L. rhamnosus HM126 was isolated from healthy human breast milk and is currently deposited at the General Microorganism Center of the China General Microorganism Culture Collection Management Committee (CGMCC No. 26167). First, the bacterial strain was cultured at 37 °C for 24 h, followed by two passages to ensure viability. The bacteria were then harvested by centrifugation (6000 rpm, 5 min, 4 °C). After washing to remove residual medium, the pellet was resuspended to the target concentration for subsequent oral administration.

All animal experiments were conducted in accordance with the Beijing Union University Guidelines for the Care and Use of Laboratory Animals and were approved by the Beijing Union University Animal Ethics Committee (JGX11-2306-4; Beijing, China). The experimental workflow is shown in Figure 1A. One hundred male BALB/c specific pathogen free (SPF) mice weighing 20–22 g were purchased from Beijing Vital River Laboratory Animal Technology Co., Ltd. (License No.: SCXK(Jing)2021-0006). After 7 days of acclimation feeding, mice were randomly divided into five groups: Control, AD model (DNFB), low-dose (L-Larh, 0.5 × 10⁸ CFU/mouse), medium-dose (M-Larh, 1 × 10⁸ CFU/mouse), and high-dose (H-Larh, 3 × 10⁸ CFU/mouse) (n = 20/group).

The Control group received sterile water for 2 weeks, whereas the DNFB group received an equivalent volume of the test sample. Modeling commenced on day 15. One day prior to modeling, hair was removed from a 2 × 2 cm² area on the back of the mouse using 6% sodium sulfide. On day 15, except for the Control group, all groups received 100 μL of 0.5% DNFB (acetone:olive oil = 3:1) applied to the depilated area for sensitization. Re-sensitization occurred on day 17. On days 21 and 28, 50 μL of 0.5% DNFB solution was applied topically to the dorsal skin to induce dermatitis. The Control group received an equal volume of the acetone-olive oil vehicle. Except for the Control and DNFB groups (which received sterile water), all dose groups received 0.1 mL/mouse of the respective test substance via continuous gavage for 28 days.

On day 28, the mice were sensitized with DNFB for 30 min. Following sensitization, the mice were placed individually and observed for 10 min to record the number of times they scratched their trunk and back with their hind paws or bit any part of their body. Consecutive scratching episodes were counted as single occurrences.

On day 29, semi-quantitative scoring of dorsal skin lesions was performed based on the following clinical criteria: erythema, edema/papules, epidermal exfoliation/scratch marks, scaling. Lesions were graded as absent, mild, moderate, or severe (0, 1, 2, or 3 points, respectively), with a maximum total score of 12 points (25).

On day 29, the mice in each group were euthanized via cervical dislocation. The spleen tissue was dissected, and the residual blood was blotted with filter paper and immediately weighed.

The skin samples were excised from the dorsal region. Portions of the skin tissue were fixed in 4% formaldehyde solution, followed by dehydration with graded ethanol, clearing with xylene, paraffin embedding, sectioning at 5 μm thickness, and examination after hematoxylin and eosin (H&E) staining.

On day 29, blood was collected from the epicardial venous plexus. Serum was obtained via centrifugation at 3000 rpm for 10 min and analyzed. Because cytokine assays require large volumes of serum to ensure accuracy, a random pairing method was employed. Serum from each pair of mice was mixed in equal volumes for cytokine detection and each group of 20 mice yielded 10 pooled serum samples (n=10/group).

Fresh fecal samples were collected and stored at -80 °C (Control, DNFB, and H-Larh groups; n=10/group). Sequencing and data analysis were performed by Shanghai Applied Protein Technology Co., Ltd. Genomic DNA was extracted using the Omega Soil DNA Kit. The V3-V4 region of the 16S rDNA gene was amplified via PCR and library preparation, followed by sequencing on the Illumina platform. Data were processed using QIIME 2 and DADA2 workflows to obtain ASV and species annotations (based on the Silva database). Subsequently, α and β diversity analyses were conducted, and LEfSe was employed to identify differentially abundant species between groups (26).

Collected feces were rapidly frozen in liquid nitrogen and stored at -80 °C (Control, DNFB, and H-Larh groups; n=10/group). Non-targeted metabolomics analysis was performed by Shanghai Applied Protein Technology Co., Ltd. (Shanghai, China) In brief, samples were extracted with cold methanol/acetonitrile/water solution (2:2:1, v/v), subjected to low-temperature ultrasonication and centrifugation, followed by vacuum drying of the supernatant. After redissolution and centrifugation, the supernatant was transferred to injection vials for analysis. After separation by a Vanquish LC ultra-high-performance liquid chromatography (UHPLC) system, mass spectrometry analysis was performed using a Q Exactive series mass spectrometer (Thermo Fisher Scientific, Waltham, MA, USA). Column temperature was set at 25 °C, with acetonitrile-water gradient elution. Mass spectrometry scanned in both positive and negative ion modes, covering a range of 80–1200 Da. Raw data were processed using ProteoWizard and XCMS software to complete peak extraction and alignment. Subsequent steps included metabolite identification, data preprocessing, and quality assessment. Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) was employed to analyze metabolic changes. Differentially expressed metabolites were screened using VIP > 1 and p < 0.05, followed by pathway enrichment analysis via MetaboAnalyst (https://www.metaboanalyst.ca/).

Statistical analysis was performed using GraphPad Prism 10.1.2 software (San Diego, CA, USA). One-way analysis of variance was conducted, followed by Dunnett’s multiple comparison test to analyze the differences in body weight, spleen weight, scratching frequency, skin lesion score, inflammatory factors, and mast cells. The data are presented as mean ± standard deviation (SD). Spearman correlation analysis between gut microbiota and differential metabolites. The significance level was set at p < 0.05.

To investigate the effects of the model establishment and HM126 intervention on body weight, we selected male mice with balanced body weights prior to experimentation. Post-experiment, significant differences in body weight were observed between the DNFB group and the Control group (p < 0.001) (Figure 1B), indicating that DNFB-induced dermatitis affects mouse body weight. Compared to the DNFB group, the L-Larh group exhibited significantly reduced body weight (p = 0.027).

The spleen is the most important immune organ in the body. Compared with the control group, the spleen weight significantly increased in the DNFB group after the experiment (p < 0.001). Compared to the DNFB group, no significant differences in spleen weight were observed among all dose groups after the experiment (p > 0.05) (Figure 1C). Reduced scratching behavior indicates the alleviation of allergic pruritus. Compared to the DNFB group, all dose groups showed a trend toward reduced scratching frequency during modeling, with the H-Larh group exhibiting the lowest frequency and significant differences (p = 0.039) (Figure 1D). Compared to the DNFB group, none of the dose groups showed significant differences in skin lesion scores after treatment (p > 0.05) (Figure 1E).

Dorsal skin tissue was collected for H&E staining to evaluate symptoms in mice with DNFB-induced AD. As shown in Figure 1F, the dorsal skin from the Control group exhibited intact structures in the epidermis, appendages, dermis, and subcutaneous tissue, with no evidence of damaging alterations or inflammatory cell infiltration. In contrast, the DNFB group exhibited typical allergic reactions, including epidermal erosion of the dorsal skin that progressed to ulceration, partial or complete loss of the epidermal cell layer, and basement membrane disruption. Extensive infiltration of inflammatory cells (including polymorphonuclear leukocytes) resulted in the formation of pustules and crusts. Individual specimens exhibited epidermal necrosis (full-thickness type) separated from the dermis, revealing a strongly eosinophilic cytoplasm within the necrotic foci. Macrophage dermal infiltration and epidermal hyperplasia are the hallmark features of animal models of AD. This demonstrates the validity of the AD mouse model employed in this experiment.

The imbalance in type 1/type 2 cell subsets, which leads to a series of cytokine alterations, constitutes the primary immunological mechanism in AD pathogenesis (27, 28). Type 1 cells, characterized by IFN-γ secretion, drive cellular immunity, cytotoxic responses, and delayed-type hypersensitivity. In contrast, type 2 cells, defined by IL-4 production, promote humoral immunity and contribute to (atopic/allergic) hypersensitivity. In AD, type 2-driven IL-4 expression activates B cells to produce excess IgE antibodies. External antigens and allergens absorbed through the skin bind to IgE molecules on mast cell surfaces, triggering mast cell degranulation and release of histamine and other mediators. This process induces pruritus and perpetuates type 2 cell dysregulation. Furthermore, Evidence suggests that TGF-β ameliorates AD by suppressing TNF-α and IgE secretion.

Figures 2A–E illustrates the effects of L. rhamnosus HM126 treatment on cytokine expression in the serum of AD mice. Compared with the Control group, serum cytokine levels in DNFB group exhibited altered patterns, with extremely significant increases in TGF-β (p < 0.001), IFN-γ (p = 0.003), IgE (p = 0.002), and IL-4 (p < 0.001). Compared with the DNFB group, the IgE levels in the L-Larh group decreased significantly (p = 0.032).All three HM126 dose groups exhibited potent, non-dose-dependent downregulation of the type 2-associated cytokine IL-4 (p < 0.001). L. rhamnosus HM126 did not affect type 1-related cytokine IFN-γ levels (p > 0.05), and TGF-β levels in all dose groups showed no significant difference compared to those in the DNFB mice (p > 0.05). The IFN-γ/IL-4 ratio in serum significantly increased across all dose groups (p < 0.001), suggesting that L. rhamnosus HM126 intervention improved the type 1/type 2 imbalance in AD mice. Compared to the Control group, the DNFB group showed significantly higher number of skin mast cells (p < 0.001). Compared to DNFB group, the number of skin mast cells in all dose groups was significantly reduced (p = 0.016, p = 0.019, p = 0.032) (Figure 2F).

To evaluate the impact of L. rhamnosus HM126 intervention on gut microbiota diversity and abundance in DNFB mice, we performed 16S rDNA gene sequencing of colonic contents. Sequencing depth assessment results indicate that the rarefaction curves for all samples have flattened (Supplementary Figure S1), demonstrating that the current sequencing depth has captured the vast majority of microbial species present in the samples. Genus-level analysis revealed a similar composition of dominant bacterial groups across all samples, including Muribaculaceae, Prevotellaceae UCG-001, Alistipes, Lachnospiraceae NK4A136 group, Clostridia UCG-014, Lactobacillus, Alloprevotella, Bacteroides, Muribaculum, and Roseburia. Compared to the DNFB group, L. rhamnosus HM126 intervention increased the abundance of Prevotellaceae_UCG-001 and Lactobacillus and decreased the abundance of Alistipes, Alloprevotella, Muribaculum, and Roseburia (Figure 3A). Collectively, these findings provide evidence that L. rhamnosus HM126 exerts a regulatory effect on the intestinal microbiota of mice.

Changes in α-diversity within gut microbiota were characterized using the Shannon index, ACE index, and Chao1 index (Figures 3C–E). Notably, the ACE (p = 0.005) and Chao1 (p = 0.004) indices were substantially lower in the L. rhamnosus HM126 group than in the DNFB group. No notable differences in the Shannon index were observed among groups. A significant disparity in microbial community structure among the Control, DNFB, and Larh groups was observed, as determined by Bray-Curtis beta-diversity (PERMANOVA, p = 0.002), (Figure 3B),indicating that L. rhamnosus HM126 intervention substantially influenced gut microbiota composition. In terms of distribution trends, the DNFB group showed a more pronounced separation from the control group, suggesting that the AD model may exacerbate disease progression by altering gut microbiota homeostasis. The sample distribution of the Larh group fell between the two extremes, partially aligning with the control group, implying that probiotics may partially reverse AD-associated dysbiosis by modulating the microbial structure.

LEfSe analysis revealed that the DNFB group showed a marked enrichment of the harmful bacterium Desulfovibrio. (Figures 3F, G). Desulfovibrio, Desulfovibrionaceae, and Desulfovibrionales, all related to the Desulfovibrionaceae family, can produce hydrogen sulfide through sulfate-reducing metabolism, disrupt intestinal barrier integrity, and lead to the activation of pro-inflammatory signaling pathways such as NF-κB, which correlates with the type 2 immune shift in AD. In contrast, L. rhamnosus HM126 intervention significantly enriched beneficial bacteria such as Faecalibacterium and Lactobacillus and partially reversed dysbiosis. In summary, L. rhamnosus HM126 may offer a novel therapeutic strategy for AD by reshaping the gut microbiota, suppressing opportunistic pathogens, and enriching beneficial metabolic bacteria.

To further investigate the effects of L. rhamnosus HM126 on fecal metabolites in mice with AD, non-targeted metabolomic analysis was performed on fecal samples from the Control, DNFB, and H-Larh groups. PCA analysis indicates that QC samples cluster closely together, demonstrating good reproducibility of the experiment (Supplementary Figure S2). We constructed an OPLS-DA model for fecal samples to characterize intergroup metabolic changes (Figures 4A, B). Differentially expressed metabolites were screened using the criteria of VIP> 1 and p < 0.05. Compared with the control group, the DNFB group showed 325 differential metabolites, including 122 upregulated and 203 downregulated ones. Additionally, 134 differential metabolites were identified between the DNFB group and the Larh group, with 53 upregulated and 81 downregulated in the Larh group relative to the DNFB group. Integrative analysis of metabolomic data from the model and probiotic intervention groups revealed that these differential metabolites were mainly categorized into lipids, organic acids, organic heterocyclic compounds, and aromatic compounds. Notably, the number of downregulated metabolites was consistently higher than that of upregulated ones across the comparative groups (Figures 4C, D).

Further screening of metabolites present in both the DNFB vs Control and Larh vs DNFB comparisons identified 33 differential metabolites that met these criteria, potentially representing the core targets of L. rhamnosus HM126 intervention (Supplementary Table S1). Following L. rhamnosus HM126 treatment, significant metabolic rebalancing was observed for certain metabolites. Specifically, asiatic acid, phytosphingosine, Ser-Glu, PGF(2α)EA levels were significantly reduced and significantly increased in the DNFB and Larh groups, respectively. Argininosuccinic acid, L-rhamnose, and gamma-L-glutamyl-L-glutamic acid were significantly upregulated in the model group and significantly downregulated after L. rhamnosus HM126 treatment. (Figures 4E–K) These substances may be associated with the improvement of L. rhamnosus HM126 in mice.

Based on the differentially identified metabolites in fecal samples, metabolic pathway analysis using MetPA revealed 25 pathways potentially related to AD metabolic dysregulation. According to the significance criteria p < 0.05 and impact > 0.1, the following pathways were most likely associated with the mechanism by which L. rhamnosus HM126 alleviated AD. Three pathways—lysine degradation, galactose metabolism, and arginine biosynthesis—were disrupted by L. rhamnosus HM126 to improve endogenous metabolism in AD (Figure 4L).

To further elucidate the associations between gut microbiota and metabolites, Spearman correlation analyses were conducted between differentially expressed metabolites and gut microbiota (Figure 5). The key gut bacterial genus Prevotellaceae UCG-001 positively correlated with phytosphingosine and negatively correlated with N-acetyl-L-tyrosine and trans-zeatin. Alistipes was positively correlated with N-acetyl-L-tyrosine, argininosuccinic acid, trans-zeatin, and 2-aminoadipic acid. A positive correlation of Lactobacillius with asiatic acid and negative correlations with L-rhamnose, D-glucosaminic acid, Glu-His, and other compounds were observed. Additionally, Roseburia was positively correlated with DL-2-aminocaprylic acid, Materia resinol, Daidzein 4’-sulfate, and Glu-His.