Impacts of host genetics on gut microbiome composition in Alzheimer’s disease

Abstract

Background

Host–microbiome interactions play essential roles in the development of Alzheimer’s disease (AD), yet the host genetic impacts on gut microbial alterations in AD remain poorly understood.

Results

Here, we simultaneously profiled host genotype and gut microbiome in 252 Chinese individuals with varying degrees of cognitive disability. Using the latent Dirichlet allocation topic model, we identified the Anaerostipes-enriched enterosignature (ES-Ana) at the microbial subgroup level as significantly negatively associated with cognitive disability, which could be recapitulated in external cohorts. With the whole-genome sequencing data, we performed microbiome genome-wide association studies for the ES-Ana relative abundance. We prioritized 41 lead genetic variants and confirmed that the high ES-Ana relative abundance showed a negative correlation with the polygenic risk score of AD, indicating its protective effect against AD. Furthermore, we identified 174 ES-Ana-associated genes, which are enriched in AD-related biological functions and phenotypes, and exhibite pervasive underexpression in glial cells during brain aging.

Conclusions

In summary, our study reveals the complex genetic effects on the gut microbiota in AD, and provides novel evidence for the roles of the gut–brain axis in AD.

Download video file ^{(35.1MB, mp4)}

Video Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s40168-026-02342-8.

Introduction

Alzheimer’s disease (AD), a prevalent neurodegenerative disorder, stands as the primary cause of dementia among the elderly [1]. AD is marked by distinctive pathological and clinical hallmarks, including amyloid-β (Aβ) deposition, neurofibrillary tangles composed of pathological tau proteins, and a gradual deterioration of cognitive function [2, 3]. In addition, AD could be influenced by both host genetic variations [4–12] and dysbiosis of gut microbiota [13–16]. Thus, revealing the genetic effects on gut microbial alterations in AD can improve our understanding of the complex host-microbiome interactions in AD pathogenesis and enhance AD prevention, early diagnosis, and targeted treatment by non-invasive means [17–20].

Gut microbiota have been supposed to modulate brain function and behavior via the microbiota-gut-brain axis, an intricate network of neural, immune, endocrine, and metabolic pathways [21, 22]. For example, an increase in gram-negative bacteria or a decrease in butyrate-producing bacteria has been shown as critical for the dysfunction of the blood–brain barrier and intestinal barrier [23, 24]. The dysfunction could further lead to the release of lipopolysaccharides and amyloids, which may trigger microglial activation in the brain, and contribute to the generation of proinflammatory cytokines related to systemic inflammation and neuroinflammation in AD pathogenesis [25–30]. However, the host genetic impacts of gut microbial alterations in AD remain unclear.

Gut microbial composition could be affected by host genetic variation [31–34]. Recently, the largest microbiome genome-wide association studies (GWAS) from the MiBioGen consortium [35] showed that the genetic variants at or near the lactase gene could influence the gut microbial abundance of Bifidobacterium, which is related to milk consumption [36, 37]. Furthermore, Rühlemann et al. showed that ABO and FUT2 genotypes could influence the gut microbial abundances of Bacteroides and Faecalibacterium spp. [36, 38, 39]. Although the effects of host genetic variation on gut microbial signatures have been studied in natural cohorts [40–42], our understanding of the GWAS analysis for gut microbial alterations in case-control cohorts [43, 44] is still limited, which is necessary for deciphering the host genetic effects on gut microbial modules in AD pathogenesis.

Most of the current studies basically revealed that host genetic variation could influence microbial signatures at the community level or single taxon level [31, 37, 38]. However, the alterations of gut microbial subgroup signatures [45–48] and their interplays with host genetic variation in human diseases have not been fully explored. Previous studies reported that host genetic variants could regulate specific keystone taxa [49], the key drivers of microbiome structure and function identified by co-occurrence network analysis [50–53]. In addition, a single genetic variant from the host was shown to have a significant association with the hubs of the gut microbial co-occurrence module, which comprised five genera; however, there was no significant association between the genetic variant and the total abundance of all thirteen genera in the co-occurrence module [54]. These observations highlight the hypothesis that host genetic variation may have effects on specific key microbes in the community instead of the whole, and inspire us to identify the microbial subgroup signatures implicated in AD.

In this study, we simultaneously profiled the gut microbiome and host genotype in a Chinese cohort of 252 individuals with varying degrees of cognitive impairment. We characterized the microbial subgroup signatures (named enterosignatures hereafter) using the latent Dirichlet allocation (LDA) topic model, and identified the Anaerostipes-enriched enterosignature (ES-Ana), which showed significantly negative correlation with cognitive disability. The key genera of ES-Ana could be used for classifying AD patients in several independent cohorts. We then performed GWAS analysis, and identified 41 lead SNPs associated with ES-Ana relative abundances. We found polygenic risk scores for AD were negatively correlated with ES-Ana relative abundances. Furthermore, we prioritized 174 ES-Ana-associated genes, and confirmed that these genes were enriched in the phenotypes and biological functions related to AD. Interestingly, these genes exhibited decreased expression levels in glial cells during brain aging, and pervasive underexpression in glial cells from AD patients. Overall, we reveal the host genetic impacts of ES-Ana relative abundance in AD, and highlight the putative implications of ES-Ana in AD pathogenesis.

Results

Overview of the study

We recruited 551 individuals from two Chinese cohorts, namely the discovery cohort ($n=252$) and the replication cohort ($n=299$), both with gut metagenome and blood whole-genome sequencing data from the same individual (Supplementary Tables S1 and S2; Fig. 1a). The individuals from the two cohorts were evaluated with neuropsychological assessments (“Methods” section), and further separated into five different diagnosis groups: Alzheimer’s disease (AD), Mild Cognitive Impairment (MCI), Subjective Cognitive Decline (SCD), Subjective Cognitive Symptoms (SCS), and Normal Control (NC). The demographic characteristics (e.g., age, gender and BMI) and cognitive scales (e.g., MMSE and MoCA_B) of the individuals from the discovery and replication cohort were summarized in Supplementary Table S3. In our cohorts, there were no dramatic difference in terms of gender and BMI across the five different diagnosis groups, while age of the individuals from different groups showed significant difference (P < 0.05, Kruskal–Wallis test) (Supplementary Table S3). As expected, the mean values of cognitive test scores (i.e., MMSE and MoCA_B) exhibited a significant decrease in SCD, MCI, and AD compared to NC (P < 0.05, Wilcoxon rank sum test) (Supplementary Table S3). In addition, proportions of APOE ε4 carriers showed a significant increase in AD compared to NC (P < 0.05, Fisher’s exact test) (Supplementary Table S3). The genotype APOE4/E4 could be only observed in the MCI (4.84%) and AD group (15.87%) (Supplementary Fig. S1a), consistent with the 8 ~ 12-fold increase of AD risk for APOE4/E4 [5].

Fig. 1.

Study overview. a The distribution of age and gender for individuals from our discovery cohort (n = 252), including five diagnosis groups: Alzheimer’s disease (AD), Mild Cognitive Impairment (MCI), Subjective Cognitive Decline (SCD), Subjective Cognitive Symptoms (SCS), and Normal Control (NC). We collected blood and fecal samples from each individual, followed by performing whole-genome sequencing and gut metagenome sequencing, respectively. The results of cognitive assessments (i.e., MMSE, MoCA_B) were also collected. b Identification of enterosignatures at the microbial subgroup level using the LDA topic model (“Methods” section). One enterosignature associated with progressed cognitive disability was identified and termed as “Anaerostipes-enriched enterosignature” (ES-Ana) hereafter. c We performed GWAS analysis for the ES-Ana relative abundance (i.e., the topic-sample probability), followed by SNP-based and gene-based analyses to reveal the biological implications and genetic relationships of ES-Ana relative abundance in AD. MMSE, Mini-Mental State Examination; MoCA_B, Montreal Cognitive Assessment-Basic version

To explore the putative associations between gut microbial composition and AD risk, we performed integrative analyses on the gut metagenome data and the matched host genotype data of the individuals from the two cohorts. Briefly, as shown in Fig. 1b, we first characterized the microbial enterosignatures using the LDA topic model, and identified the AD-associated enterosignatures at the microbial subgroup level. By incorporating the genotype data, we performed microbiome GWAS analysis to characterize the host genetic impacts of AD-associated enterosignatures, and further reveal biological associations and genetic relationships between the specific enterosignature and AD (Fig. 1c).

Anaerostipes-enriched enterosignature shows associations with cognitive disability

We first calculated gut microbial signatures in our discovery cohort at both community level and single taxon level (“Methods” section). At the community level, we did not identify significant differences in terms of alpha-diversity indexes (i.e., Richness, Shannon, and Simpson indexes) between different disease groups and the NC group (Supplementary Fig. S1b). Similarly, neither beta-diversity nor enterotype distributions (P = 0.165, Fisher’s exact test) exhibited significant differences (Supplementary Fig. S1c–e). At the single taxon level, five genera (i.e., Barnesiella, Enorma, Firmicutes_unclassified, Paraprevotella, Ruminococcaceae_unclassified) could be identified with differential abundance between AD and NC (Supplementary Fig. S2a). Notably, they cannot be replicated with the well-known taxa previously identified as showing differential abundance between AD and NC [55–57]. This low concordance at the single taxon level could be partly due to the sample heterogeneity [58] and the complex multifactorial nature of AD [19] in the gut microbiome analyses.

We then applied the LDA topic model to the gut metagenome data from the discovery cohort, aiming to identify the microbial enterosignatures (“Methods” section) at the microbial subgroup level. In total, we predicted 15 microbial enterosignatures given the smaller perplexity and larger loglikelihood of the LDA topic model (“Methods” section; Supplementary Fig. S3a). We showed their topic-sample probability (i.e., relative abundance) across the diagnosis groups with the progressed cognitive disability from NC to AD (Fig. 2a). Interestingly, we found that only Topic_8 exhibited a sustained decrease in relative abundances in the individuals diagnosed with progressed cognitive disability. Consistently, the relative abundances of Topic_8 showed significantly positive associations with both cognitive scores, i.e., MMSE (beta = 0.113; P = 0.047, Wald test) and MoCA_B (beta = 0.117; P = 0.043, Wald test) (Fig. 2b). In addition, the relative abundances of Topic_8 were significantly lower in the AD group compared to the NC group (beta = − 0.378; P = 0.042, Wald test) (Fig. 2b). These associations between Topic_8 and cognitive scores or diagnosis groups can be recaptured when adding the APOE ε4 stratification and BMI as covariates in the association analyses (Supplementary Fig. S3b). Given that Topic_8 was dominated by the genus Anaerostipes (Supplementary Fig. S2b; Fig. 3a), it was named as “Anaerostipes-enriched enterosignature” (ES-Ana) hereafter.

Fig. 2.

The enterosignature ES-Ana as a distinguishing feature of AD patients. a Median of the topic-sample probability in each diagnosis group. The topic-sample probability of Topic_8 in external validation datasets (i.e., our replication cohort and an independent Japanese cohort) is shown. The points represent the median, and the error bars represent 95% confidence intervals (CIs) of the median estimated by 1000 bootstrap iterations. b Heatmap displaying the associations between topic-sample probability and disease status or cognitive scores. The associations were estimated based on the Dirichlet regression coefficients adjusted by age and gender of the individuals. The circle size indicates the absolute value of regression coefficients. *P < 0.05. c The ROC curves for AD classifiers based on the Topic_8 relative abundances from three metagenomic datasets, with 95% CIs shaded in different colors. The AUC value of each ROC curve is shown

Fig. 3.

Association of the key genera of ES-Ana with AD. a Pie chart showing the fraction of genera probability of ES-Ana. Genera with probability < 0.01 are classified into the “Others” group. b The taxonomic tree showing the eight key genera of ES-Ana. In total, 24 different species from the key genera of ES-Ana are shown in a taxonomic tree (Supplementary Table S6). Each node represents a taxonomic level (from inside to outside: kingdom, phylum, class, order, family, genus, and species). In particular, the genera are labeled with asterisks. The strength of prevalence (blue) and mean abundance (caramel) of each species is colored in the outside circle, respectively. c Heatmap showing the differential abundance (log_2⁡(FoldChange)) of the eight key genera between each disease status compared to NC. d Schematic showing the pyruvate fermentation (BioCyc Id: CENTFERM-PWY) and the acetyl‑CoA fermentation pathways (BioCyc Id: PWY-5676) for butyrate (butanoate) biosynthesis from the MetaCyc database. The key enzymes involved in each step are indicated in green. e Scatter plot showing the positive correlation between ES-Ana relative abundance and the relative abundance of the pyruvate fermentation to butanoate pathway (left), as well as the acetyl-CoA fermentation to butanoate II pathway (right). The transformation for pathways is ${\text{log}}_{10}({10}^{-8}+\text{relative abundance})$. The associations were estimated based on the linear regression coefficients adjusted by the age and gender of the individuals. f The boxplot showing the top 20 genera with the strongest importance for the random forest classifier of AD and NC (“Methods” section). g The AUC values showing the performance of AD classifiers estimated using three independent amplicon datasets (i.e., CA1, CA2, and USA). All classifiers are trained based on the abundance of the eight key genera from ES-Ana

To confirm that ES-Ana was associated with cognitive disability, we re-estimated the ES-Ana relative abundances in our replication cohort and an independent Japanese cohort [59] (Supplementary Table S1) by fixing the genus-topic probability matrix obtained from the discovery cohort (“Methods” section; Supplementary Table S4). We found that individuals from both cohorts with lower ES-Ana relative abundances exhibited progressed cognitive disability, confirming the results from the discovery cohort (Fig. 2a). Furthermore, we used the ES-Ana relative abundances from the discovery cohort to train a logistic regression model for classifying AD and NC individuals (Fig. 2c). We found that the classification model could predict AD patients (AUC = 0.64, 95% CI: 0.52–0.75) with moderate accuracy, and showed a significantly negative relationship between ES-Ana relative abundances and AD status (beta = − 5.717; P = 0.038, Wald test) (Supplementary Table S5). We found that the classification model based on the ES-Ana relative abundance at the microbial subgroup level outperformed the model based on overall genus-level compositions at the single taxon level (AUC = 0.5). Finally, the classification model could achieve moderate performance on the replication cohort (AUC = 0.59, 95% CI 0.47–0.70) and the Japanese cohort (AUC = 0.74, 95% CI 0.47–1) (Fig. 2c), reflecting a certain level of reliability.

Key genera of ES-Ana could distinguish individuals with cognitive disability

Among all genera from the ES-Ana, we identified eight taxa (i.e., Anaerostipes, Dorea, Fusicatenibacter, Blautia, Coprococcus, Adlercreutzia, Eggerthella, Agathobaculum) with the genus-topic probability > 0.01 in the LDA topic model (Fig. 3a) (“Methods” section). To evaluate the reproducibility of the topic structure defining ES-Ana, we reconstructed microbial enterosignatures using gut metagenome data from the replication cohort. In deed, we identified one Anaerostipes-enriched topic (i.e., Topic_10) that could resemble ES-Ana. Notably, 6 out of 7 genera in Topic_10 (Supplementary Fig. S4a) overlapped with the key genera of ES-Ana, including Anaerostipes, Blautia, Fusicatenibacter, Dorea, Eggerthella, and Agathobaculum. In addition, we trained the LDA topic model using the genus-level relative abundance data based on other metagenomic profilers, including MetaPhlAn4 [60] and Kraken2 [61, 62]. As expected, the Anaerostipes-enriched topics could be identified using both additional profilers: 5 (Anaerostipes, Dorea, Blautia, Coprococcus, Eggerthella) out of 8 genera based on MetaPhlAn4 (Supplementary Fig. S4b) and 6 (Anaerostipes, Dorea, Fusicatenibacter, Coprococcus, Eggerthella, Adlercreutzia) out of 14 genera based on Kraken2 (Supplementary Fig. S4c) overlapped with the key genera from ES-Ana, respectively. Collectively, these findings demonstrate that, despite cross-cohort heterogeneity and reference database differences, the topic structure of ES-Ana is robust and generalizable.

Notably, these key genera of ES-Ana have been reported as differential abundance genera for AD or MCI [7, 13, 55, 56, 63–65]. In addition, the eight key genera belong to phyla Actinobacteria and Firmicutes (Fig. 3b), which have been reported as deficiently abundant taxa in the AD group compared to NC [63, 66, 67]. In the discovery cohort, the key genera of ES-Ana showed dramatic and pervasive abundance decrease in the AD group compared to NC (${\text{log}}_2\left(\text{FoldChange}\right)$< 0) (Fig. 3c). Most of them (e.g., Anaerostipes, Dorea, Fusicatenibacter, Blautia, Coprococcus) are butyrate-producing bacteria, which have been revealed as functional in the protective mechanisms of AD. To further investigate the putative associations between butyrate metabolism and ES-Ana, we extracted the biological pathways involved in butyrate biosynthesis (“Methods” section). As expetced, the ES-Ana relative abundance showed significantly positive associations with the primary bacterial butyrate synthesis routes: acetyl-CoA fermentation to butanoate II (beta = 0.126; P = 1.760e-16, linear regression), and pyruvate fermentation to butanoate (beta = 0.028; P = 0.012, linear regression) (Fig. 3d, e). These two pathways respectively involve two key enzymes in butyrate biosynthesis, namely the butyryl-CoA:acetate CoA-transferase (but) and the butyrate kinase (buk) [68] (Fig. 3d). Furthermore, among the 24 species belonging to ES-Ana genera, we identified eight known “but/buk-carrying” species in NCBI Gene database (Supplementary Table S6), such as Anaerostipes hadrus, Blautia producta, Coprococcus comes and Agathobaculum butyriciproducens. These results indicate that the samples with high ES-Ana relative abundance showed increased relative abundances of both the contributing taxa and the associated butyrate biosynthesis pathways, supporting previous evidence that gram-positive bacteria may exert neuroprotective effects by producing butyrate, thereby attenuating neuroinflammation [24, 69].

We next explored the co-occurrence relationships of the key genera from ES-Ana. We found that most of the eight key genera could be clustered in the same module (Supplementary Fig. S2c), where Dorea and Coprococcus were identified as the hub nodes in the AD network. Some genera from about this module showed a modular shift from the AD group to the NC group, confirming their strong AD prediction power and indicating their putative critical roles in AD pathogenesis [70], such as Agathobaculum, Coprococcus, Adlercreutzia.

Next, we identified a set of predictive genera for diagnosing AD patients, and found that the key genera Agathobaculum and Dorea showed the strongest prediction power for AD (Fig. 3f). To validate the association of the key genera of ES-Ana with cognitive disability, we analyzed three external datasets (i.e., CA1 cohort [56]; CA2 cohort [55]; USA cohort [63]) generated by amplicon sequencing (Supplementary Table S1). We developed a Lasso classification framework to evaluate the generalization performance on predicting AD based on the eight key taxa of ES-Ana using these external datasets (“Methods” section). We found that the AD patients could be identified with an average validation AUC value ranging from 0.70 to 0.76 (Fig. 3g), indicating the robustness and generalizability of the key genera from ES-Ana in predicting AD.

Genetic loci associated with ES-Ana relative abundance

We next performed microbiome GWAS to identify the genetic loci associated with ES-Ana relative abundance, followed by post-GWAS analyses to reveal the roles of ES-Ana in AD pathogenesis through host-microbiome interactions. Initially, we retained ~ 6.15 million SNPs after quality controls in our discovery cohort (“Methods” section). A total of 207 SNPs were significantly associated with the ES-Ana relative abundance under the suggestive significance level of ${10}^{-5}$ (Supplementary Table S7; Fig. 4a). Although the limited sample size could lead to reduced statistical power, both the Q-Q plot and the genomic inflation factor (${\lambda }_{\mathit{GC}}=0.987$) suggest that systematic bias and inflation has been minimized in our GWAS analysis (Supplementary Fig. S5a). Following linkage disequilibrium (LD)–based clumping (LD $r^2$ < 0.1), we finally identified 41 lead SNPs (Supplementary Table S8; Fig. 4a). Furthermore, we performed meta-analysis for ES-Ana relative abundance by integrating the GWAS summary statistics from the discovery cohort (suggestive significance level of P < ${10}^{-5}$) and replication cohort (significance level of P < 0.05 and the effect size in the same direction). We found that 61% (25 out of 41) of the lead SNPs can be recovered in the replication cohort, in which 64% (16 out of 25) SNPs showed effect size in the same direction (Supplementary Table S8). Two SNPs, rs910312 and rs72327387, were prioritized as exhibiting significantly positive associations (P < 0.05, Wald test) with ES-Ana relative abundance in the meta-analysis (Supplementary Tables S8 and S9). The relatively low replication rate could be primarily attributable to the limited sample size and the substantial cohort heterogeneity, which have been well-recognized in the microbiome GWAS meta-analyses [37, 41, 71]. Furthermore, we analyzed the difference in ES-Ana relative abundance under distinct genotypes for these two SNPs. For SNP rs910312, the individuals with the minor allele (T), including genotypes TC and TT, had increased ES-Ana relative abundances compared to the individuals who carried the homozygous major allele (CC) in both cohorts (Fig. 4b). While for SNP rs72327387, compared to the individuals carrying the homozygous major allele (GCTTT/GCTTT), those with the heterozygous (G/GCTTT) and homozygous (G/G) minor alleles were likely to have higher ES-Ana relative abundances in both cohorts (Fig. 4c).

Fig. 4.

GWAS analysis of ES-Ana. a Manhattan plot showing host genetic variants associated with ES-Ana relative abundance under the GWAS significance level of 10⁻². The red line represents the suggestive significance level of 1 x 10⁻⁵. The 41 lead SNPs are highlighted. Two SNPs, rs910312 and rs72327387, that can be replicated in the meta-analysis are shown in bold. b Violin plot showing the ES-Ana relative abundance in the individuals from different diagnosis groups, who are homozygous for the major allele (CC), heterozygous (TC), or homozygous for the minor allele (TT) at rs910312 in the discovery (left) and replication (right) cohorts, respectively. c Violin plot showing the ES-Ana relative abundance in the individuals from different diagnosis groups, carrying the homozygous major allele (GCTTT/GCTTT), heterozygous (G/GCTTT), and homozygous minor allele (G/G) at rs72327387 in the discovery (left) and replication (right) cohorts, respectively

To disentangle the SNPs that could influence AD via modulation of ES-Ana from those that affect AD independently, we examined the overlap between the 207 SNPs from GWAS of ES-Ana and those directly associated with AD via the GWAS analysis using data from the discovery cohort comprising AD and NC samples. None of these SNPs reached the suggestive significance level of ${10}^{-5}$ in the AD GWAS (Supplementary Fig. S5b), which could be due to the small sample size of the cohort. Furthermore, we categorized the overlapping SNPs using different significance thresholds (Supplementary Table S10), and still did not observe significant overlap (P > 0.05, hypergeometric test) based on the two GWAS summary results. These findings suggest that only a small subset of the SNPs associated with ES-Ana relative abundance could be directly associated with AD pathogenesis, whereas most likely exert their effects on AD indirectly through modulation of ES-Ana.

In addition, we performed GWAS analyses on the relative abundance of the eight key genera contributing to ES-Ana. Among the SNPs reached the same suggestive significance level of ${10}^{-5}$, there was no overlap between the GWAS SNPs of the ES-Ana abundance and those at the isingle taxon level (Supplementary Table S11). These results highlight that the ES-Ana could be informative in representing the intrinsic gut ecosystem at the microbial subgroup level, rather than merely reflecting single-taxon effects.

To validate the relationships between these significant GWAS SNPs for ES-Ana relative abundance and cognitive disability, we performed the phenome-wide association study (PheWAS) [72] for the 41 lead SNPs identified in the discovery cohort (“Methods” section). Thirty of the 41 lead SNPs were significantly associated with at least one phenotype (P < 0.05) (Supplementary Table S12; Supplementary Fig. S5c). The rs7142308 and rs12955099 showed the highest pleiotropy, both showing significant phenotype enrichment (P < 0.05, hypergeometric test) in the psychiatric and activities domains (Supplementary Fig. S5c). In total, 22 lead SNPs were identified as showing significant phenotype enrichment (P < 0.05, hypergeometric test) in 12 phenotype domains, such as neurological, psychiatric, metabolic, and immunological domains (Supplementary Fig. S5c). Notably, rs910312 and rs72327387, the two SNPs mentioned above (Fig. 4b, c), showed significant phenotype enrichment in the metabolic and cell domains and the neurological domain, respectively (Supplementary Fig. S5c). These results indicate that the significant GWAS SNPs for ES-Ana relative abundance are associated with multiple biological outcomes, especially the phenotypes in the neurological and psychiatric domains.

Complex genetic effects of ES-Ana relative abundance on cognitive disability

To further decipher the genetic relationship between ES-Ana relative abundance and AD, we calculated the polygenic risk scores (PRS) to estimate an individual’s genetic predisposition to AD by integrating the weighted effects of multiple genetic risk alleles. We selected the tag SNPs for ES-Ana relative abundance (Supplementary Table S7) from the AD GWAS summary statistics in our discovery cohort for the PRS calculation (“Methods” section). The PRS association analysis revealed significantly negative correlations between the PRS of AD and cognitive scores, including MMSE (beta = − 2.187; P = 2.768e−05, linear regression; Fig. 5a) and MoCA_B (beta = − 2.666; P = 4.481e−05, linear regression; Supplementary Table S13), as well as the ES-Ana relative abundance (beta = − 0.022; P = 0.003, linear regression; Fig. 5a). These associations between the PRS of AD and cognitive scores could be rediscovered in the replication cohort (Supplementary Fig. 6b; Supplementary Table S13). These results confirmed that a higher PRS of AD was associated with cognitive disability and decreased ES-Ana relative abundance (Fig. 5a; Supplementary Fig. S6a). In the same way, we validated the results using the whole-genome sequencing data rather than extracting tag SNPs for ES-Ana relative abundance (“Methods” section). As expected, we observed that the PRS of AD was significantly negatively associated with cognitive scores and ES-Ana relative abundances using the AD GWAS summary statistics in our discovery cohort (Supplementary Fig. S7a, b; Supplementary Table S13), as well as public AD GWAS summary statistics in the external European dataset (ieu-b-2 [11]; Supplementary Fig. S7c, d; Supplementary Table S13).

Fig. 5.

ES-Ana relative abundance is genetically associated with AD. a Scatter plot showing negative correlation between PRS of AD and ES-Ana relative abundance (top), as well as MMSE (bottom). b The putative SNP-microbe-AD route (left) and SNP-microbe-MMSE route (right) inferred by mediation analysis. The mediation relationships are labeled with red dotted lines. The mediation effect size and its P value are estimated by the bootstrap test with 1000 iterations. Total effect: Path C; indirect effect: Path AB; direct effect: Path C’. c MR results showing the top significantly causal relationships between ES-Ana relative abundance and host phenotypes. The forest plot shows the MR estimates and 95% CIs of the causal effects, as estimated by six different two-sample MR approaches (“Methods” section). *P < 0.05; **P < 0.01; ***P < 0.001

We hypothesized that the lead SNPs from the ES-Ana GWAS summary statistics could have effects on AD pathogenesis by affecting the ES-Ana in the gut microbiome. To this end, we explored the putative routes of SNP-microbe-AD by conducting mediation analyses (“Methods” section). In this analysis, we first focused on the lead SNP rs910312, which showed a significantly positive association with ES-Ana relative abundance in the meta-analysis (Fig. 4b). As expected, we observed increased relationship magnitude between rs910312 and AD-associated phenotypes (i.e., AD, MMSE, and MoCA_B) when incorporating ES-Ana relative abundance, suggesting that ES-Ana relative abundance had a suppression effect [73, 74] on the routes from rs910312 to AD and cognitive test scores (Fig. 5b; Supplementary Fig. S7e; Supplementary Table S14). Among the remaining lead SNPs, we identified two SNPs, rs892579213 and rs1488784096, that showed similar suppression effects on cognitive test scores (Supplementary Fig. S7e; Supplementary Table S14). Our results from the mediation analyses suggest that certain host genetic variation with potential risk may predispose individuals to develop a specific beneficial microbial composition [75, 76], which could help reduce the risk of AD through specific SNP-microbe-AD routes. These routes reflect the complex symbiotic relationships between the host and gut microorganisms. These results also confirm the functionality of ES-Ana relative abundance in the complex interplays between host genetics and AD pathogenesis.

Next, we aim to explore the potential causal effects of ES-Ana relative abundance on the phenotypes associated with AD using two-sample Mendelian randomization (MR) analyses. Notably, to ensure homogeneous genetic ancestry, we leveraged the publically available GWAS summary statistics for 118 phenotypes from the BioBank Japan Project [77] (Supplementary Table S15). Given the intricate mechanisms through which host genetics affect microbiome composition and the limited effect of genetic variants on microbial taxa variability [37], we conducted several sensitivity tests and finally extracted results with P < 0.05 in at least two out of six MR methods or P < 0.05 in the PRESSO method after correcting potential outliers (“Methods” section). We observed significant causal relationships between ES-Ana relative abundance and 10 phenotypes, including 4 diseases (type 2 diabetes, atrial fibrillation, ischemic stroke, and cataract), 2 metabolites (potassium and total bilirubin), and 3 blood pressure traits (diastolic blood pressure, mean arterial pressure, and systolic blood pressure) and body mass index (BMI) (Supplementary Table S15). Higher ES-Ana relative abundance could lead to a lower risk of type 2 diabetes (T2D) (IVW, beta = − 0.142, P = 0.014; PRESSO, beta = − 0.142, P = 0.005) (Fig. 5c). In contrast, ES-Ana relative abundance showed positive effects on diastolic blood pressure (MR-Egger, beta = 0.230, P = 0.034; IVW, beta = 0.082, P = 0.005; PRESSO, beta = 0.082, P = 0.013) (Fig. 5c). These results were in line with previous studies showing that diabetes mellitus, obesity, multiple vascular risk factors, and cardiac dysfunction risk factors (e.g., stroke and cardiovascular disease) could increase the risk of dementia [78–82], and highlighting potential common mechanisms of gut microbes among AD and T2D.

Enriched biological functions and expression dynamics of ES-Ana-associated genes

To identify the ES-Ana-associated genes, we performed positional mapping and gene annotation for the 41 lead SNPs (Supplementary Table S8) from our GWAS summary statistics of ES-Ana relative abundance (“Methods” section). As a result, 174 protein-coding genes were identified based on 440 loci in strong LD ($r^2$ > 0.8) with the lead SNPs (Supplementary Tables S16 and S17). Some well-known AD risk genes such as APOE, TOMM40 and APOC [7, 83, 84], as well as some putative AD-related genes such as FOXG1 [85, 86] were included in the ES-Ana-associated genes. In particular, FOXG1 was mapped to SNP rs910312, one of the lead SNPs that showed significantly positive associations with ES-Ana relative abundance in the meta-analysis (Fig. 4b; Supplementary Fig. S8). FOXG1 has been reported to regulate the neuronal cell cycle [85] and suppress NLRP3-mediated neuroinflammation [86] as an important transcription factor in brain development.

Next, we performed gene-set enrichment analysis to explore the enriched biological functions for these ES-Ana-associated genes. As expected, the ES-Ana-associated genes were significantly (P < 0.05, hypergeometric test) enriched in biological processes and KEGG/Reactome pathways related to sensory perception (particularly olfactory), lipid transport (e.g., cholesterol), and lipoprotein particle clearance (Fig. 6a; Supplementary Table S18). These results were consistent with existing studies showing that olfaction could be a promising sensory biomarker of AD [87], and low-/high-density lipoprotein receptors were associated with decreased Aβ deposition [88, 89]. Furthermore, we leveraged brain-wide and region-specific gene coexpression modules [90] to explore the brain regions affected by the ES-Ana-associated genes. Interestingly, the ES-Ana-associated genes were only significantly (P < 0.05, hypergeometric test) enriched in a brain-wide gene module (Supplementary Table S19). This gene module showed enriched biological functions related to protein dephosphorylation, regulation of macroautophagy, and ion transport in the pathogenesis of AD [91–93] (Supplementary Fig. S9d) and other neuropsychiatric disorders [90]. These results suggest that the ES-Ana-associated genes impact AD pathogenesis in a brain-wide manner.

Fig. 6.

Enriched biological functions and expression dynamic patterns of ES-Ana-associated genes. a Enriched GO terms (biological processes) and Reactome terms of the ES-Ana-associated genes. The black dashed line corresponds to an FDR threshold at 5% level. The colors represent the gene set overlapping estimated by Jaccard index. b Expression dynamics of ES-Ana-associated genes in four different cell types. The LOESS plots show smooth curves with 95% CIs. S8, birth ≤ age < 6 months; S9, 6 months ≤ age < 1 year; S10, 1 ≤ age < 6 years; S11, 6 ≤ age < 12 years; S12, 12 ≤ age < 20 years; S13, 20 ≤ age < 40 years; S14, 40 ≤ age < 60 years; S15, 60 years +. c Top enriched phenotypes of the ES-Ana-associated genes. The black dashed line corresponds to an FDR threshold at 5% level. The colors represent the gene set overlapping estimated by Jaccard index. d Specific ES-Ana-associated genes show differential expression (AD vs. NC) in different cell types. Upset plot shows the overlap of these ES-Ana-associated genes across cell types. e Heatmap showing the differential expression (log_2⁡(FoldChange)) of the ES-Ana-associated genes in different cell types between AD and NC. Ast, astrocyte; Mic, microglia; Oli, oligodendrocyte; OPC, oligodendrocyte progenitor cell; ExN, excitatory neuron; InN, inhibitory neuron. *FDR < 0.05,**FDR < 0.01, ***FDR < 0.001

We further characterized the spatiotemporal expression dynamics of ES-Ana-associated genes using scRNA-seq datasets from the healthy human brain [94]. We found that the ES-Ana-associated genes were significantly (P < 0.001, Wilcoxon rank sum test) highly expressed in glial and neural progenitor cells in the adolescent compared to the elder (Supplementary Fig. S9b). Consistently, the ES-Ana-associated genes showed falling expression levels in glial and neural progenitor cells during brain aging (Fig. 6b). As expected, the ES-Ana relative abundance from the NC group also exhibited a decrease during aging (Supplementary Fig. 9c). In particular, the ES-Ana-associated genes were enriched in microglia-specific down-regulated differentially expressed genes (Supplementary Fig. 9a), highlighting the intricate spatiotemporal dynamics of microglia in AD progression or brain aging [95–97]. These results indicate that ES-Ana may play roles in AD pathogenesis by influencing glial cells, supported by previous studies showing that the altered gut microbiome drives AD pathogenesis primarily through modulating microglial functions and activation [98, 99].

Finally, we showed the differential expression patterns of ES-Ana-associated genes in AD patients compared with NC. We confirmed that the ES-Ana-associated genes were significantly (P < 0.05, hypergeometric test) enriched in the gene sets associated with AD-related phenotypes (Fig. 6c; Supplementary Table S18) such as Aβ levels and pathologic tau levels. Furthermore, we explored the differential expression of the ES-Ana-associated genes in AD using a recently published snRNA-seq dataset [100]. We identified 28 ES-Ana-associated genes that were differentially expressed in more than one cell type in AD (Supplementary Table S20; Fig. 6d, e) (“Methods” section). These genes showed pervasive underexpression in glial cells from AD patients, indicating the key role of glial cells in host-microbiome interactions in AD pathogenesis. APOE, the well-known gene that influences AD risk [101, 102], showed differential expression in four out of the six cell types, most of which were glial cells (Fig. 6d). Other genes, including TENM2 [103, 104], ZNF521 [105], CAMK2D [106], that showed differential expression in more than two cell types have been reported to be functional in regulating neuronal communication, neurodevelopment, and signal transduction pathways (Fig. 6d). Furthermore, FOXG1, linked to SNP rs910312 (Fig. 4b; Supplementary Fig. S8), showed pervasive underexpression in more than half of the cell types from AD patients (Supplementary Table S20). Importantly, the observed underexpression of FOXG1 in the microglial cells is consistent with the hypothesis that FOXG1 could be a potential therapeutic target for AD [86]. In summary, these results support our observations showing that the ES-Ana-associated genes are implicated in AD pathogenesis.

Discussion

AD could be caused by both host genetic variations and gut microbial alterations. Unraveling the host genetic architecture of gut microbial alterations in AD can shed light on the complex functionality of host genetic variation on the gut microbiome in AD patients. To this end, we simultaneously profiled the gut microbiome and host genotype in a Chinese cohort of 252 individuals with varying degrees of cognitive impairment. We identified an Anaerostipes-enriched enterosignature (i.e., ES-Ana) that showed a significantly negative correlation with cognitive disability. Notably, we firstly revealed the genetic effects of gut microbial alterations in AD by performing GWAS for the ES-Ana relative abundance. We identified 41 lead SNPs and found that the ES-Ana relative abundance showed significant genetic relationships with AD. Furthermore, we prioritized 174 ES-Ana-associated genes and characterized their enriched biological functions and expression dynamics in the human brain. These genes were significantly enriched in AD-related biological functions and phenotypes. Notably, these genes were underexpressed in glial cells during brain aging and generally down-regulated in glial cells from AD patients.

In this study, the AD-associated microbial enterosignature was identified based on LDA topic model, which is more powerful in discovering the putative microbial modules instead of the individual taxon [46, 107, 108]. By applying the LDA topic model inspired by advancements in enterotypes [109–111], a total of 15 enterosignatures were identified, among which the ES-Ana displayed a consistent negative correlation with the degree of cognitive disability. This negative correlation could be rediscovered in the replication cohort and an independent Japanese cohort. In the case-control studies of AD focusing on gut microbiome, it has been widely recognized that no significant differences at the community level (e.g., alpha/beta-diversity) or few consistent differential abundance taxa at the single taxon level can be found between AD and NC [7, 64, 112]. The poor consistency across cohorts can be attributed to several factors [58], including sample heterogeneity, differences in research design and methodologies, environmental influences (e.g., diet, medications, lifestyle, geographical location), low statistical power with insufficient sample sizes, and the complex multifactorial nature of AD. These confounding factors were assumed to introduce substantial variability in microbial compositions, which could be mitigated by standardized experimental design, more comprehensive cohort metadata collection, and harmonized computational analysis [58].

The key genera of ES-Ana exhibit robust prediction power and generalizability for AD. When applied on three external amplicon-based cohorts, the models based on the eight key genera of ES-Ana can achieve an average AUC of 0.73, outperforming previously reported models (average AUC = 0.54) based on full features [19]. Future studies that combine ES-Ana with well-established AD biomarkers (e.g., amyloid-β and tau) and enable multi-omics integration within the same cohort (e.g., transcriptomics, metabolomics, and cytokine assays [44, 113–115]) can enhance predictive performance and provide novel and mechanistic insights into immune-mediated and metabolic mechanisms underlying host–microbiome interactions in AD.

As a complex ecosystem, the human gut microbiome could be involved in regulating host metabolism, immunity, and health. In our results, most of the key genera from ES-Ana (e.g., Anaerostipes, Dorea, Fusicatenibacter, Blautia, Coprococcus) are gram-positive bacteria, which can produce short-chain fatty acids (SCFAs). The SCFAs, including butyrate, acetate, and succinate, can be produced through bacterial fermentation in the gut [24], which are implicated in the inhibition of neuroinflammation [24, 69]. Interestingly, increasing the host butyrate levels through enhancing butyrate-producing bacteria has been proposed as a novel AD treatment strategy [20, 116–119]. For example, the increased abundance of Faecalibacterium prausnitzii by a long-term adherence to the Mediterranean diet [118] or direct oral administration could increase butyrate levels and improve cognitive function supported by mice models [59]. Although ES-Ana–associated genera showed biological relevance with AD pathogenesis, their clinical safety and efficacy as probiotic interventions in humans have not been fully explored. It is a great challenge to translate the microbiome-based associations into clinical applications, as several biological and clinical issues need to be resolved [120, 121], including causality inference, strain-level heterogeneity, host-dependent engraftment, and regulatory considerations.

The key genera from ES-Ana have been reported as showing strong genetic associations with dementia. Recent studies have explored the genetic relationships between individual bacteria and diverse brain diseases, including ASD [44], AD [21], depressive symptoms [122], Parkinson’s disease (PD), and amyotrophic lateral sclerosis (ALS) [123]. The PRS for Adlercreutzia has been shown as negatively associated with AD diagnosis and APOE genotype [21]. Adlercreutzia may modulate host health by secreting metabolites, including Equol, a metabolite that could mitigate microglial inflammation and downregulate apoptosis-related gene expressions [124, 125]. Anaerostipes and Fusicatenibacter have been confirmed to show protective effects against AD, PD, and ALS by MR analyses [123]. Taken together, this existing evidence supports the putative protective effects of the key genera from ES-Ana against AD.

We observed reduced expression of ES-Ana–associated genes in aging glial cells; however, the underlying molecular and cellular mechanism remains unclear. Genetic analyses indicate that higher ES-Ana abundance is associated with lower PRS of AD (Fig. 5a) and lower risk of T2D (Fig. 5c), a well-established risk factor of dementia [80, 81]. We also found that ES-Ana–associated genes are enriched in AD-related biological processes (Fig. 6a, c). Taken together, we hypothesize that ES-Ana–associated genes are implicated in a resilience-related regulatory program in glial cells that may function in regulating central nervous system homeostasis [98, 126] and AD pathogenesis [98, 127]. This regulatory program shows an age-dependent decline of functionality and could be dysregulated in AD.

Interpreting the regulatory circuitry of ES-Ana–associated genetic variants and genes remains challenging. To identify the regulatory target of rs910312, the key variant prioritized by our GWAS and post-GWAS analyses, we queried recently published bulk [128] and cell-type-specific [129–131] expression quantitative trait loci (eQTL) datasets regarding rs910312 and FOXG1. We found that rs910312 was reported as the cis-eQTL of the antisense transcript FOXG1-AS1 in bulk brain tissue [128], as well as LINC01551 in microglia from the superior temporal gyrus and thalamus [129]. Regional association analyses further showed that rs910312 resides within a well-defined linkage disequilibrium block spanning the FOXG1 locus, including FOXG1, FOXG1-AS1, and LINC01551. Together, these results support the hypothesis that rs910312 could modulate the expression activity of the non-coding transcripts near FOXG1 rather than FOXG1 itself in glial cells, where these non-coding transcripts may affect the expression level of FOXG1 transcript.

Our study, while offering valuable insights, has several limitations. First, the small sample size of our discovery cohort made it challenging in identifying GWAS SNPs associated with ES-Ana relative abundance due to low statistical power and evaluating the heritability of these microbial signatures due to large variance. We anticipate that increasing the sample size of cohorts could identify more novel and reproducible genome-wide significant associations. Second, the shotgun metagenomic sequencing technology used in this study cannot efficiently detect the reads derived from the microbes other than bacteria, such as fungi and viruses, which also have putative functions in the host-microbiome interactions in AD patients. Previous studies revealed that the alterations of specific gut fungi [132], as well as fungal-bacterial co-regulation networks [133] could be implicated in the host immunity of the individuals with cognitive disability. Although statistical analyses based on genus-level taxonomic compositions are expected to have improved generalizability and interpretability of the results [46], it is necessary to understand exactly what specific microbial strains within the ES-Ana genera can affect AD risk. Such high-resolution investigations will require the incorporation of comprehensive metagenomic assemblies with sophisticated computational frameworks modeling microbial genetic variation [39, 134, 135]. Furthermore, the associations between gut microbial traits and diverse types of host genetic variants in AD are also worth exploring, which requires a larger sample size of cohorts, more comprehensive profiling of host genetic variation (e.g., structural variation and copy number variation in host), and more powerful computing resources [31, 36, 40, 136].

Conclusions

We first reveal the AD-associated enterosignatures at the microbial subgroup level, and further reveal biological associations and genetic relationships between the enterosignature and AD by integrative analyses on the gut metagenome data and the matched host genotype data. Our study reveals the complex genetic effects on the gut microbiota in AD, and provides novel evidence for the roles of the gut-brain axis in AD.

Methods

Participant recruitment and sample collection

The cohorts in this study were derived from the Zhangjiang International Brain BioBank (ZIB, https://zib.fudan.edu.cn), a platform specializing in multimodal research on brain diseases. The individuals with cognitive disability were recruited as we described in our previous study [137, 138]. To minimize the impact of comorbidities on gut microbial dysbiosis, the individuals with a history of diabetes mellitus, liver disease, and tumors were excluded [138]. The additional recruitment criteria included (i) no antibiotic use within the past 2 months; (ii) no history of other neurological or psychiatric disorders, such as Parkinson’s disease, depression, and epilepsy; (iii) absence of severe physical illnesses. In total, we recruited 252 and 299 Chinese individuals in the discovery cohort and replication cohort, respectively. For the discovery cohort, the sample set consists of 252 fecal samples and paired blood samples from 252 individuals. For the replication cohort, the sample set consists of 299 fecal samples and paired blood samples from 299 individuals. Each individual was subjected to a comprehensive neuropsychological assessment, including an interview for evaluating cognitive decline, six standardized neuropsychological tests in three cognitive domains, global cognition tests and daily cognition evaluation. The individuals from the cohorts were further separated into five diagnosis groups based on the results from the neuropsychological assessment and internationally accepted diagnostic criteria [139]: Normal Control (NC), Subjective Cognitive Symptoms (SCS), Subjective Cognitive Decline (SCD), Mild Cognitive Impairment (MCI), and Alzheimer’s Disease (AD). The summary of the cohorts is provided in Supplementary Table S1. The metadata of each individual in the cohorts can be obtained in Supplementary Table S2. All standardized protocols for sample collection and sequencing were also described in our previous research [137, 138].

To assess the generalizability of our findings, we collected public case–control datasets involving AD patients as external validation cohorts. Finally, we downloaded one gut microbial metagenomic dataset (PRJNA679346 from Ueda et al. [59], Japan cohort), and three 16S rRNA gene amplicon datasets (PRJNA633959 from Lin et al. [56], CA1 cohort; PRJNA489760 from Li et al. [55], CA2 cohort; PRJEB51982 from Vogt et al. [63], USA cohort). Amplicon libraries of the CA1 and CA2 cohorts target the V3–V4 regions of the 16S rRNA gene, while those from the USA cohort target the V4 region. Detailed metadata for these cohorts were summarized in Supplementary Table S1.

Whole-genome sequencing and data preprocessing

Genomic DNA was extracted from 551 peripheral blood samples in the discovery cohort and replication cohort. Whole-genome sequencing (WGS) was performed on the Illumina Novaseq 6000 platform with 150-bp paired-end reads, resulting in a mean sequencing depth of 30.64 $\times$ (range from 19.87 $\times$ to 61.38 $\times$) (Supplementary Table S2). The concrete processing steps were described in our previous research [137]. Briefly, FastQC (v0.11.9) [140] was used to evaluate the quality of raw sequencing data. Subsequently, reads were aligned to the reference genome (hg19) using Sentieon (v202010.02) BWA-MEM algorithm [141, 142]. After removing PCR duplicates and recalibrating base quality scores, the Sentieon Haplotyper algorithm [143] was applied to detect single nucleotide polymorphisms (SNPs) and small insertions and deletions (INDELs) for each genome. A cohort-level vcf file was generated for subsequent analyses. The Apolipoprotein E (APOE) genotypes were determined based on the alleles of APOE SNPs rs7412 and rs429358.

The genetic variants and individuals were further filtered using the following steps [144]: (1) removing SNPs and individuals with missingness $>0.05$; (2) excluding individuals with sex discrepancy; (3) retaining autosomal SNPs; (4) retaining common variants (MAF $\ge 5\text{\%}$); (5) retaining SNPs with Hardy–Weinberg equilibrium P $>{10}^{-6}$; (6) excluding individuals with a heterozygosity rate deviating more than three standard deviations from the mean; (7) excluding related individuals with pihat $>0.2$. All the above steps were implemented using PLINK (v1.9) [145]. Finally, 6,148,383 and 7,112,836 SNPs from the discovery cohort and replication cohort were retained, respectively.

Metagenomic sequencing and data preprocessing

Sequencing libraries were generated with NEXTFLEX Rapid DNA-Seq (Bioo Scientific Corporation). Deep metagenomic shotgun sequencing of 551 fecal samples in two cohorts was performed on the Illumina NovaSeq platform, resulting in the largest Chinese metagenomic dataset with varying degrees of cognitive disability [138]. All samples were paired-end sequenced with a 150-bp read length, aiming to achieve a targeted data size of 10 GB (Supplementary Table S2). Bowtie2 [146] was applied to remove host-derived reads by mapping them to the reference genome (hg38). Trimmomatic [147] was utilized to trim sequences and remove low-quality bases. Next, reads shorter than 75 bps after trimming were removed. MetaPhlAn3 [148] was applied (with default parameters) for taxonomic profiling with the mpa_v30_CHOCOPhlAn_201901 database. Finally, relative abundances of recognizable taxa across clades from phylum to species were estimated. Most analyses in this study were based on genus-level microbial profiles unless stated. Functional profiling of the metagenomic samples was performed using HUMAnN3 [148] with default parameters, including the identification of gene families from UniRef protein databases and the quantification of pathway abundances and coverages from the MetaCyc database.

16S rRNA gene amplicon sequencing and data preprocessing

QIIME2 (version 2021.4) pipeline [149] was used to estimate the microbial profiles generated by amplicon sequencing. Briefly, the raw data were denoised to amplicon sequence variants (ASVs) using DADA2 (version 1.18.0) [150], and the taxonomic assignment was accomplished by classifying ASVs against the Greengenes database (version 13.8) [151]. Finally, the relative abundances of recognizable genera in each sample were estimated.

Identification of microbial signatures at the community level and single taxon level

For the community level, the alpha-diversity indexes (Richness, Shannon, Simpson) were estimated using the diversity function in the vegan R package. The beta-diversity based on Bray–Curtis dissimilarities was estimated using the vegdist function. The community ordination (Principal Coordinate Analysis, PCoA) was performed using the cmdscale function. Subsequently, PERMANOVA test was performed with 999 permutations using the adonis function. We conducted enterotypes classification using a PAM-based clustering approach [109], and identified two enterotypes in the discovery cohort: Bacteroides-dominant enterotype (B_Ent, $n=197$) and Escherichia-dominant enterotype (E_Ent, $n=55$).

For the single taxon level, the differential abundance analysis was performed using the linda function in the MicrobiomeStat R package [152], adjusting for the compositional effects-induced bias and confounding factors, including age and gender. The taxa with $P<0.05$ and $\left|{\text{log}}_2,\left(\text{FoldChange}\right)\right|>1$ were considered as differential abundance taxa between disease status and NC.

Identification of microbial subgroup signatures using LDA topic model

First, we filtered out low-abundance (min-abundance threshold: 0.1%) and low-prevalence (prevalence threshold: 10%) genera in our discovery cohort. Next, the latent Dirichlet allocation (LDA) topic model was used to explore latent modules of the gut microbial community in an unsupervised way. The gut microbial subgroups (or topics) of each sample and microbial reads (or words) of each subgroup were modeled using multinomial distribution in the LDA topic model. Given there were $N$ samples and $T$ latent topics of $M$ genera, the LDA topic model was shown as follows: $Z_{d,i}|{\theta }_d\sim Multinomial\left({\theta }_d\right)$ and$W_{d,i}|Z_{d,i},{\varphi }_{Z_{d,i}}\sim Multinomial\left({\varphi }_{Z_{d,i}}\right)$, where observation$d\in \left\{1,,,\cdots ,,,N\right\}$, genus$i\in \left\{1,,,\cdots ,,,M\right\}$, and topic$t\in \left\{1,,,\cdots ,,,T\right\}$. Here, ${\theta }_d$ was drawn from a Dirichlet distribution parametrized by$\alpha$:${\theta }_d\sim Dirichlet\left(\alpha \right)$; ${\varphi }_t$ was drawn from a Dirichlet distribution parametrized by$\beta$:${\varphi }_t\sim Dirichlet\left(\beta \right)$. The summation of each ${\theta }_d$ and ${\varphi }_t$ vector was 1. The process of generating genus $i$ in topic $t$ from observation $d$ could be analogous to the two-stage sampling process⁴⁷. Briefly, topic $Z_{d,i}$ was first sampled from the observation’s distribution over topics, then word $W_{d,i}$ was sampled from the topic’s distribution over words given topic$Z_{d,i}$. Finally, we utilized the parameters of the topic distribution for each observation (${\theta }_d$) and parameters of the word distribution for each topic (${\varphi }_t$) to estimate the topic-sample probability and genus-topic probability, respectively. In this study, we termed these microbial subgroups (i.e., topics) as “enterosignatures” [46].

The LDA topic model was trained using Gibbs’ sampling using the MetaTopics [107] R package. Notably, we multiplied the entries of the genus-level relative abundance matrix by 1000, followed by rounding into the nearest integers, which were treated as the input counts for the LDA topic model. The LDA topic model with a certain value of topics (i.e., K) was trained using five-fold cross-validation procedure via the selectK() function. We tested a series of K values ranging from 2 to 40 (parameter sets: 2, 5, 10, 20, 25, 30, 35, 40). Finally, we selected $K=15$ given the smaller perplexity and larger loglikelihood of the LDA topic model (Supplementary Fig. S3a), enabling a balance between model fitness and biological explainability.

Analyzing associations between microbial topic probability and host phenotypes

Associations between interested host phenotypes (disease status and cognitive scores) and microbial topic-sample probability were analyzed by the Dirichlet regression models (DirichletReg R package) [153], adjusting for the confounding factors, including age and gender. Beta coefficients and $P$ values for all the associations were estimated. In our results, the Topic_8 relative abundance was significantly related to disease status (i.e., AD and NC) and two cognitive scores (i.e., MMSE and MoCA_B). Therefore, we focused on the Topic_8 in the downstream analysis. Because the Topic_8 was Anaerostipes-enriched according to its genus probability, we termed it as “Anaerostipes-enriched enterosignature” (ES-Ana) hereafter.

Evaluating the performance of AD predictive classifiers based on ES-Ana

We built a logistic regression model for classifying AD and NC using the ES-Ana relative abundance estimated in the discovery cohort. The prediction performance was evaluated by AUC values using the pROC R package. When the external dataset has a similar taxonomic composition to the discovery cohort, the LDA topic model allows us to calculate the topic probability of new observations. Thus, the posterior probabilities of 15 topics for each sample in the replication cohort and an independent Japanese cohort (RJNA679346 from Ueda et al. [59]) were estimated using the topicmodels R package, given the genus-topic probability matrix estimated in the discovery cohort.

To confirm that the eight key genera from ES-Ana (${\varphi }_{topic\_8}$ value > 0.01) could distinguish AD patients, we collected three external amplicon datasets: (i) CA1 cohort from Lin et al. [56] (PRJNA633959), (ii) CA2 cohort from Li et al. [55] (PRJNA489760), (iii) USA cohort from Vogt et al. [63] (PRJEB51982). The LASSO algorithm was used to build the AD predictive classifiers using the SIAMCAT [154] R package, which was described in our previous study [19]. The AUC values of internal and external validation were estimated to evaluate the generalization performance on predicting AD based on the eight key taxa from ES-Ana.

Inferring microbial co-occurrence networks

For the genus-level co-occurrence network in AD and NC groups, SparCC correlation [48] was calculated by the sparcc function in the SpiecEasi R package [155]. The netConstruct function (normMethod = “clr”, sparsMethod = “threshold”, thresh = 0.3) and netAnalyze function (clustMethod = “cluster_fast_greedy”, hubPar = c(“degree”,“closeness”,“eigenvector”)) in the NetCoMi R package [156] were used to construct networks and calculate co-occurrence modules and hub nodes, respectively. The graph objects were exported and visualized by the software Cytoscape (version: 3.9.1) [157].

Microbiome GWAS analysis

The microbiome GWAS was performed on ES-Ana relative abundance in the discovery cohort by PLINK [145], via the linear regression model adjusting for the first five genetic PCs (PC1 ~ PC5), age and gender. A total of 207 SNPs were significantly associated with the ES-Ana relative abundance under the suggestive significance level of ${10}^{-5}$. Following linkage disequilibrium (LD)–based clumping (LD $r^2$< 0.1), we finally identified 41 lead variants. Finally, the Manhattan plot was created by the CMplot R package, and the regional association plot of a single SNP with ES-Ana relative abundance was visualized by LocusZoom 1.4 [158].

Meta-analysis for ES-Ana relative abundance was conducted via a random-effect model using the Restricted maximum likelihood estimator as implemented in metafor R package. The meta-analysis integrated the GWAS summary statistics from the discovery cohort (suggestive significance level of P < ${10}^{-5}$) and replication cohort (significance level of P < 0.05 and the effect size in the same direction as discovery cohort). Consequently, two SNP exhibited significantly positive associations with ES-Ana relative abundance (P $<0.05$).

PRS analysis

Polygenic risk scores (PRS) were calculated to estimate an individual’s genetic predisposition to AD by integrating the weighted effects of multiple genetic risk alleles using PRSice-2 software [159]. The PRS computation considered a range of P value thresholds for AD GWAS significance level (${5\times 10}^{-8},{1\times 10}^{-5},{5\times 10}^{-5},0.0001,0.0005,0.001,0.005,0.01,0.05,0.1,0.5,1$), and extracted the tag SNPs for ES-Ana relative abundance alongside LD clumping parameters (default: –clump-kb 250 kb –clump-p 1.0 –clump-r2 0.1). Next, the “best-fit” PRS capturing the largest average phenotypic variance $R^2$ were retained, via a linear regression model adjusting for the first five genetic PCs (PC1 ~ PC5), age and gender. Finally, scatter plots reveal the correlations between PRS of AD and ES-Ana relative abundances or cognitive scores, with a linear fit by stat_smooth function in the ggplot2 R package.

Mendelian randomization analysis

Two-sample Mendelian randomization (TSMR) was used to infer the causal relationship between ES-Ana relative abundance (exposure) and diverse AD-related phenotypes (outcomes). To ensure homogeneous genetic ancestry, i.e., East Asian populations, we leveraged the publicly available GWAS summary statistics for 118 phenotypes from BioBank Japan Project [77], including common diseases (auto-immune, cardiovascular, metabolic) and routine blood parameters. The individuals from our discovery cohort (i.e., the source of exposure) have no sample overlap with the cohort from Biobank Japan Project (i.e., the source of outcomes).

The SNPs with GWAS P $<{10}^{-5}$ and LD $r^2$< 0.01 within a 5,000 kb window in our GWAS summary statistics for ES-Ana relative abundance (exposure) were selected as the instrumental variables. We then discarded the instrumental variables if proxy (LD $r^2$ > 0.8) was unavailable in public GWAS summary statistics for one specific phenotype (outcome). TSMR analysis was performed using five methods in the TwoSampleMR R package (https://github.com/MRCIEU/TwoSampleMR): MR-Egger regression [160], weighted median [161], inverse-variance weighting (IVW), mode-based estimate (MBE) including Simple mode and Weighted mode. To exclude false positive relationships, we conducted sensitivity tests and retained the consistent results from multiple methods. Cochran Q statistics were used to assess the global heterogeneity of selected SNPs, and MR-Egger regression was used to capture horizontal pleiotropy. Furthermore, the MR pleiotropy residual sum and outlier (MR-PRESSO) [162] global test was used to estimate the presence of directional pleiotropy, and correct potential outliers via MRPRESSO R package. Finally, we excluded the relationships with significant directional horizontal pleiotropy based on either MR-Egger intercept test or MR-PRESSO test. The causal relationships were identified as those with P $<0.05$ in at least 2 out of 6 MR methods or P $<0.05$ in the PRESSO method after potential outlier correction. MR forest plots uncovering the causal effects of ES-Ana relative abundance on various diseases were generated by the forester R package.

Mediation analysis

The mediation effects of putative SNP-microbe-phenotypes (i.e., disease status, MMSE, MoCA_B) routes were estimated using the mediate function in the mediation R package [163], with 1000 bootstrap resampling iterations. Except for the indirect path, the effect size and P value of the remaining paths were estimated by the linear regression model, adjusting for the confounding factors, including the first five genetic PCs (PC1 ~ PC5), age, and gender.

SNP-based PheWAS analysis

To explore the pleiotropy of the 41 lead SNPs from GWAS summary statistics of ES-Ana relative abundance in the discovery cohort, we performed SNP-based Phenome-wide association study (PheWAS) lookups in the GWASATLAS⁷² for lead SNPs. This resource collected 4,756 GWAS summary statistics for 4155 traits across 28 diverse domains, including metabolic, psychiatric, immunological, and neurological domains. Next, enrichment analysis for each SNP was performed using hypergeometric tests, with fold enrichment to estimate the degree of overlap between two sets. Finally, only SNPs and phenotype domains with at least one significant enrichment signal (FDR < 0.05) were shown.

Identification of ES-Ana-associated genes

Positional mapping and gene annotation were performed based on 440 loci in strong LD ($r^2$ > 0.8) with the 41 lead SNPs from the GWAS summary statistics. Each locus was simply linked to its nearby genes within a maximum distance of 250 kb. As a result, 174 ES-Ana-associated genes were identified.

Functional enrichment analysis of ES-Ana-associated genes

The GENE2FUNC function performed hypergeometric tests using all 174 mapped genes with the FUMA (version: 1.5.2) [164], to assess the enrichment of all mapped genes in other gene sets curated from various sources (e.g., MsigDB [165]). We reported all the enriched terms from KEGG [166], Reactome [167], Gene Ontology [168], as well as the enriched phenotypes from GWAS catalog [169] with FDR < 0.05.

Gene expression analysis of ES-Ana-associated genes

We identified a set of differentially expressed genes (DEGs) in each of the 18 cell types from healthy human brains using the scRNA-seq data from STAB2 [94]. The DEGs were defined based on the following criteria: (i) with FDR < 0.05 in the Wilcoxon test; (ii) with $\left|{\text{log}}_2,\left(\text{FoldChange}\right)\right|\ge 1.5$.

To investigate the differential expression patterns of the ES-Ana-associated genes in AD brains, we leveraged a recently published snRNA-seq dataset of prefrontal cortex samples from 15 late-stage AD and 48 non-AD individuals [100]. We identified DEGs in seven cell types using the Scanpy [170] framework, including astrocyte, excitatory neuron, inhibitory neuron, microglia, oligodendrocyte, oligodendrocyte progenitor cell, and vascular cell. We identified a set of DEGs in each cell type based on the following criteria: (i) with FDR < 0.05 in the Wilcoxon test; (ii) with $\left|{\text{log}}_2,\left(\text{FoldChange}\right)\right|\ge 0.5$.

Authors’ contributions

X.M.Z. conceived the project; J.X.L., J.X.C., L.H.J., Z.Q.G., and F.C. performed raw data quality control and preprocessing; J.X.L. analyzed the data; X.Z.Z., A.Y.Y., and S.Y.L. contributed in data analysis; J.X.L. and Y.T.Y. interpreted the results; J.X.L., Y.T.Y., and X.M.Z. prepared the manuscript. All authors read and approved the final manuscript.

Funding

This work was partly supported by National Natural Science Foundation of China (T2225015, 62433008), Shanghai Science and Technology Commission Program (23JS1410100, 24JS2810100), Shanghai Municipal Education Commission (No.24KXZNA11), Key Science and Technology Project of Hainan Province (ZDYF2024SHFZ058), Major Project of Guangzhou National Laboratory (GZNL2024A01003), National Key R&D Program of China (2023YFF1204800). The computations in this study were performed on the CFFF platform at Fudan University.

Data availability

The gut metagenome datasets of the discovery and replication cohorts in this study have been deposited to National Genomics Data Center under the accession number PRJCA022804. The gut metagenome and amplicon datasets of external cohorts were accessed at the Sequence Read Archive under the accession number of PRJNA679346, PRJNA633959, PRJNA489760 and PRJEB51982 (Supplementary Table 1). The GWAS summary statistics for ES-Ana relative abundance in this study have been deposited to the Genome Variation Map (GVM) [171] in National Genomics Data Center, Beijing Institute of Genomics, Chinese Academy of Sciences and China National Center for Bioinformation [172] under the accession number GVP000086. The public GWAS summary statistics for AD was obtained from the IEU OpenGWAS project (https://gwas.mrcieu.ac.uk/datasets/) with GWAS ID: ieu-b-2. The codes are available in GitHub via https://github.com/ZhaoXM-Lab/microbiome-GWAS.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the Shanghai Sixth People's Hospital (Approval NO: 2019–032). Written informed consent was collected from all participants to use their clinical data for research purposes. All methods in this study were performed in accordance with the relevant guidelines and regulations.

Consent for publication

All authors reviewed the manuscript and consent to publication.

Competing interests

The authors declare no competing interests.