Parallel neuroinflammatory pathways to cerebrovascular burden and amyloid beta in Alzheimer's disease

Abstract

INTRODUCTION

Upstream neuroinflammation plays an important role in Alzheimer's disease (AD) but remains poorly understood. We tested whether two distinct neuroinflammatory markers are associated with cerebrovascular burden and amyloid beta (Aβ), and downstream, with plasma phosphorylated tau (p‐tau217), medial temporal lobe (MTL) cortical and hippocampal atrophy, and memory deficits.

METHODS

Cognitively unimpaired older adults without dementia or mild cognitive impairment were recruited from a community sample (Biomarker Exploration in Aging, Cognition, and Neurodegeneration; [BEACoN]; N = 126). We used structural equation modeling to test whether plasma chitinase‐3‐like protein 1 (YKL‐40) and glial fibrillary acidic protein (GFAP) contribute to distinct pathways.

RESULTS

Higher plasma YKL‐40 was associated with greater white matter hyperintensity (WMH), whereas higher plasma GFAP was related to increased 18F‐florbetapir (FBP) standardized uptake value ratio (SUVR). Higher plasma GFAP, WMH, and FBP SUVR were independently associated with increased p‐tau217. Plasma p‐tau217 was associated with reduced MTL cortical thickness and hippocampal volume. Reduced hippocampal volume was related to worse memory.

DISCUSSION

Future work can further investigate these neuroinflammatory pathways as potential therapeutic targets for AD.

Highlights

• Higher YKL‐40 was linked to WMH, while higher GFAP was related to FBP SUVR.

• Higher plasma GFAP, WMH and FBP SUVR were associated with pTau‐217.

• Higher pTau‐217 was linked to lower MTL cortical thickness and hippocampal volume.

• Lower hippocampal volume was related to worse memory performance.

1. BACKGROUND

Alzheimer's disease (AD) is defined by two pathologies—amyloid beta (Aβ) plaques and neurofibrillary tangles. ¹ However, less emphasis is placed on the heterogeneity of processes in AD, which impedes us from determining novel potential mechanistic pathways preceding clinical symptoms and identifying effective markers for treatment. Although Aβ is hypothesized to initiate a biological cascade that leads to accelerated deposition and spread of tau pathology and neurodegeneration, ² there are likely multiple pathways that converge to promote neurodegeneration and cognitive impairment. Studying different and multiple neuroinflammatory pathways that reduce AD and vascular pathology is an important precursor to the development of single and multi‐targeted therapies. One such pathway involves the interplay between neuroinflammation and cerebrovascular burden, which may parallel the cascading impact of Aβ pathology. ²

Reactive astrocytes and microglia are important to both Aβ and tau pathogenesis, ³ and the possible bidirectional roles of neuroinflammation with AD proteinopathy have been examined. ⁴ For example, initial glial activation is thought to be an early mechanism to eliminate Aβ, yet this process often leads to a vicious cycle of development of further chronic neuroinflammatory and neurotoxic damage. ⁵ Specific reactive astrocytic and microglial markers, such as glial fibrillary acidic protein (GFAP) and chitinase‐3‐like protein 1 (YKL‐40), contribute to a higher risk of AD. ⁶ , ⁷

YKL‐40 has emerged as an important candidate biomarker for examining the clinical progression of AD, ⁸ and can be detected at the earliest stage of pathogenesis. ⁹ Similarly, GFAP is a valuable marker in the early detection and prediction of the course of AD. ¹⁰ Although these two markers have shown promising clinical significance in tracking AD, an existing limitation of these plasma‐based markers is that they may indicate both neuroinflammatory and systemic inflammatory processes. Comparing whether GFAP and YKL‐40 are differentially linked to Aβ and cerebrovascular burden has not yet been investigated. One marker reflecting small vessel cerebrovascular disease, broadly studied in aging and dementia, ¹¹ are white matter hyperintensities (WMHs), visualized as areas with increased brightness on T2‐weighted magnetic resonance imaging (MRI).

One study found that cerebrospinal fluid (CSF) GFAP was related to Aβ, whereas CSF YKL‐40 was related to higher tau‐PET (positron emission tomography). ¹² Ferrari‐Souza et al. ¹² further found that CSF GFAP and YKL‐40 mediated the effects of Aβ and tau, respectively, on hippocampal atrophy, which was subsequently related to worse cognition. Similarly, Pelkmans et al. ¹³ reported that CSF Aβ is linked to GFAP. However, Pelkmans et al. ¹³ determined that CSF YKL‐40 mediated the effect of Aβ on tau phosphorylation and its association with neurodegeneration. Furthermore, another study suggested that GFAP mediated the association between Aβ and tau. ¹⁴ Increases in plasma GFAP levels have also been associated with WMHs in early‐onset AD, late‐onset AD, and in adults with Down syndrome (DS). ¹⁵ , ¹⁶ Similarly, there is growing evidence of the association between YKL‐40 and WMHs. ¹⁷ A longitudinal study demonstrated an association between plasma and CSF YKL‐40 levels. ¹⁸ In a previous study, both plasma YKL‐40 and CSF YKL‐40 predicted conversion from mild cognitive impairment (MCI) to dementia, independent of amyloid, tau, and neurodegeneration. ¹⁹ Plasma GFAP more accurately distinguished Aβ‐positive individuals from Aβ‐negative individuals than CSF GFAP. Plasma GFAP levels were also higher among individuals in symptomatic stages of the AD continuum, and these magnitude changes were consistently higher in plasma GFAP compared to CSF GFAP. ²⁰ Another study indicated that plasma GFAP was a sensitive biomarker that outperformed CSF GFAP in detecting amyloid pathology in early AD. ²⁰ Given the inconsistency of the reported links of GFAP and YKL‐40 with Aβ, tau, and WMHs, we tested these associations in the context of an AD pathogenic pathway.

A novel feature of the current study is the examination of cerebrovascular disease and its interface with neuroinflammation, Aβ, and tau burden. Past work indicates an association of WMH with structural imaging markers of medial temporal lobe (MTL) neurodegeneration and memory deficits in older adults. ²¹ , ²² Furthermore, related work shows that WMHs precede and promote tau. ²³ We hypothesize that one mechanism by which WMHs lead to MTL and hippocampal neurodegeneration is through tau burden. We position plasma p‐tau217 as a point of convergence of two upstream pathways (Aβ and WMHs), consistent with evidence that p‐tau217 is associated with amyloid deposition and reflects the amyloid–tau interface. Emerging evidence indicates that it also captures tau pathology at subthreshold amyloid levels. Notably, in a companion study using advanced PET quantification sensitive to early tau clusters, we found that p‐tau217 was associated with tau PET uptake in the medial temporal lobes even among amyloid‐negative individuals, ²⁴ suggesting that p‐tau217 reflects early tau pathogenesis that may not be fully dependent on established amyloid burden. Overall, we propose that cerebrovascular disease and dysfunction is a pathology that is distinguishable from amyloidosis. We further suggest that these two types of pathologies may be associated with different neuroinflammatory cascades. This is in contrast with the single pathogenic pathway proposed by the amyloid, tau, and neurodegeneration (A/T/N) framework. ²⁵ We model these pathways using structural equation modeling, which examines linear and statistically causal relationships among variables. ²⁶

RESEARCH IN CONTEXT

1. Systematic review: In our literature review, while we found that previous work investigated the roles of GFAP and YKL‐40 in the context of AD pathogenic markers, findings were conflicting. Further, there is a limited understanding of how cerebrovascular injury fits in this pathway. We describe and cite these studies in the manuscript.

2. Interpretation: We found evidence for two distinct neuroinflammatory pathways that lead to neurodegeneration and memory deficits. One path involves plasma YKL‐40 and its impact on WMH, while the other involves GFAP and its impact on A?. Both pathways converged on plasma pTau‐217, which was associated with lower medial temporal lobe (MTL) cortical thickness and hippocampal volume, and consequently, memory deficits.

3. Future directions: These observations underscore the need for a more comprehensive approach to developing an AD framework. Longitudinal analyses should assess temporal associations among markers. Future work should also include a more diverse sample and a wider range of performance on cognitive measures.

2. METHODS

2.1. Participants

One hundred twenty‐six community‐dwelling adults were included in the study, as part of the Biomarker Exploration in Aging, Cognition, and Neurodegeneration (BEACoN) study. Inclusion criteria for BEACoN consists of age ≥ 60 years and performance on cognitive assessments within age‐adjusted normal range (within 1.5 standard deviation [SD]). Exclusion criteria include diagnosis of dementia or MCI (a global Clinical Dementia Rating [CDR] score >0 or Mini‐Mental State Examination [MMSE] score <27), major health problems (e.g., uncontrolled diabetes mellitus, uncontrolled hypertension, history or presence of stroke/infarct), comorbid neurological disease, significant psychiatric disorders, use of medication for anxiety or depression or illicit drugs, and MRI or PET contraindications. All participants provided written informed consent and were compensated for their participation. Study procedures were approved by the Institutional Review Board of the University of California, Irvine. The study is in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki). Participants were included in the study if MRI, Aβ PET, plasma, or neuropsychological data of interest in our primary analyses were available. Demographic characteristics of the participants, including the sample size of each marker, are included in Table 1. Although N = 126 participants were included in the statistical analyses, not all paths within the structural equation models were tested with the same n, as the n depended on the availability of the marker included in the path. In Table 1, we report the availability of each marker.

TABLE 1.

Demographic characteristics and summary variables.

Characteristic	Values
N	126
Age, mean (SD)	70.60 (6.29)
Education, mean (SD)	16.57 (2.32)
Sex/gender, n (%), women	78 (61.90%)
Race and ethnicity, n (%)
White	102 (80.95%)
Asian	18 (14.29%)
Black	1 (0.79%)
More than one race or Other	5 (3.97%)
Hispanic	6 (4.76%)

Markers (Plasma, Imaging, Neuropsychological)	Raw values	Log‐transformed values
YKL‐40, pg/mL, n = 69, mean (SD)	51.41 (41.23)	3.73 (0.64)
GFAP, pg/mL, n = 69, mean (SD)	353.90 (331.89)	5.65 (0.63)
Total WMH, cm3,n = 112, mean (SD)	3.60 (8.14)	1.02 (0.83)
Global FBP SUVR, n = 108, mean (SD)	1.11 (0.17)	0.74 (0.08)
p‐tau217, pg/mL, n = 64, mean (SD)	0.12 (0.10)	0.11 (0.08)
MTL cortical thickness, mm, n = 122, mean (SD)	1.20 (0.06)	0.79 (0.27)
Hippocampal volume, TIV adjusted, n = 122, mean (SD)	1.96 (0.26)	1.08 (0.09)
Retroactive interference, n = 111, mean (SD)	0.84 (0.18)	‐
Delayed recall, n = 111, mean (SD)	10.72 (3.52)	‐

Abbreviations: FBP, 18F‐florbetapir; GFAP, glial fibrillary acidic protein; MTL, medial temporal lobe; p‐tau217, phosphorylated tau 217; SD, standard deviation; SUVR, standardized uptake value ratio; TIV, total intracranial volume; WMH, white matter hyperintensities; YKL‐40, chitinase‐3‐like protein 1.

2.2. Magnetic resonance imaging

MRI data were acquired on a 3.0 Tesla Siemens Prisma scanner at the Facility for Imaging and Brain Research (FIBRE) at the University of California, Irvine. (See details of scan acquisition in Supplementary Methods.)

2.3. MRI analyses

2.3.1. WMH segmentation

Image processing leveraged the open‐source ANTsX software ecosystem with a particular focus on specific deep‐learning applications developed for neuroimaging made available for both Python and R via the ANTsXNet (ANTsPyNet/ANTsRNet) libraries. Specifically, for the work described here, WMH segmentation used the ANTsPyNet function: sysu_white_matter_hypterintensity_segmentation.

The methods herein for WMH segmentation (see Figure 1A,B) and quantification have been described previously. ²² (See additional details on WMH segmentation in Supplementary Methods.) These edited global WMH volumes were used in our analyses.

FIGURE 1.

Visualizing imaging analyses. (A) Unlabeled white matter hyperintensities (WMHs) on an axial T2‐FLAIR (fluid‐attenuated inversion recovery) image. (B) Labeled WMHs in red on an axial T2‐FLAIR image. (C) Global 18F‐florbetapir (FBP) standardized uptake value ratio (SUVR) image. (D) T2 image with hippocampal and medial temporal lobe (MTL) cortical subregions labeled using Automatic Segmentation of Hippocampal Subfields (ASHS).

2.3.2. MTL subregion segmentation

MTL subregions were automatically segmented (see Figure 1D) with the Automatic Segmentation of Hippocampal Subfields (ASHS) ²⁷ software using T1‐ and T2‐weighted images.) See additional details on ASHS and extraction of measures in Supplementary Methods.)

The resulting output included thickness of the MTL cortical subregions and total MTL cortical thickness and was derived from averaging the thickness of four subregions. Because ASHS provides segmentation of individual hippocampal subfields, we summed all hippocampal subfields to preserve power and relied on past literature indicating that overall hippocampal volume is a sensitive measure of neurodegenerative change. ²⁸ All hippocampal subfield volumes were adjusted for intracranial volume (ICV) by dividing the subfield volume by the individual's ICV. Total hippocampal volume was derived by taking the sum of the ICV‐adjusted bilateral hippocampal subfield volumes. Total hippocampal volume was then multiplied by 1000.

2.4. Plasma markers

2.4.1. Plasma collection

Blood was collected without regard to prandial state, time of day, or medication timing. Blood was drawn via venipuncture from each participant into 7 mL lavender top ethylenediaminetetraacetic acid (EDTA) tubes (BD 366450). (See additional details regarding plasma collection in the Supplementary Methods.)

2.4.2. Plasma biomarker quantification

Levels of YKL‐40 were quantified in plasma samples from BEACoN participants using the U‐PLEX Human YKL‐40 ECL immunoassay MesoScale Discovery (MSD; Cat #K151VLK Gaithersburg, MD). Levels of GFAP were quantified using the R‐PLEX Neurology Panel 1. Plasma p‐tau217 levels were quantified using an immunoassay developed by ALZpath, using a previously described method. ²⁹

2.5. Amyloid PET

2.5.1. PET acquisition

18F‐Florbetapir (FBP) positron emission tomography (PET) was used to measure Aβ plaque pathology. (See additional details on PET acquisition in the Supplementary Methods.)

2.5.2. PET processing

FBP frame data were realigned, averaged, and then co‐registered to the T1‐weighted Magnetized Prepared Rapid Gradient Echo (MPRAGE) image. T1 MPRAGE images were processed with FreeSurfer v.6.0. (See details on PET processing in Supplementary Methods.) FBP images were then normalized by a whole cerebellum reference region to produce standardized uptake value ratio (SUVR) images (see Figure 1C). The mean SUVR of a validated cortical composite region was then quantified to obtain a global measure of FBP uptake for analysis.

2.6. Neuropsychological testing

Participants were administered the Rey Auditory Verbal Learning Test (RAVLT), which assesses word list learning and immediate and delayed memory. Our primary outcome measure of memory in this study was retroactive interference (RI) from RAVLT, which reflects resistance to interference during consolidation. (See additional details of RAVLT in Supplementary Methods.) In an additional analysis, we used delayed recall from RAVLT (A7; Table S1).

2.7. Vascular risk

We derived a vascular risk factor summary score as a covariate (see details in Supplementary Methods).

2.8. Statistical analysis

Statistical analyses were conducted in R (version 4.3.0) through RStudio. All imaging and plasma markers were log‐transformed to reduce the skewness of the data prior to running analyses, with a constant of 1 added before transforming to handle any zeros and to preserve non‐negative values.

We conducted a structural equation model ³⁰ analysis using the lavaan package in R in a series of linear regressions simultaneously. We used full information maximum likelihood to handle missing data. Independent variables were considered as random. In our SEM analysis, we tested whether neuroinflammatory markers, including plasma GFAP and YKL‐40, are bidirectionally associated with global WMH volume and global FBP SUVR. We assessed whether WMH volume and global FBP SUVR (Aβ) were bidirectionally associated with each other, and further whether they were associated with plasma p‐tau217. We also tested whether plasma YKL‐40 and GFAP are bidirectionally associated with plasma p‐tau217, and whether plasma p‐tau217 is associated with MTL cortical thickness and hippocampal volume. We assessed whether MTL cortical thickness and hippocampal volume are subsequently associated with performance on retroactive interference from RAVLT. In the supplementary results (Table S1), we additionally report results using RAVLT delayed recall as an alternative outcome measure. In a second supplementary analysis (Table S2), we tested whether the associations of plasma p‐tau217 with hippocampal volume and MTL cortical thickness would be statistically independent from Aβ (FBP SUVR) by adjusting for FBP SUVR in these two associations. Finally, in a sensitivity analysis (Table S3), we adjusted for age and sex, and further adjusted for a vascular risk factor summary score in associations among biomarkers.

We assessed goodness of model fit as “acceptable” if: X² p > 0.05, Comparative Fit Index (CFI ≥0.90), Tucker‐Lewis Index (TLI ≥0.90), Root Mean Square Error of Approximation (RMSEA <0.08), and Standardized Root Mean Square Residual (SRMR <0.08).

3. RESULTS

In our structural equation model, we tested associations among markers and hypothesized a theoretically driven mechanistic cascade. The overall model fit was considered acceptable: X² (13, N = 126) = 17.72, p = 0.17, CFI = 0.93, TLI = 0.86, RMSEA = 0.05, and SRMR = 0.063. Given the modest sample size, limited plasma marker availability, and use of FIML estimation, model fit was evaluated across multiple indices rather than relying on a single threshold. A bidirectional association between plasma YKL‐40 levels and plasma GFAP levels was not found (b = 0.07, 95% confidence interval [CI] –0.025 to 0.164], p = 0.147; Table 2, Figure 2). Higher plasma YKL‐40 levels were associated with greater global WMH (b = 0.158, 95% CI 0.047–0.269, p = 0.005), but not with global FBP SUVR (b = 0.019, 95% CI 0.007–0.032, p = 0.003). Meanwhile, higher plasma GFAP levels were associated with higher global FBP SUVR (b = 0.019, 95% CI 0.007–0.032, p = 0.003), but not with global WMH (b = 0.061, 95% CI –0.049 to 0.17, p = 0.277). We did not find an association between FBP SUVR and global WMH (b = 0.008, 95% CI –0.004 to 0.021, p = 0.175). Higher plasma GFAP levels were also related to higher plasma p‐tau217 levels (b = 0.012, 95% CI 0.002–0.022, p = 0.022). Both greater global WMHs and increased global FBP SUVR were independently associated with higher plasma p‐tau217 levels (WMH: b = 0.023, 95% CI 0.006–0.041, p = 0.009; FBR SUVR: (b = 0.485, 95% CI 0.266–0.705, p < 0.001). Higher plasma pTau‐217 levels were in turn associated with both reduced MTL cortical thickness (b = –0.112, 95% CI –0.178 to –0.047, p = 0.001) and lower hippocampal volume (b = –0.277, 95% CI –0.498 to –0.056, p = 0.014). Subsequently, hippocampal volume, but not MTL cortical thickness, was related to lower performance on retroactive interference (RAVLT; hippocampal volume: b = 0.692, 95% CI 0.306–1.078, p < 0.001; MTL cortical thickness: b = 0.79, 95% CI –0.499 to 2.08, p = 0.23). Simple associations among markers included in the SEM are visualized with a heat correlation map and with scatter plots (Figure 3 ; Figure S1).

TABLE 2.

Primary structural equation model (SEM) analysis.

Dependent variable	Path	Independent variable	Unstandardized B	Standardized β	p‐value	95% CI for Unstandardized B
p‐tau217	←	Total WMH	0.023	0.233	0.009	(0.006, 0.041)
		FBP SUVR	0.485	0.464	<0.001	(0.266, 0.705)
MTL cortical thickness	←	p‐tau217	−0.112	−0.347	0.001	(–0.178, –0.047)
Hippocampal volume	←	p‐tau217	−0.277	−0.255	0.014	(–0.498, –0.056)
Retroactive interference	←	Hippocampal volume	0.692	0.346	<0.001	(0.306, 1.078)
		MTL cortical thickness	0.790	0.118	0.230	(–0.499, 2.08)
		p‐tau217	0.350	0.161	0.161	(–0.14, 0.839)
YKL‐40	↔	GFAP	0.070	0.177	0.147	(–0.025, 0.164)
Total WMH	↔	FBP SUVR	0.008	0.131	0.175	(–0.004, 0.021)
		YKL‐40	0.158	0.308	0.005	(0.047, 0.269)
		GFAP	0.061	0.116	0.277	(–0.049, 0.17)
FBP SUVR	↔	YKL‐40	−0.006	−0.117	0.372	(–0.018, 0.007)
		GFAP	0.019	0.389	0.003	(0.007, 0.032)
p‐tau217	↔	YKL‐40	0.002	0.037	0.751	(–0.008, 0.011)
		GFAP	0.012	0.269	0.022	(0.002, 0.022)

Abbreviations: FBP, 18F‐florbetapir; GFAP, glial fibrillary acidic protein; MTL, medial temporal lobe; p‐tau217, phosphorylated tau 217; SUVR, standardized uptake value ratio; WMH, white matter hyperintensities; YKL‐40, chitinase‐3‐like protein 1.

All variables are log‐transformed, except for retroactive interference.

SEM analysis model with the following regression and covariance equations:.

Regressions (∼).

p‐tau217 ∼ Total WMH + FBP SUVR + YKL‐40 + GFAP.

MTL cortical thickness ∼ p‐tau217.

Hippocampal volume ∼ p‐tau217.

Retroactive interference ∼ MTL cortical thickness + hippocampal volume + p‐tau217.

Covariances: bidirectional associations (∼∼).

YKL‐40 ∼∼ GFAP.

Total WMH ∼∼ FBP SUVR.

Total WMH ∼∼ YKL‐40 + GFAP.

FBP SUVR ∼∼ YKL‐40 + GFAP.

p‐tau217 ∼∼ YKL‐40 + GFAP.

FIGURE 2.

SEM diagram indicating pathway from neuroinflammatory markers (YKL‐40 and GFAP) to downstream biomarkers and subsequent memory deficits. Solid lines demonstrate significant paths, and dotted lines represent non‐significant paths. Green arrows represent positive associations, and red arrows represent negative associations. Paths indicated as significant (*) for p‐values ≤ 0.05, (**) for p ≤ 0.01, and (***) for p ≤ 0.001. YKL‐40: chitinase‐3‐like‐protein 1; GFAP: glial fibrillary acidic protein; WMH: white matter hyperintensities; FBP: 18F‐Florbetapir; SUVR: standardized uptake value ratio; ptau‐217: phosphorylated tau 217; MTL: medial temporal lobe.

FIGURE 3.

Correlation heatmap of associations among markers that were included in the SEM analysis. Cells that are significant are denoted by (*) for p‐values ≤ 0.05, (**) for p ≤ 0.01, and (***) for p ≤ 0.001. YKL‐40: chitinase‐3‐like protein 1; GFAP: glial fibrillary acidic protein; WMH: white matter hyperintensities; FBP: 18F‐Florbetapir; SUVR: Standardized uptake value ratio; ptau‐217: phosphorylated tau 217; MTL: medial temporal lobe.

We found similar patterns of associations among biomarkers and determined that the relationship between hippocampal volume and delayed recall (RAVLT) was trending (see Table S1). In addition, we found in a separate SEM analysis that the associations of p‐tau217 with neurodegeneration of the hippocampus and MTL cortical thickness were statistically independent of FBP SUVR (Table S2). In a sensitivity analysis, adjusting for age, sex, and a vascular risk factor summary score (for associations among biomarkers), we still found evidence of two distinct and parallel neuroinflammatory pathways (see Table S3). We found that plasma YKL‐40 was still bidirectionally associated with global WMHs, and plasma GFAP was still only bidirectionally associated with FBP SUVR.

4. DISCUSSION

Our results support a model in which two glial cell activation and neuroinflammatory markers, YKL‐40 and GFAP, are differentially related to small vessel cerebrovascular disease and amyloid, respectively, suggesting at least two pathways converging to phosphorylated tau. Phosphorylated tau was subsequently associated with neurodegeneration and lower memory performance in older adults. After adjustment of age, sex, and vascular risk factors in a sensitivity analysis, main findings remained similar with evidence of two distinguishable and parallel neuroinflammatory pathways. However, we found that WMH was no longer related to p‐tau217, and there was a downstream association between p‐tau217 with only MTL cortical thickness. We postulate that controlling for age restricted the variance in p‐tau217, potentially weakening the association between WMHs and p‐tau217, as these were cognitively unimpaired individuals.

Although many studies showed a temporal ordering among amyloid, tau, and neurodegeneration, ³¹ , ³² , ³³ , ³⁴ a similar pathway has not been described with respect to cerebrovascular disease, tau, and neurodegeneration. However, previous work demonstrated associations between cerebrovascular disease and tau in both humans and mice. ²³ , ³⁵ Similarly, prior work showed that WMHs are associated with MTL atrophy in both longitudinal and cross‐sectional analyses. ²¹ , ²² Despite prior evidence of a link between Aβ pathology and longitudinal WMH progression, ³⁶ we did not find an association between Aβ and global WMHs in our study of cognitively unimpaired individuals. Furthermore, it is possible that the association between Aβ and vascular burden may be better indexed by using another measure of vascular pathology such as enlarged perivascular spaces. Of interest, we expected to observe a direct effect of phosphorylated tau on memory given past studies suggesting alternative pathways toward memory loss. ³⁷ It is possible that hippocampal volume fully mediates the association between phosphorylated tau and memory deficits. Future work should delineate whether these effects of exposures on outcomes are statistically mediated by biomarkers through longitudinal analyses to assess temporality and directionality; our cross‐sectional data and sample size limited us from performing mediational analyses.

We found that plasma YKL‐40, a glycoprotein secreted and synthesized by astrocytes, microglia, and peripheral macrophages, ¹⁷ , ³⁸ was associated with WMHs but not with Aβ. Previous work on serum YKL‐40 demonstrated its association with WMH volume, and, further, that macroscopic and microscopic white matter damage mediated the correlation between serum YKL‐40 and cognitive impairment in patients with small vessel disease. ¹⁷

In contrast to our findings with YKL‐40, we found that plasma GFAP was associated with Aβ, but not with WMHs. Prior work demonstrated that GFAP is associated with Aβ burden in unimpaired older adults and in those who are at risk for AD. ⁶ , ¹⁴ , ³⁹ Studies have also reported a link between GFAP and WMHs ¹⁶ , ³⁹ ; Shir et al. ³⁹ demonstrated that the relationship between GFAP and WMHs is only found in individuals with elevated levels of Aβ. Consistent with this understanding, the link between GFAP and WMHs shown by Edwards et al. ¹⁶ was found in adults with DS, who also demonstrate elevated levels of amyloid. ⁴⁰

There are some limitations with the current study. Due to the cross‐sectional nature of the study design, we are unable to address the temporality or causality of the correlational associations found through the SEM analysis. However, structural equation modeling allowed us to test a hypothesized “causal” model. ⁴¹ Although the current study did not include markers of microglial reactivity such as cytokines and triggering receptor expressed on myeloid cells 2 (TREM‐2), they would be highly relevant to investigate in future research, in understanding the interactive and nonlinear pathways of these neuroinflammatory markers and their contributions to AD. As this cross‐sectional work was done in community‐dwelling older adults without cognitive impairment, these findings apply only to early stages of AD, and may not be generalizable to later stages of AD, where associations of neuroinflammatory markers with pathology may change with disease progression. Although the use of plasma biomarkers is validated increasingly in AD cohorts, ⁴² challenges persist in their interpretation, such as the confounding of results from other medical comorbidities and potential peripheral inflammation. ⁴³

Given that plasma p‐tau217 is associated with both amyloid and tau pathology, ⁴⁴ another future step would be to use a more specific measure of tau aggregation, such as tau‐PET. In addition, we note that in a separate study, we have discovered that p‐tau217 is associated with tau PET uptake in the MTLs, even in amyloid‐negative individuals. Uncovering this required using a more advanced method of PET SUVR quantification that is sensitive to the early island‐like appearance of tau clusters. These data were recently presented in an Alzheimer's Association International Conference (AAIC) poster ²⁴ and are now being prepared for publication. These findings are consistent with the notion that p‐tau217 presence may be independent of amyloid. Furthermore, future work should assess the non‐linearity of these relationships between biomarkers in the context of disease progression; these findings inferred linear relationships, specific to cognitively unimpaired older adults without significant disease burden. Previous work suggested non‐linear trajectories with markers including Aβ burden and glial markers. ⁴⁵

Although the established A/T/N framework of AD suggests a single pathway, our novel findings suggest important roles of vascular and inflammatory markers that contribute to cognitive impairment. More specifically, our work points to multiple neuroinflammatory pathways, differentially associated with both amyloid and vascular disease, which independently converge onto p‐tau217, which is associated with neurodegeneration and subsequent memory impairment. While some associations between neuroinflammatory markers and other biomarkers appear to be bidirectional, there are multiple potential drivers of neuroinflammation, such as environmental and metabolic risk factors, that could trigger changes in inflammation. ⁴⁶ , ⁴⁷ In running a competing model, with unidirectional paths from GFAP and YKL‐40 to Aβ and WMHs, we had similar findings of these associations, suggesting the need for follow‐up longitudinal work.

In summary, we found that cerebrovascular burden and Aβ form two parallel and distinct pathways by which neuroinflammatory changes may alter Alzheimer's pathological features, converging on the neurodegenerative cascade leading to cognitive decline. Identifying these pathways enables the early detection of biomarkers and the targeting of therapeutic interventions to the appropriate windows of disease. This understanding can also provide traction for addressing individual differences in AD vulnerability and resilience, potentially paving the way for personalized therapeutic strategies

CONFLICT OF INTEREST STATEMENT

Michael A. Yassa has received consulting fees from Eisai Cognito Therapeutics, LLC, CuraSen Therapeutics, Inc, and Enthorin Therapeutics, LLC. Adam M. Brickman receives compensation for consultation to Cognition Therapeutics and Cognito Therapeutics; he is on the Scientific Advisory Board of CogState, is an inventor of a patent for white matter hyperintensity quantification (US Patent US9867566B2), and serves on a Data Safety Monitoring Board for the University of Illinois, Urbana‐Champaign. All other authors do not have competing interests to declare. Any author disclosures are available in the ICMJE. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All participants provided written informed consent.

Supporting information

Supporting Information File 1: dad270415‐sup‐0001‐SuppMat.docx

Supporting Information File 2: dad270415‐sup‐0002‐ICMJE.pdf