Cerebrospinal fluid sphingomyelins in Alzheimer’s disease, neurodegeneration, and neuroinflammation

Abstract

BACKGROUND:

Sphingomyelin (SM) levels have been associated with Alzheimer’s disease (AD), but the association direction has been inconsistent and research on cerebrospinal fluid (CSF) SMs has been limited by sample size, breadth of SMs examined, and diversity of biomarkers available.

OBJECTIVE:

Here, we seek to build on our understanding of the role of SM metabolites in AD by studying a broad range of CSF SMs and biomarkers of AD, neurodegeneration, and neuroinflammation.

METHODS:

Leveraging two longitudinal AD cohorts with metabolome-wide CSF metabolomics data (n=502), we analyzed the relationship between the levels of 12 CSF SMs, and AD diagnosis and biomarkers of pathology, neurodegeneration, and neuroinflammation using logistic, linear, and linear mixed effects models.

RESULTS:

No SMs were significantly associated with AD diagnosis, mild cognitive impairment, or amyloid biomarkers. Phosphorylated tau, neurofilament light, α-synuclein, neurogranin, soluble triggering receptor expressed on myeloid cells 2, and chitinase-3-like-protein 1 were each significantly, positively associated with at least 5 of the SMs.

DISCUSSION:

The associations between SMs and biomarkers of neurodegeneration and neuroinflammation, but not biomarkers of amyloid or diagnosis of AD, point to SMs as potential biomarkers for neurodegeneration and neuroinflammation that may not be AD-specific.

Background

To discover new therapeutic targets for Alzheimer’s disease (AD), we must better understand the changes caused by the disease. One approach to better understanding these changes in AD is metabolomics, a way of studying the byproducts of the body’s metabolic processes. Metabolomics grants a window into the metabolic state of the body [1] and has proven useful in studying a variety of diseases, including identifying a mechanism of insulin resistance for type 2 diabetes, developing precision medicine approaches to treating cancer, and identifying altered metabolites and their mechanism of change in central nervous system-related disorders such as simian immunodeficiency virus [2–5]. Metabolomics has also been a valuable tool for studying AD, as it has helped to identify new biomarkers and mechanisms for the disease [6].

One of the findings from AD metabolomics studies has been the association between the sphingolipid metabolic pathway and AD [7]. Sphingolipids are a family of membrane lipids that participate in diverse and fundamental cellular processes, such as cell division, differentiation, and death [8]. In mammals, sphingomyelins (SMs) are the most abundant molecules of the sphingolipid metabolic pathway (Supplemental Figure 1) [8]. Despite the identification of SMs as associated with AD, there is disagreement as to how they are associated with the disease: some studies show that SMs decrease in tissue across brain regions and in plasma during progression to AD, while others suggest that they increase [9,10]. Previous research in cerebrospinal fluid (CSF) SMs has been limited by sample size, breadth of SMs examined, and diversity of biomarkers available for AD, neurodegeneration, and neuroinflammation, which may have contributed to the lack of clarity in the role of SMs.

Here, we build on our understanding of the role of SM metabolites in AD. Leveraging two large, longitudinal cohorts with metabolome-wide CSF metabolomics, robust cognitive diagnoses, and a diversity of CSF biomarker and brain imaging measures, we analyzed the relationship between SMs, AD diagnosis, and biomarkers of pathology, and markers of neurodegeneration and neuroinflammation. The results shed light on the role of SMs in neurodegeneration and neuroinflammation.

Materials and Methods

Study Cohorts

Data were included from the Wisconsin Registry for Alzheimer’s Prevention (WRAP) and the Wisconsin Alzheimer’s Disease Research Center (ADRC) cohorts [11,12]. These longitudinal studies of preclinical and clinical AD in middle to older aged adults include CSF metabolomics, clinical diagnosis, neuroimaging, and CSF biomarkers of AD, neurodegeneration, and neuroinflammation. All participants included in the current research had at least one lumbar puncture (LP); the CSF samples for WRAP and the Wisconsin ADRC were collected and analyzed following the same protocols. Most participants (357 individuals) had only one visit, 65 individuals had two visits, 49 individuals had three visits, and 23 individuals had four visits. There was a mean of 1.2 visits (maximum = 4 visits) in Wisconsin ADRC and a mean of 2.2 visits (maximum = 4 visits) in the WRAP cohort, with a median time between visits for an individual of 2.03 years. A committee of dementia specialists determined diagnosis of dementia-AD (n = 44 individuals), mild cognitive impairment (MCI) (n = 40 individuals), or cognitively unimpaired (CU) (n = 409 individuals) (Table 1). This study was approved by the University of Wisconsin Health Sciences Institutional Review Board as part of the Generations of WRAP (GROW) study. Participants in the WADRC and WRAP studies provided written informed consent.

Table 1.

Baseline characteristics of individuals.

Characteristic	Diagnosis data set (cross-sectional) (n = 493 individuals)	Biomarker data set individuals (n = 494 individuals)	Biomarker data set visits (longitudinal) (n = 726 visits)	Overlap (n = 485 individuals)
Diagnosis, n (%)
Dementia-AD	44 (8.9%)	43 (8.7%)	44 (6.1%)	43 (8.9%)
MCI	40 (8.1%)	42 (8.5%)	45 (6.2%)	40 (8.2%)
CU	409 (82.9%)	409 (82.8%)	637 (87.7%)	402 (82.9%)
Primary Race, n (%)
White	472 (95.7%)	473 (95.7%)	693 (95.5%)	464 (95.7%)
Black or African American	16 (3.3%)	16 (3.2%)	22 (3.03%)	16 (3.3%)
American Indian or Alaska Native	3 (0.6%)	3 (0.6%)	4 (0.6%)	3 (0.6%)
Asian	1 (0.2%)	1 (0.2%)	3 (0.4%)	1 (0.2%)
Other	1 (0.2%)	1 (0.2%)	4 (0.6%)	1 (0.2%)
Sex, n (%)
Male	191 (38.7%)	191 (38.7%)	273 (39.4%)	189 (39.0%)
Female	302 (61.3%)	303 (61.3%)	453 (65.4%)	296 (61.0%)
CSF amyloid and tau status, n (%)
A−T−	319 (64.7%)	324 (65.6%)	500 (68.9%)	317 (65.4%)
A+T−	61 (12.4%)	59 (11.9%)	85 (11.7%)	59 (12.2%)
A−T+	20 (4.1%)	20 (4.0%)	28 (3.8%)	19 (3.9%)
A+T+	85 (17.2%)	84 (17.0%)	102 (14.0%)	83 (17.1%)
Not measured	8 (1.6%)	7 (1.4%)	11 (1.5%)	7 (1.4%)
Age in years mean, (SD)	64.1 (8.9)	63.1 (8.9)	63.4 (8.3)	64.0 (9.0)
PiB Global DVR, n, (SD)	179 (9.4)	N/A	N/A	179 (9.4)
PiB Positive, n, (%)	50 (27.9%)	N/A	N/A	50 (27.9%)
Number of visits
1	493	357	N/A	N/A
2	0	65	N/A	N/A
3	0	49	N/A	N/A
4	0	23	N/A	N/A

For demographics information in the sex- and amyloid-stratified groups, see Supplemental Tables 1 and 2. For data on missingness by biomarker and metabolite, see Supplemental Table 4. All individuals with PiB Global DVR measurements were present in the Diagnosis data set. For additional information on participant cardiovascular factors, see Supplemental Table 3

Abbreviations: AD, Alzheimer’s disease; MCI, mild cognitive impairment; CU, cognitively unimpaired; CSF, cerebrospinal fluid; PiB, Pittsburgh Compound B; DVR, distribution volume ratio; SD, standard deviation.

CSF samples, biomarkers, and metabolomics

The process through which CSF samples were acquired and biomarker concentrations were measured has been previously described [13]. Briefly, CSF samples were collected in the morning after fasting. Within 30 minutes of collection, samples were mixed, centrifuged, aliquoted, and then stored at −80°C. All CSF sample biomarker assays were performed at the Clinical Neurochemistry Laboratory, University of Gothenburg from March 2019 to January 2020. All biomarker data were taken from the Roche NeuroToolKit (NTK; Roche Diagnostics International Ltd, Rotkreuz, Switzerland), a panel of exploratory prototype assays designed to robustly evaluate biomarkers associated with key pathologic events characteristic of AD and other neurological disorders, as previously described [13].

For metabolomic analyses, CSF samples were shipped overnight to Metabolon, Inc. (Durham, NC), where samples were also kept frozen at −80°C until analysis [14]. The untargeted metabolomics analysis was performed using Ultrahigh Performance Liquid Chromatography-Tandem Mass Spectrometry (UPLC-MS/MS). Chemical properties, metabolite identifiers, and pathway information were provided for each metabolite. All metabolite data underwent a quality control (QC) process. Of the 412 metabolites in the initial sample, 13 were removed for being missing in ≥ 50% of the samples. Nine metabolites were removed for low variance (interquartile range = 0). Of the 1,172 CSF samples, one was removed for missing ≥ 40% of the metabolite values. We also removed 220 samples obtained as part of a clinical trial (CT: 00939822). Metabolite values were log₁₀ transformed to avoid skewness. After these QC steps, data were available on 390 metabolites from 951 CSF samples taken from 502 individuals. Of the 390 metabolites, 12 metabolites in the sphingolipid metabolic pathway were used in the analyses described. Of the 12 metabolites included in these analyses, 9 were designated by metabolon as tier 2 compounds, meaning that these compounds have a structure that has been confirmed by literature review, but they are not necessarily confirmed by a reference standard. These 9 metabolites were: behenoyl SM (d18:1/22:0), palmitoyl dihydro-SM (d18:0/16:0), SM (d18:1/14:0, d16:1/16:0), SM (d18:1/20:0, d16:1/22:0), SM (d18:1/20:1, d18:2/20:0), SM (d18:1/22:1, d18:2/22:0, d16:1/24:1), SM (d18:1/24:1, d18:2/24:0), SM (d18:2/16:0, d18:1/16:1), SM (d18:2/24:1, d18:1/24:2).

Neuroimaging

Detailed methods for radiotracer synthesis and positron emission tomography (PET) and magnetic resonance imaging (MRI) data acquisition, processing, and quantification have been previously described [15]. Briefly, anatomical MRI (T1-w and T2-w) underwent multispectral unified tissue class segmentation (SPM12) [16]. Regions of interest (ROIs) for PET analysis were defined by applying the inverse deformation field defined during tissue segmentation to the MNI152-space Automated Anatomical Labeling atlas [17] and restricting the subject-space ROIs to voxels with gray matter probabilities greater than 0.3. Reconstructed dynamic Pittsburgh compound B (PiB) PET data acquired from 0–70 minutes post nominal 555 MBq [¹¹C] PiB injection on a Siemens EXACT HR+ or Siemens Biograph Horizon PET/CT were isotopically smoothed, interframe realigned, dynamically denoised, and registered to T1-weighted MRI [15]. Amyloid burden was assessed by averaging distribution volume ratio (DVR) estimates across eight bilateral regions (Logan graphical analysis, cerebellum gray matter reference region, k2’=0.149 min⁻¹; ROIs included angular gyrus, anterior and posterior cingulate, medial orbital-frontal gyrus, precuneus, supramarginal gyrus, and middle and superior temporal gyri) [18]. The resulting measurement is referred to as the PiB Global DVR.

Data Integration

CSF data were matched to diagnosis data from the nearest clinic visit. To remove potential correlation between genetically related participants, only the oldest individual from each family group was selected (n = 32 individuals removed). Data were divided into two overlapping data sets for analysis (see Supplemental Figure 2 for a visual representation). First was the cross-sectional “Diagnosis” data set, which was focused on clinical diagnoses, comprising the oldest visit per participant (n = 493), diagnosis at that visit, and matched CSF metabolomics and PiB data for that visit. Logistic regressions were used to identify associations between diagnosis and SMs, and linear regressions were used for associations between PiB Global DVR and SMs. SM-PiB Global DVR analyses were cross-sectional because there were not enough PiB Global DVR measurements per individual to use longitudinal data. However, results for these analyses with PiB as the outcome were presented with the results from the AD biomarkers (data set described below) to better group the outcomes conceptually.

The second data set was the longitudinal “Biomarker” data set, which focused on CSF biomarkers, comprising all available visits with CSF biomarker data and their corresponding metabolomics data (n = 494 individuals and n = 726 samples). This data set included all available LP visits for each individual to maximize sample size (Table 1). The Biomarker data set was used to identify associations between CSF biomarkers and metabolites. AD biomarkers included CSF Aβ₄₂/Aβ₄₀, CSF p-tau₁₈₁/Aβ₄₂, and CSF p-tau₁₈₁ [19–21]. The neurodegeneration biomarkers used were neurogranin, neurofilament light (NfL), and alpha-synuclein (α-synuclein) [22–24]. We also included biomarkers for neuroinflammation: interleukin-6 (IL6), chitinase-3-like protein 1 (YKL40), and soluble triggering receptor found on myeloid cells 2 (sTREM2) [25–27]. All Biomarker data were checked for skewness, and NfL, p-tau₁₈₁, p-tau₁₈₁/Aβ₄₂, and IL6 were log₁₀-transformed for skewness ≥ 2 [28].

Data Analysis

All analyses were performed using R (version 4.0.2) and the “Tidyverse” packages (version 1.3.0) [29]. Logistic and linear regressions were performed using the glm and lm functions, respectively, from the “stats” package (version 3.6.2) [30]. Linear mixed effects models (LMMs) using the biomarker data set were performed using the lmer function from the “lme4” and “lmerTest” packages [31,32]. All P-values from regression models were subjected to a Bonferroni adjusted threshold of P ≤ 3.47×10⁻⁴ to determine significance (α = 0.05 / 12 metabolites / 12 primary outcomes).

Pairwise correlations and principal components analysis

To assess the extent to which SMs represented the same underlying signal, a principal components analysis (PCA) was performed on the metabolites and each of the diagnoses and biomarkers were regressed on the first two principal components (PCs) included simultaneously in the same model, with age and sex as covariates, to determine whether the PCs were more strongly associated with any of the outcomes than any of the metabolites individually. Moreover, pairwise correlations were calculated between all 72 pairs of SMs to better understand whether they represent the same underlying signal.

SM association with dementia-AD and MCI diagnoses

In the Diagnosis data set, diagnoses of dementia-AD and MCI (both relative to cognitively unimpaired [CU]) were separately regressed on each metabolite, controlling for sex and age as added covariates in the model. Pseudo R² values to evaluate model fit were calculated for each regression using the R2 function from the “semEff” package (version 0.4.0) [33].

SM association with biomarkers of AD, neurodegeneration, and neuroinflammation

For PiB Global DVR, cross-sectional associations with SMs were analyzed using the Diagnosis data set and a linear regression model adjusting for sex and age as additional covariates. Adjusted R² values were calculated for the model, to assess model fit.

LMMs were used to determine the association between SMs and biomarkers for which longitudinal data were available (Aβ_42/40, p-tau₁₈₁/Aβ₄₂, p-tau₄₂, neurogranin, NfL, α-synuclein, IL6, YKL40, and sTREM2) in the Biomarker data set. Models were adjusted for age, sex, and a random intercept for the participant ID to account for multiple observations per individual. Marginal R² values for these models were calculated using the r.squaredGLMM function from the “MuMIn” package (version 1.43.17) to assess model fit [34].

Independent signals

To determine whether models including the most strongly associated SM from our main analyses as a predictor were significantly improved by adding any other SMs, we repeated the main analyses with the top SM (stearoyl SM) as the main predictor, adding each of the other metabolites as an additional predictor.

We used an analysis of variance (ANOVA) test to assess whether the model with stearoyl SM was significantly different from the model that included both stearoyl SM and an additional SM. We used the anova function from the “stats” package (version 4.0.2) to compare each pair of nested models [30]. To ensure equal sample sizes between the two groups, samples missing any of the values necessary for either regression were dropped. We compared the R² values calculated for each of the regressions before and after removing the necessary samples to perform this analysis to ensure that they were not drastically altered by these removed samples.

Sensitivity Analyses

APOE ε2/ε3/ε4 genotype was determined using competitive allele-specific PCR-based KASP genotyping for rs429358 and rs7412 [14]. Because APOE ε4 count is strongly associated with AD and thus may feasibly influence our results, we performed a sensitivity analysis in which we repeated the main analyses but also controlled for the number of C alleles (0, 1, or 2) at the rs429358 SNP, which effectively quantifies the number of ε4 alleles [35]. Only participants whose self-reported race was “white” were retained for this analysis because of previously reported heterogeneity in effect of the APOE ε4 allele by race [36]. We added APOE ε4 count as a covariate for these analyses.

Some studies of SMs in plasma have suggested that the direction of association of SMs and AD differ based on sex [37–39]. To determine whether there are sex-specific differences in associations with SMs in CSF, the SM-diagnosis and SM-biomarker regression analyses were repeated with interaction effects modeling, where the original regression models were rerun with the addition of a term for SM*female. A second sensitivity analysis was also conducted with stratification by sex (see Supplemental Table 1 for sex-stratified sample characteristics).

Similarly, to understand whether associations between SMs and biomarkers of AD, neurodegeneration, and neuroinflammation are different with amyloid and tau positivity, we performed the SM-diagnosis and SM-biomarker regressions with additional terms for A+T+ (binary indicator), A+T− (binary indicator), SM*A+T+, and SM*A+T−. Using thresholds that were previously defined in the WRAP study, amyloid-positive individuals were defined as those with Aβ_42/40 values below 0.046 pg/mL, and tau-positive individuals were defined as those with p-tau₁₈₁ values above 24.8 pg/mL [13]. We additionally then conducted an analysis stratified by amyloid and tau status, in which the main SM-biomarker regressions were performed for all individuals who were amyloid-positive and tau-positive (A+T+), amyloid-positive and tau-negative (A+T−), and amyloid-negative and tau-negative (A−T−) separately (see Supplemental Table 2 for amyloid and tau status-stratified sample characteristics) [40].

We also performed a sensitivity analysis to determine whether our results were affected by cardiovascular factors, including systolic blood pressure, waist-hip ratio, and body mass index, by rerunning the original models with these factors as covariates (see Supplemental Table 3 for summary statistics of cardiovascular factors in both cohorts).

Each SM-diagnosis and SM-biomarker regression was repeated as in the main set of regressions with the noted change for each sensitivity analysis. Regression results were subjected to the same Bonferroni-corrected significance threshold used in the main analyses (P < 3.47×10⁻⁴). P-values and effect sizes from the sensitivity analyses were compared to the results of the main set of analyses.

Results

In both data sets, most participants were CU, white, female, and amyloid- and tau-negative as measured by LP (Table 1). In the Diagnosis data set, the mean baseline age was 64.1 (SD = 8.9). The mean age of individuals with dementia-AD was 72.3 (SD = 8.3) years, the mean age of individuals with MCI was 73.7 (SD = 8.3) years, and mean age of individuals with a diagnosis of CU was 62.5 (SD = 7.9) years. In the Biomarker data set, the mean age of all visits was 63.1 (SD = 8.9) years. The mean age of a participant visit with diagnosis of: dementia-AD was 72.2 (SD = 8.4) years, MCI was 73.2 (SD = 8.3) years, and CU was 62.1 (SD = 7.4) years. Metabolite and biomarker missingness is described in Supplemental Table 4.

Pairwise correlations and PCA

Pairwise correlations between metabolites showed that the metabolites were all closely correlated (Figure 1). No negative correlations were observed; correlation coefficients ranged from 0.46 to 0.89 (mean = 0.74); behenoyl SM (d18:1/22:0) and stearoyl SM (d18:1/18:0) were the least correlated of all metabolite pairs (Supplemental Table 5).

Figure 1. Pairwise correlations of SMs.

A heatplot showing the correlation between each metabolite pair. Darker colors indicate stronger correlations, while lighter colors indicate weaker correlations. The lowest correlation coefficient was 0.47 and the highest was 0.89.

When we performed a PCA, we observed that all of the SMs contributed to the first PC, which explained 74.9% of variance in SM levels, with similar weights. However, the second PC, which explained 9.6% of variance in SM levels, was mostly contributed to by stearoyl SM and behenoyl SM (Supplemental Table 6a, Supplemental Figure 3). When the diagnoses and biomarkers analyzed in the main set of analyses were regressed on the first two PCs, accounting for 84.5% of variance in SM levels, we observed significant associations with the same outcomes as regressions which only included stearoyl SM (d18:1/18:0). The significant associations between the outcomes and PCs had the same direction of effect when the sign of PC loadings was considered, but smaller effect sizes and larger P-values than analyses that regressed stearoyl SM (d18:1/18:0) on the same outcomes (Supplemental Table 6b, Supplemental Figure 4). For the sake of interpretability, we chose to focus the narrative of this paper on stearoyl SM, which was the most consistently and strongly associated metabolite across the outcomes.

SM associations with dementia-AD

Of the SMs analyzed, none showed a significant association with either dementia-AD or MCI diagnosis relative to CU controls (Supplemental Table 7).

SM associations with biomarkers of AD, neurodegeneration, and neuroinflammation

SM associations with AD biomarkers (PiB Global DVR, CSF biomarkers of Aβ_42/40, p-tau₁₈₁/Aβ₄₂, and p-tau₁₈₁) were substantially different between measures of amyloid and tau. None of the 12 metabolites were significantly associated with any amyloid-related biomarkers (Aβ_42/40, PiB Global DVR, and p-tau₁₈₁/Aβ₄₂) after Bonferroni correction (Table 2). Scatterplots for these outcomes plotted against metabolite levels showed large clusters of data points with no clear pattern (Figure 2B). In contrast, 12 SMs had positive, nominally significant associations with CSF p-tau₁₈₁ (P < 0.05), and 6 associations remained significant after Bonferroni correction. Scatterplots of stearoyl SM (d18:1/18:0), the most significantly associated metabolite, and p-tau₁₈₁ reflected this positive association (Figure 2B).

Table 2.

Associations between SMs and AD biomarkers.

	PiB Global DVR			Aβ42/40			p-tau181/Aβ42 (pg/mL)			p-tau181 (pg/mL)
Metabolite	β	P	R2	β	P	R2	β	P	R2	β	P	R2
behenoyl SM (d18:1/22:0)	0.02	0.85	0.09	−3.54E-04	0.86	0.13	0.02	0.54	0.16	0.02	0.07	0.13
palmitoyl dihydro-SM (d18:0/16:0)	0.07	0.60	0.09	−3.00E-04	0.93	0.13	0.01	0.84	0.15	0.04	0.04	0.14
palmitoyl SM (d18:1/16:0)	−0.19	0.23	0.10	−1.87E-03	0.68	0.14	0.03	0.64	0.16	0.17	5.59E-10	0.16
SM (d18:1/14:0, d16:1/16:0)	−0.05	0.70	0.09	−1.41E-03	0.68	0.14	0.03	0.58	0.16	0.09	8.09E-07	0.14
SM (d18:1/18:1, d18:2/18:0)	−0.15	0.38	0.10	−3.22E-03	0.50	0.14	0.07	0.30	0.16	0.22	5.10E-13	0.17
SM (d18:1/20:0, d16:1/22:0)	0.01	0.93	0.09	−2.60E-03	0.38	0.14	0.05	0.26	0.16	0.06	3.17E-04	0.14
SM (d18:1/20:1, d18:2/20:0)	0.08	0.56	0.09	−2.97E-03	0.35	0.13	0.06	0.16	0.16	0.06	3.50E-04	0.14
SM (d18:1/22:1, d18:2/22:0, d16:1/24:1)	0.15	0.19	0.10	−3.62E-03	0.13	0.14	0.05	0.18	0.17	0.03	0.01	0.13
SM (d18:1/24:1, d18:2/24:0)	0.05	0.69	0.09	−6.82E-04	0.79	0.14	0.01	0.73	0.16	0.03	0.04	0.13
SM (d18:2/16:0, d18:1/16:1)	−0.08	0.57	0.10	−4.76E-03	0.17	0.14	0.09	0.07	0.17	0.09	1.93E-06	0.14
SM (d18:2/24:1, d18:1/24:2)	0.06	0.59	0.09	−4.02E-03	0.13	0.14	0.05	0.15	0.16	0.04	0.01	0.13
stearoyl SM (d18:1/18:0)	−0.22	0.18	0.10	−1.04E-03	0.84	0.14	0.05	0.50	0.16	0.37	1.83E-23	0.23

SM: sphingomyelin; PiB Global DVR: positron emission tomography Pittsburgh compound B; Aβ_42/40: amyloid-beta 42/40 ratio; p-tau₁₈₁/Aβ₄₂: phosphorylated-tau₁₈₁ to amyloid-beta 42 ratio; p-tau₁₈₁: phosphorylated-tau; β: change in the outcome with a one unit increase of the standardized metabolite level; P: uncorrected P-value of the corresponding regression; R²: adjusted R² for PiB Global DVR, marginal R² for CSF biomarkers; bolded values indicate significant results at threshold p<3.47×10⁻⁴.

Figure 2. Associations between biomarkers, diagnosis, and stearoyl SM (d18:1/18:0).

Results from stearoyl SM (d18:1/18:0) are displayed because this SM was consistently the most significantly associated SM and all other significant associations with SMs were in the same direction. A) Boxplots showing stearoyl SM (d18:1/18:0) levels for individuals with CU, MCI, and dementia-AD diagnoses. A small but statistically insignificant increase in metabolite level can be seen moving from the CU to AD group. None of the SM-diagnosis regressions were statistically significant. B) Scatterplots with each of the four AD-specific outcomes (x-axis) and stearoyl SM (d18:1/18:0) (y-axis). Stearoyl SM was not significantly associated with the measures of amyloid, but was significantly, positively associated with p-tau₁₈₁ after Bonferroni correction (Table 2). C) Scatterplots of each of the three neurodegeneration biomarkers (x-axis) plotted against stearoyl SM (d18:1/18:0) (y-axis). Stearoyl SM was significantly, positively associated with NfL, neurogranin, and α-synuclein after Bonferroni correction (Table 3). D) Scatterplots of each of the three neuroinflammation biomarkers (x-axis) plotted against stearoyl SM (d18:1/18:0) (y-axis). Stearoyl SM was significantly, positively associated with YKL40 and sTREM2 and nominally, negatively associated with IL6 after Bonferroni correction (Table 4). A-D: The units of stearoyl SM (d18:1/18:0) as well as p-tau₁₈₁, p-tau₁₈₁/Aβ₄₂, NfL, and IL6 are standardized by log₁₀-transformation as laid out in the methods section of this paper. B-D: Longitudinal data were used to perform the regressions shown. Best fit lines constructed using linear regression models of form outcome ~ stearoyl SM with 95% confidence intervals are drawn onto each of the scatterplots.

CSF levels of SMs were clearly associated with all CSF biomarkers of neurodegeneration. For all three biomarkers (neurogranin, NfL, α-synuclein), at least 11 of the 12 SMs were nominally and positively associated with each biomarker (Table 3). This positive association was visible in scatterplots of each outcome plotted against stearoyl SM, where there was a clear association and a consistent, positive trend (Figure 2C).

Table 3.

Associations between SMs and neurodegeneration biomarkers.

	NfL (pg/mL)			α-synuclein (pg/mL)			Neurogranin (pg/mL)
Metabolite	β	P	R2	β	P	R2	β	P	R2
behenoyl SM (d18:1/22:0)	0.06	1.23E-03	0.46	31.79	8.96E-04	0.09	46.86	0.08	0.04
palmitoyl dihydro-SM (d18:0/16:0)	0.15	2.14E-06	0.46	87.32	1.16E-07	0.11	148.66	0.01	0.05
palmitoyl SM (d18:1/16:0)	0.37	1.57E-18	0.52	254.75	1.72E-35	0.26	608.66	1.34E-15	0.10
SM (d18:1/14:0, d16:1/16:0)	0.16	8.83E-08	0.47	142.70	4.19E-18	0.16	336.80	1.74E-09	0.07
SM (d18:1/18:1, d18:2/18:0)	0.37	3.99E-17	0.51	247.38	3.30E-32	0.25	626.89	2.13E-15	0.10
SM (d18:1/20:0, d16:1/22:0)	0.17	2.34E-09	0.48	103.61	1.02E-12	0.12	207.71	8.00E-06	0.05
SM (d18:1/20:1, d18:2/20:0)	0.12	6.43E-05	0.45	64.18	3.29E-05	0.10	179.55	7.21E-04	0.05
SM (d18:1/22:1, d18:2/22:0, d16:1/24:1)	0.10	5.20E-06	0.46	50.48	1.34E-05	0.09	110.05	3.23E-03	0.05
SM (d18:1/24:1, d18:2/24:0)	0.09	1.30E-04	0.47	49.82	3.79E-05	0.09	81.38	0.04	0.04
SM (d18:2/16:0, d18:1/16:1)	0.22	5.16E-12	0.49	146.47	1.36E-19	0.17	337.93	1.39E-09	0.07
SM (d18:2/24:1, d18:1/24:2)	0.14	1.02E-08	0.48	75.29	2.36E-09	0.10	111.55	0.01	0.04
stearoyl SM (d18:1/18:0)	0.45	1.94E-21	0.53	325.23	1.25E-51	0.36	994.55	2.82E-28	0.18

SM: sphingomyelin; NfL: neurofilament light; β: change in the outcome with a one unit increase of the standardized metabolite level; P: P-value of the corresponding regression; R²: marginal R² for the corresponding regression; bolded values indicate significant results at threshold p < 3.47×10⁻⁴.

Associations between the 12 SMs and neuroinflammation biomarkers (YKL40, sTREM2, IL6) showed slightly less consistent results than the associations with the neurodegeneration biomarkers. Of the SM-IL6 associations, only three were nominally significant, and none were significant after multiple testing correction (Table 4 and Figure 2D). In contrast, 10 of the 12 SMs were nominally, positively associated with YKL40; 5 of these associations remained significant after Bonferroni correction. Finally, all 12 SMs had positive, nominally significant associations with sTREM2; 9 associations remained significant after Bonferroni correction.

Table 4.

Associations between SMs and neuroinflammation biomarkers.

	sTREM2 (ng/mL)			IL6 (pg/mL)			YKL40 (ng/mL)
Metabolite	β	P	R2	β	P	R2	β	P	R2
behenoyl SM (d18:1/22:0)	0.47	0.03	0.11	−0.02	0.48	0.03	6.87	0.05	0.31
palmitoyl dihydro-SM (d18:0/16:0)	1.64	5.15E-05	0.12	0.02	0.75	0.03	19.80	3.67E-03	0.32
palmitoyl SM (d18:1/16:0)	4.55	1.06E-16	0.17	−0.10	0.12	0.03	63.27	2.92E-11	0.35
SM (d18:1/14:0, d16:1/16:0)	1.91	6.67E-07	0.12	−0.07	0.18	0.03	27.02	5.14E-05	0.33
SM (d18:1/18:1, d18:2/18:0)	4.30	2.70E-14	0.17	−0.17	7.56E-03	0.04	59.56	2.40E-09	0.35
SM (d18:1/20:0, d16:1/22:0)	1.48	7.69E-06	0.12	−0.06	0.22	0.03	15.09	0.01	0.32
SM (d18:1/20:1, d18:2/20:0)	0.88	0.01	0.12	−0.01	0.78	0.02	13.51	0.03	0.33
SM (d18:1/22:1, d18:2/22:0, d16:1/24:1)	1.03	9.70E-05	0.11	−0.03	0.40	0.03	9.82	0.03	0.32
SM (d18:1/24:1, d18:2/24:0)	0.92	7.96E-04	0.12	−0.03	0.44	0.03	5.99	0.19	0.32
SM (d18:2/16:0, d18:1/16:1)	2.26	1.12E-08	0.13	−0.14	8.31E-03	0.04	31.62	3.08E-06	0.33
SM (d18:2/24:1, d18:1/24:2)	1.28	1.27E-05	0.12	−0.05	0.23	0.03	11.71	0.02	0.32
stearoyl SM (d18:1/18:0)	6.99	3.15E-28	0.24	−0.18	5.77E-03	0.04	105.64	1.41E-18	0.39

sTREM2: soluble triggering receptor found on myeloid cells 2; IL6: interleukin-6; YKL40: chitinase-3-like protein 1; β: change in the outcome with a one unit increase of the standardized metabolite level; P: P-value of the corresponding regression; R²: marginal R² for the corresponding regression; bolded values indicate significant results at threshold p<3.47×10⁻⁴.

Behenoyl SM (d18:1/22:0) yielded no significant associations after Bonferroni correction with any biomarkers analyzed, while stearoyl SM (d18:1/18:0) had the strongest significant association with all biomarkers where any significant association was observed. Palmitoyl SM (d18:1/16:0), SM (d18:1/14:0, d16:1/16:0), SM (d18:1/18:1, d18:2/18:0), SM (d18:2/16:0, d18:1/16:1), and stearoyl SM (d18:1/18:0) were significantly associated with the same seven biomarkers (p-tau₁₈₁, NfL, α-synuclein, neurogranin, sTREM2, IL6, and YKL40) after Bonferroni correction.

Independent signals

Six models with p-tau₁₈₁ as the outcome and stearoyl SM (the most significantly associated metabolite) as the main predictor were significantly improved by the addition of another metabolite predictor using nested ANOVA models (Supplemental Table 8). Only one metabolite significantly improved the model for NfL, while there were six metabolites that significantly improved the model for α-synuclein and seven that significantly improved the model for neurogranin. Among the models for neuroinflammation biomarkers, three metabolites significantly improved the models for sTREM2, one significantly improved the model for IL6, and five metabolites significantly improved the model for YKL40.

Sensitivity Analyses

The results of the sensitivity analysis with the APOE ε4 covariate were compared with the main analysis results and were not substantially different in either significance of associations or direction of effect (Supplemental Table 9). This observation is further supported by violin plots illustrating SM levels by APOE ε4 allele count, which don’t reveal any clear trends in SM levels by APOE ε4 count (Supplemental Figure 5).

Sensitivity analyses testing for sex-specific differences in associations between SMs and the outcomes found no significant interaction effects after Bonferroni correction either when testing for the significance of an interaction term in the regression (Supplemental Table 10) or when stratifying by sex (Supplemental Table 11).

Sensitivity analyses testing for modification of the effect of SMs on the outcomes by amyloid and tau status suggested several significant differences. Simple main effects from the interaction testing suggested that effects of SMs were stronger in the A+T+ group for some biomarker and SM combinations, without a strong pattern. Five SM-neurodegeneration biomarker regressions and nine SM-neuroinflammation biomarker regressions had significant A+T+ interaction effects, in all cases having a positive interaction effect term indicating the biomarker level rising more quickly per unit increase in the SMs among the A+T+ group. Notably, 7 of the 12 SMs had a positive, significant interaction term in the models for YKL-40 (Supplemental Table 12, Supplemental Figure 6). Similarly, in the amyloid and tau (AT)-stratified regressions, there were generally larger effect sizes observed in the A+T+ group than the A−T− group for the biomarkers NfL, α-synuclein, neurogranin, sTREM2, and YKL40 when there were significant effects detected (Supplemental Table 13), though the smaller sample sizes among the subgroups likely lowered the power for detecting some of the SM-biomarker associations.

The results of the sensitivity analysis controlling for cardiovascular factors (body mass index, waist-hip ratio, and systolic blood pressure) were compared with the main analysis results and were not substantially different in either significance of associations or direction of effect (Supplemental Table 14).

Discussion

Our study provides a robust exploration of the relationship between CSF SMs and various aspects of AD. We first explored whether CSF SMs were associated with diagnosis of AD or MCI in our cohorts. We found no significant associations between SMs and diagnosis of AD or MCI, partly contrasting with previous work that found an association between CSF SMs and prodromal AD, though not with mild or moderate AD [41]. The smaller sample size (n = 37) in the previous study may account for this difference, though it could also be that the differences in the particular subgroups of AD studied (CU/MCI/AD vs controls/prodromal/mild/moderate AD) led to this particular early stage signal going undetected in our analyses. Similarly, we found that amyloid biomarkers (Aβ_42/40, PiB Global DVR, and p-tau₁₈₁/Aβ₄₂) were not significantly associated with the CSF SMs analyzed. This finding is consistent with our previous research in the WRAP cohort that found no association between any of the CSF metabolites and measures of amyloid, but did identify associations between SMs and p-tau₁₈₁ using metabolome-wide correlations and ELISA biomarkers, as opposed to the NeuroToolKit biomarker panel [14]. Solomon et. al (2022) also reported multiple SMs, including stearoyl SM (d18:1/18:0), were associated with p-tau₁₈₁ in their discovery cohort (n = 47), though the association was not observed in their replication cohort. We identified significant SM associations with p-tau₁₈₁ and several biomarkers for neurodegeneration and neuroinflammation. Five SMs (stearoyl SM (d18:1/18:0), SM (d18:2/16:0, d18:1/16:1), SM (d18:1/18:1, d18:2/18:0), SM (d18:1/14:0, d16:1/16:0), and palmitoyl SM (d18:1/16:0)) in particular were significantly, positively associated with p-tau₁₈₁ and each of the biomarkers for neurodegeneration (NfL, α-synuclein, and neurogranin) and the neuroinflammation biomarkers YKL40 and sTREM2. All these associations were similar when we controlled for the number of APOE ε4 alleles, indicating that these results were not likely to be influenced by APOE genotype. Of the twelve metabolites we analyzed, stearoyl SM (d18:1/18:0) had the strongest associations with our biomarkers when there was any statistically significant SM association observed. This particular metabolite, along with SM (18:1/18:1), has been previously identified as having a significant, positive association with AD pathology [42]. While neither of these metabolites were significantly associated with markers of amyloid in our study, they were both associated with p-tau₁₈₁ and multiple biomarkers of neurodegeneration and neuroinflammation. Our pairwise correlations indicate that many of the CSF SMs we analyzed were intercorrelated, with potential subgroups among SMs based on chain length. Nested linear models with our most strongly associated metabolite, stearoyl SM (d18:1/18:0), supported the correlated nature of the SMs. Many of these models also showed some improvement with the addition of certain SMs, indicating that SMs other than stearoyl SM (d18:1/18:0) contain additional useful information.

In the secondary analyses, we found no significant effect modifications by sex and no difference between AT groups in relationships between SMs and cognitive status. However, we found evidence suggesting stronger SM effects on several biomarker outcomes, including α-synuclein, neurogranin, sTREM2, and YKL40, within the A+T+ subgroup than in other subgroups. Larger observed effect sizes in the A+T+ groups occurred for the biomarkers generally expected to change later on in the disease trajectory (e.g. p-tau) [43]. More research with larger sample sizes will be necessary to establish associations between SMs and various stages of AD more clearly given the relatively smaller sample sizes in the A+T− and A+T+ groups in this study.

Collectively, our results implicate CSF SMs as non-specific biomarkers of neurodegeneration and neuroinflammation. Strong positive associations between many SMs and markers like p-tau₁₈₁, NfL, neurogranin, YKL-40, and sTREM2 make SMs a candidate marker of neurodegenerative and neuroinflammatory processes, but the lack of association of these SMs with amyloid measures might mean that SMs’ changes are not specific to AD. One potential mechanism for the role of SMs in neurodegeneration and neuroinflammation is that SMs only begin changing as neurodegeneration or neuroinflammation occur and neurons begin to die, but not earlier, when amyloid is first beginning to accumulate. Previous research in preeclampsia has highlighted a mechanism by which SM levels increase. Stearoyl SM (d18:1/18:0), along with palmitoyl SM (d18:1/16:0), was elevated in plasma of preeclamptic mothers, thought to be a result of lipid rafts’ (microdomains of the cell membrane, consisting of SMs and other compounds) exposure to low oxygen levels [44]. This finding may grant insight into potential mechanisms for the association of CSF SMs with neurodegeneration: as neurons die, lipid rafts might release their components, including SMs, into CSF [8,10]. Mechanistic research into the SM changes during neurodegeneration are needed to further examine these hypotheses.

As noted above, previous studies on the relationship between AD and SMs in plasma and brain tissue have found significant associations but differing directionality, making CSF SMs especially interesting [9,10]. Our lack of significant SM-amyloid associations contrasted with a previous study on the association of total SMs with Aβ in CSF, which found significant associations between SMs (including stearoyl SM (d18:0/18:1)) and various markers of amyloid including Aβ₄₂ and Aβ₄₀ [45]. Here, with a larger sample size, we found no association of SMs with Aβ markers. A few differences could explain this discrepancy, including differences in study population (e.g., our population was enriched for older individuals, mean = 64.1) or different biomarkers used (e.g., Aβ₄₂ and Aβ₄₀ vs Aβ₄₂/Aβ₄₀).

This study has certain limitations that must be taken into consideration when interpreting its results. While we were able to produce one of the largest sample sizes of AD and MCI individuals used to study CSF SMs, our samples of individuals with AD and MCI were still relatively small (n = 89) compared to cognitively unimpaired controls (n = 409) (Table 1), potentially limiting our ability to assess diagnosis-related outcomes. We also lacked diversity; over 95% of individuals included in this study were self-reported white, limiting the generalizability of this study to other populations. As sample sizes for underrepresented groups grow, we will be better able to investigate the role of SMs across a greater range of populations. The goal of this study was to test for associations between SMs and a variety of outcomes; we did not specifically attempt to predict changes in the biomarkers over time. A future analysis which aims to predict future changes in biomarker levels would be an interesting addition to the literature. We also were limited by the types of sphingolipids that we examined. Ceramides, particularly, are a type of sphingolipid that has been implicated in AD and other diseases, but we were not able to assess their role here, as they were not included in CSF for our study, potentially because of the technique used to measure the metabolites in this study [46]. There is a high correlation between p-tau₁₈₁ and neurodegeneration biomarkers in this cohort, which may differ in a sample with more non-AD dementia present. Another limitation to consider is the presence of comorbidities, especially cerebrovascular pathology, in the individuals being studied, which may influence SM levels beyond the cardiovascular risk factors we assessed in our sensitivity analysis. Finally, nine SMs analyzed were designated as tier 2 compounds by Metabolon, compounds that have a structure that has been confirmed by literature review, but they are not necessarily confirmed by a reference standard, though we note that stearoyl SM (d18:1/18:0) was not one of these tier 2 compounds [47].

Conclusion

In this study, we examined the relationship between 12 SMs, MCI/AD diagnoses, and PET and CSF biomarkers of AD, neurodegeneration, and neuroinflammation, providing a comprehensive investigation of the role of CSF SMs in AD pathology. While we found no association between SMs and dementia-AD or MCI diagnoses nor with amyloid biomarkers, we did find strong positive associations between the SMs and p-tau₁₈₁, NfL, sTREM2, neurogranin, α-synuclein, and YKL40. Based on these findings, we hypothesize that SMs are non-specific biomarkers of neurodegeneration and neuroinflammation. There is still much to be learned: identifying a mechanism for SMs’ role in neurodegeneration and neuroinflammation, understanding how SM levels differ across other neurodegenerative diseases, and characterizing them in diverse cohorts with larger numbers of AD individuals and specific lipidomics techniques to better separate and identify SM species will grant greater insights into the role of these metabolites in AD.

Abbreviations

Aβ

amyloid-beta

AD

Alzheimer’s disease

Dementia-AD

clinical diagnosis of dementia with AD as a suspected cause

ADRC

Alzheimer’s Disease Research Center

α-synuclein

alpha synuclein

ANOVA

analysis of variance

CSF

cerebrospinal fluid

CU

cognitively unimpaired

DVR

distribution volume ratio

IL6

interleukin-6

LMM

linear mixed effects model

LP

lumbar puncture

MCI

mild cognitive impairment

MRI

magnetic resonance imaging

NfL

neurofilament light

NTK

NeuroToolKit

PET

positron emission tomography

PiB

Pittsburgh compound B

p-tau₁₈₁

phosphorylated tau

p-tau₁₈₁/Aβ₄₂

p-tau₁₈₁ to Aβ₄₂ ratio

QC

quality control

ROI

region of interest

SM

sphingomyelin

sTREM2

soluble triggering receptor found on myeloid cells 2

WRAP

Wisconsin Registry for Alzheimer’s Prevention

YKL40

chitinase-3-like protein 1