Abstract

   Alzheimer’s disease (AD) is a metabolic disorder. Discovering the
   metabolic products involved in the development of AD may help not only
   in the early detection and prevention of AD but also in understanding
   its pathogenesis and treatment. This study investigated the causal
   association between the latest large-scale plasma metabolites (1091
   metabolites and 309 metabolite ratios) and AD. Through the application
   of Mendelian randomization analysis methods such as inverse-variance
   weighted (IVW), MR-Egger, and weighted median models, 66 metabolites
   and metabolite ratios were identified as potentially having a causal
   association with AD, with 13 showing significant causal associations.
   During the replication validation phase, six metabolites and metabolite
   ratios were confirmed for their roles in AD: N-lactoyl tyrosine,
   argininate, and the adenosine 5’-monophosphate to flavin adenine
   dinucleotide ratio were found to exhibit protective effects against AD.
   In contrast, ergothioneine, piperine, and 1,7-dimethyluric acid were
   identified as contributing to an increased risk of AD. Among them,
   argininate showed a significant effect against AD. Replication and
   sensitivity analyses confirmed the robustness of these findings.
   Metabolic pathway analysis linked “Vitamin B6 metabolism” to AD risk.
   No genetic correlations were found, but colocalization analysis
   indicated potential AD risk elevation through top SNPs in APOE and
   PSEN2 genes. This provides novel insights into AD’s etiology from a
   metabolomic viewpoint, suggesting both protective and risk metabolites.

Supplementary Information

   The online version contains supplementary material available at
   10.1038/s41598-024-77921-6.

   Keywords: Mendelian randomization, Plasma metabolites, Alzheimer’s
   disease, Genome-wide association study, Metabolic pathway

   Subject terms: Neurological disorders, Genetics, Neuroscience,
   Biomarkers, Neurology, Pathogenesis, Risk factors

Introduction

   Alzheimer’s disease (AD), a prevalent neurodegenerative disorder, is
   marked by gradual and covert onset, progressive exacerbation of memory
   and cognitive dysfunctions, and behavioral and motor
   disturbances^[34]1. Globally, an estimated 416 million individuals are
   within the AD spectrum, of which around 315 million are in the
   preclinical stage of AD, 69 million are experiencing prodromal AD, and
   32 million have progressed to AD dementia^[35]1. Although numerous
   studies have explored the pathogenesis of AD, its complex mechanisms
   are not yet fully understood, making it challenging to develop
   effective treatments. Thus, investigating the pathogenesis of AD and
   dependable biomarkers to establish effective prevention and treatment
   approaches is essential. Current prevailing hypotheses include the
   amyloid-beta protein (Aβ) cascade hypothesis, tau protein hypothesis,
   neuroinflammation hypothesis, oxidative stress hypothesis, cholinergic
   hypothesis, metal ion imbalance hypothesis, and mitochondrial
   dysfunction hypothesis^[36]2,[37]3.

   Optimal biomarkers would greatly facilitate the early detection of AD
   risk; however, this is currently a challenge due to the limited number
   of approved biomarkers. Recent studies have revealed that AD is a
   metabolic disorder that causes disturbance in energy metabolism through
   significant reductions in brain glucose uptake and metabolism^[38]4.
   Therefore, exploring metabolic products related to the development of
   AD can aid not only in the early screening and prevention of AD but
   also in understanding its pathogenesis and treatment. Van der Velpen et
   al. found that in the plasma of AD patients, there was an increased
   concentrations of various intermediates from the TCA cycle and
   beta-oxidation, in contrast to a notable decrease in amino acid
   levels^[39]5. Kim et al. found that amino acids and lipokines have
   significant potential in deciphering AD pathogenesis. They suggested a
   blood-based predictive insight in which these specific metabolites were
   associated with pathology as indexed by CSF AD biomarkers^[40]6. Varma
   et al., using blood samples from the prodromal phase, found a
   persistent association between elevated levels of sphingolipids in the
   blood and the extent of AD pathology^[41]7. Zhu et al.^[42]8 using
   genetic instruments, identified 69 multiple AD risk-associated
   proteins. Nonetheless, these studies are constrained by several
   factors, such as limited participant numbers, a restricted range of
   metabolites analyzed, the ability to establish correlations, and
   vulnerability to confounding elements and reverse causality. These
   factors make it challenging to establish a causal relationship between
   AD and plasma metabolites.

   Genetic epidemiology emphasizes the application of Mendelian
   randomization (MR) as an important technique. By using genetic variants
   as instrumental variables (IV), MR offers an analytical approach that
   overcomes the limitations of conventional observational studies,
   particularly those caused by unmeasured confounders. In addition, this
   technique enables us to make observational evidence
   causal^[43]9,[44]10. Simultaneously, it assigns single-nucleotide
   polymorphisms (SNPs) randomly at conception, simulating a scenario
   similar to randomized controlled experiments. This reduces the impact
   of confounding factors and minimizes bias caused by confounding factors
   or reverse causality^[45]11,[46]12. Therefore, it can be particularly
   useful to consider risk factors in terms of their long-term
   consequences^[47]13. Zhai et al.^[48]14 and Chen et al.^[49]15 employed
   486 blood metabolites to explore their causal association with AD.
   However, only 309 of these metabolites are characterized, representing
   only a subset of the vast plasma metabolome. which significantly limits
   the ability to comprehensively understand the relationship between a
   broad range of metabolites and the risk of AD onset. In 2023, Chen et
   al. published the largest and most comprehensive plasma metabolite
   genome-wide association study (GWAS) database to date^[50]16. This
   database encompasses 1,091 plasma metabolites and 309 metabolite
   ratios, providing a valuable resource for the in-depth exploration of
   potential metabolic pathways involved in AD pathogenesis. More
   extensive and systematic research is required to fully establish the
   causal associations between AD and plasma metabolites.

   Therefore, we hypothesized that a two-sample MR could be used to
   systematically investigate the causal association between AD and plasma
   metabolites. By studying the underlying markers associated with AD, new
   scientific evidence could be developed for early diagnosis, prevention,
   and treatment.

Results

Selection of IVs

   Following the elimination of LD (linkage disequilibrium) and management
   of weak IVs, 70,128 SNPs were identified as IVs for 1,400 plasma
   metabolites (Supplementary Table [51]S2). Of these, 72 suggestive
   metabolites showed potential causal links with AD (IVW P < 0.05).
   However, after validation using the weighted median (WM) and MR-Egger
   methods, six phenotypes were excluded because the directions of their
   effect estimates were inconsistent. These included serine to pyruvate
   ratio, aspartate to citrate ratio, 3-phosphoglycerate to adenosine
   5’-diphosphate (ADP) ratio, tryptophan levels, 3-methoxytyrosine
   levels, and glycosyl-N-stearoyl-sphingosine (d18:1/18:0) levels,
   leaving 66 phenotypes with a causal association with AD (P[IVW] < 0.05)
   (Fig. [52]1). The number of available IVs ranged from 35 to 242
   (Supplementary Table S9), with F-statistic values ranging from 14.476
   to 1537.795, indicating that weak IVs were not used. It included 42
   types of identified metabolites, 18 types of metabolite ratios, and six
   types of unknown metabolites. The 42 known metabolites were sourced
   from seven metabolic groups, 10 from amino acids, three from
   carbohydrates, three from cofactors and vitamins, one from energy, 14
   from lipids, two from nucleotides, and nine from xenobiotic metabolism.

Fig. 1.

   [53]Fig. 1
   [54]Open in a new tab

   Circular diagram of 66 plasma metabolites potentially causally
   associated with Alzheimer’s disease. AD Alzheimer’s disease, IVW
   inverse-variance weighted, WM weighted median.

Preliminary analysis results of MR

   Under the threshold of P < 1 × 10^−5, a total of 66 phenotypes
   exhibited potential causal association with AD (P[IVW] < 0.05). With P
   values less than 0.05 in both IVW and WM tests. We considered these
   metabolites to have a significant causal association with AD. After
   correction, 13 metabolites met these criteria (Table [55]1). These
   metabolites were 8 known metabolites, including argininate (OR 0.884,
   95%CI:0.783–0.997), 5alpha-androstan-3beta,17alpha-diol disulfate (OR
   0.884, 95%CI:0.783–0.997), pyridoxate (OR 0.899, 95%CI:0.847–0.955),
   2-oxoarginine (OR 0.876, 95%CI:0.773–0.992), and 5-methylthioadenosine
   (mta) (OR 0.862, 95%CI:0.782–0.949) that were considered protective
   factors against AD. Other metabolites, including mannose (OR 1.155,
   95%CI:1.06–1.26), cholesterol (OR 1.166, 95%CI:1.043–1.303), and
   1-stearoyl-2-oleoyl-GPI (18:0/18:1) (OR 1.149, 95%CI:1.053–1.253) had a
   significant causal association with increased risk of AD (Fig. [56]2).

Table 1.

   Significant positive results of metabolites with AD Mendelian
   randomization analysis.
   Outcome Metabolites Method NSNP Beta SE pval OR_CI
   AD(ieu-b-5067) Argininate MR Egger 51 − 0.015 0.069 0.827 0.985
   (0.860–1.128)
   WM 51 − 0.119 0.057 0.037 0.888 (0.793–0.993)
   IVW 51 − 0.09 0.039 0.023 0.914 (0.846–0.987)
   5alpha-androstan-3beta,17alpha-diol disulfate MR Egger 35 − 0.138 0.118
   0.250 0.871 (0.691–1.098)
   WM 35 − 0.152 0.074 0.039 0.859 (0.744–0.993)
   IVW 35 − 0.124 0.062 0.045 0.884 (0.783–0.997)
   Pyridoxate MR Egger 146 − 0.132 0.037 4.42E−04 0.876 (0.815–0.942)
   WM 146 − 0.115 0.046 0.013 0.891 (0.814–0.976)
   IVW 146 − 0.106 0.031 4.96E−04 0.899 (0.847–0.955)
   2-Oxoarginine MR Egger 37 − 0.073 0.134 0.590 0.930 (0.715–1.209)
   WM 37 − 0.191 0.08 0.017 0.826 (0.706–0.967)
   IVW 37 − 0.133 0.064 0.037 0.876 (0.773–0.992)
   5-Methylthioadenosine (mta) MR Egger 43 − 0.124 0.086 0.159 0.883
   (0.746–1.046)
   WM 43 − 0.173 0.068 0.011 0.841 (0.736–0.962)
   IVW 43 − 0.149 0.049 0.002 0.862 (0.782–0.949)
   Mannose MR Egger 46 0.188 0.074 0.015 1.206 (1.043–1.395)
   WM 46 0.184 0.065 0.004 1.202 (1.059–1.364)
   IVW 46 0.145 0.044 0.001 1.155 (1.060–1.260)
   Cholesterol MR Egger 38 0.093 0.083 0.272 1.097 (0.932–1.292)
   WM 38 0.173 0.08 0.030 1.189 (1.017–1.390)
   IVW 38 0.153 0.057 0.007 1.166 (1.043–1.303)
   1-Stearoyl-2-oleoyl-GPI (18:0/18:1) MR Egger 51 0.124 0.082 0.137 1.132
   (0.964–1.331)
   WM 51 0.178 0.066 0.007 1.195 (1.049–1.361)
   IVW 51 0.139 0.044 0.002 1.149 (1.053–1.253)
   Alanine/asparagine MR Egger 48 − 0.06 0.088 0.504 0.942 (0.792–1.120)
   WM 48 − 0.139 0.07 0.048 0.870 (0.758–0.999)
   IVW 48 − 0.178 0.047 1.75E−04 0.837 (0.763–0.919)
   Mannose/glycerol MR Egger 49 0.116 0.095 0.230 1.123 (0.931–1.354)
   WM 49 0.217 0.086 0.012 1.242 (1.049–1.470)
   IVW 49 0.163 0.054 0.003 1.177 (1.059–1.309)
   Mannose/S-methylcysteine MR Egger 48 0.18 0.081 0.031 1.197
   (1.022–1.402)
   WM 48 0.18 0.086 0.036 1.197 (1.011–1.417)
   IVW 48 0.161 0.053 0.002 1.175 (1.059–1.302)
   Phosphate/mannose MR Egger 50 − 0.223 0.083 1.00E−02 0.800
   (0.680–0.942)
   WM 50 − 0.302 0.072 2.89E−05 0.740 (0.642–0.852)
   IVW 50 − 0.184 0.045 4.13E−05 0.832 (0.762–0.909)
   X-11,372 MR Egger 44 0.152 0.099 0.134 1.164 (0.958–1.413)
   WM 44 0.196 0.083 0.018 1.216 (1.034–1.430)
   IVW 44 0.129 0.058 0.025 1.138 (1.016–1.274)
   [57]Open in a new tab

   AD Alzheimer’s disease, IVW inverse-variance weighted, WM weighted
   median, OR odds ratio, SE standard error, CI confidence interval.

Fig. 2.

   [58]Fig. 2
   [59]Open in a new tab

   Forest plot of 13 plasma metabolites significantly causally associated
   with AD. AD Alzheimer’s disease, IVW inverse-variance weighted.

   Detailed results of the sensitivity analysis are shown in Table [60]2.
   Following the intercept assessment of MR-Egger regression and
   MR-PRESSO, plasma metabolites with a causal association with AD showed
   no horizontal pleiotropy (P > 0.05) (Supplementary Table S8). The
   Cochran Q test did not find any heterogeneity. No single SNP was found
   to pose a bias risk to the results according to the leave-one-out test
   (Supplementary Figs. 1–13).

Table 2.

   Sensitivity analysis of significant causal relationship between
   metabolites and AD (IEU database).
   Metabolites (Exposure) Outcome Method Heterogeneity Pleiotropy Global
   test
   Q I^2 (%) Q_pval Egger intercept SE P RSSobs P
   Alanine/asparagine AD MR Egger 38.894 0.00 0.762 − 0.018 0.011 0.120
   44.499 0.687
   IVW 41.404 0.00 0.703
   Mannose/glycerol MR Egger 47.525 1.10 0.451 0.008 0.013 0.548 53.918
   0.370
   IVW 47.894 0.00 0.477
   Mannose/S-methylcysteine MR Egger 43.596 0.00 0.573 − 0.004 0.013 0.762
   48.195 0.538
   IVW 43.689 0.00 0.610
   Phosphate/mannose MR Egger 44.371 0.00 0.622 0.009 0.016 0.579 48.971
   0.594
   IVW 44.683 0.00 0.649
   X-11,372 MR Egger 33.357 0.00 0.827 − 0.003 0.012 0.781 35.543 0.839
   IVW 33.435 0.00 0.852
   Mannose MR Egger 49.862 11.76 0.252 − 0.008 0.012 0.473 57.524 0.198
   IVW 50.455 10.81 0.267
   Cholesterol MR Egger 14.748 0.00 0.999 0.012 0.012 0.328 17.254 0.999
   IVW 15.731 0.00 0.999
   5-Methylthioadenosine (mta) MR Egger 20.127 0.00 0.997 − 0.004 0.011
   0.728 21.100 0.997
   IVW 20.250 0.00 0.998
   1-Stearoyl-2-oleoyl-GPI (18:0/18:1) MR Egger 35.545 0.00 0.925 0.003
   0.014 0.839 37.065 0.951
   IVW 35.587 0.00 0.938
   2-Oxoarginine MR Egger 46.073 24.03 0.100 − 0.009 0.017 0.613 48.346
   0.131
   IVW 46.415 22.44 0.115
   Argininate MR Egger 31.323 0.00 0.977 − 0.016 0.012 0.195 34.292 0.974
   IVW 33.051 0.00 0.969
   5alpha-androstan-3beta,17alpha-diol disulfate MR Egger 56.446 41.54
   0.107 0.002 0.017 0.885 59.566 0.112
   IVW 56.482 39.80 0.109
   Pyridoxate MR Egger 96.644 0.00 0.999 0.014 0.011 0.208 99.422 1.000
   IVW 98.240 0.00 0.999
   [61]Open in a new tab

   AD Alzheimer’s disease, IVW inverse-variance weighted, SE standard
   error.

Results of the validation analysis and meta-analysis

   To substantiate our results, we performed replication MR analyses with
   AD GWAS data from the latest FinnGen database. The results based on the
   IVW method (P < 0.05), indicated that the adenosine 5’-monophosphate
   (AMP) to flavin adenine dinucleotide (FAD) ratio, n-lactoyl tyrosine,
   argininate, ergothioneine, 1,7-dimethyluric acid, and piperine passed
   the replication test with consistent Beta directions in both MR-Egger
   and WM analyses. Notably, argininate still demonstrated a substantial
   effect in the validation analysis (P[IVW+WM] < 0.05) (Supplementary
   Table S20).

   Subsequent meta-analysis of the IEU Open GWAS and FinnGen database
   further confirmed that the known metabolites argininate (OR = 0.989,
   95%CI: 0.855–1.144) and N-lactoyl tyrosine (OR = 0.936,
   95%CI:0.898–0.977) are protective factors against the risk of AD.
   Moreover, the meta-analysis clarified the protective effect of
   argininate on AD onset. The increase in metabolites 1,7-dimethyluric
   acid (OR = 1.026, 95%CI:1.010–1.043), ergothioneine (OR = 1.032, 95%CI:
   1.007–1.059), and piperine (OR = 1.080, 95%CI: 1.035–1.126) was
   associated with an increased risk of AD, a finding consistently
   validated in two-sample MR analyses (Fig. [62]3).

Fig. 3.

   [63]Fig. 3
   [64]Open in a new tab

   Meta-analysis of the causal associations between metabolites and AD. OR
   odds ratio, CI confidence interval.

Metabolic pathway analysis

   Our analysis of metabolic pathways for metabolites with potential
   positive associations indicated that AD pathogenesis was related to two
   metabolic pathways: “Vitamin B6 metabolism” (P = 0.002) and “Glyoxylate
   and dicarboxylate metabolism” (P = 0.020) (Supplementary Table S12a).
   Enrichment analysis revealed seven metabolic pathways associated with
   the risk of developing AD, which were “Beta Oxidation of Very Long
   Chain Fatty Acids” (P = 0.014), “Vitamin B6 Metabolism” (P = 0.017),
   “Transfer of Acetyl Groups into Mitochondria” (P = 0.023),
   “Glycerolipid Metabolism” (P = 0.029), “Fructose and Mannose
   Degradation” (P = 0.043), “Citric Acid Cycle” (P = 0.045), and
   “Gluconeogenesis” (P = 0.048) (Supplementary Table S12b). Two
   approaches were used to validate the metabolic pathways that could
   contribute to AD pathogenesis, with the “Vitamin B6 metabolism” pathway
   consistently emerging, implying its possible association with the
   fundamental biological mechanisms of AD. To further validate this
   hypothesis, we reanalyzed the metabolic pathways using the validated
   results from the FinnGen database, employing the same methodology as in
   the previous analysis. The validation confirmed that the “Vitamin B6
   metabolism” pathway (P = 0.003) remains closely associated with the
   pathogenesis of AD (Supplementary Table S12c,d).

Genetic correlation and direction validation

   To observe the genetic correlation between metabolites and AD from a
   broad perspective, we conducted Linkage disequilibrium score regression
   (LDSC) analysis, focusing on 13 metabolites with significant positive
   findings in MR analysis. The genetic correlation (rg) ranged from 0.518
   to 0.328, with standard deviations ranging from 0.115 to 0.456, and P
   values from 0.091 to 0.895. No significant correlation was found (P[rg]
   > 0.05), which may be due to the relatively small sample size of the
   metabolite data, potentially limiting the power to detect a meaningful
   association (Supplementary Table S11).

Colocalization analysis

   In this study, we conducted an in-depth analysis of the pathogenesis of
   AD, focusing on the common risk genes for AD: amyloid protein precursor
   (APP), presenilin-1 (PSEN1), presenilin-2 (PSEN2), and apolipoprotein E
   (APOE). Using colocalization analysis, we assessed the relationship
   between six metabolites identified in previous replication analyses and
   these genes. Bayesian methods were used to calculate posterior
   probabilities for different genetic model hypotheses (H0 to H4). The
   emphasis of our analysis was to identify the top SNPs for each
   metabolite. These SNPs displayed the highest posterior probabilities of
   association under the H4 hypothesis, indicating their potential as key
   variants in the genetic structure of AD. The analysis results revealed
   that specific top SNPs, such as rs429358 in the APOE gene and
   rs186099562 in the PSEN2 gene, demonstrated exceptionally high
   posterior probabilities of association (SNP.PP.H4 > 0.79) across
   multiple metabolites, aligning with the known significant roles of APOE
   and PSEN2 in AD risk. This finding implied that these SNPs may be
   associated with the pathological process of AD by affecting metabolic
   pathways, particularly the levels of metabolites that were closely
   associated with AD pathology. In the APP and PSEN1 genes, varying
   degrees of association were observed. Although top SNPs were
   identified, the lower SNP.PP.H4 values suggested that these
   associations are incidental findings or due to smaller effects of these
   loci in AD. This implies that although mutations in APP and PSEN1 were
   closely linked with familial AD, their impact on the levels of
   metabolites may be limited. Moreover, the interactions between these
   genes and metabolites could be more complex, possibly subject to a
   range of gene-environment interactions (Supplementary Table S10,
   Supplementary Figs. 14–37).

Discussion

   We integrated two large-scale GWAS datasets and employed a rigorous
   two-sample MR approach to explore the causal associations between 1,091
   plasma metabolites and 309 metabolite ratios with AD. In the
   preliminary MR analysis, 66 plasma metabolites and metabolite ratios
   metabolites have potential causal associations with AD, and 13 plasma
   metabolites and metabolite ratios with significant causal to AD were
   identified, including 8 known metabolites (argininate,
   5alpha-androstan-3beta,17alpha-diol disulfate, pyridoxate,
   2-oxoarginine, 5-methylthioadenosine (mta), mannose, cholesterol,
   1-stearoyl-2-oleoyl-GPI (18:0/18:1)), 1 unknown metabolite (X-11372),
   and 4 metabolite ratios (alanine to asparagine ratio、mannose to
   glycerol ratio、mannose to S-methylcysteine ratio、phosphate to mannose
   ratio). The MR replication analysis indicated that 6 plasma metabolites
   and metabolite ratios passed the replication test. N-lactoyl tyrosine,
   argininate, and AMP to FAD ratio were protective factors against AD;
   ergothioneine, 1,7-dimethyluric acid, and piperine were risk factors.
   Notably, argininate had a significant impact on the onset of AD.
   Concurrently, the colocalization analysis revealed a potential
   association between blood metabolites and the risk of AD onset through
   their linkage with the APOE and PSEN2 genes. Additionally, we
   identified one metabolic pathway, Vitamin B6 metabolism, with a
   significant relationship to the development of AD.

   Notably, the present MR analysis and corresponding validation and
   meta-analyses revealed that elevated argininate levels were
   significantly associated with a lower risk of AD. Arginine and
   argininate cation represent different ionic states of the same
   molecule. Arginine exists in various ionic forms under different pH
   conditions. Under physiological conditions, such as plasma pH of
   approximately 7.4, arginine typically exists in its cationic
   form^[65]17. Arginine contributes to various biological aspects such as
   neurotransmission, neurogenesis, neuroplasticity, cellular redox
   metabolism, oxidative stress, inflammation, and cerebral plasma flow
   regulation^[66]18–[67]20. Previous studies have provided valuable
   insights into the role of arginine in AD. Ibáñez et al. observed a
   decrease in arginine levels in the cerebrospinal fluid of AD patients
   and indicated the potential of arginine as a biomarker for monitoring
   the disease’s advancement^[68]21,[69]22. Sarı et al. also discovered
   that in patients with late-stage AD, there is a decrease in plasma
   arginine levels and a widespread reduction in the activity of arginine
   metabolism pathways^[70]23. Geravand et al. injected arginine into the
   dorsal hippocampus of AD models induced by aluminium-chloride (AlCl[3])
   and discovered that L-arginine prevented AD by activating neuronal
   nitric oxide synthase (nNOS), thereby increasing hippocampal NO
   levels^[71]24. Existing evidence also suggests that arginine metabolism
   disorder is associated with AD. A meta-analysis found that the levels
   of methylated arginine —asymmetric dimethylarginine (ADMA)—
   significantly increased in dementia patients compared with
   controls^[72]25. ADMA is an intermediate product in the metabolism of
   arginine that can inhibit the normal metabolism of arginine and NOS.
   When ADMA levels are elevated, it can compete with arginine for NOS
   binding sites, leading to uncoupling and reduced synthesis of
   NO^[73]26,[74]27. Our findings indicated that argininate may act as a
   protective factor against AD. Further experiments are needed to verify
   whether argininate can become a reliable therapeutic target for AD.

   MR replication analysis indicated that ergothioneine was a risk factor
   for AD. Numerous investigations have demonstrated that ergothioneine
   plays a protective role in the pathogenesis of AD, which contrasts with
   the findings of this study^[75]28,[76]29. Ergothioneine as a dietary
   antioxidant has been shown to improve amyloid beta clearance in the
   neuroretina of a mouse model of AD^[77]30. Wu et al.^[78]29 reported
   that ergothioneine levels decrease progressively with the worsening of
   cognitive impairment. Animal studies have confirmed that
   supplementation with ergothioneine may improve learning and memory
   abilities in AD-model mice and reduce neuropathological
   damage^[79]31–[80]33. These studies suggested that ergothioneine may be
   a potential biomarker for assessing the severity of the disease, and
   reduced ergothioneine levels may increase the risk of neurodegenerative
   diseases^[81]28,[82]29,[83]34. The exact reason for the difference
   between our results and those of others is unknown. We suggested that
   the unexpectedly high levels of ergothioneine in AD may be due to
   decreased or impaired uptake of ergothioneine by the brain. Therefore,
   further research is required to understand the role of ergothioneine in
   AD.

   Using MR analysis with two different AD GWAS, we unexpectedly found
   that piperine was a risk factor for AD. Piperine, an alkaloid isolated
   from pepper, has been proven to possess a variety of pharmacological
   actions, including anti-inflammatory, anti-tumor, antidepressant,
   neuroprotective, analgesic, and hepatoprotective effects^[84]35. Animal
   experiments indicate that piperine can improve cognitive function in
   AD-model mice, reduce oxidative stress, and promote neural
   health^[85]36. Kumar et al. found that piperine decreased the
   expression of certain related-AD genes (BACE1, PSEN1, APAF 1, caspase
   3, and catalase)^[86]37. These findings implied that piperine and
   ergothioneine may play complex and differentiated roles in disease
   progression.

   This study showed that n-lactoyl tyrosine was a protective factor
   against AD, whereas 1,7-dimethyluric acid was identified as a
   contributing risk factor to AD development. There is no study to
   directly associate these metabolites with AD. 1,7-dimethyluric acid is
   one of the primary metabolites of caffeine^[87]38, which is metabolized
   in the liver^[88]39. Numerous studies have revealed the relationship
   between caffeine and AD, with both clinical research and in vitro and
   in vivo experiments demonstrating caffeine’s neuroprotective
   effects^[89]40. An MR study by Larsson et al. demonstrated caffeine’s
   protective effects against AD^[90]41. We speculate that in patients
   with AD, caffeine is metabolized to 1,7-dimethyluric acid, leading to
   reduced levels of caffeine, thereby elevating the risk of AD.
   Nonetheless, further investigation is needed to fully elucidate the
   exact mechanism.

   Significant changes in metabolite ratios typically reflect alterations
   in metabolic pathways. Through the analysis of metabolite ratios, we
   can identify potential genetic regulatory points and metabolic
   pathways, thereby revealing shifts in metabolic flux. In our study, the
   AMP/FAD ratio was validated in replication trials, and an increased
   AMP/FAD ratio was associated with a reduced risk of AD. The changes in
   AMP/FAD ratio reflect alterations in cellular energy metabolism and
   redox states. AMP is a key molecule in regulating cellular energy
   metabolism and oxidative stress, both of which are pathological
   features of AD^[91]42,[92]43. AMP can activate AMP-activated protein
   kinase, leading to increased ATP production and reduced ATP
   consumption, thereby modulating cellular energy metabolism and
   enhancing the cell’s tolerance to energy deficits, which may protect
   neurons from damage^[93]44. Additionally, AMP can promote the
   expression of antioxidant enzymes, helping to reduce oxidative stress
   levels^[94]45. FAD is a key cofactor involved in redox reactions in
   cellular metabolism and is capable of regulating oxidative stress and
   apoptosis^[95]46. Therefore, changes in the AMP/FAD ratio may serve as
   a biomarker indicating the progression or severity of AD.

   Metabolic pathway analysis showed that vitamin B6 metabolism was
   closely related to AD onset. Vitamin B6 is involved in crucial cellular
   metabolism in the human body, with antioxidant and anti-inflammatory
   functions, and plays a vital role for treating neurological
   diseases^[96]47. Epidemiological surveys have shown that vitamin B6
   deficiency is common in the elderly, often occurring alongside vitamin
   D deficiency, and is associated with lower Mini-Mental Status
   Examination (MMSE) scores^[97]48. Vitamin B6 contributes to the
   metabolism and reduction of homocysteine levels^[98]49–[99]51. Studies
   have indicated that increased plasma homocysteine levels are a risk
   factor for AD^[100]49–[101]51. In addition, using MR analysis, Hu et
   al. proved a positive causal association between total plasma
   homocysteine and the risk of AD^[102]52. This study supports the
   potential role of vitamin B6 metabolic pathways in the occurrence and
   progression of AD, thus highlighting vitamin B6 as a therapeutic target
   for AD.

   This colocalization study revealed significant associations between
   some metabolites and the genes APOE and PSEN2. These findings suggest a
   potential relationship between these metabolites and the genetic
   factors influencing AD development. Similar to our findings, prior
   comprehensive GWAS and meta-analyses also identified APOE ε4 allele as
   the most potent genetic risk factor for AD^[103]53,[104]54. The APOE
   gene is involved in the production of a protein that helps transport
   cholesterol and other fats in the bloodstream. Difficulties in this
   process can lead to the development of AD^[105]53. It has been shown
   that APOE is closely associated with various pathological
   manifestations of AD, including Aβ plaques, neurofibrillary tangles and
   neuroinflammation^[106]55. Carriers of ApoE4 may have an increased risk
   of AD due to accelerated lipid dysregulation and energy
   deficits^[107]56. AD patients often exhibit significant alterations in
   their lipid profiles, such as abnormal levels of cholesterol,
   phospholipids, and fatty acids^[108]57. Certain lipids, such as Omega-3
   fatty acids, are believed to have neuroprotective effects and may slow
   the progression of AD^[109]58. Conversely, the intake of trans fatty
   acids may increase the risk of AD^[110]59. Changes in lipid profiles
   may serve as potential biomarkers for AD, particularly in ApoE4
   carriers. By analyzing lipid profiles, it is possible to identify AD
   risk at an early stage and intervene. For example, Omega-3 therapy in
   ApoE4 carriers can reduce neuronal integrity loss^[111]60. The PSEN2
   gene is closely associated with early-onset AD^[112]61. Mutations in
   the PSEN2 gene influence the onset and progression of AD through the
   production and metabolism of Aβ proteins, calcium homeostasis, and
   autophagy^[113]62,[114]63. However, studies focusing on APOE and PSEN2
   as therapeutic targets remain in the preliminary phases. The present
   findings suggest a potential interaction between plasma metabolites and
   the APOE and PSEN2 genes in the context of AD development.

   Our study has several strengths. First, to the best of our knowledge,
   this is the first two-sample MR study to evaluate the association
   between plasma metabolites and AD risk. Such a design can mitigate
   limitations associated with confounding factors often encountered in
   conventional observational studies and can offer more robust evidence
   of causality between exposure and outcomes. Therefore, the main
   strength of this study lies in the increased statistical power achieved
   using summarized data from a GWAS consortium in a two-sample MR. Also,
   unlike one-sample MR, the weak instrument bias in two-sample MR is
   almost non-existent. This unbiased the effect toward the results of the
   confounding multivariable regression. With the largest sample size ever
   recorded, this MR study is the most comprehensive in terms of causal
   relationships between plasma metabolites and AD. In addition, the SNPs
   related to plasma metabolites were obtained from the most extensive and
   comprehensive GWAS conducted so far. This enables the establishment of
   a robust instrumental variable with unbiased MR estimates. Third, this
   study included several exposure factors that were plasma metabolites
   and metabolite ratios. This generates a large amount of analytical work
   and poses major challenges to the analysis. Fourth, to ensure the
   reliability and stability of the results, we employed multiple MR
   models, implemented stringent quality control measures, and conducted a
   series of sensitivity analyses. This approach can control the bias
   caused by pleiotropic effects and validate the associations observed in
   the IVW analysis. Furthermore, we performed further re-analyses of
   another AD dataset, meta-analyses, LDSC models, and colocalization
   studies to demonstrate the validity of our results across different
   perspectives, dimensions, and geological processes. Therefore, this
   study provides new ideas for investigating the influence of plasma
   metabolites on the risk and pathogenesis of AD.

   Nonetheless, it is important to recognize the limitations of this
   study. First, for the MR analysis, we only used GWAS data from European
   ancestry populations to reduce the impact of racial differences, but
   this may limit the reliability of extrapolating our findings to other
   ethnic groups. Second, although the MR analysis method has been proven
   to be effective in etiological studies, further research is necessary
   to understand the underlying mechanisms. Finally, the metabolite and
   metabolic ratio data were measured at a single time point for each
   participant, which limits our ability to assess longitudinal changes or
   their correlation with AD progression. As a result, the KEGG pathway
   enrichment analysis, while providing insights into the biological
   significance of identified metabolites, reflects only a cross-sectional
   view based on static data.

   In summary, MR analysis revealed 13 plasma metabolites and metabolite
   ratios with a significant causal association with AD. After replication
   analysis, six of these metabolites and metabolite ratios were validated
   through replication tests, with argininate remaining significant.
   Following the colocalization analysis, we found that metabolites could
   increase the risk of AD through their association with top SNPs in the
   APOE and PSEN2 genes. Therefore, it is possible to use plasma
   metabolomics for the clinical prevention and screening of AD. As the
   number of studies directly linking these plasma metabolites to AD is
   limited, further experimental and clinical studies are needed to
   confirm these results and investigate the underlying mechanisms.

Methods

Data source

   The SNPs associated with metabolic products originated from a recent
   comprehensive GWAS investigation obtained from the Canadian
   Longitudinal Study on Aging (CLSA) cohort, involving 8,299 elderly
   participants (62.4 ± 9.9 years, all from European descent).This
   dataset, generated using the UPLC-MS/MS (Metabolon HD4) platform,
   included 1,091 plasma metabolites and 309 metabolite ratios. GWAS
   identified associations between 690 metabolites and 248 genetic loci,
   as well as 143 metabolite ratios and 69 genetic loci. Plasma
   metabolites refer to metabolites measured in plasma samples, most of
   which are substrates or products generated by enzyme-catalyzed
   reactions. These metabolites reflect the metabolic state of the whole
   body and are used to identify potential biomarkers for various
   diseases. Metabolite ratio is the ratio of metabolite levels of
   metabolite pairs sharing an enzyme or transporter using
   metabolite-protein associations recorded in the Human Metabolome
   Database (HMDB)^[115]64, which is the ratio of metabolites with higher
   biocompatibility. This made it the most comprehensive study of plasma
   metabolites genome-wide data to date^[116]16. The known 871 plasma
   metabolites were classified into 8 metabolic groups, including amino
   acids, cofactors and vitamins, lipids, energy, carbohydrates,
   nucleotides, peptides, and xenobiotic metabolism, as detailed in the
   KEGG database^[117]65. A portion of the metabolites originated from
   partially characterized molecules. Two hundred and twenty metabolites
   starting with ‘X’, represent unidentified metabolites. The detailed
   names of the 1,400 metabolites and metabolite ratios are given in
   Supplementary Table [118]S1. SNPs related to AD were sourced from IEU
   Open GWAS ([119]https://gwas.mrcieu.ac.uk/), with the registry number
   ieu-b-5067, involving 488,285 participants, including 954 cases and
   487,331 controls, all of European ancestry, comprising 12,321,875 SNPs
   in total.

MR assumptions and selection of IVs

   Three fundamental assumptions are essential for IVs in MR studies: (1)
   a strong correlation between IVs and the exposure factor (relevance
   assumption), (2) IVs are independent of confounding factors
   (independence assumption), and (3) IVs affect the outcome solely
   through the exposure factors (exclusion restriction assumption). We
   have provided results using tables adapted from clinicaltrials.gov as
   per instructions, including the CONSORT flow chart (Fig. [120]4). In
   this study, metabolic products were considered as the exposure factors,
   and AD as the outcome. To meet these core assumptions and ensure the
   rigor and authenticity of the design, after multiple trials, this study
   ultimately established the optimal threshold for SNPs significantly
   related to metabolic products as P < 1 × 10^−5. The F-value of the
   instrumental variable should be greater than 10 to avoid the effects of
   weak instruments. The formula for calculating F is F = R^2
   (N-2)/(1-R^2). To avoid inaccuracies in the results due to allelic
   ambiguity and genotyping errors, palindromic SNPs were removed. To
   ensure the independence of SNPs, it was necessary to remove LD, with a
   window set at 10,000 kb and r^2 < 0.001. Using Phenoscanner V2
   ([121]http://www.phenoscanner.medschl.cam.ac.uk/), SNPs linked to known
   AD risk factors were identified and excluded before analysis.

Fig. 4.

   [122]Fig. 4
   [123]Open in a new tab

   Study flow chart CONSORT diagram. MR Mendelian randomization, GWAS
   genome-wide association study, LD linkage disequilibrium, IVW
   inverse-variance weighted, WM weighted median. ^#Potential causal
   association; *significant causal associations.

Statistical analysis

   In this study, we adopted the IVW method as our primary analytical
   approach. Characterized by its use of regression without an intercept
   and employing the inverse of outcome variance as weights, this method
   integrates each SNP’s wald estimate, which is instrumental in reducing
   the influence of confounding factors and enhancing the accuracy of the
   results^[124]13. To validate the directionality of the IVW analysis, we
   used the MR-Egger and WM methods. A potential causal association was
   considered plausible if the IVW analysis resulted in a p-value < 0.05
   and if these results aligned with the directional trends observed in
   other analytical methods.

   To ensure the robustness of our analysis, a series of sensitivity
   analyses were performed. These included evaluations for heterogeneity,
   horizontal pleiotropy, and leave-one-out analysis. Horizontal
   pleiotropy, in particular, was assessed through the intercept obtained
   from the MR-Egger regression. This regression method is notable for
   including an intercept term. A substantial deviation of this intercept
   from zero indicates potential horizontal pleiotropy among IVs^[125]66.
   Heterogeneity was assessed using the Cochran’s Q test, observing the Q
   value and Q p-value in IVW and MR Egger methods. A larger Q value
   indicates a smaller P value, and a P < 0.05 signifies the presence of
   heterogeneity. For results exhibiting heterogeneity or horizontal
   pleiotropy (P < 0.05), we applied the MR-PRESSO method to iteratively
   remove the identified IVs, conducting a global test on the remaining
   SNPs, removing SNPs that caused aberrant results and repeating this
   process recursively, until the P > 0.05 of the global test^[126]67. The
   leave-one-out test was used to assess whether the results were
   influenced by a single SNP^[127]68. To avoid false positives resulting
   from multiple testing, when validating the causal associations between
   metabolites and AD, if the direction of the IVW test and WM were
   consistent and both had P < 0.05, a significant causal association
   between the metabolite and AD was confirmed^[128]69. In addition,
   steiger filtering was applied to prevent reverse causal associations.
   All analyses were carried out using R software (version 4.3.2) with the
   following packages: TwoSampleMR, MRPRESSO, fastMR, and Coloc.

Metabolic pathway analysis

   Following the collation of KEGG IDs for plasma metabolites with IVW
   (P < 0.05), we employed pathway analysis and enrichment analysis
   modules. MetaboAnalyst 6.0 ([129]https://www.metaboanalyst.ca/) was
   used to analyze metabolic pathways and conduct enrichment analysis
   using the Small Molecule Pathway Database (SMPDB) and the KEGG database
   for reference. This established a significance threshold of 0.05 for
   the pathway analysis to investigate the underlying mechanisms through
   which the identified plasma metabolites influenced the risk of AD
   development.

Confirmatory analysis and meta-analysis

   To ensure the reliability of our results, we further used the latest AD
   GWAS data from the FinnGen database (Public release: Dec 18, 2023).
   Then, the preliminary findings were re-evaluated using data from 10,520
   cases and 401,661 controls, all of European ancestries. After
   conducting MR analysis on both databases, a meta-analysis (Review
   Manager 5.3) was conducted on the random-effects IVW model of both sets
   of analyses to ultimately identify plasma metabolites with a causal
   effect on AD.

Colocalization analysis

   Colocalization analysis is commonly employed to assess whether a single
   genetic variant in a specific region is potentially responsible for
   both phenotypes, reinforcing the associative evidence between
   them^[130]70,[131]71. Colocalization analysis involves five mutually
   exclusive model assumptions (H0-H4), covering all possible association
   scenarios, and the sum of the posterior probabilities of these five
   models equals 1. In particular, the H4 model hypothesizes the presence
   of a common variant within the region affecting two traits, but it is
   unclear which variant it is^[132]72. In the study of AD, several
   well-recognized genes (such as APP, PSEN1, PSEN2, and APOE) are known
   to influence its onset^[133]53,[134]73–[135]75. In our study, genetic
   tools were developed by sourcing QTL loci from the gene database
   ([136]https://www.eqtlgen.org/), focusing solely on cis-eQTLs found
   within a 1 Mb radius around the genes. The aim of using colocalization
   analysis was to identify specific genetic variant loci potentially
   associated with disease outcomes through their influence on
   metabolites, serving as a crucial supplementary analysis^[137]76.

Genetic correlation and direction validation

   Although SNPs related to AD have been rigorously excluded in the
   process of instrumental variable selection, the remaining SNPs may
   still affect the genetic risk of outcomes due to false positives caused
   by genetic correlation in MR studies^[138]77,[139]78. LDSC is a
   statistical measure for assessing the degree of LD association between
   an SNP and its surrounding SNPs. This study used this method to
   investigate whether the causal association between metabolites and AD
   was potentially influenced by a shared genetic structure.

Electronic Supplementary Material

   Below is the link to the electronic supplementary material.
   [140]Supplementary Material 1^ (27.9MB, pdf)
   [141]Supplementary Material 2^ (20.1MB, xlsx)

Acknowledgements