Cortical Gray Matter Proteins Associated With Cerebral Amyloid Angiopathy in Community-Dwelling Older Adults

Abstract

Background and Objectives

Cerebral amyloid angiopathy (CAA) is the accumulation of β-amyloid (Aβ) in the walls of small vessels in the leptomeninges and cortex and is a risk factor of intracerebral hemorrhage and dementia, but its underlying mechanism is unknown. We examined cortical proteins in relation to CAA to elucidate the molecular mechanisms underlying CAA.

Methods

Data were collected from participants of 5 community-based cohorts of older adults. Participants were recruited from personal dwellings or retirement centers, were without known dementia at enrollment, and signed an Anatomic Gift Act for brain donation. At death, autopsy was performed and cortical proteins were quantified from the dorsolateral prefrontal cortex using mass spectrometry–based proteomic analysis, and pathologic indices of brain pathologies including CAA and Alzheimer disease (AD) were assessed during neuropathologic evaluations. Targeted mass spectrometry–based proteomic analysis was performed for the quantification of total Aβ protein and Aβ38 peptide. Ordinal logistic regression models were used to test the association between the proteins and CAA.

Results

A total of 887 participants were included, with a mean age at death of 89.0 (SD = 6.8) years, and 67.2% (n = 596) were women. Eighty proteins were related to CAA, of which 12 remained associated with CAA after controlling for AD pathology. However, only 4 proteins remained associated with CAA when all 12 proteins were examined in a single model: secreted modular calcium-binding protein 1 (SMOC1), secreted frizzled-related protein 1 (SFRP1), APOE, and APOE4. Examining the 4 proteins together with the 3 Aβ measures (Aβ load, total Aβ protein, and Aβ38 peptide) in a factor analysis and a structure equation model suggested 2 factors and paths: a factor including Aβ38, SFRP1, and APOE protein, which had the larger effect size in relation to CAA (standardized estimate = 0.459, SE = 0.050, p < 0.001), and a second factor including Aβ load, total Aβ protein, SMOC1, and APOE4 that had a smaller effect size in relation to CAA (standardized estimate = 0.253, SE = 0.083, p < 0.001).

Discussion

This study suggests 2 molecular pathways underlying CAA, with a larger effect size for the pathway including SFRP1 and APOE protein and C-terminally truncated Aβ before position 40. However, the study is an observational cross-sectional study that limits causal inference from the findings.

Introduction

Cerebral amyloid angiopathy (CAA) involves the accumulation of β-amyloid (Aβ) in the walls of small arteries, arterioles, veins, and even capillaries in the leptomeningeal and cortical vessels.¹ CAA results in either blood supply shortage and ischemia or loss of vessel integrity and hemorrhage.² Clinical consequences of CAA-related ischemia are cognitive impairment and dementia³ because of microinfarcts and loss of myelin or axonal damage,⁴ and clinical and imaging consequences of CAA-related hemorrhage are lobar intracerebral hemorrhage, microhemorrhages, superficial siderosis, and transient focal neurologic deficits.⁵ Because CAA is common in older adults,⁶ identifying the molecular mechanisms underlying CAA and development of CAA treatments is a public health priority.

The current anti-Aβ immunotherapy for Alzheimer disease (AD) has brought extra attention to CAA as a result of its association with the immunotherapy side effect amyloid-related imaging abnormality (ARIA).⁷ Similarities that exist between CAA and ARIA include APOE ε4 allele as the genetic risk factor,^2,7 the ability of anti-Aβ antibodies to induce ARIA and an inflammatory type of CAA,⁸ and similar MRI findings. Therefore, understanding molecular mechanisms underlying CAA is not only imperative for the development of CAA-targeted treatments but may also help in decreasing side effects of anti-Aβ antibody treatment in patients with AD.

Large-scale proteomic analysis of brain tissue is an invaluable approach to clarify molecular mechanisms underlying brain diseases because proteins form the backbone of cells and tissue structure and are the key regulators of cellular functions. In previous studies, quantified proteins were interrogated in relation to AD pathology.⁹ In this study, we examined the quantified proteins in relation to CAA.

Methods

Participants

Participants were from one of the 5 ongoing cohort studies conducted at Rush Alzheimer's Disease Center: the Religious Orders Study (ROS),¹⁰ the Rush Memory and Aging Project (MAP),¹⁰ the Minority Aging Research Study (MARS),¹¹ the Rush Alzheimer's Disease Center African American Clinical Core (AA Core),¹¹ and the Rush Alzheimer's Disease Center Latino Core (Latino Core).¹² Older adults were enrolled in the 5 cohorts if they did not report a diagnosis of dementia at enrollment and consented to annual clinical examination. While consenting to brain donation was a prerequisite for enrollment in ROS and MAP, brain donation was optional in MARS, AA Core, and Latino Core. ROS began in January 1994 by enrolling older clergy women and men across the United States. MAP began in September 1997 by enrolling older adults living in personal dwellings, retirement centers, and subsidized housing across northeast Illinois. MARS and AA Core began in August 2004 and January 2008, respectively, by recruiting older African Americans, and the Latino Core began in September 2015 by recruiting older Latino adults, to complement the ROSMAP cohort, which was less diversified and mostly (approximately 90%) comprised non-Latino White. Like MAP, the participants of MARS, AA Core, and Latino Core were community dwellers recruited across the greater Chicago metropolitan area.

By December 2023 when the analyses began, 2,020 participants had died, undergone autopsy, and completed neuropathologic assessments including assessment of CAA. From the 2,020 decedents with valid CAA data, the analytic sample comprised the first 887 who had completed proteomic assessment of the brain cortex. Residential status of the participants in the last visit before death is provided in eTable 1.

Postmortem Pathologic Assessment

The median time from death to brain removal was 6.5 hours (interquartile range [IQR] 5.0–9.1). After brain removal, one hemisphere was frozen at −80°C for further studies including quantification of proteins and the other hemisphere was fixed in 4% paraformaldehyde. The fixed hemisphere was cut into 1-cm slabs, and thin tissue sections were prepared for the collection of pathologic indices from predetermined regions that were not close to areas of gross tissue injury due to large infarcts, hemorrhages, or trauma. Further details are provided elsewhere.¹³

Cerebral Amyloid Angiopathy

As described previously,¹⁴ tissue sections were prepared from 4 regions: midfrontal, midtemporal, angular, and calcarine cortices. Immunohistochemical methods using monoclonal antibodies (4G8 dilution 1:9,000; Covance Labs, Madison, WI, 6F/3D dilution 1:50; Dako North America Inc., Carpinteria, CA, and 10D5 dilution 1:600; Elan Pharmaceuticals, San Francisco, CA) against Aβ were used for scoring Aβ deposition in the small vessels and CAA severity.

For scoring CAA, we followed consensus protocols and criteria.¹⁵ CAA severity in each of the 4 regions was scored using a semiquantitative scale indicating the degree of Aβ deposition in each vessel and the number of affected vessels.¹⁴ Regional CAA scores were averaged to compute an ordinal 4-level CAA summary score (none, mild, moderate, and severe; Figure 1). The average of the regional CAA score was also used as a continuous CAA measure in a sensitivity analysis.¹⁴ For descriptive purposes, we used a binary variable indicating the presence of moderate-severe CAA.

Figure 1. Cerebral Amyloid Angiopathy in the Brain.

4G8 immunohistochemistry performed in the calcarine cortex represents none (A), mild (B), moderate (C), and severe (D) level of cerebral amyloid angiopathy. Black arrows indicate meningeal vessels with amyloid beta deposition. Red arrows indicate parenchymal vessels with amyloid beta deposition. Blue arrow represents parenchymal vessels with no amyloid beta deposition. Scale bar = 200 μm.

Alzheimer Disease

Modified Bielschowsky silver stain was used for staining sections and visualizing the AD pathologic hallmarks: diffuse and neuritic plaques and neurofibrillary tangles. Regional counts of the hallmarks, including those from the midfrontal region, were used to yield the global AD pathology score.¹⁶ The stained sections were also reviewed for adjudicating an AD pathologic diagnosis. In addition, anti-Aβ immunohistochemical staining of the brain regions, including the midfrontal region, was used for quantification of Aβ load, as described in eMethods.

Tandem Mass Tag Mass Spectrometry–Based Proteomic Analysis

For quantification of proteins, the dorsolateral prefrontal cortex (DLPFC), including midfrontal region as the main composer, was chosen to be the donor region because of its major contribution to diverse cognitive functions and its vulnerability to the accumulation of AD and other pathologies. As described previously,¹⁷ 100 mg of DLPFC tissue from the frozen hemisphere was homogenized, sonicated, and centrifuged, and the supernatant was separated for protein quantification. After enzymatic digestion, the peptides were labeled, separated, and quantified using liquid chromatography and mass spectrometry and by alignment against the canonical UniProtKB human proteome database. Particularly, the APOE4 protein, coded by the APOE ε4 allele, was identified by the presence of LGADMEDVR sequence in the APOE spectral assignment.^18,19 Several quality control measures were applied, including use of global internal standard and regression of several possible confounders such as protein batch.²⁰ The TMT proteomic analysis of DLPFC quantified 10,030 proteins, with most proteins quantified in all the participants (median of the number of missing of the proteins: 0.0 [IQR 0.0–61.0]). The median time between death and the protein quantification in the brain tissues kept frozen was 11.0 years (IQR 7.0–15.0).

Targeted Proteomic Analysis

Targeted mass tag mass spectrometry analysis is an approach where thousands of proteins can be quantified in a single analysis. By contrast, targeted liquid chromatography–selection reaction monitoring (SRM) proteomic analysis provides the opportunity to quantify different proteoforms of the same protein by designing the assays for quantifying specific, also known as proteotypic, peptides.²¹ Approximately 20 mg of frozen DLPFC tissue was homogenized and digested. Corresponding to peptides of interest, stable isotope-labeled synthetic peptides were created, mixed with the digested peptides, and examined using liquid chromatography SRM mass spectrometry. The quantified levels of Aβ peptides (total Aβ: LVFFAEDVGSNK and Aβ38: GAIIGLMVGG) were used in this study.

Other Covariates

Annually, 21 neuropsychological tests were administered. In a systematic centralized process, a neuropsychologist first reviewed the test scores and inferred the presence of impaired cognition and then a dementia-expert clinician, either a neurologist, neuropsychologist, or a family physician, reviewed the neuropsychologist's impression together with other selected clinical data available and adjudicated the presence of dementia according to established criteria.²² In addition, a neurologist reviewed the annual clinical and neuropsychological data and clinical diagnoses of deceased participants and rendered a summary clinical diagnosis of dementia at death.

Other covariates were age at death; sex; race/ethnicity; years of education; vascular risk factors including hypertension, diabetes mellitus, smoking, and blood lipids (low-density and high-density lipoprotein cholesterol and triglyceride); the presence of APOE ε4 allele; and the presence of dementia and Lewy bodies, which are described in eMethods. In addition, eMethods includes the methods for obtaining other omics data related to the genes encoding the proteins that were independently related to CAA, including single-nucleotide genetic variations, bulk tissue and single nuclei RNA expression, and DNA methylation.

Statistical Analysis

Ordinal logistic regression models were used to examine the associations of the proteins with CAA, and the proportional odds assumption was assessed using the Score test; no violation of this assumption was observed. Proteins were tested in separate models, and p values were adjusted using the Bonferroni method to address inflation of type I error due to multiple comparisons. Participants with missing data for a protein were excluded only from the model analyzing that specific protein. Proteins that remained associated with CAA after further controlling for AD pathology were examined in a single model using backward elimination to identify proteins independently related to CAA. In supplementary analyses, we examined single-nucleotide variants, RNA expression, and DNA methylation levels of the genes encoding the proteins related to CAA using ordinal logistic regression, with CAA as the outcome. We first examined each of the single-nucleotide variants and CpG sites in relation to CAA using separate ordinal logistic regression models. Then, p values of the associations of the single-nucleotide variants or CpG sites in each genic region were combined using a Cauchy variant of the Fisher product method^23,24 to provide an overall omnibus p value that indicates whether the genetic variation or DNA methylation in a gene was related to CAA.²⁵ If the overall omnibus p value of a gene was significant, we next considered the p values of the single-nucleotide variants or CpG sites of the gene in the association with CAA to identify significant genetic loci of the gene, based on p values < 0.05 after Bonferroni correction.

Functional enrichment analysis was performed using the gProfiler web server.²⁶ Analysis was performed using default parameters and applying the g:SCS algorithm for multiple test correction. Terms were considered significant if the adjusted p value was <0.05. Specifically, for Gene Ontology (GO) enrichments, a 2-stage algorithm, implemented in gProfiler, was used to highlight GO terms and remove redundancy between terms rather than simply selecting the term with the highest significance level.

We used causal mediation to test whether CAA-related proteins mediated the association of APOE4 with CAA. We used an exploratory factor analysis to examine the factor structure of CAA-related proteins and Aβ measures (immunohistochemically derived Aβ load, total Aβ protein, and Aβ38 peptide), with varimax rotation. Factors with eigenvalues greater than 1 were retained. An absolute factor loading above 0.4 was used as the threshold for assigning a variable to a factor.

We used structural equation modeling (SEM) to compare the 2 pathways of Aβ accumulation in relation to CAA. The goodness-of-fit index (GFI) and the root mean square error of the approximation (RMSEA) were the model fit indices. All the models including the SEM were controlled for age at death and sex. The analyses were performed using SAS version 9.4.

Standard Protocol Approvals, Registrations, and Patient Consents

Each study was separately approved by a Rush University Medical Center Institutional Review Board. Written informed consent was collected from all the participants, and an Anatomic Gift Act was signed by the participants for brain donation. The institutional review board approval numbers are L91020181 (ROS) and L86121802 (MAP).

Data Availability

The analyzed data are available on request submitted at radc.rush.edu.

Results

The characteristics of the participants (n = 887) are summarized in Table 1. CAA was not distributed uniformly by its severity: 22.9% (n = 203) did not have CAA, 43.7% (n = 388) had mild CAA, 21.1% (n = 187) had moderate CAA, and 12.3% (n = 109) had severe CAA. Compared with those with none-mild CAA, participants with moderate-severe CAA were older and were more likely to have APOE Ɛ4 allele, dementia, and AD pathology (Table 1).

Table 1.

Characteristics of Study Participants (n = 887)

Characteristicsa	None-mild CAA (n = 591)	Moderate-severe CAA (n = 296)	Test statistics, p valuesa	All (n = 887)
Age at death, y	88.6 (6.9)	90.0 (6.4)	t = −3.0, 0.003	89.0 (6.8)
Years from baseline to death	8.3 (4.7)	8.4 (4.6)	t = −0.09, 0.93	8.3 (4.7)
Sex			χ2 = 0.10, 0.75
Men	196 (33.2)	95 (32.1)		291 (32.8)
Women	395 (66.8)	201 (67.9)		596 (67.2)
Race/ethnicity			χ2 = 0.66, 0.42b
Latino	20 (3.4)	21 (7.1)		41 (4.6)
Non-Latino Black	54 (9.1)	24 (8.1)		78 (8.8)
Non-Latino White	513 (86.8)	251 (84.8)		764 (86.1)
Other	4 (0.7)	0		4 (0.5)
Education, y	15.7 (3.6)	15.8 (3.6)	t = −0.31, 0.75	15.7 (3.6)
APOE Ɛ4 allele	106 (18.2)	121 (41.7)	χ2 = 55.8, <0.001	227 (26.0)
APOE Ɛ2 allele	85 (14.6)	52 (17.9)	χ2 = 1.64, 0.20	137 (15.7)
Dementia	203 (34.4)	158 (53.4)	χ2 = 29.6, <0.001	361 (40.7)
Dementia at baseline	26 (4.4)	28 (9.5)		54 (6.1)
Dementia during follow-up	177 (29.9)	130 (43.9)		307 (34.6)
Years from baseline to incident dementia	6.6 (4.6)	5.9 (4.2)		6.3 (4.5)
Hypertension	416 (70.5)	198 (66.9)	χ2 = 1.21, 0.27	614 (69.3)
Diabetes mellitus	142 (24.0)	48 (16.2)	χ2 = 7.1, 0.008	190 (21.4)
History of smoking	200 (34.1)	89 (30.4)	χ2 = 1.25, 0.26	289 (32.9)
No. of vascular risk factors	1.0 (1.0)	1.0 (2.0)	K-W χ2 = 5.2, 0.022	1.0 (1.0)
Low-density lipoprotein cholesterol (mg/dL)c	98.3 (30.1)	101.1 (28.1)	t = −1.16, 0.25	99.3 (29.5)
High-density lipoprotein cholesterol (mg/dL)c	55.8 (16.1)	56.3 (17.0)	t = −0.42, 0.67	55.9 (16.4)
Triglyceride (mg/dL)c	145.2 (69.2)	143.2 (71.3)	t = 0.36, 0.72	144.5 (69.8)
Pathologic diagnosis of Alzheimer disease	324 (54.8)	244 (82.4)	χ2 = 65.3, <0.001	568 (64.0)
Square root of global Alzheimer disease pathology score	0.69 (0.39)	0.93 (0.37)	t = −8.7, <0.001	0.77 (0.40)
Presence of Lewy bodies	139 (23.5)	88 (29.8)	χ2 = 4.11, 0.043	227 (25.6)

Abbreviation: CAA = cerebral amyloid angiopathy.

^aContinuous variables are presented by mean (SD) and compared by the t test, categorical variables are presented by count (percent) and compared by the χ² test, and the number of vascular risk factors is presented by median (interquartile range) and compared by the Kruskal-Wallis test.

^bFor comparison of race/ethnicity, cells of groups other than non-Latino White were merged because of few individuals in the corresponding cells.

^cLow-density lipoprotein cholesterol was available in 678 participants; high-density lipoprotein cholesterol and triglyceride were available in 684 participants.

Associations of Proteins With CAA vs AD

Of the 10,030 quantified proteins, 80 proteins were related to CAA (eTable 2; eFigure 1). We excluded 1 protein, keratin type II microfibrillar component 7C (P15241), because its source organism was identified as sheep in the canonical UniProtKB human proteome database. Among the 79 proteins, 53 proteins were associated with more severe CAA and 26 proteins were associated with less severe CAA. Examining the proteins using GO analysis indicated that the proteins positively associated with CAA were functionally enriched in 20 pathways, with the strongest enrichment observed in the extracellular matrix, followed by heparin binding, locomotion, and cell differentiation (Figure 2). The proteins negatively correlated with CAA were enriched in only 3 pathways: endosomal sorting complexes required for transport, regulation of secretion by cell, and regulation of vesicle-mediated transport.

Figure 2. GO Analysis of Pathways Enriched by CAA-Associated Proteins.

In 2 separate GO analyses, we examined 53 (A) and 26 (B) proteins that were upregulated and downregulated in relation to CAA in protein-wide association analyses. The findings indicate that the strongest enrichments by the upregulated proteins were observed in the extracellular region, followed by collagen-containing extracellular matrix, locomotion, heparin-binding, and cell differentiation pathways. However, only 3 pathways were enriched by the downregulated proteins, including the endosomal sorting complexes required for transport, regulation of secretion by cell, and regulation of vesicle-mediated transport. CAA = cerebral amyloid angiopathy; GO = Gene Ontology.

For comparison, we examined the 10,030 proteins in relation to the global AD pathology score. We found 923 proteins related to AD pathology (eFigure 2), excluding keratin type II microfibrillar component 7C whose source organism was identified as sheep. A total of 338 proteins were positively associated with AD pathology, enriched in 38 pathways, and 585 proteins were negatively associated with AD pathology, enriched in 37 pathways (eFigure 3). The analyses given above indicated that some proteins were related to both CAA and AD, suggesting a possible confounding effect of AD in identifying CAA-related proteins.

Some previous studies reported an association between CAA and Lewy bodies.²⁷ In our study, neither the presence of Lewy bodies, detected by the pathologic examination, nor the level of alpha-synuclein, derived from the proteomic analysis, was associated with CAA (eTable 3).

Associations of the Proteins With CAA Controlled for AD

Because some proteins were associated with both CAA and AD, which are correlated pathologies,¹⁴ we repeated the protein-wide association analyses with CAA while controlling for the global AD pathology score to remove a possible confounding effect of AD. Of the 10,030 proteins, 12 proteins remained associated with CAA (eTable 4), and their higher values were positively associated with more severe CAA (Figure 3). The associations of the 12 proteins with CAA did not change when the models were further controlled for education or vascular risk factors (eTable 4, models 2–4). Replacing age at death with age at baseline (and time from baseline to death) or with age at dementia/age at the last visit (and time from dementia/last visit to death) for those with/without dementia to account for the diversity in age at enrollment or age at dementia diagnosis did not change the association of the 12 proteins with CAA (eTable 4, models 5–6). We also tested whether the associations of the 12 proteins with CAA varied by age by adding an interaction term between age at death and each of the 12 proteins in separate models. None of the interaction terms was significant (eTable 5), indicating no moderation effect by age.

Figure 3. Proteins Related to CAA After Controlling for AD Pathology.

In a protein-wide association analysis examining 10,030 proteins in relation to CAA controlling for the global AD pathology score, 12 proteins remained associated with CAA. All the models were also controlled for age at death and sex. p Values were adjusted by Bonferroni correction. The y-axes are the relative log₂(abundance) of the proteins. AD = Alzheimer disease; APCS = serum amyloid P-component; CAA = cerebral amyloid angiopathy; SFRP1 = secreted frizzled-related protein 1; HTRA1 = high-temperature requirement A serine peptidase 1; MDK = midkine; NTN1 = netrin-1; OLFML3 = olfactomedin-like protein 3; PTN = pleiotrophin; SMOC1 = secreted modular calcium-binding protein 1; TIMP3 = metalloproteinase inhibitor 3.

Almost all 12 proteins were correlated (eTable 6). Therefore, we examined the 12 proteins in a single model in relation to CAA to identify proteins that were driving the association. Using backward elimination to remove proteins with the highest p values (i.e., weakest association) in subsequent steps, we observed that secreted modular calcium-binding protein 1 (SMOC1), secreted frizzled-related protein 1 (SFRP1), APOE, and APOE4 remained in the model and were the proteins driving the association with CAA (Table 2). Replacing the backward elimination with Lasso or elastic net penalty procedures did not change the selection of the 4 proteins (eTable 7).

Table 2.

Associations of 12 Proteins With CAA When Examined Together in 1 Model

Proteins remained in the model after backward elimination
Uniprot ID	Gene	Protein name	Association with CAA
			OR (95% CI)	p Value
Q9H4F8	SMOC1	Secreted modular calcium-binding protein 1	2.63 (1.93–3.59)	<0.001
Q92765	FRZB	Secreted frizzled-related protein 1	2.76 (1.69–4.50)	<0.001
P02649	APOE	Apolipoprotein E	2.35 (1.43–3.87)	<0.001
APOE4	APOE4	Apolipoprotein E4	1.22 (1.09–1.36)	<0.001

Proteins removed from the model during backward elimination
Uniprot ID	Gene	Protein name	Step of removal	p Value
Q92743	HTRA1	High-temperature requirement A serine peptidase 1	1	0.97
Q9NRN5	OLFML3	Olfactomedin-like protein 3	2	0.83
P21246	PTN	Pleiotrophin	3	0.74
P21741	MDK	Midkine	4	0.68
O95631	NTN1	Netrin-1	5	0.75
P35625	TIMP3	Metalloproteinase inhibitor 3	6	0.64
P02743	APCS	Serum amyloid P-component	7	0.56
P06727	APOA4	Apolipoprotein A-IV	9a	0.095

Abbreviations: CAA = cerebral amyloid angiopathy; OR = odds ratio.

In a single ordinal logistic regression with backward elimination, 12 proteins we included as the model terms and CAA was the outcome. The model was controlled for age at death and sex.

^aDuring step 8, sex was removed from the model. Age at death remained in the model together with the 4 proteins in the top 4 rows.

We used the ordinal 4-level measure of CAA (none, mild, moderate, and severe) because of its familiarity as a small vessel disease severity scale. However, continuous measures provide more power. Therefore, in a sensitivity analysis, we replaced the 4-level measure of CAA with a continuous measure of CAA and repeated the analyses, which identified 5 proteins associated with the continuous measure of CAA (eTables 8 and 9). The 5 proteins included 3 previously identified using the ordinal 4-level CAA measure, SMOC1, frizzled-related protein, and APOE4, and did not explain CAA better than the 4 proteins identified using the ordinal 4-level measure of CAA, supporting use of the ordinal 4-level measure of CAA.

RNA Expression

We examined whether the associations observed between the proteins and CAA were also replicable at the RNA level. First, we examined the bulk tissue RNA expression levels of SMOC1, SFRP1, and APOE, available in 771 participants, and found that none were related to CAA (eTable 10). Next, we examined single nuclei RNA expression levels of the genes encoding the 3 proteins, available in 332 participants (eFigure 4). We did not observe an association of single nuclei RNA expression of the 3 genes with CAA (eTable 11). The analysis given above indicated that the associations between the 3 proteins and CAA were not because of increased expression of the genes at the mRNA level. We also examined the associations of single-nucleotide genetic variations (available in 804 participants) and DNA methylation levels (available in 666 participants) in the genetic regions of SMOC1 and SFRP1 with CAA. Genetic variations of neither SMOC1 (p = 0.78) nor SFRP1 (p = 0.18) were related to CAA. Although the DNA methylation level of SMOC1 (p = 0.049) was related to CAA and of SFRP1 had a borderline association with CAA (p = 0.059), the DNA methylation levels in neither of the CpG sites in the 2 genetic loci were related to CAA after Bonferroni adjustment.

APOE ε4 Allele, Proteins, and CAA

APOE ε4 allele is a genetic risk factor of CAA.²⁸ We examined whether the association between APOE ε4 allele and CAA was mediated by APOE4 protein or other proteins. In this study, APOE ε4 allele was a risk factor of CAA (odds ratio 3.46, 95% CI 2.59–4.61, p < 0.001) and was related to the 4 CAA-related proteins (eTable 12). We used separate mediation analyses to examine whether APOE4, APOE, SMOC1, and SFRP1 mediated the association between APOE ε4 allele and CAA. The mediation analysis indicated that SMOC1 mediated 46.1% (SE = 6.6%, p < 0.001), SFRP1 24.0% (SE = 5.9%, p < 0.001), and APOE protein only 14.3% (SE = 4.8%, p = 0.003) of the association between APOE e4 allele and CAA, but APOE4 protein did not mediate the association between APOE ε4 allele and CAA (mediation = 23.8%, SE = 32.5%, p = 0.46) although participants with APOE ε4 allele had higher levels of APOE4 protein (1.75 [SD = 0.91] vs −0.68 [SD = 0.49], p < 0.001), supporting the validity of APOE4 quantification.¹⁸ Therefore, APOE ε4 allele is a risk factor of CAA, but not through a direct effect of APOE4 protein.

Aβ, the Proteins, and CAA

Accumulation of Aβ in the small vessel walls is the main cause of CAA. We examined the associations of Aβ and the 4 major CAA-related proteins with CAA to identify whether the associations changed when both Aβ and the proteins were included in the same model, compared with when they were analyzed in separate models. We used 3 measures of Aβ (eTable 13): Aβ load, derived from immunohistochemical measurement of Aβ; total Aβ protein; and Aβ38 peptide, both derived from SRM proteomic analysis. When we examined Aβ load or total Aβ protein and SMOC1 in relation to CAA, the association of Aβ with CAA was attenuated, suggesting mediation or confounding by SMOC1. Such attenuations were not observed when we examined Aβ38 peptide or proteins other than SMOC1 (Table 3).

Table 3.

Associations of Aβ With CAA When Examined Without or With the 4 CAA-Related Proteins

Source of Aβ	Series	Model terms	Associations with CAA: estimate (SE), p value
			SMOC1	SFRP1	APOE	APOE4
Aβ load, a pathologic measure	(A) Aβ and the proteins are examined in separate models	Aβ load	0.830 (0.086), <0.001	0.818 (0.088), <0.001	0.830 (0.086), <0.001	0.849 (0.087), <0.001
		Protein	1.552 (0.135), <0.001	2.018 (0.224), <0.001	1.797 (0.220), <0.001	0.412 (0.052), <0.001
	(B) Each model includes Aβ and one of the proteins	Aβ load	0.246 (0.121), 0.042	0.669 (0.091), <0.001	0.741 (0.087), <0.001	0.725 (0.089), <0.001
		Protein	1.281 (0.188), <0.001	1.586 (0.225), <0.001	1.484 (0.222), <0.001	0.302 (0.053), <0.001
Total Aβ protein, a proteomic measure	(A) Aβ and the proteins are examined in separate models	Total Aβ protein	0.249 (0.023), <0.001	0.247 (0.024), <0.001	0.249 (0.023), <0.001	0.249 (0.024), <0.001
		Protein	1.499 (0.139), <0.001	2.033 (0.237), <0.001	1.810 (0.228), <0.001	0.394 (0.053), <0.001
	(B) Each model includes Aβ and one of the proteins	Total Aβ protein	0.142 (0.032), <0.001	0.212 (0.025), <0.001	0.218 (0.024), <0.001	0.221 (0.024), <0.001
		Protein	0.904 (0.192), <0.001	1.539 (0.239), <0.001	1.295 (0.231), <0.001	0.265 (0.055), <0.001
Aβ38 peptide, a proteomic measure	(A) Aβ and the proteins are examined in separate models	Aβ38	0.559 (0.061), <0.001	0.557 (0.061), <0.001	0.559 (0.061), <0.001	0.558 (0.061), <0.001
		Protein	1.498 (0.139), <0.001	2.031 (0.237), <0.001	1.805 (0.227), <0.001	0.393 (0.053), <0.001
	(B) Each model includes Aβ and one of the proteins	Aβ38	0.486 (0.062), <0.001	0.507 (0.063), <0.001	0.464 (0.064), <0.001	0.543 (0.061), <0.001
		Protein	1.363 (0.141), <0.001	1.857 (0.245), <0.001	1.293 (0.248), <0.001	0.380 (0.054), <0.001

Abbreviations: Aβ = β-amyloid; CAA = cerebral amyloid angiopathy; SFRP1 = secreted frizzled-related protein 1; SMOC1 = secreted modular calcium-binding protein 1.

In a series of ordinal logistic regression models, we examined the associations of 3 measures of Aβ (Aβ load, derived from immunohistochemical assessment of brain sections; total Aβ protein and Aβ38 peptide, both derived from proteomic analysis of 20 mg of the dorsolateral prefrontal cortex) with CAA without (series-A) and with (series-B) the 4 CAA-related proteins in the models. In each model CAA was the outcome, and model terms were one of the Aβ measures alone (series-A) or with (series-B) one of the 4 proteins listed in the right columns. All the models were controlled for age at death and sex. The analyses indicated that the associations of 2 of the Aβ measures (Aβ load and total Aβ proteins) with CAA were attenuated when SMOC1 was in the model.

To confirm the aforementioned findings, we examined the 3 Aβ measures and the 4 CAA-related proteins in a factor analysis. Two factors had eigenvalues above 1 and explained 62.9% of the total variance. The first factor included SMOC1, total Aβ protein and Aβ load, and APOE4 protein, and the second factor included Aβ38, SFRP1, and APOE protein (eTable 14).

Based on the findings above, we leveraged a SEM approach to compare the 2 pathways of Aβ accumulation, suggested by the aforementioned factor analysis, in relation to CAA, given SEM's ability to concurrently estimate both directional and nondirectional relationships among variables²⁹ (Figure 4). The first pathway, termed the short Aβ pathway, included a latent factor with Aβ38 peptide as an exogenous variable, representing short Aβ peptides with end position (C-terminal position) <40. The second pathway, termed the long Aβ pathway, included total Aβ protein and immunohistologically measured Aβ, and previous studies indicated that Aβ peptides with end position >40 (such as Aβ1–42) constitute most of the Aβ in the brain and are required for nucleation of Aβ and formation of Aβ plaques.^30,31

Figure 4. A Structural Equation Model Illustrating 2 Aβ Pathways Underlying CAA.

The short Aβ pathway indicate Aβ peptides with end position (C-terminal position) < 40, and the long Aβ pathway indicates Aβ peptides with end position greater than 40. ***p < 0.001; *p < 0.05. Aβ = β-amyloid; CAA = cerebral amyloid angiopathy; SMOC1 = secreted modular calcium-binding protein 1

The fit statistics of the SEM analysis indicated that the model fit the data well (RMSEA = 0.099 [90% CI 0.088–0.111], GFI = 0.939). The standardized coefficients (Figure 4) indicated that the effect size of short Aβ pathway on CAA was 2 times greater than the effect size of long Aβ pathway, suggesting a dominant role for short Aβ peptides in causing CAA. We observed no association between age at death and short Aβ pathway in contrast to the positive association between older age and long Aβ pathway.

Discussion

In a large autopsy study of approximately 900 brains, we quantified levels of more than 10,000 proteins in the cortical gray matter and identified 80 proteins associated with CAA, of which 12 proteins showed associations with CAA independently of AD pathology and 4 proteins remained associated with CAA when all 12 proteins were examined in 1 model. Using factor analysis and SEM, we highlighted 2 possible pathways associated with CAA: a short Aβ pathway and a long Aβ pathway. The short Aβ pathway was composed of short Aβ peptides, with an end before position 40, and SFRP1 and APOE proteins while the long Aβ pathway was composed of long Aβ peptides and SMOC1 and APOE4 proteins. The magnitude of the contribution of the short Aβ pathway was twice that of the long Aβ pathway, and older age was a risk factor for the long, but not the short, Aβ pathway. Further studies are required to determine whether modification of the proteins can prevent or slow down the building up of CAA in older adults.

CAA is common, with its moderate-to-severe stages observed in at least one-fourth of older adults.⁶ Moreover, it is not benign as it is associated with lacunar strokes,³² intracerebral hemorrhage,² dementia,³ epilepsy,³³ and transient focal neurologic episodes.⁵ Deposition of Aβ fibrils in the walls of small arteries and arterioles is the first step in the development of CAA,¹ but the mechanism underlying the Aβ deposition is unclear. Our study extended previous similar studies^34-37 by including a much larger sample size; examining more proteins; controlling for the confounding effect of AD pathology; and analyzing the genetic variants, epigenetic changes, and RNA levels of the genes encoding CAA-related proteins. In addition, we examined the CAA-related proteins together with 3 different measures of Aβ in relation to CAA, which identified 2 separate molecular pathways associated with CAA accumulation.

One identified molecular pathway associated with CAA included Aβ38 peptide that might represent short Aβ peptides C-terminally truncated before position 40 and not differentially measured in this study, such as Aβx-37 and Aβx-38. The other pathway included total Aβ protein and Aβ load that might represent long Aβ peptides C-terminally truncated at positions after 40, which were not directly measured in this study, such as Aβx-42, because these long Aβ peptides are the major Aβ peptides extracted from the brain^30,38 and are the major constituents of Aβ plaques,^30,38 immunohistochemically measured as Aβ load in this study. The SEM analysis indicated that the contribution of the short Aβ pathway including Aβ38 to CAA accumulation was twice that of the long Aβ pathway including total Aβ protein and Aβ load. Therefore, our findings are in line with previous studies that identified Aβ peptides such as Aβx-37, Aβx-38, and Aβx-40 to be enriched in CAA.^36,39 In addition, our findings highlight a role for the long Aβ peptides that are also deposited in the Aβ plaques^30,31,36,39 and suggest the molecular mechanisms underlying the correlation between AD and CAA.

SMOC1 together with APOE4 and other correlated proteins were involved in the association of long Aβ peptides with CAA. SMOC1 is also one of the matrix proteins associated with AD,⁹ which is expected considering the role of long Aβ peptides in plaque formation. SMOC1 is a matrix protein that is located in the basement membrane of almost all tissues including blood vessels and the brain⁴⁰ and colocalizes with Aβ fibrils in CAA and Aβ plaques.³⁷ A recent study reported co-immunoprecipitation between SMOC1 and Aβ fibrils including monomer and dimers of Aβ.⁴¹ Therefore, we hypothesize that SMOC1 is the anchor for the precipitation of Aβ in the vessel walls and the development of CAA, a process to which APOE4 and other proteins also contribute.

SFRP1 is a secreted protein with multiple functions, including modulating Wnt pathway⁴² and inhibition of α-secretase and consequently overproduction of Aβ.⁴³ It is less probable in our analytic sample that higher SFRP1 is linked to CAA by inhibiting ADAM10 and more production of Aβ, because this pathway indicates a mediating effect for Aβ in the association between SFRP1 and CAA, not supported by our data (Table 3). A more plausible pathway is through Wnt–β-catenin signaling where SFRP1 together with APOE⁴⁴ contributes to inhibiting Wnt-β-catenin pathway triggered by short Aβ peptides,⁴⁵ which results in neuroinflammation modulation, blood-brain barrier dysfunction, and CAA.⁴⁵ Further studies should identify genetic variants, RNA levels, and other omics related to SMOC1 and SFRP1 to clarify risk-susceptible loci and molecular pathways from genes to the proteins and to CAA.

Of interest, older age at death was a risk factor for the long, but not the short, Aβ pathway, consistent with the association of age and AD⁴⁶ or Aβ plaques that are composed of the long Aβ peptides. In addition, the common long Aβ pathway may underlie the observed lower frequency of AD⁴⁶ and CAA^6,47 in the oldest old autopsies. The short Aβ pathway was not related to age possibly because it is more driven by other pathways rather than by aging. Short Aβ peptides are produced not only by gamma secretase but also during removal of the long Aβ deposits in the brain by peptidases such as metalloproteinases or neprilysin.⁴⁸ The latter is also supported by the correlation between the short and long Aβ pathways in the current SEM method (Figure 4).

Our findings suggest promising targets for the prevention or alleviation of CAA. Furthermore, they potentially can be leveraged to lower the risk of ARIA among persons receiving immunotherapies for AD. The SEM approach suggests that anti-Aβ treatments that target Aβ plaques may also cause ARIA by targeting Aβ that has been already present as CAA in the vessel walls through the long Aβ pathway. To mitigate the risk of ARIA, the SEM method suggests preconditioning the anti-Aβ recipients with compounds that make inactive complexes with SMOC1 or prevent interaction between SFRP1 and short Aβ peptides.

Several limitations must be noted. The findings are generated from an observational cross-sectional study, which limits causal inferences. The participants were mostly White with more years of education than the average American. Among 25⁴⁹ or more proteoforms of Aβ, we could quantify only Aβ38 with an unknown N-terminal position. Future studies with more advanced proteomic analysis are required to quantify all Aβ proteoforms and elucidate their role in CAA. The study did not include the imaging markers of CAA such as microbleeds or pathologic indices of CAA complications such as microhemorrhages, superficial siderosis, or white matter alterations. The study strengths included a harmonized structural postmortem pathologic assessment, quantifying more than 10,000 proteins in the frontal cortex of approximately 900 community-dwelling participants and availability of genetic variants and other omics data from the same participants and the same brain region, enabling cross-examination of the protein findings using other omics.

Glossary

AA Core

Rush Alzheimer's Disease Center African American Clinical Core

Aβ

β-amyloid

AD

Alzheimer disease

ARIA

amyloid-related imaging abnormality

CAA

cerebral amyloid angiopathy

DLPFC

dorsolateral prefrontal cortex

GFI

goodness-of-fit index

GO

Gene Ontology

IQR

interquartile range

Latino Core

Rush Alzheimer's Disease Center Latino Core

MAP

Rush Memory and Aging Project

MARS

Minority Aging Research Study

RMSEA

root mean square error of the approximation

ROS

Religious Orders Study

SEM

structural equation modeling

SFRP1

secreted frizzled-related protein 1

SMOC1

secreted modular calcium-binding protein 1

SRM

selection reaction monitoring

Author Contributions

S. Oveisgharan: drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data. L. Yu: drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data. J. Yang: drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data. S. Agrawal: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. R. Vialle: drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data. K. de Paiva Lopes: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. S. Tasaki: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. Y. Wang: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. V.A. Petyuk: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. T.L. Young-Pearse: drafting/revision of the manuscript for content, including medical writing for content. V. Menon: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. J. Zhao: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. G. Miao: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. L.L. Barnes: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. J.A. Schneider: major role in the acquisition of data. P.L. De Jager: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. N. Seyfried: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data. D.A. Bennett: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design; analysis or interpretation of data.

Study Funding

The study is supported by NIH grants: P30AG10161, P30AG72975, R01AG15819, R01AG17917, R01AG022018, U01AG46152, U01AG61356, and RF1NS139948.

Disclosure

The authors report no relevant disclosures. Go to Neurology.org/N for full disclosures.