All patients who were assessed aged ≥65 years old. MCI was defined according to the following criteria: 1) cognitive concern or complaint by the subject, informant, nurse, or physician, with Clinical Dementia Rating (CDR) <0.5; 2) objective impairment in at least one cognitive domain based on performance of 1.5 SD below the mean using the norms obtained in the pilot study; 3) essentially normal functional activities, determined by the CDR and the Activities of Daily Living evaluation; and 4) absence of dementia, decided by Diagnostic and Statistical Manual of Mental Disorders IV. Cognitive function assessment was done to enroll all the local residents aged ≥65 who participated in the physical examination based on a voluntary basis and informed consent. Two years later, follow-up was also managed on a voluntary basis. Eight physicians, well trained by Shanghai Mental Health Center to be qualified in cognitive evaluation, performed the cognitive function assessment. Cognitive impairment caused by the following conditions were excluded: a history of stroke, Parkinson’s disease, infection, poisoning, trauma, severe hearing impairment, mental illness, abnormal cardiac function with a brain natriuretic peptide test, and drug-substance abuse. The other diseases associated with cognitive impairment were excluded by routine hematological examination and CT/MRI, and finally serum Aβ and tau (Simoa platform) were tested voluntarily, so that the last patients with MCI enrolled were MCI due to AD. Collect stool specimens at this point in time (within 1 week). After 2 years, MCI progression was judged by Global Deterioration Scale; those who had scores ≥1 were defined as progressive MCI group (P-G), and those who had scores <1 were defined as unprogressive MCI group (U-G) (Reisberg et al., 1982; Corbi and Burgos, 2022). Our study was approved by the Medical Ethics Committee of Shanghai Pudong New Area People’s Hospital, Shanghai, China (K44). Written informed consent was obtained from all participants or their legally acceptable representatives.

Accurately weigh an appropriate amount of sample into a 2-mL centrifuge tube, add 600 µL of MeOH [containing 2-amino-3-(2-chloro-phenyl)-propionic acid (4 ppm], and vortex for 30 s; add steel balls and place in a tissue grinder for 120 s at 50 Hz at room temperature and subject to ultrasound for 10 min; centrifuge for 10 min at 12,000 rpm and 4°C, filter the supernatant by 0.22 μm membrane, and transfer into the detection bottle for Liquid Chromatograph-Mass Spectrometer (LC-MS) detection.

The LC analysis was performed on a Vanquish UHPLC System (Thermo Fisher Scientific, USA). Chromatography was carried out with an ACQUITY UPLC® HSS T3 (2.1 mm × 100 mm, 1.8 µm) (Waters, Milford, MA, USA). The column maintained at 40°C. The flow rate and injection volume were set at 0.3 mL/min and 2 μL, respectively. For LC-Electrospray Ionization Mass Spectrometry(ESI)(+)–MS analysis, the mobile phases consisted of (B2) 0.1% formic acid in acetonitrile (v/v) and (A2) 0.1% formic acid in water (v/v). Separation was conducted under the following gradient: 0~1 min, 8% B2; 1~8 min, 8%~98% B2; 8~10 min, 98% B2; 10~10.1 min, 98%~8% B2; and 10.1~12 min, 8% B2. For LC-ESI(−)–MS analysis, the analytes were carried out with (B3) acetonitrile and (A3) ammonium formate (5mM). Separation was conducted under the following gradient: 0~1 min, 8% B3; 1~8 min, 8%~98% B3; 8~10 min, 98% B3; 10~10.1 min, 98%~8% B3; and 10.1~12 min, 8% B3.

Mass spectrometric detection of metabolites was performed on Orbitrap Exploris 120 (Thermo Fisher Scientific, USA) with ESI ion source. Simultaneous MS1 and MS/MS (full MS-ddMS2 mode, data-dependent MS/MS) acquisition was used. The parameters were as follows: sheath gas pressure, 40 arb; aux gas flow, 10 arb; spray voltage, 3.50 kV and −2.50 kV for ESI(+) and ESI(−), respectively; capillary temperature, 325°C; MS1 range, mass-to-charge ratio (M/Z) 100–1,000; MS1 resolving power, 60,000 full width at half maximum (FWHM); number of data dependent scans per cycle, 4; MS/MS resolving power, 15,000 FWHM; normalized collision energy, 30%; and dynamic exclusion time, automatic.

The raw data were firstly converted to mzXML format by MSConvert in ProteoWizard software package (v3.0.8789) and processed using R XCMS (v3.12.0) for feature detection retention time correction and alignment.

Two different multivariate statistical analysis models, unsupervised and supervised, were applied to discriminate the groups [Principal Component Analysis (PCA), Partial least squares Discriminant Analysis (PLS-DA), and Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA)] by R ropls (v1.22.0) package. The statistical significance of P-value was obtained by statistical test between groups. Finally, combined with P-value, Variable Importance for the Projection (VIP) (OPLS-DA variable projection importance) and fold change (FC) (multiple of difference between groups) were included to screen biomarker metabolites. By default, when P-value < 0.05 and VIP value > 1, the metabolite was considered to have significant differential expression.

Differential metabolites were subjected to pathway analysis by MetaboAnalyst, which combines the results from powerful pathway enrichment analysis with the pathway topology analysis. The identified metabolites in metabolomics were then mapped to the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway for biological interpretation of higher-level systemic functions. The metabolites and corresponding pathways were visualized using KEGG Mapper tool.

All the analyses were performed with the genes cloud tools, a platform online for data analysis (https://www.genescloud.cn).

Cohort 1 included 87 subjects and was divided into three groups (normal group, 30; MCI group, 22; and AD group, 35). Cohort 1 included additional 87 patients with MCI who were followed up for 2 years and finally were divided into two groups in which P-G included 32 patients and U-G included 55 patients.

Table 1 shows the demographic characteristics of the patients of three groups (NC, MCI, and AD), and Table 2 shows the demographic characteristics of the patients of two groups (P-G and U-G). As indicated by the results, the education times in the AD group were lower than those in the MCI group, and those in the MCI group were lower than those in the normal group (P < 0.001), and serum Aβ1-40 levels were lower in the cognitively impaired group than that in the normal group (P = 0.04). As indicated by the results of P-G and U-G, serum Aβ1-40 levels in the P-G group was lower than that in the U-G group (P = 0.04).

As shown in Figure 1 , the results of fecal metabolites analysis in the AD, MCI, and normal groups were reliable. There were 21 different metabolites in the three groups ( Figure 1B ), in which positive ion included Pelletierine, 3-Amino-4-hydroxybenzoate, o-Hydroxylaminobenzoate, 2-Deoxystreptamine, p-Synephrine, alpha-Curcumene, Stearidonic acid, Piperine, 3-Ketosphingosine, Allopregna, 9(S)-HPOT, Tetrahydropalmatine, Docosatetraenoyl Ethanolamide, and Erythrodiol; and negative ion included 4-Aminophenol, 4-Hydroxycinnamic acid, D-Glucuronic Acid, Syringic acid, Dihydrotestosterone, All-trans-Retinoic acid, and All-trans-Retinoic acid). 3-Amino-4-hydroxybenzoate was positively correlated with p-Synephrine and o-Hydroxylaminobenzoatev ( Figure 1E ). Alpha-Curcumene was positively correlated with Erythrodiol ( Figure 1E ).

As shown in Figure 2 , there were eight metabolites that differ significantly among the three groups, specifically, alpha-Curcumene, Erythrodiol, Biocytin, 9(S)-HPOT, Dihydrotestosterone, D-Glucuronic Acid, p-Synephrine, and o-Hydroxylaminobe. In addition, 9(S)-HPOT was helpful to identify people with cognitive impairment. Furthermore, D-Glucuronic Acid and Dihydrotestosterone have high sensitivity and specificity for determining AD, but o-Hydroxylaminobenzoate and p-Synephrine for MCI.

The differential metabolites of the three groups were analyzed, and the model was established by the machine learning method. The results are shown in Figure 3 , in which the MCI diagnosis model established according to the top seven characteristics of weight before the difference substances had the best effect (Area Under Curve (AUC) = 0.7007). The first seven metabolites in the weight order were M425T668 (KEGG C20945 named Erythrodiol), M203T497 (KEGG C09649 named alpha-Curcumene, which was classed in Prenol lipids and subclassed in Sesquiterpenoids), M168T320 (KEGG C04548 named p-Synephrine, which was classed in Phenols and subclassed in 1-hydroxy-2-unsubstituted benzenoids), M154T87 (KEGG C16235 named o-Hydroxylaminobenzoate, which was with the pathway of aminobenzoate degradation and microbial metabolism in diverse environments), M154T107 (KEGG C12115 named 3-Amino-4-hydroxybenzoic acid), M163T416 (KEGG C02627 named 2-Deoxystreptamine, which was with the pathway of biosynthesis of other secondary metabolites), and M311T501 [KEGG C16321 named 9(S)-HPOT, which was with the pathway of lipid metabolism].

The result of differential material enrichment analysis of the three groups is shown in Figure 4 . The metabolic pathways of different substances are mainly enriched in five pathways including pathways in cancer and intestinal immune network for Immunoglobulin A (IgA) production and alpa-Linolenic acid metabolism and small-cell lung cancer and interleukin17 (IL-17)–producing T helper (Th17) cell differentiation.

Between P-G and U-G, there were 19 differential metabolites that are shown in Figures 5A–E , including positive ion (L-Proline, Mesaconate, Lysine, Perillyl alcohol, Indoleacetaldehyde, Iminoarginine, Caryophyllene alpha-oxide, Indolelactic acid, Chlorpheniramine,Palustradienal, Sphingosine, and Dicyclomine) and negative ion (2-Iminobutanoate, R-3-Hydroxybutyric acid, Betaine, 2-Oxo-4-methylthiobutanoic acid, Alpha-Oxo-benzeneacetic acid, N-Acetyl-L-aspartic acid, and Probenecid). Perillyl alcohol was positively correlated with L-Proline and Caryophyllene alpha-oxide ( Figure 5F ).

The characteristics of metabolites were analyzed by machine learning method, and the prediction model of MCI progress was established, as shown in Figure 6 . The best model results were established by the first seven different substances (AUC = 0.7402). The first six metabolites in the weight order were M206T310 (KEGG C02043 named Indolelactate, which was classed in Indoles and derivatives and subclassed in Indolyl carboxylic acids and derivatives), M160T246 (KEGG C00637 named Indole-3-acetaldehyde, which was classed in Indoles and derivatives and subclassed in Indoles, which was with the pathway of Tryptophan metabolism), M116T53 (KEGG C00148 named L-Proline, which was classed in Carboxylic acids and derivatives and subclassed in amino acids, peptides, and analogs, which was with the pathway of arginine and proline metabolism; central carbon metabolism in cancer; membrane transport; digestive system; and translation), M153T564 (KEGG C02452 named Perillyl, which was classed in Prenol lipids and subclassed in Monoterpenoids alcohol, which was with the pathway of Limonene degradation and Monoterpenoid biosynthesis), M130T44_2(KEGG C01732 named Mesaconate, which was with the pathway of glyoxylate and dicarboxylate metabolism, carbon metabolism, and C5-branched dibasic acid metabolism), and M300T451 (KEGG C00319 named Sphingosine, which was classed in Organonitrogen compounds and subclassed in Amines, which was with the pathway of Sphingolipid metabolism, Sphingolipid signaling pathway, Apoptosis, and Necroptosis).

The result of differential material enrichment analysis of the two groups is shown in Figure 7 . The metabolic pathways of different substances are mainly enriched in four pathways: apoptosis (cellular processes), necroptosis (cellular processes), glyoxylate and dicarboxylate metabolism (metabolism), and sphingolipid signaling pathway (environmental information processing).

There were seven characteristic metabolites of MCI and six characteristic metabolites of progressive MCI analyzed by machine learning. The relationship between these 13 characteristic metabolites and serum biomarkers is shown in Table 3 . Nine of the 13 characteristic metabolites were associated with serum biomarkers and the concentration of D-Glucuronic acid was negatively correlated with APOE4 (B = −3.15, P = 0.003). Notably, Perillyl alcohol was negatively correlated with all of the five biomarkers (P-tau181, NF-light, Aβ1-42, Aβ1-40, and GFAP), but Indoleacetaldehyde was positively correlated with three biomarkers (P-tau181, Aβ1-42, and GFAP). Three characteristic metabolites (3-Amino-4-hydroxybenzoate, 2-Deoxystreptamine, and p-Synephrine) were positively correlated with Aβ1-42. 2-Deoxystreptamine, 9(S)-HPOT, Indoleacetaldehyde were positively correlated with GFAP. L-Proline and Indoleacetaldehyde were positively correlated with NF-light.