Electroacupuncture may improve memory problems in Alzheimer's by affecting the gut-brain connection

Senescence-Accelerated Mouse Prone 8 (SAMP8) and Senescence-Accelerated Mouse Resistant (SAMR1) aged 24 weeks were purchased from Peking University Health Science Center (Beijing, China). All the animals were housed at a temperature of 22 ± 1 °C, relative humidity of 50 ± 1%, and a light/dark cycle of 12/12 h; they were fed with sterile food and water ad libitum. Prior to formal experiments, mice with cognitive impairment were excluded through screening via the novel object recognition test.

The animals were randomly divided into four groups, and each group contained 12 male mice: (1) Control group: Consisted of SAMR1 that received simultaneous restraint with the EA group, without any additional intervention (n = 12). (2) Model group: Consisted of SAMP8 that received simultaneous restraint with the EA group, without any additional intervention (n = 12). (3) EA group: Consisted of SAMP8 underwent electroacupuncture treatment for 4 weeks (n = 12). Specific protocols were as follows: underwent electroacupuncture at bilateral GV20 and BL23 according to a widely recognized mouse acupoint atlas [1]. The location of the three acupoints was also clearly described in a systematic review of animal-based studies on the effects of acupuncture on AD [2]. For the GV20 acupoint (Baihui), acupuncture needles (specification: 0.16 mm × 7 mm) manufactured by Universal Acupuncture Medical Equipment Co., Ltd. (Suzhou, China) were inserted horizontally at a 15-degree angle to a depth of 1–2 mm beneath the skin/acupoint. For the BL23 acupoint (Shenshu), needles were inserted perpendicularly (90 degrees) to a depth of 2–3 mm. Subsequently, the needles were connected to the Han's electroacupuncture apparatus (HANS-100A, Nanjing Jisheng Medical Technology Co., Ltd., Nanjing, China). Electroacupuncture parameters were set to continuous wave mode with a current intensity of 1 mA and a frequency of 2 Hz. The stimulation intensity was adjusted to induce a slight tremor of the needle handle while ensuring the mouse exhibited no struggling or vocalization. Daily electroacupuncture intervention was administered once per day by operators who had received standardized training prior to the study. Each treatment course lasted 6 consecutive days. Mice in the treatment group received a total of four treatment courses, with one-day intervals between courses. (4) EA+Antibiotic group: Antibiotics were administered to SAMP8 via drinking water (ampicillin 1 mg/mL, metronidazole 1 mg/mL, neomycin 1 mg/mL, vancomycin 0.5 mg/mL. All the above antibiotics were purchased from Sigma-Aldrich Co., Ltd., St. Louis, MO, USA) for 2 weeks before EA intervention in the EA+Antibiotic group, and discontinued during EA treatment (n = 12). The control, Model, and EA groups received normal sterile water without antibiotics.

Mice were housed in individual ventilated cages with 3 mice per cage, and bedding was changed every 3 days. Restraint was applied for 15 min daily (consistent with EA treatment duration) in the Control and Model groups to match the stress level of the EA and EA+Antibiotic groups.

Water intake was recorded daily by measuring the volume of water consumed per cage, and body weight was measured weekly. No significant differences in water intake (mean 4–5 mL/mouse/day) or body weight (stable at 28–32 g) were observed among groups during the experiment, indicating that antibiotics did not affect water intake or general health of the mice.

The detailed information including background, research purpose, study design, and experimental methods were submitted during application process. The experimental protocol was approved (HUCMS-00312778) by the Animal Ethics Committee of the Hubei University of Chinese Medicine, and animal experiments were performed in compliance with the guidelines of the Animal Ethics Committee of the University.

All mice underwent the Morris water maze test (MWM) on the day following the completion of the aforementioned interventions to assess their spatial learning and memory abilities. All mice were required to complete both the place navigation test and the spatial probe test. Specifically, formal place navigation training was conducted once daily for 5 consecutive days. During each hidden platform trial, a mouse was placed into the water facing the wall of the circular black pool (diameter: 120 cm, height: 50 cm) from one of four entry points located in different quadrants. The time taken for the mouse to find and climb onto the hidden platform in the fixed quadrant within 1 min was recorded as the escape latency (in seconds), serving as the behavioral performance indicator. If a mouse failed to reach the target platform within 1 min, the escape latency was recorded as 60 s by default. Four platform trials were performed per mouse each day to obtain an average escape latency value. On the day following the completion of place navigation training, the spatial probe test was conducted. The target platform was removed, and the mouse was placed into the water from any entry point in a quadrant other than the original platform quadrant. The number of times the mouse crossed the original platform site within 1 min was recorded. Throughout the entire experiment, data recording and analysis were performed using the WMT-100 Morris video tracking system (Chengdu Taimeng Technology Co., Ltd., Chengdu, China).

After the MWM test, the novel object recognition test (NOR) was performed. The NOR test was used to evaluate short-term memory and performed as previously described. At first, animals were placed in an open arena along with two identical objects situated 15 cm from the side walls in two opposite parts of the maze. Animals were left to freely explore the arena and the object for 6 min before being placed back in their cage. After 60 min, a novel object was placed instead of one of the two identical initial objects. The animals were then given an additional 6 min to explore the new arena. For each animal, the discrimination index (the percentage of the time exploring the novel object among the total time spent exploring both objects) was determined.

At the end of the behavioral test, animals were first anesthetized with sodium pentobarbital (50 mg/kg). Six animals per group were then euthanized by cervical dislocation, followed by immediate dissection of fresh hippocampal and cortical tissues on ice. The remaining six animals per group underwent transcardial perfusion with ice-cold 0.9% sodium chloride, followed by perfusion with 4% paraformaldehyde (PFA) via left ventricular catheter until cardiac arrest occurred to complete the euthanasia procedure. The mid-sagittally cut hemispheres underwent paraffin embedding, and 5 μm thick serial sections were prepared using the HM355S rotary microtome (Thermo Fisher Scientific, Walldorf, Germany), which features a waterfall-based section transfer system and Peltier cooling elements. Initial sections from each animal series were prepared and analyzed following routine Nissl staining, while consecutive serial sections were processed for immunohistochemistry.

Paraffin sections were dewaxed to water. Staining the slices with toluidine blue (G1032-500 mL; Servicebio, Wuhan, China), soaking the slices in dyeing solution for 5 min, washing with water, slightly differentiating with 1% glacial acetic acid (G10000218-500 mL; Servicebio, Wuhan, China), washing with tap water to stop the reaction, controlling the degree of differentiation under microscope, and drying the slices in oven after washing with tap water. Then put the slices in clean xylene and transparent for 5 min, and seal the slices with neutral gum. Observe and take pictures under a microscope. Nissl bodies of the hippocampus and cortex were dark blue with a light blue background. The counts of Nissl bodies per area (cell counts/mm²) were recorded.

Paraffin sections were dewaxed and rehydrated after drying at 37 °C overnight. For antigen retrieval, the hydrated sections were placed in sodium citrate buffer (P0081; Beyotime, Shanghai, China) and heated with a microwave oven (Midea, Foshan, China) until boiling for 15 min, followed by natural cooling at room temperature. The sections were then incubated with 3% hydrogen peroxide solution at 37 °C for 25 min to quench endogenous peroxidase activity, followed by washing with phosphate-buffered saline buffer solution 3 times. Afterward, the sections were incubated with primary antibodies against Tau (10274-1-AP; Proteintech, Wuhan, China) with the dilution ratio of 1:50 overnight at 4 °C, followed by incubation with the horseradish peroxidase-conjugated goat anti-rabbit IgG antibody with the dilution ratio of 1:200 for 2 h. The sections were stained with the 3,3′-diaminobenzidine substrates (P0202; Beyotime, Shanghai, China) and counterstained with hematoxylin, and dehydrated with ethanol and xylene to prepare for slide mounting. The images were viewed by a bright-field microscope (Olympus VS200, Tokyo, Japan), and the staining intensity of each image was quantified using Image J 2.0 software. The integrated optical density (IOD) and its values of the positive area (IOD/area = mean optical density (MOD)) of each image were determined and obtained as representative TAU staining density. Higher mean optical density indicates stronger positive expression.

TNFα, IL1β and Aβ_1-42 concentrations in the mouse hippocampus and cortex were quantified using a sandwich enzyme-linked immunosorbent assay (ELISA). Briefly, dissected hippocampal and cortical tissues were homogenized in ice-cold phosphate-buffered saline (PBS) containing protease inhibitors. The resulting homogenates were centrifuged (10,000× g, 4 °C, 15 min), and the clarified supernatants were collected. Total protein concentration in each supernatant was determined using BCA assay. Subsequently, equal amounts of total protein from each sample, alongside standards for TNFα, IL1β, and Aβ_1-42, were loaded in duplicate into their respective mouse-specific ELISA kit wells (TNFα/IL1β/Aβ_1-42 kits, Shanghai Enzyme-linked Biotechnology Co., Ltd., China) according to the manufacturers' protocols. Following incubation, washing, and color development steps, absorbance was measured at the appropriate wavelength. TNFα, IL1β, and Aβ_1-42 concentrations in the samples were interpolated from the standard curve and expressed as picograms per milligram of total protein (pg/mg protein).

Genomic DNA was extracted from stool samples using the TGuide S96 Magnetic Soil/Stool DNA Kit (Tiangen Biotech, Beijing, China) following the manufacturer's protocol. The bacterial 16S rRNA gene hypervariable V3-V4 region was amplified using the primer pair 338F (5′-CCTACGGGNGGCWGCAG-3′) and 806R (5′-GACTACHVGGGTATCTAATCC-3′). PCR products were verified on an agarose gel and purified with the Omega DNA Purification Kit (Omega Inc., Norcross, GA, USA). The purified products were sequenced on the Illumina NovaSeq 6000 platform (San Diego, CA, USA) with paired-end reads (2 × 250 bp).

Bioinformatics analysis was primarily conducted using USEARCH 8 (v8.1.1861) [23]. Briefly, primers were identified and removed using USEARCH. Sequences were filtered based on single nucleotide quality with a maximum expected error rate of 0.01. The Unoise3 algorithm was then applied to denoise the data and generate amplicon sequence variants (ASVs) [24], with ASVs containing fewer than 10 reads being discarded. Species annotation of ASVs was carried out using the Silva v123 database [25] with the SINTAX algorithm implemented by USEARCH (v8.1.1861) [26] and a confidence threshold of 60%. Alpha diversity was calculated using the R package vegan (v2.8-0) [27], while beta diversity distance matrices were computed using the R package phyloseq (v1.52.0) [28]. Principal coordinates analysis (PCoA), heatmaps, and UPGMA clustering were employed for analyzing beta diversity. Additionally, the Wilcoxon rank-sum test and Linear Discriminant Analysis Effect Size (LEfSe, v1.0.8.) [29] were utilized to detect significant taxonomic differences between groups, with a significance threshold set at p-value < 0.05 and FDR-adjusted p-value < 0.2 for feature distinction. Microbial functional prediction was performed using PICRUSt2 (v2.6.2) [30].

The metabolomics analysis was conducted using a Waters ACQUITY UPLC I-Class system (Waters Corporation, Milford, MA, USA) PLUS ultra-high-performance liquid chromatography system coupled with a Waters Xevo G2-XS QTOF high-resolution mass spectrometer (Waters Co., Milford, MA, USA), equipped with a Waters ACQUITY UPLC HSS T3 column (1.8 μm, 2.1 × 100 mm). The column temperature was maintained at 40 °C with a flow rate of 0.40 mL min⁻¹, and the injection volume was 2 μL. In both positive and negative ion modes, the mobile phases comprised 0.1% formic acid in water (phase A) and 0.1% formic acid in acetonitrile (phase B). The gradient program (time, %B) was: 0.0 min, 5%; 0.5 min, 5%; 5.5 min, 50%; 9.0 min, 95%; 10.5 min, 95%; 12.0 min, 5% (re-equilibration at initial conditions).

For LC-MS/MS analysis, the Xevo G2-XS QTOF was operated in MSE mode using MassLynx v4.2 for simultaneous collection of precursor and fragment ion data. Each acquisition cycle alternated low-energy (off) and elevated-energy scans (10–40 eV) at 0.2 s per spectrum over m/z 50–1200. ESI source parameters were: capillary voltage +2500 V/−2000 V (positive/negative), cone voltage 30 V, source temperature 100 °C, desolvation temperature 500 °C, desolvation gas 800 L h⁻¹, and cone/backflush gas 50 L h⁻¹. Samples were randomized. A pooled QC sample was injected every 10 injections, and 2-chloro-L-phenylalanine (20 mg L⁻¹) served as an internal standard to monitor retention-time and response stability.

Raw data generated by MassLynx v4.2 were processed in Progenesis QI (v2.2, Nonlinear Dynamics, Newcastle, UK) for peak picking, retention-time alignment, and deconvolution. Features failing standard QC criteria (e.g., poor peak shape or inconsistent QC response) were excluded prior to statistics. Metabolite identification relied on the METLIN database (https://metlin.scripps.edu/↗, accessed on 4 June 2025) and an in-house spectral library with theoretical fragment matching. Peak areas were normalized to the total ion current (total peak area). Principal component analysis and Spearman correlation were used to assess within-group repeatability and QC performance. Identified compounds were categorized and mapped to pathways using KEGG (www.genome.jp/kegg↗, accessed on 4 June 2025), HMDB (http://www.hmdb.ca/↗), and LIPID MAPS (www.lipidmaps.org/↗, accessed on 4 June 2025). Differential analysis combined fold-change calculations, two-sided t-tests (p < 0.05), and VIP > 1 from OPLS-DA modeling (R package ropls) with 200 permutation tests for model validation. KEGG pathway enrichment significance was evaluated using the hypergeometric test.

Metabolite information, including the CID (Compound ID) and SMILES notation, was retrieved using the PubChem (https://pubchem.ncbi.nlm.nih.gov/↗, accessed on 13 June 2025) RESTful API (Application Programming Interface) [31]. The ChEMBL ID for each metabolite was obtained via the UniChem (https://www.ebi.ac.uk/unichem/↗, accessed on 13 June 2025) API [32] using the corresponding CID. To gather evidence of chemical–target interactions, the PubChem API was employed to obtain interaction evidence based on the metabolite's CID. Subsequently, the ChEMBL [33] database was utilized to retrieve the ChEMBL ID for the metabolite's targets. The target information was further obtained from ChEMBL, filtering for targets derived from Homo sapiens (Taxonomy ID: 9606). Additionally, target predictions were made using the ChEMBL SEA (https://sea.bkslab.org/↗, accessed on 13 June 2025) [34] (Similarity Ensemble Approach) API based on the SMILES notation of the metabolite. Finally, the targets from these three sources were combined to form a comprehensive list of all potential targets for the metabolites under study.

To identify functional categories of genes, we employed the clusterProfiler package (v4.6.2) [35], which enabled us to determine Gene Ontology (GO) terms and KEGG pathways.

Principal Component Analysis (PCA) was performed using the R package factoextra (https://cloud.r-project.org/package=factoextra↗, accessed on 8 June 2025) to visualize the clustering of samples based on metabolomic and microbiomic data, with the first two principal components (PC1 and PC2) explaining the main variation in the dataset. The pheatmap package (https://cran.r-project.org/web/packages/pheatmap/index.html↗, accessed on 8 June 2025) in R was used to perform the clustering based on Euclidean distance. The Wilcoxon rank-sum test was used for pairwise comparisons of non-normally distributed continuous variables between groups, with a significance threshold of p < 0.05.

To evaluate the restorative effects of EA on learning and memory functions in AD model, this study employed classical behavioral tests, including the Morris water maze and novel object recognition test, to assess changes in learning and memory performance before and after EA treatment. The experimental design and protocol for EA in AD model are shown in Figure 1A. As expected, EA effectively alleviated cognitive impairment, including declines in learning and memory abilities as well as working memory. In addition, to explore whether EA could maintain their neuroprotective effect in the absence of gut microbiota, antibiotics were utilized before EA treatment administration to deplete the bacteria. Pretreatment with antibiotics was performed following the protocol described in the previous research [36]. Our findings revealed that EA had no significant impact on cognitive function in animals pretreated with antibiotics. This result suggests that the gut microbiota is indispensable for EA's cognitive protective effect, as antibiotic-induced gut microbiota depletion abolished EA's beneficial effects, consistent with the gut–brain axis mediating EA's neuroprotection (Figure 1B–E).

As Aβ, tau, and neuroinflammation have been previously regarded as the core pathological basis of AD [37], we subsequently examined the neuronal damage in the brains of AD model mice. Nissl staining results showed significant neuronal damage in the hippocampal and cortical regions of Model group mice, while EA intervention alleviated this damage (Figure 1F). ELISA technology was further used to detect the levels of pro-inflammatory cytokines and Aβ in the hippocampal and cortical tissues, and immunohistochemical staining was combined to observe the expression of tau protein (Figure 1G). The results showed that compared with the control group, the levels of pro-inflammatory cytokines, Aβ content, and tau protein in the hippocampal and cortical tissues of the Model group were significantly higher, while EA treatment significantly reduced the levels of the above indicators (Figure 1H).

Notably, TNF-α levels exhibited a similar variation pattern to Aβ_1-42: both were significantly elevated in the Model group compared to the Control group, and significantly reduced after EA treatment. This consistency suggests a synergistic pathological role of TNF-α and Aβ_1-42 in AD: Aβ aggregation can activate microglia to secrete TNF-α, while TNF-α further promotes Aβ deposition by inhibiting Aβ clearance [38]. In contrast, IL-1β showed an opposite trend in the cortex, which may be attributed to the dual role of IL-1β in AD: in the early stage of neuroinflammation, IL-1β can promote the activation of neuroprotective pathways, while in the late stage, it exacerbates neuronal damage [39]. EA may modulate IL-1β in a region-specific manner to balance its pro-inflammatory and neuroprotective effects. Collectively, EA reduces neuroinflammation and Aβ pathology simultaneously, breaking the "Aβ-inflammation" positive feedback loop and exerting neuroprotective effects.

The above experimental results confirm that the AD model used in this study has clear neuronal damage and core pathological changes in AD, and EA can exert neuroprotective effects by reducing neuronal damage and improving related pathological features. However, these neuroprotective effects were markedly diminished after broad-spectrum antibiotic administration, suggesting that the gut microbiota play an indispensable mediating role in EA therapy.

To investigate the impact of EA treatment on the diversity of gut microbiota in a mouse model of AD, we conducted 16S rRNA gene high-throughput sequencing analysis on fecal samples from the control group (Ctrl), AD model group (Model), and EA treatment group (EA). Through OTU clustering, a total of 703 operational taxonomic units (OTUs) shared among the three groups were detected, while unique OTUs were also identified in each group: 35 unique OTUs in the Ctrl group, 93 in Model group, and 16 in the EA group (Figure 2A). The OTU clustering results indicated certain differences in microbial communities among the three groups, with the Model group having the highest total number of OTUs and the most unique OTUs.

Taxonomic annotation revealed the dominant microbial phyla, classes, orders, families, and genera in each group. At the phylum level, the dominant phyla among the three groups were Bacteroidetes and Firmicutes. The proportion of Bacteroidetes was highest in the Ctrl group, averaging 57%, but decreased to about 55% in the Model group. In the EA group, the proportion of Bacteroidetes significantly increased to 65%. Additionally, the relative proportion of Proteobacteria showed significant differences among the three groups. Proteobacteria significantly increased in Model group (from 1% to 5%) but recovered to 0.8% in the EA group (Figure 2B, left). At the genus level, the proportion of Lactobacillus was higher in the EA group compared to Ctrl and Model groups. Furthermore, the abundance of Alistipes decreased from 9.7% in Model group to 5.9% in Model group, but increased to 8.0% in the EA group. Similarly, Odoribacter decreased from 4.7% in Ctrl group to 2.8% in Model group, but increased to 3.7% in the EA group. Other genera such as Alloprevotella and Rikenellaceae_RC9_gut_group also showed restorative effects following EA treatment (Figure 2B, right). These results suggest that EA treatment can modulate the proportions of key gut microbiota.

Alpha diversity indices are commonly used to describe species richness and evenness within a specific region [40]. We used Alpha diversity indices (including ACE, Chao1, and Shannon indices) to assess the richness and diversity of the microbiota. The results showed that the Alpha diversity of the EA group was significantly lower compared to Ctrl and Model groups (Figure 2C), indicating that EA treatment may selectively enrich beneficial microbiota by reducing the overall diversity of gut microbiota, thereby optimizing the gut microecology.

Additionally, beta diversity analysis, through principal coordinate analysis (PCoA), demonstrated significant differences in microbiota composition among the three groups (Figure 2D). The first two principal coordinates, PC1 and PC2, explained 23.41% and 14.34% of the community variation, respectively, accounting for a total of 37.75% of the variation. The distribution of samples from different groups in the PCoA space reflected significant differences in their microbial community structures. The samples within each group clustered closely in the PCoA plot, indicating good consistency in gut microbiota structure within groups. The Model group (yellow) samples were distinctly separated from the control group along the PC1 axis, suggesting that the AD state led to significant changes in gut microbiota structure, indicating further exacerbation of community structure diversity. After EA treatment (orange), the samples partially returned to the distribution range of the Ctrl group along the PC1 axis, although they still differed from the Ctrl group, they were significantly closer compared to Model group, indicating that EA treatment partially restored the gut microbiota structure. The beta diversity analysis results suggest that EA treatment can partially reverse the microbial community dysbiosis induced by the AD state by influencing microbiota composition.

Subsequently, we further identified microorganisms with significantly different abundances among the three groups through linear discriminant analysis effect size (LEfSe) analysis (Figure 2E,F). The results showed that Model group (green) was significantly enriched with Proteobacteria and Enterobacteriaceae, suggesting that these microbiota are closely related to a pro-inflammatory state. The Ctrl group (red) was enriched with Lachnospiraceae within the Firmicutes phylum, which are typically associated with a healthy state and anti-inflammatory effects. The EA group (blue) was significantly enriched with the order Gastranaerophilales and the class Melainabacteria, indicating that EA treatment can reduce pro-inflammatory microbiota in the AD model.

In summary, the above results demonstrate that EA treatment can significantly modulate the diversity of gut microbiota in a mouse model of AD, particularly by reducing microbial species richness and optimizing microbial community structure, providing support for EA treatment of AD at the gut microbiota level. Further research will delve into the specific changes in microbiota and their potential functional roles.

To more accurately evaluate the restorative effects of EA treatment on gut microbiota dysbiosis in AD models, we first filtered bacterial phyla and genera with a relative abundance greater than 0.001% to ensure the biological significance of the analysis. Subsequently, the Wilcoxon rank-sum test was used to statistically analyze differences in microbial abundance between Ctrl group and Model group (p < 0.05). We further defined the "restoration trend" as a median abundance pattern of "low–high–low" or "high–low–high" changes, identifying microbiota that showed significant alterations in Model group and exhibited restoration trends after EA treatment.

At the phylum level, we found that the relative abundances of Verrucomicrobia, Acidobacteria, Cyanobacteria, Actinobacteria, and Proteobacteria were significantly increased in Model group, while their abundances markedly decreased after EA treatment, approaching levels observed in the Ctrl group (Figure 3A). These changes suggest that EA treatment may restore gut microbial ecological balance by modulating the abundance of key phyla.

At the genus level, genera with significant restoration after EA treatment included Sphingomonas, Massilia, Ruminococcaceae_UCG-013, Escherichia-Shigella, and Erysipelatoclostridium (Figure 3B). Among these, Ruminococcaceae_UCG-013 was significantly reduced in the Model group but rebounded after EA treatment. Other genera showed abnormal enrichment in the Model group, followed by a significant decline post-EA treatment, aligning with the Ctrl group levels. These results further support that EA therapy selectively regulates the abundance of health-associated genera, alleviating microbial dysbiosis.

To investigate the functional impact of EA treatment on the microbiota, we compared differences in functional pathways across the three groups using principal component analysis (PCA). The results revealed clear separation between the Model group and the Ctrl group along the PC1 (26.3%) and PC2 (16.9%) dimensions, while the EA group partially clustered closer to the control group (Figure 3C). This indicates that EA treatment may partially restore functional pathway states associated with normal gut ecology.

The antibiotic experiment in our study clearly elucidates the dual mechanisms underlying EA therapy: it eliminates or reduces harmful microorganisms and their toxic metabolites, while concurrently promoting the enrichment of beneficial microorganisms and their associated metabolites. This latter process is precisely the critical pathway that antibiotics fail to replicate and may even disrupt. These findings fully demonstrate that EA exerts its effects through the simultaneous action of two indispensable dimensions: "inhibiting pathogenic microorganisms" and "enriching probiotic microorganisms," with neither being dispensable.

Additionally, to explore the regulatory effects of EA on functional pathways, we conducted KEGG pathway analysis and selected pathways with median abundance changes displaying "low–high–low" or "high–low–high" patterns across the three groups (Figure 3D). The analysis showed that pathways related to neurodegenerative diseases, infectious diseases, and cancer were significantly enriched in the Model group, whereas their abundances decreased notably in the EA group, approaching the Ctrl group levels. These findings suggest that EA treatment may ameliorate pathological gut microecological states by suppressing disease-associated functional pathways.

Microbiota-derived metabolites influence host physiology and behavior through multiple pathways, and their dysregulation may play a critical role in the pathogenesis of neurological disorders [41]. Our study investigated how EA treatment modulates gut microbiota and their metabolites to affect host metabolic function using metabolomic analysis. Principal component analysis (PCA) revealed significant separation between the Model group and the Ctrl group along the PC1 (46.3%) and PC2 (8.6%) dimensions, indicating pronounced metabolic disturbances in the AD model. The EA group clustered closer to the Model group rather than the Ctrl group (Figure 4A), suggesting limited regulatory effects of EA on host metabolism, possibly through partial modulation of specific metabolites. A comparison of differential metabolites across groups showed a larger number of differences between the Model group vs. Ctrl group and EA group vs. Ctr group, whereas fewer differences were observed between the EA group and Model group (Figure 4B), further supporting that EA may act via selective regulation of key metabolites. The heatmap (Figure 4C) demonstrated that most metabolites were significantly up-regulated in the Model group compared to the Ctrl group, with minor adjustments in metabolite abundance after EA treatment.

We further visualized differential metabolite changes between groups (Model vs. Ctrl, EA vs. Ctrl, and EA vs. Model) based on fold change magnitude (Figure 4D). Compared to the Ctrl group, metabolites such as PE(22:2(13Z,16Z)), Deoxycholylariginine, and PS(PGJ2/18:4(6Z,9Z,12Z,15Z)) were significantly up-regulated in the Model group, while down-regulated metabolites like Melatonin and Coproporphyrin I reflected AD-associated metabolic dysregulation [42]. The EA group exhibited a metabolite profile similar to the Model group when compared to the Ctrl group, indicating incomplete restoration of pathological metabolites to normal levels. However, metabolites such as 25-Hydroxyvitamin D3-26,23-lactol, N-Acetylneuraminate 9-phosphate, and carvone oxide were significantly up-regulated in the EA group compared to the Model group, despite being among the most down-regulated metabolites in the Model vs. Ctrl comparison (Top 10). This suggests that EA may partially restore metabolic function by targeting critical metabolites.

To explore the functional significance of differential metabolites, we performed pathway enrichment analysis using the KEGG database (Figure 5A). Compared to the Ctrl group, the Model group showed significant downregulation of pathways such as phenylpropanoid biosynthesis and purine metabolism, which are closely linked to neuronal function and cerebral metabolism. The downregulation of phenylpropanoid biosynthesis (associated with antioxidant activity) may exacerbate oxidative stress, while impaired purine metabolism (critical for energy production and nucleic acid synthesis) could lead to energy deficits and neuronal dysfunction [43]. EA treatment partially restored these pathways, indicating its potential to alleviate AD-related metabolic disorders by recovering key metabolic functions.

The Venn diagram further illustrates the intersections of differential metabolites. The overlap between up-regulated metabolites in the Model group and down-regulated metabolites in the EA group may represent potential pathological metabolites suppressed by EA treatment. Conversely, the intersection between down-regulated metabolites in the Model group and up-regulated metabolites in the EA group likely signifies key health-associated metabolites restored by EA (Figure 5B). These intersections provide clues to the metabolic regulatory effects of EA. However, in the detailed metabolite analysis, the heatmap reveals that EA has limited restorative capacity for most metabolites. Notably, metabolites with more pronounced recovery include Microcystin LR, PC(DiMe(11,3)/MonoMe(11,5)), Lucidone B, reduced riboflavin, 4-Coumaroylshikimate, 5-alpha-Cholesta-7,24-dien-3beta-ol, and Juzirine (Figure 5C), suggesting that these metabolites may serve as potential therapeutic targets or functional biomarkers of EA.

To further explore the dependencies between gut microbiota and metabolites, we performed Spearman correlation analysis on EA-restored microbes and metabolites, selecting significant pairs with absolute correlation coefficients >0.7 and (p < 0.01) (Figure 5D). The results showed that metabolites such as Lucidone B, reduced riboflavin, 4-Coumaroylshikimate, and 5alpha-Cholesta-7,24-dien-3beta-ol were significantly positively correlated with three bacterial genera (Sphingomonas, Massilia, and Escherichia-Shigella), while Microcystin LR, PC(DiMe(11,3)/MonoMe(11,5)), and Juzirine exhibited significant negative correlations. These metabolites may accumulate due to abnormal microbial activity in the pathological state of AD, leading to metabolic imbalances in the nervous system. For example, although Microcystin LR possesses antioxidant properties, its excessive accumulation may increase oxidative stress [44]. EA treatment, by reducing the abundance of these three bacterial genera, may regulate its levels appropriately, thereby mitigating neuronal damage. Similarly, the restoration of metabolic balance in Lucidone B and reduced riboflavin may improve mitochondrial function and energy metabolism, promoting neuroprotection. Additionally, 5alpha-Cholesta-7,24-dien- 3beta-ol, a key intermediate in cholesterol metabolism, plays a crucial role in maintaining neuronal membrane stability.