Microglial immunometabolism endophenotypes contribute to sex difference in Alzheimer's disease

Abstract

INTRODUCTION

The molecular mechanisms that contribute to sex differences, in particular female predominance, in Alzheimer's disease (AD) prevalence, symptomology, and pathology, are incompletely understood.

METHODS

To address this problem, we investigated cellular metabolism and immune responses (“immunometabolism endophenotype”) across AD individuals as a function of sex with diverse clinical diagnosis of cognitive status at death (cogdx), Braak staging, and Consortium to Establish a Registry for AD (CERAD) scores using human cortex metabolomics and transcriptomics data from the Religious Orders Study / Memory and Aging Project (ROSMAP) cohort.

RESULTS

We identified sex‐specific metabolites, immune and metabolic genes, and pathways associated with the AD diagnosis and progression. We identified female‐specific elevation in glycerophosphorylcholine and N‐acetylglutamate, which are AD inflammatory metabolites involved in interleukin (IL)‐17 signaling, C‐type lectin receptor, interferon signaling, and Toll‐like receptor pathways. We pinpointed distinct microglia‐specific immunometabolism endophenotypes (i.e., lipid‐ and amino acid‐specific IL‐10 and IL‐17 signaling pathways) between female and male AD subjects. In addition, female AD subjects showed evidence of diminished excitatory neuron and microglia communications via glutamate‐mediated immunometabolism.

DISCUSSION

Our results point to new understanding of the molecular basis for female predominance in AD, and warrant future independent validations with ethnically diverse patient cohorts to establish a likely causal relationship of microglial immunometabolism in the sex differences in AD.

Highlights

• Sex‐specific immune metabolites, gene networks and pathways, are associated with Alzheimer's disease pathogenesis and disease progression.

• Female AD subjects exhibit microglial immunometabolism endophenotypes characterized by decreased glutamate metabolism and elevated interleukin‐10 pathway activity.

• Female AD subjects showed a shift in glutamate‐mediated cell‐cell communications between excitatory neurons to microglia and astrocyte.

1. BACKGROUND

Alzheimer's disease (AD) is a devastating neurodegenerative disease, and AD‐related dementia (AD/ADRD) will affect 16 million Americans and 152 million people worldwide by 2050. ¹ Many more will have mild cognitive impairment (MCI) due to AD or preclinical AD. ² Notably, age‐matched women comprise a higher proportion of AD cases (∼ 70%). ³ , ⁴ , ⁵ Women also show faster cognitive decline after diagnosis of MCI or ADRD. ⁶ , ⁷ Interestingly, levels of amyloid‐beta measured with positron emission tomography (PET) brain imaging and biochemical analysis of cerebrospinal fluid (CSF) have shown limited sex differences. ⁸ Therefore, clinically actionable, sex‐specific biomarkers that could potentially improve our understanding of the pathogenesis of the predominance of AD in women are urgently needed. These biomarkers could guide treatment and have practical applications in AD clinical trials, such as monitoring therapeutic effects or selecting participants for inclusion.

The underlying genetic basis and molecular mechanisms of AD sex differences are active areas of investigation. ⁵ Indeed, multiple factors have been implicated in driving female predominance of AD, such as (a) genetic ⁹ including incomplete X‐inactivation leading to overexpression of ubiquitin specific peptidase 11 in women, ¹⁰ (b) aberrant S‐nitrosylation, ¹¹ (c) sex hormones ¹² like follicle‐stimulating hormone, ¹³ (d) variations in brain structure, ¹⁴ (e) psychosocial stress responses, ¹⁵ (f) cellular metabolism, ¹⁶ (g) mitophagy, ¹⁷ , ¹⁸ and (h) local and systemic inflammation. ¹⁹ Importantly, AD is no longer considered exclusively a central nervous system (CNS) disease; rather, recent studies implicate important roles in AD for neuro‐ and systematic inflammation ²⁰ , ²¹ and cellular metabolism. ¹⁶ , ²² This intersection between the immune system and metabolism is termed the immunometabolism endophenotype. ²³

Immune cells, such as macrophages, T cells, and dendritic cells, undergo significant metabolic changes in response to different stimuli. ²⁴ Reciprocally, metabolic signals, such as glucose, fatty acids, and amino acids, modulate immune cell functions. ²⁵ Immunometabolism has recently emerged as a promising area of research in neurodegenerative diseases, such as AD. ²⁶ For example, microglia—the resident immune cells of the brain long known to play a role in AD ²⁷ —undergo metabolic changes that affect their function in AD. Similarly, alterations in glucose metabolism are observed in the brains of AD patients and may contribute to the disease pathology. ²⁸ It has also been suggested that immune‐mediated neuroinflammation in AD may be modulated by metabolic pathways, ²⁹ such as the kynurenine pathway and the inflammasome pathway. ³⁰

Recent advances in genomic tools, including metabolomics and single‐nucleus transcriptomics (snRNA‐seq), are facilitating the exploration of neuro‐immune networks. Such studies have identified pathological immune cell subpopulations in AD ³¹ and demonstrated a higher proportion of disease‐associated cell subpopulations in female individuals with AD pathology. ³² However, it is not known whether female and male coordination of the immune and metabolic systems are similar in AD. To address this question, we integrated human brain metabolomics and snRNA‐seq transcriptomics data from subjects with well‐characterized AD phenotypes. As detailed below, this effort has provided new insights into the basis of sex differences in AD.

2. METHODS

2.1. Metabolomics and transcriptomes in the ROSMAP cohort

2.1.1. ROS/MAP cohort

The Religious Order Study and Rush Memory and Aging Project (ROSMAP) cohort were utilized for both metabolomics ³³ (Synapse ID: syn26401311) and transcriptomics (Synapse ID: syn21241740). Samples were isolated from the dorsolateral prefrontal cortex (DLPFC) of 473 human subjects, along with clinical metadata (Table S1), including sex, age at death, education, post mortem interval (PMI), and apolipoprotein E (APOE) genotype status. Three clinical scores were used to define AD outcome in the context of tau and amyloid phenotypes: (1) clinical diagnosis of cognitive status at time of death ³⁴ (cogdx), which defined control (cogdx = 1), MCI (cogdx = 2 and 3), AD (cogdx = 4 and 5), and other dementia (cogdx = 6); ³³ (2) Braak staging, which is a semiquantitative measure of neurofibrillary tangles, with individuals assigned to one of three Braak subgroups from light to severe, I‐II, III‐IV, and V‐VI; and (3) Consortium to Establish a Registry for AD (CERAD) score, ³⁵ which is a semiquantitative measure of neuritic plaques, with all individuals assigned to two subgroups: CERAD I‐II (denoting severe and moderate neuritic plaques) and CERAS III‐IV (mild or none neuritic plaques). Brain metabolomics profiles were acquired from the Synapse website, with accession number syn26401311.

2.1.2. Metabolomics data preprocessing

Data preprocessing was performed following established methods. ³³ The raw data included 514 human subjects with metabolomic data. Metabolites with 25% or more missing values were removed. To correct for sample‐wise variation, probabilistic quotient normalization was applied. A k‐nearest‐neighbor algorithm was used to impute the remaining missing values. Outlier samples were identified using the local outlier factor method, and subsequently removed. To impute the new missing values, an extra round of the k‐nearest‐neighbor algorithm was employed. All preprocessing steps were implemented by maplet packages (v 1.1.2) in R version 4.1.1. After the preprocessing step, we also matched human subjects' IDs with bulk RNA‐seq data in ROSMAP to ensure that all the individuals had both metabolomics and bulk RNA‐seq data. Finally, 668 metabolites were included in the subsequent analysis, based on data from 469 human brain samples. This included 203 individuals with AD (female n = 146, male n = 57), as well as 117 individuals with MCI (female n = 77, male n = 40), 142 controls (no cognitive impairment, female n = 99, male n = 43), or 7 other forms of dementia (female n = 2, male n = 5).

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literatures using traditional sources. Sex differences in Alzheimer's disease (AD) have been recognized, with women showing a higher prevalence and more neuropathological severity than men. Investigations have highlighted possible underlying mechanisms, such as genetic factors, partial X chromosome inactivation, and hormonal influences. The hypothesis of immunometabolism—the cross‐talk between immune responses and cellular metabolic processes—has emerged as a promising avenue for understanding AD sex differences.

2. Interpretation: Our multi‐omics analysis unveiled sex‐specific alterations in immunometabolism in AD. We identified sex‐specific metabolites (i.e., phosphatidylethanolamine), immune and metabolic genes, networks and pathways (i.e., interleukin [IL]‐10 and IL‐17 signaling) associated with the AD pathogenesis and disease progression. These findings suggested that sex‐specific differences in immune responses, intracellular and extracellular metabolism, and microglial immunometabolism may contribute to the sex‐specific differences in AD. We showed that sex‐specific pathways may offer potential therapeutic approaches for prevention and treatment of AD in sex‐specific manners.

3. Future directions: The distinct immunometabolism profiles between sex in AD call for precision medicine biomarker and therapeutic intervention discovery. Further investigations in ethnically diverse patient cohorts are essential to establish a likely causal relationship of microglial immunometabolism in the sex differences in AD.

2.1.3. Differential analysis of metabolites

We investigated the relationship between metabolite concentration and various binary outcomes (Y) using three complementary tests: (1) Test 1 focused on identifying differential metabolites based on AD pathologic scores in different female and male subgroups: (i) AD versus (vs.) control, (ii) Braak V‐VI vs. Braak I‐II, and (iii) CERAD I‐II vs. CERAD III‐IV, after adjusting potential confounding factors, including age, PMI, education level, and APOE4 allele status (formula 1). (2) We tested sex by metabolite interaction across entire cohort using the same binary outcomes of test 1 (formula 2). (3) We also investigate the sex‐specific differential metabolites comparing female vs. male individuals within AD pathologic groups separately (formula 3). The logistic regression model was implemented using the glm function in a binomial distribution linked with a logit function in R (v4.1.1):

$$\mathrm{Y}(\mathrm{AD}\mathrm{vs}.\mathrm{control})\sim \mathrm{metabolite}+\mathrm{age}+\mathrm{PMI}+\mathrm{education}+APOE4$$ (1)

$$\begin{array}{ccc}& & \mathrm{Y}(\mathrm{AD}\mathrm{vs}.\mathrm{control})\sim \mathrm{sex}+\mathrm{metabolite}+\mathrm{sex}\times \mathrm{meta}\mathrm{boli}\mathrm{te}\hfill \\ & & +\mathrm{age}+\mathrm{PMI}+\mathrm{education}+\mathrm{APOE}4\hfill \end{array}$$ (2)

$$\mathrm{Y}\left(\mathrm{Female}\mathrm{vs}.\mathrm{Male}\right)\sim \mathrm{metabolite}+\mathrm{age}+\mathrm{PMI}+\mathrm{education}+APOE4$$ (3)

The adjusted coefficient of metabolite concentration indicated its association with each outcome. A positive coefficient (> 0) implied that an increase in metabolite concentration was associated with a higher likelihood of an outcome; while a negative coefficient (< 0) suggested that an increase in metabolite concentration was linked to a lower likelihood of an outcome. Differential metabolites were identified as having a significant association with each outcome after considering the influence of the confounding variables. Benjamini and Hochberg's (BH) method ³⁶ was used to compute false discovery rate (FDR) adjusted p‐values.

2.1.4. Bootstrap with propensity score matching

We have utilized 10,000 bootstrap iterations and 1:1 propensity score matching (PSM) for females with males to further interrogate sex difference in AD, which may be influenced by an unbalanced ratio between female and male individuals. This was executed in a three‐step process: (1) one‐to‐one PSM of males to females from the entire cohort after adjusting for age, PMI, education years, and APOE4 allele status; (2) replicating the data analysis as described in Sections 2.1.3; and (3) conducting 10,000 bootstrap iterations with replicated experiments based on steps 1 and 2. We analyzed the density distribution of count value, statistical p value, and coefficient across various AD pathology groups, including cogdx score, Braak stage, and CERAD score.

2.1.5. Bulk RNA‐seq analysis

Human brain bulk RNA‐seq analysis was conducted on samples from 469 human subjects from the ROSMAP cohort with matched metabolomics data (Section 2.1.3). The raw count file for bulk RNA‐seq data was obtained from Synapse with accession number syn21241740. Differential gene expression profiles were analyzed using the edgeR package (v3.34.0) with default parameters on the R platform. Comparison groups for bulk transcriptome data analyses were consistent with those used in the metabolomics data analyses. Metabolite‐gene expression correlation analyses were conducted by Spearman's rank correlation coefficient (ρ), with multiple testing correction (FDR < 0.05).

2.2. Single‐nucleus transcriptome (snRNA‐seq) integration and analysis

2.2.1. Datasets collection and processing

We obtained four publicly available snRNA‐seq datasets (Table S2) generated from brain cortex samples. For Dataset Lau_2020, ³⁷ the count file was downloaded from the Gene Expression Omnibus (GEO, ID GSE157827) database, and the metadata file was retrieved from the supplementary table of the original paper. We provided a detailed summary for selecting AD and control groups across the 4 snRNA‐seq datasets (Table S2): (1) Among 4 datasets, 3 datasets recruited human subjects with age‐matched AD and controls; and (2) the selection of AD cases and controls based on both pathologic criteria (all datasets) and clinical diagnosis (3 of 4 datasets). In particular, AD cases across all datasets were counted by Braak stages (≥IV or V) as a critical criterion for AD group allocation. And (3) of these datasets, three of them (GSE157827, GSE174367, GSE138852) have the control subjects with no cognitive impairment and with a Braak stage of ≤II. Two datasets (syn18485175 and GSE138852) were confirmed that the controls have no or very little pathology, such as occasional diffuse plaque in cortex.

We identified 169,506 cells/nuclei. The standard snRNA‐seq processing pipeline was applied using the Seurat 4.0 package in R (v 4.1.0). First, the dataset was loaded into a Seurat object, and cell filtering was performed based on prespecified criteria, including mitochondria expression percentage, nFeature_RNA, and nCount_RNA. The dataset was then normalized, integrated with batch effect correction, scaled, and subjected to principle component analysis (PCA) and dimensional reduction using (uniform manifold approximation and projection) UMAP. Marker genes from original publications were used to annotate the cell types. Dataset Grubman_2019 ³⁸ (GEO ID, GSE138852) has a processed version provided by the author. We obtained these data, along with the metadata and annotated cell types, from adsn.ddnetbio.com. After filtering nuclei using criterion consistent with other datasets, the dataset Grubman_2019 included 11,884 nuclei for next step analysis. Dataset Morabito_2021 ³⁹ was acquired from the AD Knowledge Portal at Synapse (https://www.synapse.org, ID syn22079621). We used the same number of nuclei and cell types as the original publications (n = 61, 472). Dataset Mathys_2019 ⁴⁰ was also downloaded from Synapse (ID syn18485175). In total, 70,643 nuclei were used for integration analyses after preprocessing, consistent of what was conducted with the other three datasets.

2.2.2. snRNA‐seq data integration

Prior to data integration, metadata harmonization was performed. All scRNA‐seq datasets were normalized by applying log+1 transformation to the count data. As the four preprocessed datasets were well‐annotated, we initially trained our data integration model using the scVI. ⁴¹ In this step, the original datasets had already undergone batch effect correction. Thus, we corrected the batch effect based on the datasets during the integration process. We then performed semi‐supervised single‐cell annotation using the variational inference (scANVI ⁴² ) model to improve latent representation of the data, which resulted in a more accurate integration outcome. ⁴³ The integration pipeline described above was implemented using scvi‐tools in the Python (v3.9.5). Next, we saved latent matrix for standard snRNA‐seq analysis using the Seurat 4.0 package in R (details were described in the previous section). The following cell markers were used for cell type identification (Table S3): (1) P2RY12, CSF1R, CX3CR1, CD74 for microglia; (2) GFAP, AQP4, GJA1 for astrocytes; (3) MOBP, PLP1, ST18 for oligodendrocytes; (4) PDGFRA, VCAN, OLIG1 for oligodendrocyte precursor cells (OPC); (5) CLDN5, FLT1 for endothelial cells; (6) RBFOX3, GRIN1, LDB2, RORB for excitatory neurons; and (7) CCK, CNR1, GAD1, GAD2 for inhibitory neurons. In summary, the fully integrated dataset included 99 human cortex samples comprising a total of 313,505 single nuclei. Of these, 53 samples were from individuals with AD (female n = 24, male n = 29) and 46 samples were from controls (female n = 19, male n = 27). Before cell proportion analysis, we removed the two samples with cells lower than 100 to reduce the outlier influence. For each individual, we scaled the cell proportion by dividing the number of cells in each cell type by the sum of cells in the individual. Then, we performed the Mann–Whitney U test to detect the normalized cell proportion changes between AD and control in female and male subpopulations (Table S4).

2.3. Immune‐metabolic pathway activity score (PAS) analysis

2.3.1. Alzheimer's pathway activity score (alzPAS) calculation

We evaluated the activity of specific immune and metabolic pathways as described previously ⁴⁴ using the integrated snRNA‐seq data (Method Section 2.2). We calculated the average expression level of each immune or metabolic gene within a specific cell type, comparing it to the average expression level across all cell types. This provided a relative expression (RE) level for each gene across each cell type. For each pathway in a specific cell type, we computed the alzPAS using the weighted average of the RE levels of all genes in that pathway. The weighting factor was based on the number of pathways containing each gene. We then normalized alzPAS by applying a log₂ transformation. To ensure accurate alzPAS, we excluded genes with extremely high (RE levels > 3 times 75th percentile) or low relative expression levels (RE levels < 1/3 times 25th percentile). We assessed the statistical significance of alzPAS in a specific cell type using a random permutation test, in which we shuffled cell type labels 1000 times to create a null distribution of pathway activity scores. By calculating a p‐value for each alzPAS, we determined if the activity of a specific pathway was significantly higher or lower than the mean value in each cell type. Thirty‐nine immune and 83 metabolic pathways were as defined by the Reactome. ⁴⁵

2.3.2. Statistical analysis for alzPAS

We used the comparison group of AD vs. control as an example to illustrate the statistical analysis for alzPAS. We applied a linear regression model to fit the trend between AD and control in both female and male subgroups to compare the changes in alzPAS. We determined the absolute value of the slope (|slope|) of the linear regression model to scale the changes between AD and control. We used a cutoff of |slope| > 0.1 and p < 0.05 to identify differentially activated pathways. A positive slope (> 0) indicated increased activity for a specific pathway in AD vs. control, while a negative slope (< 0) indicated reduced activity for a specific pathway in AD vs. control. A slope equal to 0 indicated no changes in alzPAS for a specific pathway in AD vs. control.

2.4. Metabolic cell‐cell communication network

We utilized the MEBOCOST ⁴⁶ tool to estimate metabolite‐mediated cell‐cell communication for snRNA‐seq data. MEBOCOST calculates communication scores between donor and receiver cell types by utilizing inferred metabolites and prior knowledge of metabolite‐sensor partners. We used our integrated snRNA‐seq (Section 2.2) to calculate cell‐cell communications based on the default metabolite‐sensor partners networks in MEBOCOST. The metabolic enzyme or receptor genes were filtered when they detected less than 10 cells and the cell‐cell connections should be found in higher than 10% of the cells. ⁴⁶ The 1000 times of permutation testing were performed with shuffled snRNA‐seq data to assess statistical significance (p < 0.05). The final network visualization was through Cytoscspe v3.8.2 (https://cytoscape.org/).

3. RESULTS

3.1. Study design

We performed multimodal integration of clinical (i.e., neuropathological diagnosis scores, APOE4 genotypes) and multi‐omics data (including metabolomics and bulk and snRNA‐seq data) to investigate sex differences in immunometabolism endophenotype of AD (Figure 1). The ROSMAP cohort used in this study included 469 human subjects with matched clinical, brain metabolomics, and brain bulk transcriptome data (Table S1): 324 females aged 88 ± 4 and 145 males aged 87 ± 4, respectively. AD phenotypes were defined by cogdx based on original study. ³³ The Braak staging and CERAD score were also used to define AD progression. We found that females exhibited more severe burdens of AD pathology in 10,000 times of 1:1 PSM cohort observations (Figure S1), consistent with previous studies. ⁴⁷ , ⁴⁸ , ⁴⁹

FIGURE 1.

The multimodal omics analytic framework. Integrative analysis of immunometabolism endophenotype in Alzheimer's disease (AD) by sex. (1) We utilized the ROS/MAP cohort's metabolomics and bulk RNA‐seq data from 469 human subjects (Table S1), comprising 146 females and 57 males with AD, 77 females and 40 males with mild cognitive impairment (MCI), 99 females and 43 males without cognitive impairment, and two females and five males with other dementia types. (2) The second integrative single‐nucleus RNA‐sequencing (snRNA‐seq) dataset was derived from four distinct data sources (Table S2), including 53 AD patient samples (24 females and 29 males) and 46 control patient samples (19 females and 27 males). Immune and metabolic pathway activity score (alzPAS) analysis was conducted to elucidate the immunometabolic endophenotype across various AD pathology stages. The metabolic cell‐cell communication network was used to uncover molecular mechanisms at the single‐cell level

We next developed a network‐based multi‐omics analytic framework (Figure 1) to elucidate the sex differences in cellular metabolism and regulation of the immune systems in human brains, particularly in microglia‐ and astrocyte‐specific manners. We investigated immunometabolism alteration using multi‐omics datasets: (1) bulk brain metabolomics and transcriptomics data to inspect sex differences in metabolite‐immune gene association networks; and (2) integrated snRNA‐seq data to evaluate sex difference in the immune‐metabolic pathway activity and metabolic cell‐cell communication networks. To comprehensively characterize the immunometabolism endophenotype in each cell type, we utilized deep generative models (scVI ⁴¹ and scANVI ⁴² ) to integrate four publicly available snRNA‐seq datasets ³⁷ , ³⁸ , ³⁹ , ⁴⁰ (Table S2), while reducing batch effects and increasing sample size and cell subpopulations, such as microglia. Before integration, we harmonized clinical metadata and annotated cell types in each separate dataset. We utilized the scVI training model and scANVI to obtain a latent matrix and annotate the cell types based on prior knowledge from harmonized data. In total, the integrated dataset consists of 99 human cortex samples with 313,505 single nuclei. Among these, there were 53 samples from individuals with well‐characterized AD neuropathology (24 females and 29 males) and 46 samples from matched controls (19 females and 27 males). In addition, we designed a PAS ⁴⁴ to assess alterations in activity across 39 immune and 83 cellular metabolic pathways. Taken together, this comprehensive bioinformatics framework (Figure 1) presents an opportunity for novel insights into immunometabolism endophenotypes, metabolic characteristics, and mechanisms of immune‐metabolic pathway crosstalk underlying sex differences in AD pathogenesis and progression.

3.2. Sex‐specific differential metabolic profiles in AD cases vs. controls

We identified 48 sex‐specific differentially expressed metabolites (FDR < 0.05, Figure 2 and Table S5) between AD and controls, after adjusting for age at death, APOE4 genotyping, years of education, and PMI (see Methods). These differential metabolites belong to (1) amino acid pathways (e.g., glutamate metabolism; glycine, serine, and threonine metabolism; histidine metabolism; and urea cycle, arginine and proline metabolism); (2) lipid pathways (e.g., fatty acid metabolism; phosphatidylcholine [PC]; phosphatidylethanolamine [PE]; phospholipid and bile acid metabolism); (3) other pathways, such as co‐factors and vitamins, and carbohydrate (Figure 2A). Among these differential metabolites, urea (AD vs. controls, p = 7.17 × 10⁻⁴; Braak V‐VI vs. III‐IV, p = 0.001), dimethylarginine (ADMA + SDMA) (AD vs. controls, p = 3.71 × 10⁻⁴; Braak V‐VI vs. III‐IV, p = 1.81 × 10⁻⁴), glycerophosphorylcholine (GPC, AD vs. controls, p = 4.39 × 10⁻¹⁰; Braak V‐VI vs. III‐IV, p = 0.001) showed elevated abundance in females with AD compared to controls, as well as females with Braak staging V‐VI relative to those with Braak staging III‐IV. We identified that the abundances of 1‐stearoyl‐2‐oleoyl‐GPE (18:0 / 18:1) in PE (AD vs. controls, p = 3.47 × 10⁻⁴; Braak V‐VI vs. III‐IV, p = 0.002) and N‐acetylglutamate (Braak V‐VI vs. III‐IV, p = 7.21 × 10⁻⁷) in glutamate metabolism were reduced in female AD compared to controls, and in female subjects with Braak staging V‐VI compared to staging III‐IV. In co‐factors metabolic pathways, we found that nicotinamide riboside (NR) was specifically reduced in female AD compared to controls (p = 2.13 × 10⁻⁴, Figure 3). NR, an NAD+ precursor with reported protective effect for AD, ⁵⁰ exhibited sex‐specific metabolic changes in 3xTg AD mouse brains. ⁵¹ This finding is notable in the context of recent evidence that treatment with a pharmacologic stabilizer of NAD⁺/NADH has also been shown to prevent the onset of disease in a rat model of AD. ⁵²

FIGURE 2.

Sex‐specific metabolic alterations across different Alzheimer's disease (AD) pathological groups. (A) Metabolomics data were downloaded from the AD Knowledge Portal with accession ID syn26401311. Metabolite levels were compared across various disease phenotypes in female and male subgroups respectively, such as AD versus (vs.) control, Braak stage V‐VI vs. Braak stage I‐II, and other comparison groups. The linear model was used to calculate the effect of metabolite on outcomes in female and male subgroups, including the formula: glm (AD vs. control) ∼ metabolites + age + education + APOE4 genotypes + pmi (post mortem interval). (B) The coefficient and density distribution for two highly reproducible metabolites (glycerophosphorylcholine [GPC] and N‐acetylglutamate are derived from comparisons between AD and control within 10,000 PSM cohorts. The boxplot illustrates the metabolite coefficients when comparing AD vs. controls across 10,000 PSM cohorts for both females (F) and males (M). The density plot denotes ‐log10(P) distribution across 10,000 PSM cohorts. The y‐axis represents the density of PSM cohorts within a specific range, while the x‐axis highlights the distribution of p‐values

FIGURE 3.

Sex‐specific metabolite‐gene association network in AD. Spearman's rank correlation coefficient (ρ) was used to create the metabolite‐gene association network. The colorful edges connecting metabolite‐gene pairs indicates that the absolute value |ρ| > 0.3 with multiple testing correction (FDR < 0.05). The gray edges represent the gene‐encoding enzymes enriched in the specific pathways. Circle nodes represent differential metabolites when comparing AD vs. control in female subgroups. Square nodes represent differential genes when comparing AD vs. control in female subgroups. The diamond‐shaped nodes represent the metabolic and immune pathways

We conducted the differential analysis of metabolites in 1:1 (female vs. male) balanced cohorts (10,000 randomizations) using a PSM approach after we adjusted age, PMI, education years, and APOE4 genotypes (see the Methods section). We observed that majority of differential metabolites (FDR < 0.05) were replicated in PSM cohorts (Figure S2). For instance, differential metabolites, GPC and N‐acetylglutamate, displayed higher reproducibility (Figure 2B) between AD and controls in the female individuals. Specifically, in 97% of the PSM cohorts, females with AD have elevated GPC levels compared to female controls. Yet, for male subgroups, only 66% of PSM matched cohorts identified significantly elevated GPC levels in those with AD. In addition, our initial results indicated a decreased N‐acetylglutamate levels in AD compared to control in females, but not in the males. From 10,000 PSM cohorts, 92% cohorts re‐identified the decreased N‐acetylglutamate in females, whereas only 46% re‐identified decreased N‐acetylglutamate in males.

3.3. Sex‐specific differential metabolic profiles across AD pathological groups

We further inspected the metabolic differences between females and males within distinct AD pathologic categories. Females exhibited distinct metabolomic profiles in the CERAD I‐II score group, representing severe [I] and moderate [II] neuritic plaques (Figure S3 and Table S6). These female‐specific metabolic changes in the pathways of glutamate metabolism, fatty acid, phosphatidylcholines, and phosphatidylethanolamine, aligned with the AD‐specific metabolic pathways observed in females (Figure 2). Notably, these differential metabolites reveal consistency across 10,000 PSM cohorts (Figure S4A). For instance, within these cohorts, 1‐linoleoyl‐2‐arachidonoyl‐GPC (18:2/20:4n6) was elevated in females compared to males in the CERAD I‐II group in 97% PSM cohorts (Figure S4A,C). The fatty acid metabolites of 13‐HODE and 9‐HODE were significantly elevated in female AD patients compared to males, but not in MCI and NCI groups (Figure S4A,B).

3.4. Sex‐by‐metabolite interactions in AD

Among 48 sex‐specific differential metabolites between AD and controls (Figure 2A), 16 metabolites showed significant sex‐by‐metabolite interaction across different comparison groups (Figure S5 and Table S7): (i) AD versus controls, (ii) Braak V‐VI vs. Braak III‐IV, and (iii) CERAD I‐II vs. CERAD III‐IV after adjusting potential confounding factors, including age, PMI, education level, and APOE4 allele status. For example, N‐acetylglutamate showed a stronger sex‐by‐metabolite interaction in individuals with Braak V‐VI compared to Braak III‐IV. In particular, we observed that 92% PSM cohorts re‐identified the decreased N‐acetylglutamate in females with Braak V‐VI (Figure 2B). Taken together, these sex‐by‐metabolite interaction analyses suggest that the glutamate and lipid metabolism in AD exhibit female‐specific effects, especially for N‐acetylglutamate and 1‐linoleoyl‐2‐arachidonoyl‐GPC (18:2 / 20:4n6). Further validation using a larger, male and female balanced cohorts are highly warranted.

3.5. Differential immunometabolism between males and females with AD

We further employed the metabolite‐gene association network to assess immunometabolic alterations in female and male patients using matched bulk RNA‐seq and metabolomic data (Figure 1). We found that female and male AD show differential correlation patterns in metabolite–gene association networks connecting metabolites and enzyme‐coding genes (Figure 3). For example, a decrease in N‐acetylglutamate has a negative correlation with TRIM5 expression in females with AD ($\rho$ = −0.35). Tripartite motif containing 5 (TRIM5) is an innate immune sensor gene in interferons pathway. ⁵³ The methionine sulfoxide is an oxidation of methionine and shows potential role in aging. ⁵⁴ We found that increased methionine sulfoxide in female AD is negatively associated with pyruvate kinase M1/2 (PKM) expression ($\rho$ = −0.41) and voltage dependent anion channel 1 (VDAC1) expression ($\rho$ = −0.39) in females. VDAC1 regulates mitochondrial function in the TCA cycle. ⁵⁵ PKM play crucial roles in glucose metabolism. ⁵⁶ We found that the methionine sulfoxide abundance was positively associated with immune gene IKBKB expression ($\rho$ = 0.37) in female AD patients. Altogether, these observations suggest that the methionine sulfoxide may play roles in immunometabolism by regulating mitochondrial function and innate immune pathways in a female‐specific manner.

We also found that elevation of 1‐stearoyl‐2‐oleoyl‐GPE (18:0 / 18:1) in female AD is positively associated with IL1RL2 ($\rho$ = 0.33) and p21 activated kinase 1 (PAK1, $\rho$ = 0.33) gene expression. IL1RL2 plays crucial roles in interleukin 1 (IL‐1) family signaling, ⁵⁷ , ⁵⁸ while PAK1 is involved in both the C‐type lectin receptors ⁵⁹ (CLRs) pathway and Toll ‐like receptor (TLR) signaling. ⁶⁰ , ⁶¹ An elevated level of GPC correlates with IL17RB expression ($\rho$ = 0.55) ⁶² and RAF1 expression ($\rho$ = 0.53), potentially regulating IL‐17 signaling and the CLR pathways in male AD. We also found that GPC correlates with multiple immune and metabolic genes in female AD. In particular, GPC exhibits positive associations with IKBKB ($\rho$ = 0.35) and IRAK2 ($\rho$ = 0.43) in TLR signaling, as well as with P2RX7 ($\rho$ = 0.36) in nucleotide‐binding domain leucine‐rich repeat containing receptors (NOD‐like receptors [NLRs]) signaling in female AD. Conversely, GPC is negatively associated with PTPN5 ($\rho$ = −0.32) in IL‐1 signaling and PAK1 in the CLRs pathway in female AD. GPC also correlates with metabolic genes, such as MTMR10 ($\rho$ = 0.33) in phosphatidylinositol (PI) metabolism in female AD. Taken together, these observations show that GPC has a strong sex‐specific immunometabolic influence on multiple metabolic and immune pathways in females with AD.

3.6. Microglial immunometabolism between males and females with AD

To improve the statistical power ⁶³ and reduce technical variability from the snRNA‐seq data, ⁴³ , ⁶⁴ we integrated four publicly available snRNA‐seq datasets to create a combined dataset using two deep generative models: scVI ⁴¹ and scANVI ⁴² (Figure 1). We identified seven major cell types with effective batch correction (Figure 4A and Figure S6), including microglia, astrocytes, excitatory neurons (Ex‐neuron), inhibitory neurons (Inh‐neuron), oligodendrocytes, endothelial cells, and OPCs. In particular, the abundance of microglia was suggestively elevated in female AD patients compared to controls (p = 0.034 and FDR = 0.275, Mann–Whitney U test, Figure 4B), while no significant changes were observed in other cell types between females and males (Figure S7 and Table S4).

FIGURE 4.

Immunometabolism endophenotype alterations in microglia, astrocytes and excitatory neurons (Ex‐neuron). (A) Single‐nucleus RNA‐sequencing data UMAP plot are shown across eight well‐characterized cell types by marker genes. (B) The box plot illustrates the changes in cell percentage between Alzheimer's disease (AD) vs. control in female and male subgroups. The y‐axis represents scaled cell percentages calculated by dividing the cell number by the total cells in each sample. Each dot in the box plot represents an individual sample. (C) Cell‐type specific Alzheimer's pathway activity score (alzPAS) alterations are quantified in male and female subgroups. Each circle in the diagram represents an immune or metabolic pathway, and the filling color represents AD and control respectively. Red lines indicate elevated alzPAS value from control to AD, while blue lines indicate decreased alzPAS value from control to AD

PAS ⁴⁴ was further used to quantitative the pathway activity across each cell type. We performed alzPAS analysis with simultaneous evaluation of 83 metabolic and 39 immune pathways across each cell type (Table S8). In female AD subjects, 69% (57 / 83) of metabolic pathways in microglia showed reduced alzPAS value compared to female controls, in particular for (1) three lipid metabolism pathways (cholesterol biosynthesis, metabolism of steroid hormones, bile acid and bile salt metabolism) and (2) two amino acid pathways (threonine catabolism, and glutamate and glutamine metabolism) (Figure 4C and Figure S8). The astrocytes and excitatory neurons in female AD subjects showed reduced PAS values in both carbohydrate and lipid metabolism pathways, such as glucose metabolism, fructose metabolism, bile acid and bile salt metabolism, and glycerophospholipid biosynthesis. In contrast, the male AD subjects showed different metabolic pathway activities in multiple brain cell types, such as decreased threonine catabolism in microglia, astrocytes, ex‐neurons, and OPCs. Additionally, astrocytes and OPCs in male AD subjects showed reduced urea cycle activity (Figure S8). We observed uniquely reduced alzPAS levels of threonine catabolism in microglia between female and male AD compared to control. Males (AD, alzPAS = 0.00) exhibited higher activity in the threonine catabolism pathway than females (AD, alzPAS = −0.36). These observations reveal that microglia in female AD brain cortex may have a slower rate of threonine decomposition compared to males, consistent with elevated abundance of threonine in female brains relative to male brains (Figure 2).

We next turned to investigate alzPAS across 39 immune pathways in each brain cell type, including microglia, astrocytes, excitatory neurons (Figure 4C), inhibitory neurons, oligodendrocytes, and endothelial cells (Figure S8). For microglia, female AD patients exhibit decreased immune alzPAS in the CLRs pathway compared to controls; whereas male AD subjects showed an elevated CLRs pathway activity. We found that microglia from female AD subjects displayed increased IL‐10 signaling activity and reduced pathway activities (measured by alzPAS) across Rap‐1 signaling, complement cascade pathway, and IL‐6 family signaling compared to female controls. However, there were no significant changes in alzPAS for these pathways in microglia between male AD and controls. In addition, male AD subjects showed elevated IL‐17 signaling activity in microglia compared to controls, while no differential pathway activities were observed in microglia from female AD vs. controls. Similar pathway activity differences between females and males were also observed in astrocytes for IL‐10 and IL‐7 signaling. More specifically, astrocytes in female AD subjects showed elevated IL‐10 and IL‐7 signaling activities measured by alzPAS, relative to female control individuals. In male AD subjects, astrocytes show reduced IL‐10 and IL‐7 signaling pathway activities when compared to male control individuals. In summary, these observations reveal distinct microglia‐specific immunometabolism patterns between female and male AD subjects, including both immune pathways (i.e., CLRs and IL‐10 signaling) and metabolic pathways (i.e., lipid metabolism and glutamate and glutamine metabolism).

3.7. Glutamate‐mediated cell‐cell communication in female AD subjects

Using a modified MEBOCOST ⁴⁶ tool, we evaluated metabolite‐mediated cell‐cell communications inferred from the integrated snRNA‐seq data of 313,505 single‐nucleus transcriptomes. We then calculated communication scores between paired cell types based on inferred metabolites and known sensor partners (see the Methods section). We found strong metabolite‐mediated cell‐cell communications from excitatory neurons to astrocytes, inhibitory neurons, and OPCs in both female patients with AD (Figure 5A). However, microglia derived from female AD subjects lose several key metabolite‐mediated cell‐cell communication markers from excitatory neurons when compared to controls. By contrast, males show loss of cell‐cell communication markers from excitatory neurons to microglia in both AD and controls. We also observed that astrocytes in female AD subjects expressed glutamate ionotropic receptors (GRID1, GRID2, GRIA2) and glutamate metabotropic receptors (e.g., GR), as the sensors to high abundance of the extracellular metabolite glutamic acid secreted by excitatory neurons (Figure 5B). The female patients had decreased glutamic acid and glutamate ionotropic receptor mediated communications from excitatory neurons to microglia, which may link severe AD pathogenesis in female individuals. ⁶⁵ As shown in Figure 4, we found decreased glutamate and glutamine metabolic pathway activities in microglia of female AD subjects compared to controls. We also observed a positive female‐specific association between glutamate and P2RX7 expression in AD subjects (Figure 3). P2RX7 is a marker gene of microglia ⁶⁶ and plays a crucial role in NOD‐like receptors signaling. ⁶⁷ In summary, these findings suggest that glutamate may serve as a potential immunometabolic sensor in AD severities and disease progression in a female‐specific manner.

FIGURE 5.

Metabolic cell‐cell communication network alterations show sex differences in AD. (A) Metabolic cell‐cell communications networks. The pie plots display the proportion of metabolite signals or sensors generated by each cell type. The edges connecting the cell types indicate metabolite‐mediated cell‐cell communications with a significant level of p < 0.05 (permutation test 1000 times, see Methods), and the thickness of the edges reflects the total number of metabolite–sensor connections between paired cells (number of communications > 10). The check mark in the lower right corner of the table indicates the inferred metabolite‐sensor communications in those subgroups, and the colored arrows on the right side of the table represent the corresponding edges in the network. The colors of the edges indicate metabolite‐mediated cell‐cell communications in AD or control groups in a sex‐specific manner. (B) The metabolite‐sensor communications were estimated among microglia, astrocytes, and ex‐neurons. The color of edge corresponds to the same subgroups as shown in panel A. The direction of edge from cell (red diamond node) to metabolite (red circle node) indicates that cells secrete metabolites inferred by the combined snRNA‐seq data (see Methods). The direction of edges from cell to sensor (green circle node) indicates that cells express sensor‐coding genes calculated by the combined snRNA‐seq data. The connection between metabolites and sensors based on prior knowledge of metabolite‐sensor partners in the MEBOCOST (see the Methods section)

4. DISCUSSION

We comprehensively investigated the differences in sex‐specific immunometabolism in AD using a network‐based multi‐omics analytic framework (Figure 1). Sex‐specific alterations were observed at both whole brain and single‐cell levels, including: (1) sex‐biased differential metabolites (e.g., N‐acetylglutamate, GPC, 1‐stearoyl‐2‐oleoyl‐GPE [18:0 / 18:1]) and associated immune genes (P2RX7, PTPN5, PAK1,TRIM5) in the brain cortex region (Figures 2 and 3); (2) sex differences associated with metabolic and immune pathway activity across seven cell types (Figure 4), such as elevated CLRs pathway activity and reduced glutamate and glutamine metabolism in microglia; (3) female‐specific glutamate‐mediated cell‐cell communications in AD (Figure 5). For example, we observed decreased markers of communications between ex‐neurons and microglia facilitated by glutamic acid in female AD subjects. These findings obtained from metabolomics and bulk‐RNA analysis are further supported by snRNA transcriptomes, as well as existing evidence concerning sex differences associated with AD pathology.

We also found that microglia showed sex differences in immunometabolism in AD through snRNA‐seq data. For example, female AD cortex showed a higher proportion of microglia compared to controls (p = 0.034 and FDR = 0.275) although it is not highly significant. One possible explanation by lack of highly significant may be caused by overall low microglial populations in the currently available datasets. ⁶⁸ Microglia and astrocytes in female AD subjects exhibited elevated IL‐10 signaling activity quantified by alzPAS compared to controls. However, male AD subjects have decreased IL‐10 signaling activity compared to male controls. IL‐10 exhibits an anti‐inflammatory effect in macrophages. ⁶⁹ This effect is mediated by metabolic reprogramming of glycolysis and tricarboxylic acid cycle, facilitating mitophagy to remove dysfunctional mitochondria. ¹⁸ , ⁶⁹ Blocking IL‐10 anti‐inflammatory pathway has been shown to improve phagocytic microglia activity and limit cerebral amyloidosis in AD mouse models. ⁷⁰ Therefore, it is crucial to consider sex‐specific differences in IL‐10 anti‐inflammatory responses when developing future therapeutic strategies for AD.

The immunometabolism endophenotype heterogeneities also showed a sex‐specific pattern in AD from metabolites, pathways, and cell types. We found that decreased abundance of N‐acetylglutamate and N‐acetyl‐aspartyl‐glutamate (NAAG) were observed in consistent with 10,000 PSM cohorts in females with AD (Figure 2 and Figure S2). The microglia displayed decreased alzPAS value in glutamate and glutamine metabolism in females with AD, which potentially influences cell‐cell communications between microglia and ex‐neuron (Figure 4). The fluoroglutamine PET imaging analyses has showed male gliomas significantly higher glutamine uptake, and male murine astrocytes exhibited more sensitive to glutaminase 1 (GLS1) inhibition. ⁵⁹ , ⁶¹ These findings provide preliminary evidence that N‐acetylglutamate may offer a potential sex‐specific biomarker for AD.

GPC levels significantly increased in both female (p = 4.39 × 10⁻¹⁰) and male (p = 0.005) AD subjects compared to controls. However, the association of GPC with immune genes and pathways displayed a sex‐specific pattern in AD. In female AD subjects, GPC was associated with IKBKB, IRAK2, NFKBIA, PTPN5, P2RX7, and PAK1 expression. These genes are crucial to regulate innate immune pathways, including the nuclear factor kappa‐B (NF‐κB) related Toll‐like receptor pathway, NOD‐like receptors signaling, IL‐1 signaling, and CLRs pathways (Figure 3). The previous finding has demonstrated that the sex hormone estrogen modulated NF‐κB induced neuroinflammation in AD progression. ⁷¹ , ⁷² , ⁷³ Our finding indicates that estrogen mediates the communication between ex‐neuron and microglia in AD subjects (Figure 5B). In particular, microglia in female AD subjects showed a reduction in alzPAS related to the metabolism of steroid hormones and innate immune pathways, such as CLRs pathways and IL‐6 family signaling, but no alzPAS alteration were observed in males. However, further research is needed to investigate the relationship between estrogen and sex‐specific metabolites such as GPC and 1‐stearoyl‐2‐oleoyl‐GPE (18:0 / 18:1).

We acknowledge several potential limitations. The ROSMAP cohort included almost twice the number of female subjects compared to males, a ratio closely mirroring the epidemiological data on AD. ³ To rigorously examine the sex differences within this cohort, we conducted 10,000 bootstraps with 1:1 female to male PSM to examine the distribution of sex differences across different AD pathological subgroups after adjusting age, PMI, education years, and APOE4 genotypes. Our analysis revealed that females exhibit a more severe burden of AD pathology in PSM balanced cohorts as well (Figure S1). Our used ROSMAP cohort is consistent with previous studies that women exhibit more neuropathological markers than men, including higher Braak stages for tangles and elevated amyloid scores. ⁴⁷ , ⁴⁸ , ⁴⁹ We also note that all datasets utilized in this study were generated from human samples, and experimental validation is needed in future studies of sex‐specific metabolites, like metabolites the GPC, and glutamate metabolism, in independent human cohorts or AD transgenic mouse models. Last, it is known that genetic factors, including partial X chromosome inactivation and sex hormones, impact sex differences in AD, ⁸ , ⁷⁴ and we were not able to measure these parameters in the current datasets.

In conclusions, differences in immunometabolism regulation appear to play an important role in biological sex‐specific manifestations of AD, in particular in microglia‐specific manner. This new understanding could help improve development of therapeutics, clinical trials, and treatment of patients with AD.

CONFLICT OF INTEREST STATEMENT

Dr Cummings has provided consultation to AB Science, Acadia, Alkahest, AlphaCognition, ALZPathFinder, Annovis, AriBio, Artery, Avanir, Biogen, Biosplice, Cassava, Cerevel, Clinilabs, Cortexyme, Diadem, EIP Pharma, Eisai, GatehouseBio, GemVax, Genentech, Green Valley, Grifols, Janssen, Karuna, Lexeo, Lilly, Lundbeck, LSP, Merck, NervGen, Novo Nordisk, Oligomerix, Ono, Otsuka, PharmacotrophiX, PRODEO, Prothena, ReMYND, Renew, Resverlogix, Roche, Signant Health, Suven, Unlearn AI, Vaxxinity, VigilNeuro pharmaceutical, assessment, and investment companies. Dr Leverenz has received consulting fees from Vaxxinity and grant support from GE Healthcare. The other authors have declared no competing interests. Author disclosures are available in the supporting information.

CONSENT STATEMENT

All human subjects used in this study have been adequately informed consent by the Religious Orders Study / Memory and Aging Project (ROSMAP) cohort. The consent was not necessary for this study.

Supporting information

Supporting information

Supporting information

Supporting information

Hou Y, Caldwell JZK, Lathia JD, et al. Microglial immunometabolism endophenotypes contribute to sex difference in Alzheimer's disease. Alzheimer's Dement. 2024;20:1334–1349. 10.1002/alz.13546