Identification of plasma biomarkers in lipid metabolism for accurate prediction of Alzheimer’s disease

Abstract

Metabolomics may reveal non-invasive biomarkers for early diagnosis in Alzheimer’s disease (AD) and provide new insights into the disease mechanisms to develop effective treatments. Here, we comprehensively analyzed the blood plasma metabolomes from a Chinese cohort of 447 individuals, including 188 AD, 181 MCI (mild cognitive impairment), and 78 NC (normal control). Differential analysis identified altered metabolites, followed by forward feature selection to prioritize a panel of key metabolites, and construction of a diagnostic model using logistic regression. Key metabolite-enriched pathways were identified and quantified for comparison across different groups, which was then validated through external datasets. We observed extensive metabolic dysregulation in AD compared to age-matched NC, with 25% of the differential metabolites also significantly dysregulated in MCI in the same directions. A panel of 22 key metabolites was prioritized, where triglycerides (TG) and phosphatidylethanolamines (PE) ranked top in importance. With these key metabolites, we trained a diagnostic model that classified AD from NC accurately (Area Under the Curve [AUC] = 0.935 in the replication cohort). Pathway quantification analysis showed significant changes in lipid metabolism in AD, which were validated in two external cohorts. We presented a precise and robust blood metabolic diagnostic model for AD, which may help promote early diagnosis and deepen the understanding of AD mechanisms.

Introduction

Alzheimer’s disease (AD) is a neurodegenerative disorder that is the most common cause of dementia in the elderly [1]. This intractable disease is characterized by progressive cognitive decline including memory loss, language difficulty, and poor executive function, making it one of the major challenges for the public health [2, 3]. Due to lack of therapies to effectively treat the disease, there is an urgent need to better understand the pathological mechanisms and trajectory of AD, so as to diagnose and intervene it at an early stage. For decades, the amyloid cascade hypothesis, based on the core pathological hallmarks of abnormal accumulation of extracellular beta-amyloid (Aβ) and hyperphosphorylated tau in neurofibrillary tangles in the brain, remains a predominant pathogenic theory [4]. However, the treatments targeting Aβ clearance to improve cognitive functions in clinical trials have not yet achieved the desired results [5, 6], which highlights the importance of exploring other pathological pathways. Recent studies have showed the prevalence of oxidative stress [7, 8], inflammation [9–11], mitochondrial dysfunction [12–14], and metabolic disturbances [15, 16] in Alzheimer’s disease. These findings indicated promise not only in enhancing our understanding of Alzheimer’s disease, but also in discovering potential biomarkers for early diagnosis and disease monitoring.

Metabolomics, with global quantification of metabolites, offers a promising tool for unraveling the pathogenic pathways of Alzheimer’s disease and identifying novel biomarkers [17, 18]. By measuring the concentrations of circulating metabolites—the end products of cellular regulation and complex biochemical processes, it can reflect the interplay of genetics and environment as well as metabolic changes in human tissues and organs, thereby serving as indicators of disease processes and relevant phenotypes [19, 20]. Recent metabolomics research has demonstrated significant associations between numerous blood metabolites and the risk of AD [21–26]. Notably, a recent large-scale plasma metabolomics study [27] involving 1397 participants identified plasma ammonia as a potential biomarker for AD. However, despite its large sample size, the restricted metabolite coverage and the use of a top-down analytical framework limited the mechanistic interpretation of the findings. Additionally, several previous studies have explored the potential of using peripheral metabolites as predictive biomarkers for Alzheimer’s disease. A landmark study by Mapstone et al. [28] identified a panel of 10 plasma phospholipids that predicted conversion from cognitively normal status to amnestic MCI or AD within 2–3 years, achieving an overall accuracy of 90% (AUC = 0.92). Similarly, Jia et al. [29] reported a panel of 11 metabolites that distinguished AD from normal controls and successfully differentiated AD from other dementias with an AUC = 0.96. More recently, large-scale prospective metabolomics analyses have further demonstrated the predictive value of circulating metabolites for AD. For instance, a study [30] on 274,160 participants from the UK Biobank developed a Metabolic Risk Score (MetRS) based on eight selected metabolites, achieving an AUC = 0.861 for predicting future AD status when combined with demographic and cognitive variables. Collectively, these studies highlighted the feasibility of blood metabolites as biomarkers for predicting AD status or progression. Although these studies have provided valuable insights, many have been limited by relatively small sample sizes or restricted coverage of metabolites, making their findings insufficient to fully elucidate the relationship between metabolome and AD. Moreover, there lacks a consensus view, as some identified metabolites failed to be independently validated externally, even with contrary results [31–33]. Although these findings underscore potential biomarkers, they still demonstrate suboptimal performance in early diagnosis. Despite these limitations, which hinder the widespread clinical application of blood metabolomics in early screening and diagnosis, the potential of blood metabolomics in elucidating the pathogenic mechanisms of AD and identifying novel biomarkers remains significant due to its convenience and non-invasiveness [34].

In this work, we conducted a comprehensive analysis on liquid chromatography-tandem mass spectrometry (LC–MS/MS)-based metabolomics of blood plasma samples from a Chinese cohort of 447 participants. We identified a key panel of 22 differential metabolites from 1190 quantified metabolites as a metabolic biomarker for AD, based on which a logistic regression classifying model was built that accurately discriminated AD from normal controls (NC). This AD diagnostic model achieved high performance in our replication cohort. By quantifying metabolic pathways associated with this panel of metabolites, we characterized the metabolic differences at the pathway level and identified significant changes in lipid metabolism in Alzheimer’s disease and MCI compared to NC. Our findings were largely validated in multiple external datasets at the quantified pathway level, indicating the robustness, interpretability, and generalization power of our metabolic biomarker. Our results may enhance the understanding of mechanisms in Alzheimer’s disease, and support future metabolome-based diagnostic and therapeutic endeavors.

Methods

Participants

This study was approved by the Research Ethics Committee of the Sixth People’s Hospital of Shanghai, China (Approval No: 2019-032). Written informed consent was obtained from each participant. Participants in disease groups were recruited from outpatients visiting the Department of Geriatric Medicine of the Shanghai Sixth People’s Hospital between March 2019 and September 2022. The control group recruited age-matched elderly people living in the vicinity with normal cognitive function. Each participant underwent comprehensive neuropsychological assessments, which included interviews assessing subjective cognitive decline, six neuropsychological tests across three cognitive domains, global cognitive tests, and daily cognitive assessments. Participants in disease groups were classified into MCI and AD based on neuropsychological assessment results and international diagnostic criteria by professional clinicians. The diagnosis of AD followed the guidelines of the National Institute on Aging-Alzheimer’s Association (NIA-AA) [35]. MCI was diagnosed according to the Jak/Bondi criteria [36]. A total of 477 individuals were primarily recruited for the study, including 195 AD, 196 MCI, and 86 NC. The additional recruitment criteria include: (1) without a disease history or family history of other neurologic or psychiatric diseases, such as Parkinson’s disease, depression, epilepsy, and neuron developmental delay; (2) without disease histories of hematological disease or tumor; (3) without serious somatic diseases; (4) with adequate vision and hearing. To avoid the effect of drug use on blood metabolome profiles, especially the lipidome, samples that used cholesterol-lowering or other lipid-lowering drugs within the past six months were also excluded. After excluding individuals according to the above criteria, 447 participants, comprising 188 AD, 181 MCI, and 78 NC samples, were finally selected in this analysis. According to the batches of data collection, they were divided into the discovery cohort and the replication cohort. The first batch, comprising 239 participants (96 AD, 90 MCI, and 53 NC) and collected during 2021–2022, served as the discovery cohort. The second batch, comprising 208 participants (92 AD, 91 MCI, and 25 NC) and collected during 2019–2020, served as the replication cohort. Further details on cohort information, participant recruitment, neuropsychological assessments, and diagnostic criteria are available in Supplementary Methods.

External validation cohort

We utilized two independent external datasets from publicly available metabolomics studies on AD as the validation cohorts. The details of these datasets can be found in the respective papers [37, 38]. Briefly, Wang et al. included 57 AD, 58 MCI, and 57 NC participants, while Teruya et al. included 8 dementia patients, 8 age-matched healthy elderly individuals, and 8 healthy young subjects.

Plasma collection and metabolite extraction

Plasma was prepared from whole blood by the procedures as described below: The blood was centrifuged at 1000 × g for 10 min at 4 °C. The supernatant was collected and centrifuged at 2000 × g for 5 min at 4 °C. The prepared plasma was then aliquoted and stored at −80 °C in a refrigerator until further analysis. The quality control (QC) sample was prepared by combining equal aliquots of the supernatants from all samples, and served as a reference for quality control (QC) purposes throughout the analysis. To detect metabolites as much as possible, the hydrophilic and hydrophobic metabolites were respectively extracted and analyzed, and details have been listed in Supplementary Methods.

Plasma metabolomics and data preprocess

Untargeted metabolomics was conducted on an ultra-performance liquid chromatography (UPLC) system, ExionLC AD, coupled to a Quadrupole-Time of Flight mass spectrometer (TripleTOF™ 6600, SCIEX). The detailed analytical procedures can be found in Supplementary Methods.

The mass spectrum obtained from metabolomics experiments were processed using Analyst software (version 1.6.3) from SCIEX. The metabolite identification was conducted by referencing standards in self-built database and public databases. Specifically, metabolites were first identified using an in-house metabolite library, i.e., Metware Database (MWDB). Then, further metabolite identification was carried out based on accurate mass, isotope pattern, and MS/MS spectra according to the public databases such as METLIN (http://metlin.scripps.edu/index.php), HMDB (http://www.hmdb.ca/), and KEGG (https://www.genome.jp/kegg/). Metabolite quantification was performed using the multiple reaction monitoring (MRM) mode on a triple quadrupole mass spectrometer. After acquiring the raw LC-MS data, extracted ion chromatographic peaks for all metabolites were integrated to calculate the peak areas. The peak area (Area) of each chromatographic peak represented the relative abundance of the corresponding metabolite. Finally, all integrated peak area data were exported and saved for subsequent analysis.

Quality control

A quality control (QC) sample was inserted after every 10 analytical samples to monitor instrument stability and reproducibility of the analytical process under the same treatment conditions. Thus, QC samples can serve as an additional QC measure for evaluating analytical performance and as a reference for standardizing raw metabolomics data. We calculated coefficient of variation (CV) values of the metabolites in QC samples, and metabolites with CV larger than 30% were removed. Metabolites with missing values in more than 10% of the samples were also removed, and the remaining missing values were imputed with a very small constant (0.1 in this study) [39]. Metabolites common to both cohorts were used, resulting in 1190 metabolites. Finally, the data underwent log transformation and z-score scaling for subsequent analysis [40].

Feature selection

A machine learning feature selection algorithm was developed to screen differential metabolites further, aiming to achieve the optimal classification accuracy in AD/MCI versus NC with as few metabolic combinations as possible. For the candidate feature set, a classifier with Area Under the Curve for the Receiver Operating Characteristic curve (ROC-AUC) as the evaluation metric was constructed, and forward stepwise feature selection was used to identify an optimal set of features. Before each iteration, the current optimal subset of features was fixed, and the remaining features were added stepwise until the optimal AUC was reached, ending the cycle. Finally, the selected features and the resulting AUC value were then represented as the optimal feature set and AUC. Forward selection was constructed using Scikit-learn (version 0.24.1) packages [41] in Python.

Diagnostic model

After adjusting for age, sex, BMI, and education years, a logistic regression algorithm was used to develop a classification model for AD/MCI diagnosis. The logistic regression model was constructed based on the key panel of metabolites to classify AD/MCI versus NC, with hyperparameter grid search determined through 10-fold cross-validation in the discovery set. The model’s generalizability was validated in the replication set, and AUC was used to evaluate the performance of the model. The 95% confidence interval of AUC was calculated using bootstrap with 1000 runs. Machine learning algorithms were implemented using Scikit-learn (version 0.24.1) packages [41] in Python.

Metabolic pathway analysis

The KEGG API (https://www.kegg.jp/kegg/rest/keggapi.html) was used to download the KEGG metabolic pathways and their associated metabolites and enzyme reactions. Based on the key panel of metabolites, we first identified the reactions they participated in, and then selected the enzymes involved (EC numbers in KEGG). Subsequently, we mapped these enzymes to genes and converted them to UniProt IDs. Finally, we performed a joint enrichment analysis with the obtained UniProt IDs and the key panel of metabolites, with the tool MetaboAnalyst 6.0 [42] (https://www.metaboanalyst.ca/MetaboAnalyst/home.xhtml).

We referred to the definition of the DA score as described in previous studies [43, 44] to capture the tendency for a pathway to have significantly increased/decreased levels of metabolites, relative to a control group. The differential abundance (DA) score was calculated by applying a differential abundance test (FDR corrected Student’s t tests) between AD and NC to all metabolites in a pathway. After determining which metabolites were significantly altered, the DA score was calculated as:

DA score = (No. of metabolites increased significantly - No. of metabolites decreased significantly)/No. of significantly differential metabolites in the pathway.

Thus, the DA score varies from −1–1. A score of −1 indicates that all significantly differential metabolites in a pathway decreased in abundance, while a score of 1 indicates that all significantly differential metabolites increased.

Individual-level metabolic pathway quantification

The KEGG metabolic pathways and their associated metabolites as hierarchical structures were obtained from KEGG API. The metabolites were mapped into the KEGG database to determine the pathways to which they belong based on the hierarchical relationship, and then the abundance of metabolites within the same pathway was summed for each sample to quantitatively characterize the activity levels of metabolic pathways. Subsequently, the data was log-transformed and z-score scaled across all samples for subsequent analysis.

Statistical analysis

Analysis of variance (ANOVA) and chi-square tests were used to compare continuous variables and categorical variables, respectively. Partial least squares discriminant analysis (PLSDA) was utilized to visualize metabolic differences between the AD group, MCI group, and normal control group. Fold change (FC) was calculated as the ratio of the mean metabolic abundance between the disease groups (AD or MCI) and the control group. Significantly differential metabolites between the disease group (AD, MCI) and the control group were identified using T-tests. After adjusting for the false discovery rate (FDR), a significance threshold of 0.05 was applied to all comparisons. Linear regression model was utilized and adjusted for covariates (i.e. age and sex). The debiased sparse partial correlation (DSPC) network was constructed using the panel of key metabolites with the tool MetaboAnalyst 6.0 [42]. The partial correlation coefficients with p-value < 0.05 were considered statistically significant, and the visualization of the DSPC network was realized by the Python package networkx (version 2.6.3) [45]. All the above analysis was performed using Python version 3.7. Statistical analysis was carried out with the Python packages scipy.stats (version 1.7.3) and statsmodels (version 0.13.2). PLSDA was constructed using the Python package Scikit-learn (version 0.24.1) [41].

Results

Study design and cohort description

The overall design of this study was illustrated in Fig. 1. Specifically, as part of the Zhangjiang International Brain Biobank (ZIB, https://zib.fudan.edu.cn), a total of 447 individuals were recruited for the study, including 188 AD, 181 MCI, and 78 NC. They were divided into the discovery cohort and the replication cohort, according to the batches of data collection (Fig. 1a). The demographic characteristics of those participants were summarized in Table 1. Compared to NC, the AD cases were significantly older and had fewer years of education. There were no sex and body mass index (BMI) differences between the AD and NC groups. As expected, cognitive scores based on Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment Basic (MoCA_B) and Addenbrooke’s Cognitive Examination III (ACEIII) were significantly lower in AD cases than NC. Both the discovery and the replication cohorts exhibited similar patterns of demographic characteristics (Table 1).

Fig. 1. Study design.

a This study covered a total of 447 individuals, who were divided into discovery cohort (n = 239) and replication cohort (n = 208). b An untargeted metabolomics approach based on LC–MS/MS was applied to plasma samples to get metabolic MS profiles. c The metabolic profiles of Alzheimer’s disease (AD) cases, mild cognitive impairment (MCI) cases, and normal controls (NC) were compared in the discovery cohort to obtain the differential metabolites. Then, a key panel of metabolites was further determined as the metabolic biomarker for AD using the forward feature selection algorithm. d Based on the biomarker, a logistic regression classifier for AD diagnosis was trained in the discovery cohort with 10-fold cross-validation and then evaluated in the replication cohort. e Metabolic pathways were quantified by summarizing the abundance of the measured metabolites within the pathway to represent metabolic changes at the pathway level. f Significantly altered pathways were validated in external cohorts by principal components analysis (PCA) and differential tests of pathway abundance.

Table 1.

Demographic characteristics of participants for metabolomic study.

	The Discovery Cohort				The Replication Cohort
	NC	MCI	AD	p-value (AD vs. NC)	NC	MCI	AD	p-value (AD vs. NC)
Subject,N	53	90	96	.	25	91	92	.
Age,mean ± SD(years)	62.9 ± 7.1	67.6 ± 6.9	72.8 ± 6.1	<0.001	64.8 ± 7.7	68.6 ± 6.3	70.6 ± 8.0	0.0014
Sex,N(male/female)	15/38	31/59	32/64	0.536	7/18	36/55	37/55	0.376
Edu_yrs,mean ± SD(years)	12.8 ± 3.0	10.7 ± 3.2	9.8 ± 3.6	<0.001	12.4 ± 3.1	11.6 ± 3.2	9.6 ± 4.3	0.003
BMI,mean ± SD(kg/m²)	23.0 ± 2.8	23.7 ± 3.0	23.4 ± 3.5	0.478	23.8 ± 3.7	23.6 ± 2.9	23.7 ± 3.1	0.933
MMSE,mean ± SD	28.2 ± 2.3	26.3 ± 2.0	16.8 ± 4.4	<0.001	28.8 ± 1.2	26.6 ± 1.7	17.1 ± 5.3	<0.001
MoCA_B,mean ± SD	26.0 ± 2.6	20.7 ± 4.1	11.7 ± 4.5	<0.001	27.0 ± 1.8	21.3 ± 3.2	12.8 ± 4.1	<0.001
ACEIII,mean ± SD	84.0 ± 6.2	68.6 ± 9.1	46.2 ± 13.1	<0.001	84.6 ± 7.3	71.2 ± 7.9	47.9 ± 13.6	<0.001

ANOVA (analysis of variance) test and Chi-squared test were applied to compare the differences between AD and NC for continuous variables and categorical variables, respectively. P-values in bold indicate statistical significance.

AD Alzheimer’s disease, NC normal control, MCI mild cognitive impairment, MMSE mini-mental state exam, Edu_yrs Years of education, BMI body mass index, MoCA_B montreal cognitive assessment basic, ACEIII, addenbrooke’s cognitive examination III.

The plasma metabolomics profile for each participant was obtained using an untargeted metabolomics approach based on LC–MS/MS (Fig. 1b). After MS/MS identification and data preprocessing, 1190 metabolites reproducibly detected in two batches were reserved for subsequent analyses, covering a broad spectrum of metabolic categories (Supplementary Table S1). The quality control (QC) samples in each cohort were closely clustered in the center of the principal components analysis (PCA) display (Supplementary Fig. 1a-b), and the correlations among QC samples in the discovery cohort (Supplementary Fig. 1c) and the replication cohort (Supplementary Fig. 1d) were both high, suggesting the stability of the output data.

Dysregulated metabolites in AD and MCI

To compare the plasma metabolic landscape of AD, MCI, and NC, we first performed the partial least squares discrimination analysis (PLSDA) to the metabolic profiles. As shown in Fig. 2a, a clear and obvious separation of metabolic profiles was observed between AD and NC groups, indicating that metabolic changes were widely present in AD. MCI samples were scattered among AD and NC samples, suggesting an intermediate progress. Next, we performed differential analysis and identified 72 differential metabolites in AD vs NC and 39 in MCI vs NC, respectively (T-tests, false discovery rate (FDR) < 0.05 and fold change >1.2 or <0.83) (Fig. 2b, c, Supplementary Table S2–3). The two sets of differential metabolites had a significant overlap (Fig. 2d; permutation test, p-value = 0.0001), primarily consisting of phosphatidylethanolamines (PEs) (Supplementary Table S4), indicating that some metabolic changes could persist across the pre-AD and AD stages. In fact, some lipidomics studies have revealed the significant associations between circulating PEs and cognitive decline [46], APOE genotypes [47], CSF tau [48], suggesting PEs as potential detection biomarkers for early Alzheimer’s disease [49]. Additionally, considering the variation of age and gender across different diagnosis groups that might be influential to metabolites, we also performed linear regression between disease status and metabolites by adjusting for the two covariates (i.e. age and sex). After adjusting for the covariates, most of the identified differential metabolites remained significant (Supplementary Table S5). The differential metabolites identified through linear regression showed significant overlap and strong concordance with those identified using the T-tests, which confirmed the reliability of the results (hypergeometric tests, p-value = 4.29e-55).

Fig. 2. Differential analysis.

a The partial least squares discrimination analysis (PLS-DA) of the global metabolomes of 239 plasma samples in the discovery cohort, including normal controls (NC, n = 53), mild cognitive impairment (MCI, n = 90), and Alzheimer’s disease (AD, n = 96). b-c Significantly differential metabolites between AD and MCI compared to NC (T-tests with the p-value adjusted by false discovery rate, FDR < 0.05) are shown in red (up-regulated, with fold change > 1.2) and blue (down-regulated, with fold change < 0.83) in the volcano plot. The top significant metabolites are labeled. d The Venn diagram shows the overlap of differential metabolites between AD vs NC and MCI vs NC. e. A hierarchical clustering heatmap illustrating the abundance (transformed to z-scores) of significantly differential metabolites between AD and NC.

Consistent with the results of PLSDA, the heatmap based on the differential metabolites showed distinct patterns between AD/MCI and NC samples (Fig. 2e, Supplementary Fig. 2). Most of the differential metabolites were down-regulated in the AD and MCI stages compared to NC, indicating that the relevant metabolic activity might be weakened. The differential metabolites encompassed a variety of metabolic classes. The most predominant class was “Lipids” (Supplementary Fig. 3a-b), suggesting an important role of lipid metabolism in the pathophysiological pathways of AD. In sub-class level categories (Supplementary Fig. 3c-d), the predominantly differential lipid metabolite categories were PEs and triglycerides (TGs) for AD vs NC, while those for MCI vs NC were PEs and phosphatidylcholines (PCs). These differences suggested that the metabolic dysregulation might have partially changed over the progression of AD pathology, culminating in substantial disturbances in triglycerides, which was consistent with previous reports that lower triglyceride levels were associated with a higher risk of Alzheimer disease [50–53]. Additionally, phosphatidylethanolamines (PEs) and phosphatidylcholines (PCs) were the most abundant phospholipids in biological membranes, and their importance in cellular transport, signal transduction, and energy metabolism has been demonstrated in numerous studies [54–58]. Our results showed that PEs and PCs were closely linked to disease progression, especially in the early stage of AD. Nevertheless, it is worthwhile to keep in mind that only a subset of MCI would actually progress to AD, thus some MCI profiles do not necessarily represent the status of early-stage AD.

Overall, our findings showed significant changes in the plasma metabolomes of individuals with AD and MCI compared to NC, suggesting that metabolic alterations could potentially serve as disease biomarkers.

Accurate detection of AD through key metabolites

Based on the differential metabolites, next we wanted to prioritize a key panel of metabolites for AD. As depicted in Fig. 1c and Supplementary Fig. 4a-b, we performed a forward feature selection algorithm to obtain an optimal set of metabolites for AD classification, from the 72 identified differential metabolites from the discovery cohort. We tested the performances of 6 widely used machine learning classifiers, including logistic regression, decision tree, random forest, support vector machine, k-nearest neighbors and naive bayes in the feature selection process. The AUC trajectories of these classifiers showed that the logistic regression model was stable and efficient (Supplementary Fig. 4c). As a result, we chose logistic regression to perform feature selection and prioritize the metabolites. We incrementally added features (metabolites) to train the logistic regression models while adjusting for the covariates--age, sex, education years, and body mass index (BMI). The optimal feature set was selected when the logistic regression model yielded the largest AUC distinguishing AD and NC (Supplementary Fig. 5a). As a result, the selected feature set comprised 22 metabolites as the key panel (Supplementary Table S6).

Employing this key panel of 22 metabolites, we constructed an AD-diagnostic model by training a new and robust logistic regression classifier with 10-fold cross-validation (Fig. 1d, Supplementary Fig. 4a). This model showed an AUC value of 0.935 (95% confidence interval (CI): 0.890–0.973) in the replication cohort (Fig. 3a). To estimate the contribution of covariates as additional predictors besides key metabolites, we also assessed the predictive performances of two additional logistic regression models using the covariates alone (as the baseline model), or using the 22 key metabolites alone. In the replication cohort, the logistic regression model using only the key metabolites achieved an AUC of 0.884, compared to 0.779 by the baseline model using only the covariate indicators (Fig. 3a). The presented diagnostic model combining key metabolites with covariate indicators provided optimal predictive performance. We also performed ANOVA to compare the two models (with and without covariates), which revealed a statistically significant difference (p = 1.4743e-06), suggesting the importance of including covariates in AD diagnosis. To further evaluate the impact of the covariates, we constructed a balanced test subset with similar distributions of covariates between AD and NC groups from the replication cohort, on which our diagnostic model yielded an AUC of 0.908, compared to 0.58 by the baseline model (Supplementary Fig. 5b-c). The significant difference in the predictive power of the two models in the balanced test subset confirmed the predictive value of key metabolites beyond covariate factors. We further validated our diagnostic model by assessing its classification performance separately in the male and female subsets. As a result, accuracy was higher in females (AUC = 0.944) than in males (AUC = 0.911), yet both groups showed strong discrimination between AD and NC, underscoring the robustness of the model (Supplementary Fig. 6). These results indicated that while covariates contributed to the discrimination between AD and NC, the majority of the predictive power was accounted for by the metabolite-based features. Additionally, we tried selecting the top 5, top 10, and top 15 metabolites (with which the original feature selection models also performed well, as shown in Supplementary Fig. 5a), respectively, to train reduced models for AD diagnosis. We found that the performances of these reduced models were also impressive (Supplementary Fig. 5d-e), indicating that the first few specific metabolites may serve as reliable biomarkers for AD diagnosis, such as TG(16:1_16:1_22:6), PE(P-18:1_18:2), Mandelic acid, Trigonelline, and Cer(d18:0/22:0) as top 5 metabolites. As these reduced models still performed slightly worse than the 22-metabolite model, we decided to adopt the model with 22 metabolites as our final AD-diagnostic model.

Fig. 3. A panel of metabolites identified for AD diagnosis.

a Performance of the logistic regression-based AD-diagnostic full model, the baseline model, and the metabolites-alone model in the replication cohort. b The debiased sparse partial correlation (DSPC) network of the key panel of 22 metabolites. Here, each node represented a metabolite, and each edge represented a significant partial correlation (p-value < 0.05) between the two nodes. Edge weight represented the partial correlation coefficient. Only significant partial correlations were illustrated. c Correlations of the abundance of the metabolites with the demographic characteristics (top) and the fold changes of each metabolite in AD compared with NC (bottom) in the discovery cohort. *p $\le$ 5e-02; **p $\le$ 1e-02; ***p $\le$ 1e-03. PCC, Pearson correlation coefficient; Edu_yrs, Years of education; MMSE, Mini-Mental State Exam; BMI, Body mass index; MoCA_B, Montreal Cognitive Assessment Basic; ACEIII, Addenbrooke’s Cognitive Examination III.

Similarly, we constructed a metabolic diagnostic model for MCI. Through feature selection, the model achieved the highest AUC when 20 differential metabolites were selected (Supplementary Fig. 7a). Using these 20 metabolites, we trained a logistic regression model on the discovery cohort and validated its performance in the replication cohort, achieving an AUC of 0.829 (Supplementary Fig. 7b). Furthermore, we also tested the diagnostic performance of the identified 22-metabolite panel for MCI, which yielded an AUC of 0.8 in the replication cohort (Supplementary Fig. 7c). These results suggested that the metabolite-based diagnostic models can also recognize MCIs accurately, but with lower accuracy compared to AD, possibly due to that only a small portion of MCIs would actually develop into AD.

To analyze the interactions among the 22 metabolites of the final AD-diagnostic model, we calculated the partial correlations among them and constructed a debiased sparse partial correlation (DSPC) network (Supplementary Table S7), as shown in Fig. 3b. Dense interactions were observed among the metabolites of the same class, with interactions among lipid metabolites being the most intensive. Most of the lipid-class metabolites were positively correlated with each other, indicating a highly coordinated metabolite regulatory network.

We also investigated the correlations of these metabolites with demographic characteristics (Fig. 3c, Supplementary Table S8). These metabolites all showed significant correlations with at least one of the three cognitive scales, with 19 of them significantly correlated with all three. These findings further supported the relevance of the 22 metabolites with AD, making them a possible biomarker for this disease and its early stage.

Dysregulated lipid pathways identified for AD

Groups of metabolites functioning in organized pathways may help elucidate the underlying functional changes in AD better than individual metabolites. As a result, we then investigated the altered metabolic pathways, which are collections of multiple metabolic reactions representing a higher level of biochemical activity. We conducted a joint enrichment analysis with the 22 metabolites and the corresponding enzymatic genes (see Methods), and identified 7 significantly enriched metabolic pathways (Supplementary Table S9). The enrichment bubble chart, as shown in Fig. 4a, highlighted pathways such as linoleic acid metabolism (FDR = 3.89e-56), glycerophospholipid metabolism (FDR = 9.72e-56), and ether lipid metabolism (FDR = 1.2e-43) pathways.

Fig. 4. Dysregulation of multiple metabolic pathways related to AD.

a Kyoto Encyclopedia of Genes and Genomes (KEGG) metabolic pathway enriched by the key panel of 22 metabolites. Hypergeometric Test followed by Benjamin-Hochberg false discovery rate (FDR) correction was used, and only pathways with FDR < 0.05 were presented. The differential abundance (DA) score represents the average gross change for all differential metabolites measured in a pathway. The size of the dot represents the number of significant metabolites in the pathway. b Pearson correlations between the significantly enriched pathways. c Boxplots of the individual-level pathway activities across AD, MCI, and NC samples in the discovery cohort and the replication cohort. The p-values were calculated with a two-sample T-test. ns: non-significant; *p $\le$ 5e-02; **p $\le$ 1e-02; ***p $\le$ 1e-03.

Some of these pathways have been reported as relevant for AD-related symptoms or neuropathological processes. Linoleic acid and alpha-linolenic acid are two essential fatty acids reported to be involved in memory loss, neuroinflammation, and neuronal apoptosis [59–62]. Alpha-linolenic acid may attenuate AD-related symptoms by reducing neuroinflammation through its antioxidant and anti-inflammatory actions [63]. Linoleic acid is a precursor of arachidonic acid, while arachidonic acid and its derivatives were reported to play an important role in neuroinflammation and synaptic functions [64, 65]. The levels of these fatty acids, whether in blood or brain were strongly associated with the progression of AD [66–68], suggesting their importance in the pathogenesis of AD. Additionally, growing evidence has indicated that lipid dysregulation is involved in the development of AD [69]. Sphingolipids were highly concentrated in the nervous system and played a crucial role in the formation of membranes and myelin sheaths [70]. Lipidomics evaluations of sphingomyelin (SM) levels in the AD brain have revealed global dysregulation of the SM pathway [71–73]. Previous studies have also reported changes in glycerophospholipids and their derivatives in the brains of individuals with AD [74, 75]. Glycerolipids have been reported to be associated with neurodegeneration [76, 77]. The accumulation of brain glycerolipids has been found in both aging and AD [78, 79].

We calculated the differential abundance (DA) score to capture the average overall change for all significantly altered metabolites within a pathway, highlighting the general tendency for a pathway with increased/decreased levels of metabolites compared to the control group. The DA score of the significantly altered pathways (Fig. 4a) indicated that most of these pathways decreased in activity. These results were consistent with the findings at metabolite level, where the most of the differential metabolites were down-regulated in the AD compared to NC. It should be noted, however, that the pathway-level results represented global patterns of metabolite aggregates within pathways, and that a few individual lipids in these pathways were significantly upregulated (Supplementary Fig. 8). Moreover, certain lipid classes were not captured in our study. Thus, our conclusions were drawn only from the lipid classes quantified here, and that the term “downregulation” applied specifically to the overall pathway level, not to individual lipid species.

To conclude, we identified significantly altered metabolic pathways, which aligned with known AD pathology, and showed that they were mainly down-regulated.

Lipid pathway dysregulation confirmed in external datasets

We next aimed to further confirm the usefulness of our key panel of 22 metabolites and the enriched pathways for AD diagnosis. Due to issues such as the measurement noises or the differences in metabolomics platforms, individual key metabolites often exhibited low reproducibility [80], making it difficult to replicate across multiple cohorts [81, 82]. Indeed, we found that the tested metabolites varied largely across different studies, so it was not feasible to directly apply our 22 metabolites to diagnose AD in external cohorts. However, pathways, being a higher-level entity, may be more stable. Consequently, we then quantified metabolic pathway activities of our identified pathways, and tried to validate them across external cohorts.

We quantified the individual-level expression of these pathways, by summarizing the abundances of the detected metabolites within each pathway for each sample, and then performing log transformation and Z-transformation across all samples (Supplementary Table S10, Methods). In this way, the pathway activities can be compared across samples. As in Fig. 4b, we found that metabolic pathways in the lipid metabolism group exhibited high correlation across samples in the discovery cohort. They formed a highly correlated cluster consisting of linoleic acid metabolism, alpha-linolenic acid metabolism, and arachidonic acid metabolism, indicating their consistent effects on AD [67]. We then estimated the variations in these metabolic pathways among different sample groups. Significant differences in pathway activities were observed for all the pathways among almost all comparison group pairs (Fig. 4c). Moreover, consistent with the group-level pathway activity changes indicated by the DA scores (Fig. 4a), all 7 pathways showed globally downregulated patterns in AD samples compared to NC samples (Fig. 4c).

For the validation, we tried to evaluate whether the quantified metabolic pathway activities of our 7 pathways can distinguish AD from NC in other cohorts. We used two completely independent external datasets from other AD metabolomics studies, Wang et al [37]. and Teruya et al [38]. For these two datasets, we similarly quantified the abundance of the metabolic pathways (Supplementary Table S11–12), and subsequently applied PCA on the pathway abundance to see whether AD groups can be separated from NC groups. For the Wang et al. cohort, metabolites of 5 among the 7 metabolic pathways were measured in their data, except for glycerolipid metabolism and ether lipid metabolism. As shown in Fig. 5a, we observed a clear separation between the disease groups (MCI, AD) and the NC participants, although the AD and MCI groups were not well distinguishable. The abundance levels of these pathways were down-regulated in AD and MCI compared to NC, with 3 of which were significant, consistent with what we observed in our dataset (Fig. 5b). For the Teruya et al. cohort, four pathways were available to use. Still, as depicted in Fig. 5c, PCA separated dementia subjects from healthy elders. Furthermore, 3 of 4 pathways showed down-regulated activities in dementia subjects compared to healthy elders, where the ether lipid metabolism pathways were significant (Fig. 5d).

Fig. 5. External validation of identified AD-related pathways.

a Principal Component Analysis (PCA) of the pathway activities comparing AD (red), MCI (orange), and normal controls (blue) in the Wang et al. cohort. b Comparisons of the pathway activities among AD, MCI, and normal controls in the Wang et al. cohort. c Principal Component Analysis (PCA) of the pathway activities comparing dementia subjects (pink) and healthy elders (HE, light blue) in the Teruya et al. cohort. d Comparison of the pathway activities of dementia subjects and healthy elders in the Teruya et al. cohort. ns: non-significant; *p $\le$ 5e-02; **p $\le$ 1e-02; ***p $\le$ 1e-03; ****p $\le$ 1e-04.

Taken together, these results confirmed that the metabolic pathways identified through our key panel of metabolites could distinguish AD from NC in external cohorts, indicating the robustness of our metabolic biomarker for AD diagnosis across different cohorts. Also, the consistent alterations in these pathways may actually be essential for AD/dementia.

Discussion

In this study, we employed clinical cohort analysis to investigate the changes in plasma metabolites throughout the progression of Alzheimer’s disease (AD) and identified circulating metabolites with potential diagnostic value. Specifically, by utilizing differential analysis and forward feature selection algorithm, we identified a panel comprising 22 metabolites. Furthermore, by quantifying the pathways associated with these metabolites, we consistently verified the robustness of AD-related pathways. Overall, our research demonstrated the extensive metabolic disruption in plasma metabolic profiling, which may enhance precise detection of AD and its early-stage (MCI), thereby offering potential for future clinical treatments.

Our differential analysis revealed significant metabolic disruptions during the AD stage, notably including a marked downregulation of lipids. Consistent with previous reports, our findings revealed widespread dysregulation of lipid metabolism in Alzheimer’s disease. In particular, alterations in phosphatidylcholines (PCs), phosphatidylethanolamines (PEs), and triglycerides (TGs) observed in this study were consistent with those reported by Liu et al. [21] using a smaller cohort specific for lipidomics. However, there were also some discrepancies. Liu et al. reported upregulation of certain diacylglycerols (DGs) and phosphatidylinositols (PIs) in AD, whereas most DGs and PIs in our study did not reach statistical significance. These discrepancies likely reflected differences in cohort characteristics and analytical platforms--Liu et al. analyzed an Australian cohort using a targeted lipidomics approach, while our study employed untargeted metabolomics in a Chinese cohort, encompassing 1190 metabolites including both lipids and non-lipids. Our analyses have, additionally, expanded previous observations by identifying additional sporadic lipid species—including DHA ethyl ester and 7-Oxateasterone—that exhibit stronger and more consistent associations with disease severity and cognitive decline. It is noteworthy that we also identified non-lipid metabolites associated with AD, including L-cystine and Phenylacetyl-L-glutamine, which may reflect disturbances in the metabolism of amino acids. These results have collectively not only replicated the central lipid dysregulation reported by Liu et al., but also greatly expanded the findings in both lipid and non-lipid metabolites that were involved in AD pathogenesis. Phosphatidylethanolamine (PE) exhibited significant differences in both MCI and AD stages, suggesting it may represent a metabolic change throughout the progression of AD. Several studies have also highlighted the critical role of PE in AD pathology [83, 84].

Beyond the primary changes observed in lipid species, our predictive panel also included Trigonelline and 1-Methylxanthine that were significantly decreased in AD. 1-Methylxanthine is a primary metabolite of caffeine and serves as a reliable biomarker for coffee consumption, which has been longitudinally associated with a reduced risk of cognitive decline [85, 86]. Specifically, 1-Methylxanthine contributes to the known benefits of caffeine by modulating adenosine receptors and reducing both amyloid-beta accumulation and tau hyperphosphorylation [87]. Meanwhile, Trigonelline--a structural analog of nicotinic acid--is a critical precursor in nicotinamide adenine dinucleotide (NAD + ) metabolism [88]. It has been shown that NAD+ metabolism involved essential cellular processes that protect neurons from the oxidative stress and mitochondrial dysfunction to counteract Aβ and p-Tau pathologies [89]. These results suggested that the progression of AD may involve not only systemic dysregulation of lipid metabolism but also a decline in protective, diet-derived metabolites.

We identified a panel of 22 metabolites and constructed a robust AD diagnostic model, which demonstrated strong performance in our replication cohort. Pathway enrichment analysis identified a set of pathways associated with AD symptoms or neuropathological processes, which were shown to be significantly and consistently downregulated at the overall pathway-level in AD. These AD-related pathways highlighted the importance of lipid metabolism (mainly included linoleic acid metabolism, alpha-linolenic acid metabolism and arachidonic acid metabolism), which may be involved in pathological mechanisms of AD. These findings suggested that future in-depth research on these related pathways may facilitate the discovery of useful targets for therapeutic strategies for Alzheimer’s disease.

In recent years, machine learning-based predictive models [21, 90, 91] often struggle with generalization, and the identified biomarkers have poor reproducibility due to the small overlap in the sets of metabolites measured in different metabolomics studies [80]. In this study, we identified a set of metabolites to construct an accurate diagnostic model for Alzheimer’s disease, and then innovatively designed a pathway quantification method to validate the metabolites at pathway levels across several external datasets. Based on this approach, we validated the consistent downregulation of the identified panel of metabolites in AD at the pathway level, thereby avoiding the issue of irreproducibility of metabolite-level findings across metabolomics studies. Our results consistently emphasized the importance of lipid metabolism in Alzheimer’s disease in in-house and external datasets.

The limitations of our study should also be noted. First, the key set of metabolites we identified comprised a significant proportion of lipids, particularly triglycerides (TG). This was likely due to the forward feature selection process, which prioritized the optimal subset of features for the best classification performance while disregarding potential interrelationships between features. We emphasized that the combination of metabolic features can achieve superior diagnostic performance, even though the main components (triglycerides) of this feature set may share similar biological functions, which was shown in DSPC network analysis (Fig. 3b).

Second, for MCI, our diagnostic models showed lower performance (Supplementary Fig. 7c). In fact, metabolome still demonstrated reasonable accuracies in early diagnosis (Supplementary Fig. 7a-b). The reason for lower performance with MCI could be that (1) during the MCI stage, the metabolome was less disturbed (consistent with the results of PLSDA), resulting in the identification of only a small number of MCI-related metabolites; (2) only a small portion of the MCI cases were actually going to progress to AD. It would be beneficial if future follow-up data could demonstrate whether the correctly predicted cases among MCI would develop into AD.

Last, we summarized the expression values of metabolite levels on the same pathway as the activity of this pathway. This was a rough estimation method that might not be accurate, because the metabolic reactions should include both the consumption and the production of metabolites. Development of more accurate quantification methods could potentially improve the accuracy of the results. Nevertheless, we showed that these pathways were consistently significant across different datasets.

In summary, we presented a comprehensive analysis based on high-coverage metabolomics for a Chinese AD cohort. Our approach delineated the landscapes of metabolic dysregulation in AD and MCI, and highlighted the potential role of lipid metabolism in AD mechanisms. Our findings may help advance the understanding of metabolic-related mechanisms of AD, and facilitate future clinical applications.

Author contributions

X. M. Zhao conceived the idea, designed and supervised the project. X. Luo conducted the data processing and analysis. L. Jia and J. Cao provided numerous constructive suggestions during the data analysis process. Q. Guo monitored the collection of samples and the construction of the cohort. X. Luo wrote the manuscript. X. M. Zhao and J. Chen helped revise the manuscript. All authors read and approved the final version of the manuscript.

Funding

This work was partly supported by Key Science and Technology Project of Hainan Province (ZDYF2024SHFZ058), National Natural Science Foundation of China (T2225015, 61932008, 62433008, 32200537), Shanghai Science and Technology Commission Program (23JS1410100), Foundation of the Shanghai Municipal Education Commission (No.24KXZNA11), National Key R&D Program of China (2023YFF1204800, 2020YFA0712403), Lingang Laboratory & National Key Laboratory of Human Factors Engineering Joint Grant (LG-TKN-202203-01) and Lingang Laboratory (LG-GG-202401-ADA010100 and LG-GG-202401-ADA050100). The funders had no role in the study design, data collection, data analysis, interpretation, writing of this manuscript, or decision to publish this study.

Data availability

All analyzed results supporting the findings of this study are available within the Supplementary Information files. All other data can be requested from the authors.

Code availability

In-house scripts used in this study are freely available at https://github.com/ZhaoXM-Lab/Metabolome_AD.

Competing interests

The authors declair no competing interests.

Consent statement

The procedures of this study were approved by the Ethics Committee of Shanghai Sixth People’s Hospital (approval number: 2019-032). All participants or their legal guardians provided written consent for research programs including this study. All methods were performed in accordance with the relevant guidelines and regulations.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-026-03933-7.