies examining blood biomarkers using Quanterix SiMOA technology report associations of age with total tau^37–40, Aβ42/40³⁷, NfLight^37,38, Aβ42^37,39–41, Aβ40^40,41, and TNFα³⁹, though direction of associations was not always reported. While not consistent in direction or magnitude of association with biomarkers in the present study, age may influence temporal changes in biomarkers. Future research will examine temporal changes in biomarkers and cognition, in tandem with age and other covariates.

Notably, reported relationships for plasma biomarkers are not consistent across the literature, suggesting that sample variability and analysis methods may influence results. While relationships of biomarkers with age, gender, APOE, race, education, and BMI have been shown, the majority of these studies reported on subsets of these factors or as supporting information, secondary to a main analysis^37–42, or such information was not reported at all^13,43,44.

While the use of the two different Quanterix machines may be considered a limitation in this study, these data call attention to possible impacts of technologic upgrades within a single platform. We note that most SiMOA studies in the comparable literature used the HD1 analyzer^13,37–44. To address this potential limitation, we assessed reproducibility of results using data from two batches (2013 and 2014) from the same machine (HDX). Many of the batch coefficients were still significant, suggesting that the batch coefficients in the main analysis were perhaps not driven entirely by the different machines. The significance of batch even after the standardization of machine highlights the importance of including such a coefficient to account for temporal variability in data collection in longitudinal studies.

We did not see consistency in significant relationships from our main analysis to the post hoc sensitivity analysis [Figure 1, Supplemental Figure 5]. While direction and significance of coefficients was typically consistent within data sets, some variables, such as B12 deficiency and cerebrovascular disease (CB), produced estimates differing in direction based on dataset, leaving uncertainty in these associations. Notably, B12 deficiency, CB, COPD, diabetes, hypercholesterolemia, and gender did not produce any significant coefficients across all 4 analyses. We note, however, that other SiMOA studies have found diabetes to be associated with Neuro/AD biomarkers^36,45.

Our sample size is comparable to those reported in similar studies. Studies using larger, more diverse samples would have greater statistical power to confirm associations found here or to detect other associations, such as relationships involving p-tau or race, which were not examined here. However, the availability of a longitudinal, well-characterized cohort with uniform biospecimen collection is a strength of the present study, as are the sensitivity and post hoc analyses which assessed reproducibility and provided insight about batch effects.

There are some limitations to the study. We did not assess racial differences due to the limited representation in the UKADRC cohort, but more diverse and/or larger data may provide insights, as⁴² and⁴⁵ have done with Mexican Americans and Non-Hispanic Whites. Extant research on other fluid biomarkers, like CSF, has reported ethnoracial differences and additional research is needed⁴⁶. We also lacked data on potentially relevant health conditions, like chronic kidney disease or renal function, but other investigators have found them to be associated with multiple biomarkers. Chronic kidney disease was significantly related to Aβ40, Aβ42, Aβ42/40, total tau, and NfLight^36,45, while renal function was significantly associated with serum NfLight³⁴. We also did not have information on fasting status for the blood collection.

Having all participants consent to autopsy introduces selection bias, as these individuals were more likely to be highly educated and motivated to participate in research, and participants may also be motivated by family history of dementia⁴⁷. This likely limits generalizability. We also assumed that there were no major, unmeasured social or environmental events that could have substantialy affected our measurements or results; in particular, all data examined herein were collected prior to the onset of the COVID-19 pandemic.

Overall, the SiMOA plasma biomarkers considered herein do not appear strongly associated with medical conditions or demographic characteristics among cognitively normal research participants. However, this does not preclude that such plasma biomarkers, or their temporal changes, may predict cognitive decline or detect the presence of ADRD and VCID.

Vascular Biomarker Model Results
Outcome:	log(PlGF)			log(MMP9)
Predictor	Est.	SE	P-val	Est.	SE	P-val
Age	0.014	0.016	0.371	0.030	0.012	0.015
Batch	0.525	0.271	0.054	1.371	0.211	<0.001
CB	−0.545	0.429	0.205	−0.042	0.319	0.896
CV	0.249	0.245	0.312	−0.245	0.188	0.194
Diabetes	−0.248	0.283	0.381	0.054	0.218	0.805
Gender	0.091	0.200	0.648	0.036	0.153	0.815
Hypercholesterol	0.383	0.212	0.072	0.100	0.163	0.541
Hypertension	−0.304	0.213	0.155	0.028	0.162	0.864
Smoking	−0.275	0.192	0.154	0.133	0.147	0.365
R-squared	0.063			0.174
Effective N	226			232

NOTE: Model results for vascular biomarkers. Est: coefficient estimate; SE: Standard Error; PlFG: Placental growth factor; MMP9: Matrix Metallopeptidase 9; CB: Cerebrovascular conditions; CV: Cardiovascular conditions.

Inflammatory Biomarker Model Results
Outcome:	log(TNFa)			log(IL6)			log(IL8)			log(IL10)			log(IL1b)
Predictor	Est.	SE	P-val	Est.	SE	P-val	Est.	SE	P-val	Est.	SE	P-val	Est.	SE	P-val
Age	0.002	0.009	0.845	0.015	0.010	0.131	0.017	0.011	0.137	0.021	0.008	0.006	−0.003	0.017	0.866
B12 Deficiency	0.017	0.183	0.924	0.327	0.206	0.115	−0.104	0.242	0.667	0.260	0.169	0.124	−0.617	0.380	0.106
Batch	−0.294	0.146	0.046	0.314	0.166	0.059	0.625	0.181	0.001	0.110	0.128	0.389	−0.628	0.310	0.044
BMI	0.003	0.011	0.782	0.020	0.013	0.128	−0.017	0.014	0.237	0.003	0.010	0.770	0.018	0.024	0.466
Cancer	0.146	0.125	0.244	0.330	0.140	0.019	−0.090	0.158	0.569	0.196	0.108	0.071	0.332	0.255	0.195
CB	−0.084	0.233	0.721	0.034	0.274	0.901	0.000	0.286	0.999	0.184	0.213	0.387	0.349	0.453	0.442
COPD	0.030	0.203	0.882	−0.260	0.229	0.258	0.163	0.251	0.516	−0.122	0.178	0.493	−0.203	0.412	0.622
CV	0.199	0.132	0.133	0.195	0.148	0.190	−0.033	0.170	0.845	−0.002	0.115	0.986	−0.017	0.265	0.950
Depression	−0.164	0.135	0.226	−0.122	0.153	0.427	−0.046	0.168	0.785	0.159	0.119	0.181	−0.515	0.276	0.064
Diabetes	0.106	0.150	0.481	−0.068	0.169	0.687	0.000	0.191	0.998	0.069	0.131	0.599	0.023	0.298	0.938
Gender	0.051	0.106	0.628	−0.032	0.120	0.788	0.016	0.136	0.907	−0.133	0.093	0.157	−0.023	0.217	0.915
Hypercholesterol	0.140	0.113	0.215	−0.093	0.127	0.463	−0.054	0.145	0.707	−0.071	0.099	0.472	0.007	0.228	0.974
Hypertension	0.012	0.113	0.919	0.080	0.127	0.529	0.029	0.143	0.840	−0.069	0.098	0.481	−0.369	0.228	0.107
Smoking	−0.138	0.103	0.180	0.000	0.116	0.997	0.139	0.130	0.287	−0.043	0.090	0.630	−0.050	0.207	0.810
R-squared	0.075			0.082			0.088			0.1			0.09
Effective N	225			226			204			225			201

NOTE: Model results for inflammatory biomarkers. Est: coefficient estimate; SE: Standard Error; TNFα: Tumor necrosis factor alpha; IL: Interleukin; BMI: Body mass index; CB: Cerebrovascular conditions; CV: Cardiovascular conditions; COPD: Chronic obstructive pulmonary disease.

Highlights.

• Among N=237 cognitively normal adults, we studied candidate ADRD plasma biomarkers

• Biomarkers were largely not associated with demographic or health factors

• APOE was associated with Aβ42/Aβ40 ratio

• These results support hypotheses that plasma biomarkers are informative for ADRD

Research In Context.

1. Systematic Review: We reviewed the literature for studies assessing plasma biomarkers for ADRD, particularly for adults with normal cognition. We then reviewed the literature to identify known determinants of plasma biomarkers.

2. Interpretation: We observed no strong associations between biomarkers and demographic or health factors in two overlapping cohorts of 237 and 239 cognitively normal research volunteers. Our findings support the use of these plasma biomarkers in dementia research and suggest changes in biomarker levels may be attributed to ADRD processes.

3. Future Directions: As biomarker utility rests on ability to provide diagnosis and prognosis, additional studies will assess whether these biomarkers predict future cognitive decline.