A total of 162 participants, over 50 years old, including 57 AD patients, 43 MCI patients and 62 CN subjects were included in this cross-sectional study. All participants were enrolled in China-Japan Friendship Hospital from April 2022 to November 2022. Apolipoprotein E (APOE) genotype testing, a battery of cognitive tests, and medical history gathering were all performed on each participant. The majority of participants underwent quantitative electroencephalography (qEEG) and Magnetic Resonance Imaging (MRI).

AD is clinically diagnosed with the 2011 National Institute on Aging-Alzheimer’s Association (NIA-AA) criteria (McKhann et al., 2011). MCI is also defined by the 2011 NIA-AA diagnostic criteria (Albert et al., 2011). CN controls were those who performed normally on the standardized neuropsychological tests and with or without cognitive complaints or concerns during the structured interview.

Listed below are exclusion criteria: (1) Cognitive dysfunction caused by severe psychiatric disorders or mental retardation, (2) Cognitive decline resulting from other nervous diseases, such as trauma, stroke, tumor, Parkinsonism, encephalitis or epilepsy or other types of dementia, such as vascular dementia (VaD), frontotemporal dementia (FTD), and Lewy body dementia (LBD), (3) Cognitive decline resulting from diseases of other systems such as severe anemia and thyroid disorders, (4) A history of malignant tumor, severe diseases, or other conditions affecting the urinary system, (5) A refusal to participate during neuropsychological testing; or incomplete clinical data.

The neuropsychological test battery included measures of global cognition such as Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) and cognitive performance in the domains of memory, executive function, attention, language and visuospatial ability. Activity of Daily Living Scale (ADL) was used for accessing the function ability during daily life. The specific scales have been applied in clinical practice and published in previous articles from our team (Qiao et al., 2022).

The urine samples (200 μL) were placed in the centrifuge tubes and re-suspended with prechilled 80% methanol by well vortex. The samples were then centrifuged at 15,000 g for 20 min at 4°C after being incubated on ice for 5 min. A portion of the supernatant was diluted with LC–MS grade water to a final concentration that contained 53% methanol. The samples were then moved to a brand-new centrifuge tube and centrifuged there for 20 min at a speed of 15,000 g at 4°C. The supernatant was then added to the analysis of the LC–MS/MS system (Want et al., 2006; Barri and Dragsted, 2013).

In Novogene Co., Ltd. (Beijing, China), UHPLC–MS/MS analyzes were carried out utilizing a Vanquish UHPLC system (Thermo Fisher, Germany) paired with an Orbitrap Q Exactive™HF mass spectrometer (Thermo Fisher, Germany). A 17-min linear gradient was used to inject samples onto a HypesilGoldcolumn (100 × 2.1 mm, 1.9 ) at a flow rate of 0.2 mL/min. Eluents A (0.1% formic acid (FA) in water) and B (methanol) were used in the positive polarity mode. Eluents A (5 mM ammonium acetate, pH 9.0) and B (Methanol) were used in the negative polarity mode. The following settings were made for the solvent gradient: 2% B, 1.5 min; 2–85% B, 3 min; 85–100% B, 10 min; 100–2% B, 10.1 min; 2% B, 12 min. With a spray voltage of 3.5 kV, capillary temperature of 320°C, sheath gas flow rate of 35 psi, aux gas flow rate of 10 L/min, S-lens RF level of 60, and aux gas heater temperature of 350°C, the QExactive^™HF mass spectrometer was operated in positive/negative polarity mode.

Peak alignment, peak selection, and quantification for each metabolite were carried out using Compound Discoverer 3.1 (CD3.1, Thermo Fisher) to handle the raw data files produced by UHPLC–MS/MS. The following primary parameters were set: 0.2 min for the retention time tolerance; 5 ppm for actual mass tolerance; 30% for the signal intensity tolerance; 3 for the signal/noise ratio; and minimum intensity, etc. Peak intensities were then normalized to the total spectral intensity. The molecular formula was predicted using the normalized data based on additive ions, molecular ion peaks, and fragment ions. The peaks were then compared with the MassList, mzVault, and mzCloud databases to generate accurate qualitative and relative quantitative findings. When data were not normally distributed, standardize using the following method to produce relative peak areas: sample raw quantitation value/(The sum of sample metabolite quantitation value/The sum of QC1 sample metabolite quantitation value); Finally, the findings of the relative quantification and metabolite identification were determined after molecules with relative peak areas in QC samples with CVs more than 30% were eliminated. The Kyoto Encyclopedia of Genes and Genomes (KEGG)¹, Human Metabolome Database (HMDB),² and Lipid metabolites and pathways strategy (LIPIDMaps)³ databases were used to annotate these metabolites.

The statistical software R (version 3.4.3) and Python (version 2.7.6) were used to conduct the statistical analysis. Partial least squares discriminant analysis (PLS-DA) were performed at metaX (Wen et al., 2017). We used t-test to determine the statistical significance (p-value). The metabolites with Variable Importance in the Projection (VIP) > 1 and value of p<0.05 and fold change (FC) ≥ 1.2 or FC ≤ 0.833 were regarded as differential metabolites. The functions of these metabolites and metabolic pathways were researched using the KEGG database. SPSS 23.0 was used for statistical analysis. The Shapiro–Wilk test was used to examine the normality of quantitative data. Mean (x ± s) was used for the description of normal data while non-normal data used median (P25, P75). Analysis of Variance (ANOVA) was used for normal data mean comparison while the Kruskal-Wallis H test was utilized for non-normal data distribution comparison. For post hoc comparisons, value of ps were Bonferroni-corrected. Besides, Pearson’s chi-square test or Fisher’s exact probability were used for comparision of the proportions of categorical variables. Statistical significance was defined as a two-tailed value of p<0.05.

Machine learning was utilized to identify the optimal multivariate signatures, which took both metabolites and demographic data (age and APOE ε4 status) as input parameters, in order to discriminate AD from CN and MCI from CN. The classifier was made up of feature selection and classifiers (Shi et al., 2019). The dataset was separated into a training set (0.7) and a test set (0.3). The “n” top input variables with the lowest mean square error (MSE) that best distinguished AD or MCI diagnostic groups were chosen using the least absolute shrinkage and selection operator (LASSO). Support vector machine (SVM) classifiers were constructed in order to anticipate the outcome under 10-fold cross-validation on top of these “n” features. The kernel functions of linear, polynomial, radial, and sigmoid were contrasted. When evaluating the model in the test set, accuracy and area under the curve (AUC) (Receiver Operating Characteristic, ROC curve) were applied to assess the diagnostic value.

Basic information and clinical characteristics of enrolled participants were included in Supplementary material. A total of 2,140 metabolites were identified. Compared to the CN group, significantly differential metabolites were filtered in the AD group and MCI group by setting VIP > 1.0, FC > 1.2 or < 0.833, and value of p<0.5. The expression of the differential metabolites in AD group was displayed as a volcano plot and the top 10 regulated metabolites either upregulated or downregulated were displayed in a stem plot (Figures 1A,B) while the expression of the differential metabolites in MCI group was shown in Figures 1C,D. There were 125 significantly differential metabolites between the AD and CN groups, including 46 upregulated ones and 79 downregulated ones. In parallel, there were 93 significantly differential metabolites between the MCI and CN groups, including 23 upregulated ones and 70 downregulated ones. 23 metabolites were significantly regulated in both AD and MCI group, including 6 upregulated ones and 17 downregulated ones. Atropine was the most upregulated metabolite in both AD and MCI group while riboflavin-5-phosphate and N1-isopropyl-2-(phenylthio)benzamide was the most downregulated metabolite in AD and MCI, respectively. All differential metabolites were shown in a heatmap in Supplementary Figure S1. Z-scores of the top 30 differential metabolites were shown in Supplementary Figure S2.

Among the top 20 differential metabolites ranged by value of p, many metabolites were correlated with each other, most positively. A strong correlation (r > 0.7) between metabolites was shown in the chord diagram (Figure 2). In AD-CN group, robinetin was found to be positively correlated with 1,7-bis(4-hydroxyphenyl) heptan-3-one, 4-hexyloxyaniline was found to be positively correlated with 6,7,8-trimethoxy-2-[3-(trifluoromethyl) phenyl]-4H-3,1-benzoxazin-4-one and desmethylcitalopram was found to be positively correlated with 2-[(3S)-1-(3-Methoxybenzyl)-3-pyrrolidinyl]-1,3-benzothiazole. In MCI-CN group, 3-hydroxydecanoic acid, decanoic acid, undecanoic acid and N-(p-Coumaroyl) serotonin were found to be positively correlated with each other. Decanoic acid, N-(p-Coumaroyl) serotonin and capric acid were also found to be positively correlated with each other. There was also strong positive relationship between metabolites which could not be categorized. The overall correlation heatmap was shown in Supplementary Figure S3 when the red indicated a positive relationship and the blue indicated a negative relationship.

According to KEGG enrichment analysis results, caffeine metabolism was enriched in AD-CN group (p < 0.05). In parallel, vitamin B6 metabolism and fructose and mannose metabolism pathway were enriched in MCI-CN group (p < 0.05). The relative pathways were shown in Supplementary Figure S4. The metabolic network showed the relationship among compounds, pathways, modules, enzymes and reactions to facilitate the presentation of the overall metabolic response. In AD-CN group, 32 compounds were mapped with the KEGG database while 17 compounds were differential metabolites as indicated by green squares. Neuroactive ligand-receptor interaction, galactose metabolism, dopaminergic synapse and other 8 metabolism pathways were enriched as shown in Figure 3A. In MCI-CN group, 29 compounds were mapped with the KEGG database while 10 compounds were differential metabolites as indicated by green squares. Degradation of flavonoids and vitamin B6 metabolism, and other 4 metabolism pathways were enriched as shown in Figure 3B. Antigen processing and presentation, vitamin digestion and absorption and lysosome were enriched in both AD-CN and MCI-CN group.

Based on previous analysis, we extracted differential metabolites plus age and APOE ε4 status to construct the LASSO model. Based on LASSO results, we built SVM classifiers with 10-fold cross-validation to investigate the ideal multivariate signatures that distinguished AD from CN. After training in training sets, we compared the results of test sets using different kernel functions in SVM. For AD-CN model, 30 metabolites, age and APOE ε4 status were identified when MSE reached minimum with the value of lambda (min) equaling 0.03614 (Figure 4A). The linear kernel function achieved the highest predictive value with an accuracy of 0.9143 in AD-CN group. Figures 4B,C showed the ROC curves in training set and test set in AD-CN group. Similarly, for MCI-CN model, 45 metabolites, age and APOE ε4 status were identified when MSE reached minimum with the value of lambda (min) equaling to 0.02197 (Figure 4D). Linear kernel function achieved the highest predictive value with an accuracy of 0.6452 in MCI-CN group while Figures 4E,F showed the ROC curves in MCI-CN group. The optimal model achieved an AUC of 0.9575 in AD-CN group and an AUC of 0.7333 in MCI-CN group in test sets. The specific metabolites included in the diagnostic panel were shown in Supplementary Table S3. The evaluation of diagnostic models was shown in Supplementary Table S4.

Atropine, S-Methyl-L-cysteine-S-oxide, D-Mannose 6-phosphate (M6P), Spiculisporic Acid, N-Acetyl-L-methionine, 13,14-dihydro-15-keto-tetranor Prostaglandin D2, Pyridoxal 5’-Phosphate (PLP) and 17(S)-HpDHA were considered valuable for both AD and MCI diagnosis. They were also differential metabolites for both AD-CN and MCI-CN group and were defined as hub metabolites. The boxplots showed the log2 transformed quantitative value (Figure 5).

Hub metabolites were found to be correlated with cognitive tests, although most weakly (Figure 6). Significant labels were shown on the dots. Among 8 diagnostic metabolites, atropine, M6P and N-Acetyl-L-methionine were significantly correlated with more than half of cognitive tests while S-Methyl-L-cysteine-S-oxide, 13,14-dihydro-15-keto-tetranor Prostaglandin D2, PLP and 17(S)-HpDHA were significantly correlated with less than half cognitive tests. Nevertheless, none of the correlations between Spiculisporic Acid and cognitive domains reach significance. The relative ρ and p were shown in Supplementary Table S5 and scatter dot plots were shown in Supplementary Figures S5–S11.