Trimethylamine-producing microbe Bacillus megaterium KCTC 3007 promotes antitumor immunity in endometrial cancer via type I interferon response pathways.

In this retrospective cohort study, we identified patients from the institution's ECa Cohort who met the following inclusion criteria: (1) aged ≥ 20 years at initial diagnosis of ECa; (2) received primary surgical treatment at Seoul National University Hospital (SNUH) between January 2014 and December 2018; (3) whose formalin-fixed paraffin-embedded (FFPE) blocks were stored at Department of Pathology; and (4) donated their blood samples, obtained on day before surgery, and fresh frozen primary ECa tissues, obtained at the time of surgery, for scientific purposes under the written informed consent. In addition, patients were excluded if they (1) received hormone therapy, chemotherapy or radiation prior to surgical treatment; (2) lost to follow-up during primary treatment; or (3) had insufficient clinical data. For the control group, we identified patients with benign uterine diseases who met the following inclusion criteria: (1) aged ≥ 20 years; (2) received primary surgical treatment at SNUH between January 2014 and December 2018; (3) donated their blood samples, obtained on day before surgery, and fresh frozen endometrial tissues, obtained at the time of surgery, for scientific purposes under the written informed consent. However, patients were excluded if they had any malignancies or had insufficient clinical data. We reviewed the medical records of the patients and collected clinicopathologic, treatment-related, and survival data. For patients with ECa, risk stratification and molecular classification were conducted according to the ESGO/ESTRO/ESP risk stratification and ProMisE classification, respectively. The detailed methods on ProMisE classification are described in our previous study (PMID: 37,236,032). Patients underwent surveillance with physical examination and serum CA-125 levels every three to four months for the first two years, every six months for the next two years, and annually thereafter. Imaging studies were conducted as per physicians' preference or when symptoms or examination findings were suspicious of recurrence. RFS and OS were defined as the time intervals from the date of surgery to the date of disease progression and to the date of cancer-related death or last visit, respectively. For the subsequent whole transcriptomic sequencing and metabolomics study, the fresh-frozen, primary ECa tissues (for ECa patients) or endometrial tissues (for patients with benign uterine diseases) and matched blood samples were retrieved from Seoul National University Hospital Human Biobank. RNA extraction was performed using TRIzol reagent (#15,596,026, Invitrogen), adhering to the manufacturer's protocol. Quality control of the RNA was assessed based on the A260/A280 and A260/A230 ratios, with samples having an RNA Integrity Number (RIN) greater than 7 proceeding to library preparation. This was conducted using the TruSeq RNA Access Library Prep Kit. Sequencing was performed on an Illumina HiSeq2500 (Illumina, San Diego, USA) at Macrogen (Seoul, Korea). Detailed methodology is described in the Supplementary Data. Metabolite concentrations were assessed in sera samples using the MxP® Quant 500 kit (Biocrates Life Sciences AG, Innsbruck, Austria) with an ACQUITY UPLC system (Waters Corporation, Milford, USA) coupled with Triple Quad™ (5500 +, AB Sciex). The analysis encompassed the determination of small molecules via UPLC separation and lipid class metabolites through flow injection analysis (FIA). The targeted metabolomic analysis employing the Biocrates kit adhered to the provided protocol. Sample preparation involved adding 10 μL of the plasma sample to a 96-well-based device containing an internal standard (IS), phenyl isothiocyanate (PITC) solution for specific metabolite derivatization. Following derivatization, the target analytes underwent extraction with an organic solvent, succeeded by a dilution step. Identification of the obtained extracts relied on isotopically labeled standards and multiple reaction monitoring (MRM) utilizing optimized MS conditions as provided by Biocrates. To cover all metabolites, each measurement necessitated two UPLC runs and two FIA runs. Analyte quantification was based on either a seven-point calibration curve or one-point calibration, with normalization via IS, computed as concentrations in μM (μmol/L). For data quality control, all samples underwent pooling for each sample type, positioned at the start, middle, and end of each kit, and analyzed in triplicate. Detected data underwent generation and processing through Analyst® (version 1.7.1, AB Sciex, Darmstadt, Germany) and MetIDQ™ software (version Oxygen-DB110-3005, Biocrates Life Sciences AG). We regrouped three clinical parameters; age, stage, and risk. Considering the age of menopause and hormone level change, which are important in ECa characteristics, we divided the cohort into those aged over (Group ‘aged’) and below (Group ‘young’) 50 years and conducted further analyses. In addition, due to the small numbers included in each stage of subgroups, we created a modified staging system; early stage includes stage 1 and 2, and advanced stage includes stage 3 and 4. Lastly, given the limited sample size of our cohort, we employed a modified risk stratification by combining high and high-intermediate groups into a 'modified high-risk' category, and low and low-intermediate groups into a 'modified low-risk' category. This modification was necessary because the original four-tier ESGO/ESTRO/ESP classification yielded unbalanced group sizes (high = 32, high-intermediate = 13, low-intermediate = 4, low = 11), rendering the intermediate-risk groups statistically insufficient for robust clinical and metatranscriptomic analyses. Sequencing data were subjected to filtration and quality control utilizing the fastp version 0.23.2 [ 117 ]. Subsequently, the cleaned data were aligned to the human genome database GRCh38 via HISAT2 version 2.2.1 [ 118 ], and data not mapped were isolated using Samtools version 1.19 [ 119 ]. These unmapped sequences were then aligned to the microbiome database employing Kraken2 version 2.1.3 [ 120 ], specifically utilizing the minikraken database (version released in March 2020), to derive microbiome compositional data, or metatranscriptomics data. The metatranscriptomics data obtained underwent analysis employing a suite of tools within R version 4.3.2 (Vienna, Austria). This analysis integrated read count data, Operational Taxonomic Unit (OTU) information, and metadata to construct a comprehensive phyloseq object via the phyloseq [ 121 ] package. To facilitate the exclusion of contaminants, computational decontamination was conducted using the decontam [ 122 ] package, targeting bacterial taxa within tissue samples and excluding any identified as contaminants. Both alpha and beta diversities were computed and visually represented using the ‘phyloseq’ [ 123 ] and ‘microViz’ [ 124 ] package in R version 4.2.3 (Vienna, Austria). Computational decontamination process using DNA quantification data was performed using R package ‘decontam’ [ 125 ]. R package ‘pheatmap’ [ 126 ] and Graphpad Prism version 10.0.1 (Boston, USA) were used to visualize differences in microbial composition between groups. The R package ‘survival’ was employed for survival analysis [ 127 ]. For comparing the abundance of specific bacterial taxa across groups, MaAsLin2 was employed to carry out differential abundance testing [ 128 ]. Baseline characteristics such as age, BMI, and underlying disease, which showed significant differences between the groups, were included as covariates in the adjustment process. As described in the above section, sequencing data were subjected to filtration and quality control utilizing the fastp tool [ 117 ]. Subsequently, the cleaned data were aligned to the human genome database GRCh38 via HISAT2 [ 118 ]. Mapped data was processed by featureCounts (Subread version 2.0.6) [ 129 ]. Using DESEq2 package in R software [ 130 ], we idenfied differentially expressed genes and normalized the read counts. Normalized data was used for Gene Set Enrichment Analysis (GSEA) version 4.3.2 (Cambridge, USA) with the MsigDB hallmark gene set [ 131 ]. And Tumor Immune Estimation Resource (TIMER) version 2.0 (Boston, USA) was used to perform in silico immune cell profiling [ 132 ]. R package ‘pheatmap’ [ 126 ] and Graphpad Prism were used to visualize the result of survival analysis, GSEA, and TIMER. Targeted metabolomics data were processed using MetaboAnalyst version 6.0 (Boston, USA) [ 133 ]. The statistical analysis was conducted in 'one factor' mode to identify differences between BM-enriched and depleted groups. Data filtering included a reliability threshold of 25%, variance based on the interquartile range, and abundance based on the mean intensity value. Data transformation was performed using a cube root transformation, and scaling was conducted with range scaling. For meaningful metabolite selection, a volcano plot, based on p -value and fold change, Principal Component Analysis (PCA) [ 134 ], Partial Least Partial Least Squares Discriminant Analysis (PLS-DA) [ 135 ], and Orthogonal Partial Least Squares – Discriminant Analysis (OPLS-DA) [ 136 ], were employed. We utilized an open RNA sequencing dataset from Mariathasan et al., 2018 [ 42 ]. Genentech provided authorization for our access to the dataset with the accession number EGAD00001003977 at the European Genome-Phenome Archive. We analyzed data from 298 patients who have immunotherapy response information. We processed the raw FASTQ files using the method previously described and successfully extracted data on both the tissue microbiome and host gene expression. To elucidate the mechanism of action and the impact of specific microbes, we conducted survival analysis, GSEA [ 131 ], and TIMER [ 132 ] analyses. Tissue microbiome data (Metatranscriptomics) are available in Supporting Data 2. Peripheral blood was collected and immediately diluted with an equal volume of phosphate-buffered saline (PBS). This diluted blood sample was then carefully layered over a pre-added Ficoll medium (Ficoll-Paque™ PLUS #17,144,003, Cytiva) in a centrifuge tube, maintaining a blood to Ficoll ratio of 35:15. The tube was centrifuged at 400 g for 30 min with acceleration and deceleration settings adjusted to zero to ensure a smooth gradient formation without disturbing the layers. After centrifugation, distinct layers were observed: plasma, PBMCs, Ficoll, and RBCs from top to bottom. The PBMC layer was carefully aspirated and transferred into a 50 ml conical tube. The collected PBMCs were then washed with PBS, and resuspended with 1 × RBC lysis buffer (#420,301, BioLegend) and incubated for 10 min at room temperature. The samples were filtered through a 70 μm cell strainer and the number of cells was counted. Purified PBMCs were then allocated to 6-well plates and co-cultured for 3 or 4 days at a 5% concentration with MTM media, nutrient agar (NA) media, and the supernatants from KCTC 2178 and KCTC 3007. MTM was used as a control because it supports PBMC proliferation, and NA was used as it is the culture medium for the selected microbes. The supernatants from KCTC 2178 and KCTC 3007 were collected after culturing the bacteria in NA broth media until reaching an OD600 of 1. We received bacterial strains from the Korean Collection for Type Cultures (KCTC), which were initially cultured in NA broth under aerobic conditions and subsequently subcultured to ensure vitality. The cultures were preserved by creating a glycerol stock solution with a 20% concentration for storage. This stock was thawed and reintroduced into the broth for two additional rounds of subculturing. After this, 1 mL of the culture was collected and centrifuged at 5000 rpm for 15 min. The supernatant above the pellet was carefully collected. This supernatant was then passed through a 0.2 µm filter (Sartorius, S6434-FMOSK*0.2) to ensure purity before being stored. We incubated 100 µL of HEC-1B cells (1 × 10 4 cells) per well in a 96-well plate at 37 °C in a 5% CO2 incubator for 24 h. After incubation, PBMCs that had been cultured for 3 or 4 days were added to each well at a concentration of 2 × 10 5 cells. Following 24 h of co-culture, we treated the cells with 10 µL of Quanti-Max (#QM1000, BIOMAX) for 30 min. The optical density was then measured at 450 nm (OD450). Each experiment was performed in three technical replicates. For experiments with TMAO, HEC-1B and AN3CA cells (1 × 10 4 cells/well) were seeded in 96-well plates with 100 µL of medium and incubated at 37 °C in a 5% CO₂ incubator for 24 h. After incubation, peripheral blood mononuclear cells (PBMCs; 1 × 10 5 cells/well) that had been pre-cultured for 4 days were added to each well. The co-culture was maintained for 4 days in MTM medium containing 300 µM or 600 µM TMAO (# T72761 , Sigma-Aldrich). Control wells were co-cultured in MTM medium without TMAO under the same conditions. After treatment, 10 µL of Quanti-Max reagent (#QM1000, BIOMAX) was added to each well and incubated for 30 min. The optical density was measured at 450 nm (OD450). Each experiment was performed in technical replicates. For experiments with TMA, we incubated 100 µL of HEC-1B and AN3CA cells (1 × 10 4 cells) per well in a 96-well plate at 37 °C in a 5% CO₂ incubator for 24 h. After incubation, the cells were then treated with TMA (#317,594, Sigma-aldrich) in MTM medium at final concentrations of 300 µM and 600 µM for 3 days. After treatment, 10 µL of Quanti-Max (#QM1000, BIOMAX) was added to each well and incubated for 30 min. The optical density was measured at 450 nm (OD450). Each experiment was performed in three technical replicates. Total RNA from incubated PBMCs was extracted using TRIzol reagent (#15,596,026, Invitrogen) following the manufacturer’s protocol. Subsequently, cDNA synthesis was performed using the TOPscript™ RT DryMIX kit (#RT200, Enzynomics), followed by qPCR using the TOPreal™ qPCR 2X PreMIX kit (#RT501M, Enzynomics) and a StepOnePlus (#4,376,600, Applied Biosystems). Gene expression levels were normalized to MTM group, and relative expression levels were compared across other experimental groups. Primer sequences are provided in Table S2. For experiments with TMAO, total RNA was extracted from PBMCs incubated in MTM medium containing TMAO for 4 days. using TRIzol reagent (#15,596,026, Invitrogen) according to the manufacturer’s protocol. Subsequently, cDNA synthesis was performed using the TOPscript™ RT DryMIX kit (#RT200, Enzynomics), followed by qPCR using the TOPreal™ qPCR 2X PreMIX kit (#RT501M, Enzynomics) and a StepOnePlus (#4,376,600, Applied Biosystems). Gene expression levels were normalized to MTM group, and relative expression levels were compared across other experimental groups. Primer sequences are provided in Table S2. For experiments with TMA, total RNA from HEC-1B and AN3CA cells treated with 300 µM and 600 µM TMA was extracted using TRIzol reagent (#15,596,026, Invitrogen) according to the manufacturer’s protocol. cDNA was synthesized using the TOPscript™ RT DryMIX kit (#RT200, Enzynomics), followed by quantitative PCR performed with the TOPreal™ qPCR 2X PreMIX kit (#RT501M, Enzynomics) on a StepOnePlus Real-Time PCR System (#4,376,600, Applied Biosystems). Gene expression levels were normalized to those of the MTM-treated group, and relative expression was compared among the other experimental groups. Primer sequences are provided in Table S3. Pellets were obtained from PBMCs co-cultured with media or bacterial supernatants for 3 and 4 days. To analyze the immune cell population composition, we performed fluorescence-activated cell sorting (FACS) on these pellets. Specifically, we aimed to identify T cell, dendritic cell (DC), and macrophage populations. Antibody mixtures were then added to the samples for surface marker staining and incubated for 1 h at 4℃. In the case of macrophages, the samples were fixed and permeabilized using a fixation/permeabilization buffer set (BioLegend, #424,401), and an intra Ab mixture was added for intra-marker staining and incubated for 1 h at 4℃. The stained cells were then acquired using CANTO II (BD Biosciences) and BD-FACS Diva software v.8.0.2 (BD Bioscience) and analyzed using FlowJo software (v.10, TreeStar). The staining markers used for each cell type were as follows: T cells were stained with CD45 (#368,510, BioLegend), CD3 (#300,328, BioLegend), CD4 (#317,414, BioLegend), CD8 (#344,732, BioLegend), CD45RA (#304,130, BioLegend), CD62L (#304,810, BioLegend); DCs were stained with CD45 (#368,510, BioLegend), CD11b (#301,324, BioLegend), CD11c (#337,210, BioLegend), CD123 (#306,010, BioLegend); and Macrophages were stained with CD45 (#368,510, BioLegend), CD11b (#301,324, BioLegend), MHC2 (#327,022, BioLegend), CD163 (#333,614, BioLegend), CD86 (#105,006, BioLegend). T cell population composed as follows: CD4; CD45 + CD3 + CD4 + , CD4 naive; CD45 + CD3 + CD4 + CD45RA + CD62L + , CD4 effector memory (CD4 EM); CD45 + CD3 + CD4 + CD45RA − CD62L − , CD4 effector (CD4 Eff); CD45 + CD3 + CD4 + CD45RA + CD62L − , CD4 central memory (CD4 CM); CD45 + CD3 + CD4 + CD45RA − CD62L + , CD8; CD45 + CD3 + CD8 + , CD8 naive; CD45 + CD3 + CD8 + CD45RA + CD62L + , CD8 EM; CD45 + CD3 + CD8 + CD45RA − CD62L − , CD8 Eff; CD45 + CD3 + CD8 + CD45RA + CD62L − , CD8 CM; CD45 + CD3 + CD8 + CD45RA − CD62L + . DC population composed as follows: plasmacytoid DC (pDC); CD45 + CD11b + CD11c − CD123 + , conventional DC (cDC); CD45 + CD11b + CD11c + CD123 − . Macrophage population composed as follows: M1 macrophage; CD45 + CD11b + MHC2 + CD86 + , M2 macrophage; CD45 + CD11b + MHC2 + CD163 + . For experiments with TMAO, pellets were obtained from mouse splenocytes co-cultured with MTM or MTM medium containing 300 µM TMAO for 3 days. To analyze the immune cell population composition, we performed FACS on these pellets. Specifically, we aimed to identify T cell. Antibody mixtures were then added to the samples for surface marker staining and incubated for 1 h at 4℃. The stained cells were then acquired using CANTO II (BD Biosciences) and BD-FACS Diva software v.8.0.2 (BD Bioscience) and analyzed using FlowJo software (v.10, TreeStar). The staining markers used for each cell type were as follows: T cells were stained with T cells were stained with CD45 (#103,116, BioLegend), CD3 (#100,218, BioLegend), CD4 (#100,422, BioLegend), CD8a (#563,068, BD Bioscience), CD44 (#563,970, BD Bioscience), CD62L (#104,412, BioLegend). Pellets were collected from PBMCs co-cultured with NA media and KCTC 3007 for four days. RNA extraction was performed using TRIzol reagent (#15,596,026, Invitrogen), adhering to the manufacturer's protocol. Quality control of the RNA was assessed based on the A260/A280 and A260/A230 ratios, with samples having an RNA Integrity Number (RIN) greater than 7 proceeding to library preparation. This was conducted using the TruSeq Stranded Total RNA Library Prep Gold Kit. Sequencing was performed on an Illumina NovaSeq X (Illumina, San Diego, USA) at Macrogen (Seoul, Korea). The obtained raw FASTQ files were processed using fastp [ 117 ], HISAT2 [ 118 ], and featureCounts [ 129 ]. GSEA [ 131 ] and TIMER [ 132 ] were utilized to explore the functional activation of PBMC by KCTC 3007. Furthermore, Ingenuity Pathway Analysis (IPA; Qiagen, Hilden, Germany) [ 48 ] was employed to elucidate the upstream regulators, their downstream pathways, and regulator effects. UPLC-MS/MS analysis was conducted to perform untargeted metabolomics. Supernatants were collected from PBMCs after four days of co-culturing with NA media, and supernatants from KCTC 2178 and KCTC 3007. Each group was analyzed in triplicate ( N  = 3). In the LC chromatographic analysis, data from the initial 0 to 2 min were excluded due to baseline stabilization. The LC system used a Waters BEH C18 column (Waters Corporation, Milford, USA). Mobile phases consisted of water with 0.1% formic acid (Buffer A) and acetonitrile with 0.1% formic acid (Buffer B). The gradient was carefully programmed to transition through several phases over the course of the analysis: initially, the mobile phase started with 95% Buffer A and 5% Buffer B for the first minute, followed by a sharp transition to 1% Buffer A and 99% Buffer B from 1 to 20 min. This composition was maintained from 20 to 25 min. At 25 min, the gradient quickly returned to 95% Buffer A and 5% Buffer B for a brief period until 25.1 min, and then these conditions were maintained until the end of the run at 30 min. Mass spectrometry was performed using a Bruker Impact II q-ToF instrument (Bruker Corporation, Billerica, USA) in positive ion mode. Settings included a capillary voltage of 4500 V, a nebulizer pressure of 1.2 bar, and an m/z range of 50 to 1500. Metabolite identification relied on comparing m/z differences (typically within 5 ppm) and MS/MS scores, which were derived using Metaboscape software (Bruker Corporation, Billerica, USA) to compare the observed spectral data against a spectral library database. Samples with significant mass discrepancies or inconclusive MS/MS pattern matches were excluded from further analysis. Following qualitative analysis, all identified metabolites in the samples were quantified based on their peak areas. These quantitative results were then subjected to statistical analysis for comparison. The bacterial strains were subcultured twice over two days in NA media broth. Afterward, 1 mL of broth at OD 600 = 1 was collected and centrifuged to extract the pellet. The pellet was resuspended in 100μL of distilled water (DW), boiled for 10 min at 100 °C, and then mixed with 900μL of DW, followed by centrifugation at 12,000 rpm for 10 min. PCR was performed on the supernatant using CutC primers and 1 µL of DNA with 2X PCR premix (#BN430, Dyne Bio). The PCR products were analyzed on a 1.5% agarose gel prepared with TAE buffer (#Y50004, Young Science) and stained with SYBR Safe DNA Gel Stain (# S33102 , Invitrogen). Gel electrophoresis of the PCR products utilized the ivDye 1 Kb DNA ladder (#V1003-025, GenDEPOT). Imaging was conducted using the ChemiDoc Touch Imaging System (Bio-Rad, USA). Pellets were collected from PBMCs co-cultured with NA media and KCTC 3007 for four days. Live cell imaging was performed using the Tomocube HT-X1 (Daejeon, Korea). HEC-1B cells were incubated in 6-well plates one day before imaging. On the day of imaging, live KCTC 3007 or incubated PBMCs were added to each well. Imaging was conducted for 4 h, with location differences corrected and three technical replicates per well. The goal was to identify significant changes in the time-lapsed images. To ensure consistency, PBMCs added to each well had the same initial cell count at the start of the co-culture. The counts of PBMCs attached to a single cancer cell were then compared, as this reflects PBMC proliferation and activity. To further evaluate immune cell recruitment in the presence of KCTC 3007, pellets were collected from PBMCs co-cultured with NA medium and live KCTC 3007 bacteria (OD = 1, 5% v/v) for one day. Live cell imaging was performed using the Tomocube HT-X1 (Daejeon, Korea). HEC-1B and AN3CA cells were seeded in 12-well plates and incubated for 24 h prior to imaging. On the day of imaging, pre-incubated PBMCs were added to each well. Imaging was conducted for 4 h with positional drift correction applied, and three technical replicates were included per well. Three observation points were randomly designated per well, and PBMCs were added immediately after point selection. The goal was to identify significant changes in the time-lapse images. To ensure consistency, PBMCs added to each well had the same initial cell count at the start of co-culture. The number of PBMCs attached to individual cancer cells was then compared, as this reflects PBMC proliferation and activity. For pellets collected from PBMCs co-cultured in MTM medium containing 600 µM TMAO, a four-day incubation was performed prior to imaging. Live cell imaging was performed using the Tomocube HT-X1 (Daejeon, Korea). HEC-1B cells were seeded in 12-well plates and incubated for 24 h prior to imaging. On the day of imaging, the pre-incubated PBMCs were added to each well. Three observation points were randomly designated per well, and PBMCs were added immediately after point selection. Imaging was conducted for 4 h, with positional drift correction applied, and three technical replicates were included per well. To evaluate whether TMAO, the metabolite produced by BM, directly induces cancer cell pyroptosis, we performed live cell imaging experiments. PBMCs were stimulated with 600 μM TMAO for 4 days, then co-cultured with HEC-1B cells and imaged for 4 h. Pyroptosis progression was quantified based on previous pyroptosis imaging studies [ 137 – 139 ]. Simultaneously, apoptosis was assessed by FACS using Annexin-V staining (#V13241, Invitrogen). Similar to the previous microbial supernatant-based live cell imaging experiments, three locations per well with three technical replicates were analyzed for quantification. The primary goal was to identify significant changes in time-lapsed images. To ensure consistency, PBMCs added to each well had the same initial cell count at the start of the co-culture. The counts of PBMCs attached to individual cancer cells were compared, as this reflects PBMC proliferation and activity. Additionally, the number of cancer cell deaths was quantified to evaluate cytotoxicity, with cell death determined based on characteristic morphological changes observed during imaging. All statistical evaluations were executed in R, MetaboAnalyst, and Graphpad Prism. Alpha diversity was assessed via 'estimate_richness' in the package ‘phyloseq’ [ 123 ], with group comparisons made using the Wilcoxon test. Beta diversity was determined by PERMANOVA based on Bray–Curtis dissimilarity, utilizing ‘adonis2’ in the 'vegan' package [ 140 ]. To compare the abundance of specific bacterial taxa across groups, we utilized MaAsLin2 (Microbiome Multivariable Associations with Linear Models), a robust statistical framework extensively benchmarked to address the inherent challenges of microbiome data, such as sparsity, zero-inflation, and non-normality [ 1 ]. To account for variations in sequencing depth, Total Sum Scaling (TSS) normalization was applied, followed by log transformation to stabilize variance and mitigate the influence of extreme outliers. MaAsLin2 employs generalized linear models (GLMs) that are specifically engineered to handle high-dimensional, sparse taxonomic count data without requiring excessive pruning, thereby preserving biological information. Baseline characteristics including age, BMI, and underlying diseases, which showed significant differences between groups, were included as covariates in the multivariable models to adjust for potential confounding effects. Multiple testing correction was performed using the Benjamini-Hochberg (BH) method, with a false discovery rate (FDR) threshold of q  < 0.05 defining statistical significance. This comprehensive approach ensures that identified microbial associations are robust and rigorously controlled for clinical confounders. In GSEA, gene sets were classified as significant if they achieved a false discovery rate (FDR) of less than 0.05 and a family-wise error rate (FWER) p -value of less than 0.05. For the identification of significant metabolites, PCA, PLS-DA, and OPLS-DA was utilized [ 134 – 136 ]. For experimental data analysis, including WST assays, qPCR, FACS, and live cell imaging, either paired t-tests or repeated measures ANOVA were used for paired data. Adjusted p value below 0.05 was deemed statistically significant. Comparison methods were chosen based on the results of a normality test. Samples that demonstrated normal distribution were analyzed using parametric methods, while those without normal distribution were analyzed using nonparametric methods. The detailed statistical methodologies employed in this study are thoroughly delineated in the legends accompanying each figure.

A total of 78 patients were included in this study. Among them, 18 were diagnosed with benign diseases including myoma, adenomyosis, endocervical polyp, endometrial hyperplasia, and endometrial polyp (Table S1). Patients with benign diseases were significantly younger than those in the ECa group (median age 47.5 vs. 56.9 years, p  < 0.001). The median BMI of the benign group was 24.8 kg/m 2 , which was not significantly different from that of the ECa group ( p  = 0.826). A total of 60 patients were diagnosed with ECa, including 55 endometrioid (91.7%) and 5 serous (8.3%) histologies (Table 1 ). ECa patients showed a median age of 56.9 years, BMI of 24.9 kg/m 2 , and CA125 of 26.3 U/mL. Early-stage disease comprised 2009 International Federation of Gynecology and Obstetrics (FIGO) stage I ( N  = 34; 56.7%) and II ( N  = 10; 16.7%), while advanced-stage disease comprised 2009 FIGO stage III ( N  = 14; 23.3%) and IV ( N  = 2; 3.3%). Grade 3 disease was observed in 19 patients (31.7%). According to the ESGO/ESTRO/ESP risk stratification [ 3 ], 32 (53.3%), 13 (21.7%), 4 (6.7%), and 11 (18.3%) patients were assigned to the high, high-intermediate, low-intermediate, and low-risk groups, respectively. For statistical purposes, we categorized the high- and high-intermediate-risk groups as the “modified high-risk group” and the low-intermediate- and low-risk groups as the “modified low-risk group”. A total of 40 patients (66.7%) received adjuvant treatment, including 11 with radiotherapy only (18.3%), 12 with chemotherapy only (20.0%), and 17 with concurrent chemoradiation therapy (CCRT) (28.3%). During a median follow-up of 64.7 months, 12 patients experienced recurrence (20.0%) with a median recurrence-free survival (RFS) of 60.9 months. ProMisE classification assigned 10 (16.7%), 14 (23.3%), 8 (13.3%), and 28 (46.7%) patients to POLE -mutated, MMR-deficient, p53- abnormal, and p53 wild-type subgroups. Table 1 Characteristics of endometrial cancer (ECa) patients ( N  = 60) ECa ( N  = 60) Endometrioid ( N  = 55) Serous ( N  = 5) Age (yrs), median [IQR] 56.9 [48.9;63.2] 56.7 [48.3;62.0] 65.3 [54.3;71.9] BMI (kg/m 2 ), median [IQR] 24.9 [22.8;26.8] 24.9 [22.9;26.8] 20.2 [19.0;27.4] CA125 (U/ml), median [IQR] 26.3 [15.1;49.0] 26.4 [14.9;48.7] 21.8 [12.4;488.4] Underlying disease, n (%)  Hypertension 23 (38.3%) 20 (36.4%) 3 (60.0%)  Diabetes mellitus 14 (23.3%) 13 (23.6%) 1 (20.0%)  Dyslipidemia 15 (25.0%) 11 (20.0%) 4 (80.0%) Histologic subtypes, n (%)  Endometrioid 55 (91.7%) 55 (100%) 0  Serous 5 (8.3%) 0 5 (100%) 2009 FIGO Stage, n (%)  I 34 (56.7%) 32 (58.2%) 2 (40.0%)  II 10 (16.7%) 10 (18.2%) 0  III 14 (23.3%) 13 (23.6%) 1 (20.0%)  IV 2 (0.3%) 0 2 (40.0%) Grade, n (%)  Grade 1 18 (30.0%) 18 (32.7%) 0  Grade 2 23 (38.3%) 23 (41.8%) 0  Grade 3 19 (31.7%) 14 (25.5%) 5 (100%) ESMO-ESGO-ESTRO risk, n (%)  High 32 (53.3%) 27 (49.1%) 5 (100%)  High-intermediate 13 (21.7%) 13 (23.6%) 0  Low-intermediate 4 (6.7%) 4 (7.3%) 0  Low 11 (18.3%) 11 (20.0%) 0 Adjuvant therapy, n (%) 40 (66.7%) 35 (63.6%) 5 (100%)  Radiotherapy only 11 (18.3%) 11 (20.0%) 0  Chemotherapy only 12 (20.0%) 8 (14.5%) 4 (80.0%)  CCRT 17 (28.3%) 16 (29.1%) 1 (20.0%) Recurrence, n (%) 12 (20.0%) 10 (18.2%) 2 (40.0%)  Recurrence at 2 yr 8 (13.3%) 6 (10.9%) 2 (40.0%)  Recurrence at 3 yr 8 (13.3%) 6 (10.9%) 2 (40.0%) Recurrence-free survival (Months), median [IQR] 60.9 [36.6;74.2] 61.1 [45.8;74.3] 6.6 [4.1;44.9] Molecular classification, n (%)  POLE EDM 10 (16.7%) 10 (18.2%) 0  MMR-D 14 (23.3%) 14 (25.5%) 0  p53abn 8 (13.3%) 3 (5.5%) 5 (100%)  p53wt 28 (46.7%) 28 (50.9%) 0 Follow-up period (Months), median [IQR) 64.7 [56.7;74.7] 64.9 [58.5;76.2] 13.9 [4.9;44.9] Characteristics of endometrial cancer (ECa) patients ( N  = 60) The study scheme, summarized in Fig. 1 A, involved collecting endometrial tissue and serum from patients. RNA sequencing was conducted on tissue samples from all patients ( N  = 78), while targeted metabolomics was performed on the serum samples of the ECa group ( N  = 60). The RNA sequencing data facilitated both metatranscriptomics and transcriptomics analyses, while the targeted metabolomics data supported metabolomics analysis. The findings from these multi-omics analyses informed experimental validation. Analyses of multi-omics data are summarized in Figure S1A. Metatranscriptomics data were processed into diversity analysis, composition analysis, differential abundance testing, and survival analysis. Figure 1 B outlines the cohort composition. The benign group was divided into aged ( N  = 6) and young ( N  = 12) categories. The ECa group was segmented based on age (ECa_aged and ECa_young) and pathological subtype (Endometrioid and Serous). The endometrioid subtype, with a sufficient number of patients, allowed for subgroup analyses and was further divided into aged (Endo_aged) and young (Endo_young) groups. Fig. 1 Comparative metatranscriptomic composition of endometrial tissue between benign and ECa groups, demonstrating distinct microbial profiles. A Schematic overview of the study design. Endometrial tissue and serum were collected from the patients. RNA sequencing was performed on tissue samples ( N  = 78), facilitating metatranscriptomics and transcriptomics analyses. Targeted metabolomics was conducted on serum samples from the ECa group ( N  = 60), supporting metabolomics analysis. These multi-omics findings led to further experimental validation. B Classification of patients in the study. The benign group was categorized into aged and young subgroups. The ECa group was divided by age (ECa_aged and ECa_young) and pathological subtype (Endometrioid and Serous). C Analysis of alpha diversity, measured by the Chao1 index for species richness and the Shannon index for evenness, analyzed using the Wilcoxon test. D Principal Coordinates Analysis (PCoA) plot showing differences in beta diversity between the benign and ECa groups, analyzed using Permutational Multivariate Analysis of Variance (PERMANOVA). E Stacked barplot illustrating the microbial composition at the class level in benign versus ECa groups. F Heatmap depicting differentially abundant bacterial species between the benign and ECa groups. Species with p  < 0.05 and FDR q  < 0.05, as calculated by MaAsLin2, were included Comparative metatranscriptomic composition of endometrial tissue between benign and ECa groups, demonstrating distinct microbial profiles. A Schematic overview of the study design. Endometrial tissue and serum were collected from the patients. RNA sequencing was performed on tissue samples ( N  = 78), facilitating metatranscriptomics and transcriptomics analyses. Targeted metabolomics was conducted on serum samples from the ECa group ( N  = 60), supporting metabolomics analysis. These multi-omics findings led to further experimental validation. B Classification of patients in the study. The benign group was categorized into aged and young subgroups. The ECa group was divided by age (ECa_aged and ECa_young) and pathological subtype (Endometrioid and Serous). C Analysis of alpha diversity, measured by the Chao1 index for species richness and the Shannon index for evenness, analyzed using the Wilcoxon test. D Principal Coordinates Analysis (PCoA) plot showing differences in beta diversity between the benign and ECa groups, analyzed using Permutational Multivariate Analysis of Variance (PERMANOVA). E Stacked barplot illustrating the microbial composition at the class level in benign versus ECa groups. F Heatmap depicting differentially abundant bacterial species between the benign and ECa groups. Species with p  < 0.05 and FDR q  < 0.05, as calculated by MaAsLin2, were included Comparisons between the benign and ECa groups were conducted. Alpha diversity measures, based on richness (Chao1 index) and evenness (Shannon index), indicated significant differences between the benign and ECa groups ( p  = 0.021 and 0.002, respectively; Fig. 1 C). Beta diversity also showed significant differences between the groups ( p  = 0.002, R 2  = 0.049; Fig. 1 D). Compositional barplot analysis revealed distinct microbial patterns between benign and ECa groups (Fig. 1 E and S2). The benign group demonstrated remarkable microbiome homogeneity with minimal inter-sample variation. At the phylum level, Firmicutes maintained consistent dominance across nearly all benign samples, corresponding to the predominance of Bacilli at the class level and Bacillus at the genus level. This compositional stability was evident in the uniform distribution patterns across benign samples. In contrast, the ECa group exhibited substantial compositional heterogeneity and inter-sample variation. While Firmicutes and Bacilli remained present, their dominance was disrupted across the cohort, with many samples showing marked expansion of alternative lineages. Notably, Proteobacteria (represented at the class level as Gammaproteobacteria) and Actinobacteria (Actinobacteria) demonstrated increased relative abundance in numerous ECa samples. At the genus level, this resulted in the emergence of taxa such as Pseudomonas and Klebsiella . The visual heterogeneity in the ECa barplot—characterized by variable compositional patterns across samples—reflected this shift from a stable, Bacilli -dominated low-diversity microbiome in benign tissues to an unstable, compositionally diverse microbial community associated with malignancy. Beyond the taxa displaying compositional heterogeneity in the barplots, differential abundance analysis identified additional significantly different taxa between groups, as detailed in Supporting Data 1. At the phylum level, Chlamydiae showed significantly higher abundance in benign samples (Figure S2A). At the class level, Chitinophagia demonstrated elevated presence in the benign group with high coefficient values (Fig. 1 E). At the genus level, Hydrogenophaga was notably enriched in benign tissues, while Desulfosarcina was the only genus showing statistically significant enrichment in the ECa group (Figure S2B). At the species level, differential abundance testing revealed higher levels of Desulfosarcina alkanivorans, Lactobacillus fermentum (LF), Nostoc sp. NIES-4103, Sphaerochaeta coccoides, Yersinia rohdei, and Nocardiopsis gilva in the ECa group. In contrast, several species including Hydrogenophaga sp. NH-16, Staphylococcus aureus, Clostridium tetani, and Rhodopseudomonas palustris were more abundant in the benign group, along with numerous other species (Figs. 1 F and S1B). We analyzed metatranscriptomics data from the following groups: ECa, ECa_aged, Endo, and Endo_aged. Initially, we evaluated survival metrics based on clinical parameters. Survival analysis revealed marginal associations ( p  < 0.058) between prognosis and pathologic subtype, ESGO/ESTRO/ESP risk group, and modified risk group (Figs. 2 A-C and S3). Figure 2 A suggests a trend toward poorer prognosis in the serous subtype compared to the endometrioid subtype; however, this observation should be interpreted cautiously due to the limited number of serous cases ( N  = 5). In contrast, no significant differences were observed across groups in modified stage, adjuvant therapy, or molecular classification (Figures S4 and S5), including in molecular subtype analyses. These comparisons included small subgroups such as the POLE-mutated group ( N  = 10; Figures S4C, S4F, S5C, S5F), which may limit the interpretation of these null findings. Fig. 2 Multi-omics analyses highlighting Bacillus megaterium as a beneficial microbe. A Kaplan–Meier curve showing significantly different RFS between endometrioid and serous subtypes in the ECa group, analyzed using the log-rank test. B Kaplan–Meier curve displaying significantly different RFS according to ESGO/ESTRO/ESP risk stratification in the ECa group, with the log-rank test for trend applied. C Kaplan–Meier curve demonstrating significantly different RFS according to modified risk stratification in the ECa group, analyzed using the log-rank test. D PCoA plot illustrating differences in beta diversity between patients with and without adjuvant chemotherapy in the Endo_aged group, analyzed using PERMANOVA. E Differentially abundant microbes between patients with and without adjuvant chemotherapy in the Endo_aged group, including microbes with p  < 0.05 and FDR q  < 0.05, as calculated by MaAsLin2. F Kaplan–Meier curve showing significantly different RFS between Bacillus megaterium (BM)-enriched and depleted patients in the ECa group, analyzed using the log-rank test. G Barplot depicting enriched hallmark gene sets in BM-enriched and depleted patients in the ECa group, with gene sets meeting FWER p  < 0.05 and FDR q  < 0.05 as determined by Gene Set Enrichment Analysis (GSEA). H Volcano plot depicting differential abundance of immune cell populations based on BM enrichment in the ECa group. This analysis was estimated using Tumor Immune Estimation Resource (TIMER) 2.0, with significant differences identified by a t-test at p  < 0.05. I Kaplan–Meier curve showing significant differences in RFS according to the presence of BM in an external dataset from Mariathasan et al., 2018 [ 42 ]. Differences were analyzed using the log-rank test. J Volcano plot based on p -value and fold change for BM enrichment in the ECa group. Features with p   2 are labeled. K Orthogonal Partial Least Squares – Discriminant Analysis (OPLS-DA) analysis for BM enrichment in the ECa group. The top 15 differentially abundant metabolites are included based on VIP scores. L Correlation plot showing the relationship between relative BM abundance and serum TMAO concentration in the ECa group, analyzed using Pearson correlation Multi-omics analyses highlighting Bacillus megaterium as a beneficial microbe. A Kaplan–Meier curve showing significantly different RFS between endometrioid and serous subtypes in the ECa group, analyzed using the log-rank test. B Kaplan–Meier curve displaying significantly different RFS according to ESGO/ESTRO/ESP risk stratification in the ECa group, with the log-rank test for trend applied. C Kaplan–Meier curve demonstrating significantly different RFS according to modified risk stratification in the ECa group, analyzed using the log-rank test. D PCoA plot illustrating differences in beta diversity between patients with and without adjuvant chemotherapy in the Endo_aged group, analyzed using PERMANOVA. E Differentially abundant microbes between patients with and without adjuvant chemotherapy in the Endo_aged group, including microbes with p  < 0.05 and FDR q  < 0.05, as calculated by MaAsLin2. F Kaplan–Meier curve showing significantly different RFS between Bacillus megaterium (BM)-enriched and depleted patients in the ECa group, analyzed using the log-rank test. G Barplot depicting enriched hallmark gene sets in BM-enriched and depleted patients in the ECa group, with gene sets meeting FWER p  < 0.05 and FDR q  < 0.05 as determined by Gene Set Enrichment Analysis (GSEA). H Volcano plot depicting differential abundance of immune cell populations based on BM enrichment in the ECa group. This analysis was estimated using Tumor Immune Estimation Resource (TIMER) 2.0, with significant differences identified by a t-test at p  < 0.05. I Kaplan–Meier curve showing significant differences in RFS according to the presence of BM in an external dataset from Mariathasan et al., 2018 [ 42 ]. Differences were analyzed using the log-rank test. J Volcano plot based on p -value and fold change for BM enrichment in the ECa group. Features with p   2 are labeled. K Orthogonal Partial Least Squares – Discriminant Analysis (OPLS-DA) analysis for BM enrichment in the ECa group. The top 15 differentially abundant metabolites are included based on VIP scores. L Correlation plot showing the relationship between relative BM abundance and serum TMAO concentration in the ECa group, analyzed using Pearson correlation In the metatranscriptomics analysis, we observed no significant differences in alpha diversity across any of the groups. However, we identified significant differences in beta diversity associated with modified stage ( p -values ranging from 0.017 to 0.030; Figures S5A-S5D) and adjuvant chemotherapy ( p -values ranging from 0.019 to 0.031; Figs. 2 D, S5E-S5G) across the ECa, ECa_aged, Endo, and Endo_aged groups. These findings suggest a consistent pattern of microbial community variations related to clinical outcomes. In the context of differential abundance testing, several microbes were identified based on clinical parameters (Table 2 and Supporting Data 1). Notably, only Cupriavidus taiwanensis (CT), Marinomonas primoryensis (MP), and Bacillus megaterium (BM) demonstrated significant associations with RFS (Figs. 2 F, S6, S7). Through an exploratory comparison between subtypes within the ECa group, a potential enrichment of CT was observed in the serous subtype ( p  < 0.001, FDR = 0.006), although these findings are based on a limited number of samples ( N  = 5). Furthermore, CT showed suggestive associations with poor prognosis in the ECa and ECa_endo groups ( p  = 0.025 and p  = 0.062, respectively; Figures S6A, S6B), implying a possible unfavorable role of this microbe in tumor progression. There was no significant correlation between CT and RFS in the Endo and Endo_aged groups ( p  = 0.149 and p  = 0.226, respectively; Figures S6C, S6D). MP emerged from the analysis of recurrence at two and three years in the ECa group ( p  < 0.001 for both, FDR = 0.016 and 0.024, respectively). It was more abundant in patients who experienced recurrence. Furthermore, MP was linked to poor prognosis across the ECa, ECa_aged, Endo, and Endo_aged groups ( p  < 0.001, p  < 0.001, p  = 0.029, and p  = 0.084, respectively; Figures S6E-S6H), indicating its harmful effect. Table 2 Summary of differential abundance at the species level ECa ( N =60 ) ECa_aged ( N =44 ) Endo ( N =55 ) Endo_aged ( N =39 ) Histology ** ** NA NA Modified risk * - - * Adjuvant chemotherapy - - - * Recurrence * - - - Molecular classification - * - - *, q < 0.05; **, q < 0.01; NA, Not applicable; Dash (-), Not significant No significant associations were found for Grade, Grade (Endometrioid), Modified stage, Adjuvant treatment, radiotherapy, and chemoradiotherapy ( q >0.05) Detailed species lists are provided in Supporting Data 1 Summary of differential abundance at the species level *, q < 0.05; **, q < 0.01; NA, Not applicable; Dash (-), Not significant No significant associations were found for Grade, Grade (Endometrioid), Modified stage, Adjuvant treatment, radiotherapy, and chemoradiotherapy ( q >0.05) Detailed species lists are provided in Supporting Data 1 Conversely, BM was identified through differential abundance testing based on adjuvant chemotherapy in the Endo_aged group (Fig. 2 E). It was the sole microbe associated with prolonged RFS ( p  = 0.047, p  = 0.088, p  = 0.032, and p  = 0.064 in the ECa, ECa_aged, Endo, and Endo_aged groups, respectively; Figs. 2 F, S7). Given that BM was associated with favorable prognosis across all subgroups regardless of treatment status, we concluded that BM is likely a beneficial microbe, with its effects possibly mediated through stimulation of intrinsic immune pathways that might function independently of adjuvant chemotherapy. In conclusion, our metatranscriptomics analyses have elucidated significant microbial differences through diversity analysis and differential abundance testing across various parameters. Using these approaches coupled with survival analysis, we identified CT and MP as potentially harmful microbes, while identifying BM as beneficial. These insights prompted further investigations to thoroughly explore the roles of these microbes. For further analysis of the relationship between microbial presence and host gene expression, we conducted Gene Set Enrichment Analysis (GSEA). In the ECa group, CT and MP were associated with gene sets linked to carcinogenesis, cancer cell proliferation, and poor prognosis (CT, Figures S8A-S8D; MP, Figures S8E-S8H). Conversely, the BM-enriched group demonstrated enrichment in hallmark gene sets associated with immune responses, including complement, inflammatory response, and interferon α and γ responses. This suggests antitumor immunity through the activation of pro-inflammatory responses (Fig. 2G). Similar patterns were also observed in the ECa_aged, Endo, and Endo_aged subgroups (Figure S10). Therefore, we concluded that CT and MP are harmful microbes associated with detrimental gene enrichment and poor prognosis, whereas BM is beneficial, associated with gene sets that promote antitumor immunity and prolonged RFS. To verify immune activation in the BM-enriched group, we performed in silico immune cell profiling using Tumor Immune Estimation Resource (TIMER) 2.0. We identified enrichment of various immune cell populations including T cells, macrophages, neutrophils, and dendritic cells (DCs) in the BM-enriched group compared to the BM-depleted group (|log 2 (Fold change)|> 1, -log 10 ( p value) > 1; Figs. 2 H, S11A-S11C), suggesting various immune cell proliferation. In contrast, CT-depleted group showed minor differences or enrichment of immune cells compared to the CT-enriched group (|log 2 (Fold change)|> 1, -log 10 ( p value) > 1; Figures S11D-S11G). The MP-depleted group also showed enrichment of immune cells, too (|log 2 (Fold change)|> 1, -log 10 ( p value) > 1; Figure S12). In conclusion, in silico immune cell profiling showed that BM was associated with various immune cell proliferation, while CT and MP were associated with immune cell depletion. Next, we utilized an external validation cohort to assess the prognostic relevance of these bacterial species [ 42 ]. This cohort comprised RNA sequencing data from bladder tissue of 298 metastatic urothelial cancer patients who received atezolizumab, with data on overall survival (OS) and immunotherapy response data. Survival analysis, GSEA, and TIMER were performed. However, analyses could only be conducted on CT and BM as MP was present in only three samples out of 298. The Kaplan–Meier curve indicated that the presence of CT did not correlate with OS ( p  = 0.517; Figure S13A), suggesting that its association was specific to our in-house cohort. Conversely, the presence of BM significantly correlated with prolonged OS ( p  = 0.049; Fig. 2 I). Although BM enrichment did not significantly prolong OS, it showed a similar trend ( p  = 0.111; Figure S13B). As BM was the only microbe demonstrating significance in this external dataset, further analyses focused exclusively on this species. GSEA confirmed the enrichment of interferon α and γ responses, inflammatory response, and complement in both BM-present and -enriched groups (FWER p  < 0.05, FDR q  < 0.05; Figures S13C, S13D), consistent with our in-house findings. In silico immune cell profiling indicated enrichment of various immune cells, including T cells, B cells, macrophages, and DCs, suggesting immune cell proliferation in the BM-present or -enriched groups (|log2(Fold change)|> 1, -log10( p value) > 1; Figures S13E, S13F). Using multiple in silico analyses of survival data and mechanisms of action, we validated the beneficial effects of BM, unlike CT and MP. We subsequently analyzed the targeted metabolomics data, focusing on BM enrichment. By comparing metabolomic profiles between groups enriched and depleted in BM, we identified trimethylamine N-oxide (TMAO) as a consistent and significant marker, as evidenced by univariate t-tests ( p   2; Figs. 2 J, S14A-S14C) and Orthogonal Partial Least Squares—Discriminant Analysis (OPLS-DA) across all the subgroups (VIP score = 2.209, 2.290, 2.289, and 2.259, respectively; Figs. 2 K, S14D-S14F). Distinct metabolome patterns were observed in Figures S14A-S14C. To validate the robustness of group separation and metabolite discrimination, we performed complementary multivariate analyses including Principal Component Analysis (PCA) and Partial Least Squares Discriminant Analysis (PLS-DA) in addition to OPLS-DA. All three methods showed considerable overlap between the benign and ECa groups, with the unsupervised method (PCA) demonstrating the most substantial overlap, making clear group separation challenging (Figure S15A). However, supervised methods (PLS-DA and OPLS-DA) revealed improved separation patterns between the two groups, enabling the identification of significantly discriminating metabolites (Figure S15B). Notably, these supervised approaches consistently identified TMAO as an influential metabolite with high VIP scores (2.85 in OPLS-DA and 2.21 in PLS-DA), alongside several other differentiating metabolites, confirming its meaningful contribution to the observed group discrimination despite the overall compositional overlap (Figures S15C and 2 K). Furthermore, we observed a positive correlation between BM tissue levels and serum TMAO concentrations in each subgroup ( p  = 0.0629, 0.0174, 0.0287, and 0.0278, respectively; Figs. 2 L, S16E-S16G). Comparison of serum TMAO concentrations between benign and ECa patients revealed a trend toward higher levels in the ECa group, although this difference did not achieve statistical significance (benign: 5.7 [3.1; 9.1] μM vs. ECa: 8.9 [3.6; 25.9] μM; p  = 0.087; Supporting Data 3). To examine the relationship between age, BM abundance, and TMAO levels, we performed additional correlation analyses. Age showed no linear correlation with BM abundance (Figure S17). However, age stratification revealed distinct patterns: participants aged ≥ 50 years demonstrated a significant positive correlation between BM abundance and TMAO levels, while young participants showed no significant correlation (Figure S18). Notably, in the aged group, this BM-TMAO correlation remained significant regardless of malignancy status (benign: p  = 0.025; ECa: p  = 0.017; Figures S18D-S18F), whereas young patients exhibited no significant correlation in either group. We further investigated the potential association between cardiovascular disease (CVD) and TMAO levels, as some research suggests a correlation between TMAO and CVD [ 43 , 44 ]. Medical record review identified four ECa patients with documented CVD history. Serum TMAO levels tended to be higher in the ECa group with CVD history compared to those without; however, this difference did not reach statistical significance ( p  = 0.087; Table S2). Within our cohort, no significant correlation was observed between CVD history and TMAO levels ( p  = 0.4901, 0.7209; Figures S14H-I). Based on our analysis, we hypothesized that BM is directly related to TMAO. Preliminary studies indicated the cutC and cutD genes encode the enzyme choline trimethylamine-lyase (CutC) and its activating enzyme (CutD), which catalyze the conversion of choline into trimethylamine (TMA), the precursor to TMAO [ 45 , 46 ]. TMA is subsequently converted into TMAO by the liver enzyme flavin-dependent monooxygenase 3 (FMO3) in the human body [ 47 ]. We confirmed the presence of the cutC gene in BM genome data deposited in the NCBI gene database (Figure S19A), and the CutD protein in the NCBI protein database (Figure S19B). Therefore, our multi-omics analysis suggests that BM contributes to beneficial effects in ECa tissue by producing TMA, increasing serum TMAO levels, and stimulating antitumor immunity. We then proceeded to validate this hypothesis experimentally. To experimentally validate the efficacy of BM, we obtained two strains, KCTC 2178 and KCTC 3007, from the Korean Collection for Type Cultures (KCTC). We utilized fresh peripheral blood mononuclear cells (PBMCs) from healthy volunteers and the HEC-1B cell line. The PBMCs were co-cultured with culture media (MTM for the MTM group, NA for the NA group) or microbial supernatants (KCTC 2178 for the 2178 group, KCTC 3007 for the 3007 group) for 3 and 4 days. Subsequently, we performed a series of experiments, including the WST assay, quantitative PCR (qPCR), and fluorescence-activated cell sorting (FACS). Additionally, the supernatant collected on day 4 was used for untargeted metabolomics, and the pellet was used for RNA sequencing (Fig. 3 A). Fig. 3 Enhanced antitumor immunity via type I interferon response in human immune cells treated with supernatants from KCTC 3007. A Schematic diagram of the in vitro experiments. Peripheral blood was collected from healthy volunteers, and PBMCs were extracted. These cells were co-incubated with control media and microbial supernatants for 3 or 4 days. Post-incubation, cells were centrifuged to separate supernatant and pellet for subsequent untargeted metabolomics, WST assay, quantitative PCR (qPCR), immune cell profiling, live cell imaging, and RNA sequencing. B Results of a WST assay measuring relative cell viability across different treatment groups after four days of incubation. The left panel utilized Dunnett’s multiple comparisons test following a repeated-measures ANOVA. The right panel employed a paired t-test ( N  = 10 for each group). C IL1β expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test in the right panel, with N  = 10 for each group. D IL6 expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. E IFNA1 expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test, with N  = 10 for each group. F IFNB1 expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test, with N  = 10 for each group. G IFNG expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test, with N  = 10 for each group. H CD4 + T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Dunnett’s multiple comparisons test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. I CD4 + Effector T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test in the right panel, with N  = 10 for each group. J Combined populations of total CD4 + effector, effector memory, and central memory T cells, normalized to the inactivated naïve population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. K CD8 + T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Dunnett’s multiple comparisons test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. L CD8 + Effector T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test in the right panel, with N  = 10 for each group. M Combined populations of total CD8 + effector, effector memory, and central memory T cells, normalized to the inactivated naïve population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Dunnett’s multiple comparisons test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group Enhanced antitumor immunity via type I interferon response in human immune cells treated with supernatants from KCTC 3007. A Schematic diagram of the in vitro experiments. Peripheral blood was collected from healthy volunteers, and PBMCs were extracted. These cells were co-incubated with control media and microbial supernatants for 3 or 4 days. Post-incubation, cells were centrifuged to separate supernatant and pellet for subsequent untargeted metabolomics, WST assay, quantitative PCR (qPCR), immune cell profiling, live cell imaging, and RNA sequencing. B Results of a WST assay measuring relative cell viability across different treatment groups after four days of incubation. The left panel utilized Dunnett’s multiple comparisons test following a repeated-measures ANOVA. The right panel employed a paired t-test ( N  = 10 for each group). C IL1β expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test in the right panel, with N  = 10 for each group. D IL6 expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. E IFNA1 expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test, with N  = 10 for each group. F IFNB1 expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test, with N  = 10 for each group. G IFNG expression data obtained by qPCR for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test, with N  = 10 for each group. H CD4 + T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Dunnett’s multiple comparisons test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. I CD4 + Effector T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test in the right panel, with N  = 10 for each group. J Combined populations of total CD4 + effector, effector memory, and central memory T cells, normalized to the inactivated naïve population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. K CD8 + T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Dunnett’s multiple comparisons test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group. L CD8 + Effector T cell population normalized to CD3 + population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Friedman test in the left panel, and a Wilcoxon matched-pairs signed rank test in the right panel, with N  = 10 for each group. M Combined populations of total CD8 + effector, effector memory, and central memory T cells, normalized to the inactivated naïve population obtained by FACS for different treatment groups after four days of incubation. Each panel was analyzed using Dunnett’s multiple comparisons test in the left panel, and a paired t-test in the right panel, with N  = 10 for each group For the 3-day co-culture, PBMCs from four volunteers were used. The 3007 group demonstrated superior antitumor efficacy and a clearer mechanism of action compared to the 2178 group. The 3007 group showed significant antitumor efficacy ( p  = 0.0151; Figure S20A) and elevated gene expression of IL1β and IL6 , which are involved in pro-inflammatory activities ( p  = 0.0422 and p  = 0.0309, respectively; Figures S20B and S20C). However, no significant differences were observed in interferon gene expression (Figures S20D-S20F). Additionally, we observed increased proliferation of immune cell populations including CD8 + T cells and CD8 + effector T cells ( p  = 0.0253 and p  = 0.0151, respectively; Figures S21A and S21B), and a decrease in the CD4 + naïve T cell population ( p  = 0.0761; Figure S21E). There was no significant change in the DC and macrophage populations (Figure S22). In contrast, the 2178 group showed an increasing trend in WST assay ( p  = 0.1043 and p  = 0.0503, respectively; Figures S20A, S23A) and IL6 gene expression ( p  = 0.2568 and p  = 0.0830, respectively; Figures S20B, S23C) compared to controls and showed no significant changes in immune cell populations (Figures S23G-S23Q). On day 3, we observed a concurrent increase in both interleukin gene expression and the CD8 + effector T cell population, while the frequencies of macrophage and DC populations remained stable. This suggests that by day 3, the immune response had progressed to a robust adaptive phase, characterized by significant expansion and activation of T cells. We interpreted this as a downstream consequence of an earlier, essential antigen presentation phase mediated by antigen-presenting cells (APCs). Considering this immunological cascade, we anticipated no further changes in macrophage and DC populations and therefore focused our day 4 analysis exclusively on T cells, examining interferon gene expression profiles within these expanding populations. The 4-day co-culture data revealed slightly different patterns from the 3-day data. The WST assay indicated that KCTC 3007 was once again the only strain demonstrating antitumor efficacy ( p  = 0.0278; Fig. 3 B). Interestingly, a significant increase in IL1β was observed in the 3007 group ( p  = 0.0059; Fig. 3 C). IL6 expression did not show significant changes ( p  = 0.6363; Fig. 3 D), differing from day 3. Notably, genes related to type I interferons were significantly elevated ( IFNA1 ; p  = 0.0059, IFNB1 ; p  = 0.0039; Fig. 3 E, F), although IFNG expression remained unchanged ( p  = 0.4922; Fig. 3 G). Though no significant change was observed in CD4 + T cells ( p  = 0.8118; Fig. 3 H), CD4 + effector T cells in the 3007 group showed an increasing trend ( p  = 0.0645; F ig. 3I), along with the combined populations of total CD4 + effector, effector memory, and central memory T cells when normalized to the inactivated naïve population ( p  = 0.0855, Fig. 3 J). Similarly, increasing trends were noted for the total CD8 + effector, effector memory, and central memory T cells when similarly normalized ( p  = 0.1166, Fig. 3 K-M). While the 2178 group showed a significant increase in IL1β expression ( p  = 0.0281 and p  = 0.0039, respectively; Figs. 3 C and S24B) and significant decreases in IFNA1 ( p  = 0.0281 and p  = 0.0020, respectively; Figs. 3 E and S24D) and IFNB1 expression ( p  = 0.0168 and p  = 0.0020, respectively; Figs. 3 F and S24E), no significant changes in immune cell populations were observed (Figs. 3 H-M, S24G-S24L). In summary, we concluded that KCTC 3007 is an effective strain that promotes antitumor immunity. Based on our experiments, KCTC 3007 stimulates the secretion of IL1β and IL6, and induces the proliferation and activation of CD8 + T cells. This, in turn, leads to the secretion of type I interferon. To complement our experimental findings, we conducted RNA sequencing. The analysis included four randomly selected control samples (NA group) and their matched treatment samples (3007 group) from a total of ten pairs after a 4-day co-culture. The differentially expressed genes (DEGs) between these groups showed distinct patterns of enrichment (Figure S25A). GSEA revealed that the 3007 group exhibited upregulation in gene sets such as Mitotic Spindle, Interferon Alpha Response, and Notch Signaling, as defined in the MSigDB hallmark gene sets (Fig. 4 A-C). Additionally, Gene Set Variation Analysis (GSVA) indicated a similar, though more extensive, pattern of significant enrichment and suggested enhanced immune cell proliferation and activation of type I interferon responses (Figure S25B). Specifically, genes crucial for type I interferon responses showed significantly different expression patterns in the 3007 group compared to the NA group (Fig. 4 D). Among the interferon or interferon receptor genes, IFNAR1 was the only one that exhibited a notable elevation ( p  = 0.0397; Fig. 4 E). Fig. 4 RNA sequencing data showing that KCTC 3007 promotes antitumor immunity through type I interferon responses. A - C Enrichment plots of hallmark gene sets enriched in PBMCs treated with supernatants from KCTC 3007 compared to those treated with NA media, identified by GSEA. Gene sets meeting FWER p  < 0.05 and FDR q  < 0.05 were selected ( N  = 4 for each group). D Heatmap depicting differentially expressed genes within the interferon alpha response gene set. Genes included show significance with p  < 0.05 according to DESeq2 ( N  = 4 for each group). E Expression levels of IFNAR , IFNAR1 , and IFNAR2 in PBMCs treated with NA or KCTC 3007. Statistical analyses were conducted using paired t-tests or Wilcoxon matched-pairs signed-rank tests ( N  = 4 for each group). F Barplot depicting upstream regulators identified in Ingenuity Pathway Analysis (IPA). Only regulators with p  < 0.05 were included. G Plot focusing on the relationship among genes based on IFNAR1 and type 1 interferon, identified by upstream regulator analysis conducted via IPA. H - I Regulator effector networks produced by IPA, highlighting immune cell proliferation and antitumor mechanisms RNA sequencing data showing that KCTC 3007 promotes antitumor immunity through type I interferon responses. A - C Enrichment plots of hallmark gene sets enriched in PBMCs treated with supernatants from KCTC 3007 compared to those treated with NA media, identified by GSEA. Gene sets meeting FWER p  < 0.05 and FDR q  < 0.05 were selected ( N  = 4 for each group). D Heatmap depicting differentially expressed genes within the interferon alpha response gene set. Genes included show significance with p  < 0.05 according to DESeq2 ( N  = 4 for each group). E Expression levels of IFNAR , IFNAR1 , and IFNAR2 in PBMCs treated with NA or KCTC 3007. Statistical analyses were conducted using paired t-tests or Wilcoxon matched-pairs signed-rank tests ( N  = 4 for each group). F Barplot depicting upstream regulators identified in Ingenuity Pathway Analysis (IPA). Only regulators with p  < 0.05 were included. G Plot focusing on the relationship among genes based on IFNAR1 and type 1 interferon, identified by upstream regulator analysis conducted via IPA. H - I Regulator effector networks produced by IPA, highlighting immune cell proliferation and antitumor mechanisms Further analysis using Ingenuity Pathway Analysis (IPA) identified IFNAR1, RNY3, and IFNB1 as significant upstream regulators in the 3007 group (Fig. 4 F) [ 48 ]. Notably, IFNAR1 and type I interferons co-regulated downstream genes, forming networks that led to the activation of apoptosis in cancer cell lines (Fig. 4 G) and increased quantity of actin filaments (Fig. 4 H), while inhibiting replication of viral replicon (Fig. 4 I). Given that an increased quantity of actin filaments is consistent with cell proliferation and that antiviral responses are very similar to antitumor responses [ 49 ], these findings highlight the potential role of KCTC 3007 in promoting immune cell proliferation and antitumor immunity through the action of type I interferons. We next explored the mediators linking KCTC 3007 to antitumor immunity. We next tested our hypothesis that TMA and TMAO mediate this connection. To test this, we performed UPLC-MS/MS analysis on the supernatants from the NA group and 3007 group. Consistent with our hypothesis, we observed a significant decrease in compounds within the choline category in the 3007 group compared to the NA group ( p  = 0.0249; Fig. 5 A), suggesting an enhanced conversion of choline to TMA and TMAO [ 47 ]. However, despite technical limitations, direct identification of TMA and TMAO was unsuccessful. To further support our hypothesis, we experimentally confirmed the presence of the cutC gene in KCTC 3007 using primers designed based on the cutC gene sequence from Klebsiella pneumoniae , a well-established TMA-producing bacterium (Figure S26) [ 46 ]. Fig. 5 TMA-producing microbe KCTC 3007 acts through CutC gene expression to promote immune cell proliferation and enhance antitumor immunity. A Diagram and results of metabolomics analysis showing the conversion of choline into TMA and TMAO, analyzed using UPLC-MS/MS. The relative abundance of choline in each group was compared using paired t-test ( N  = 3 for each group). B Summary of findings from the study by Mirji et al., 2022 [ 50 ]. TMAO activates various immune cells, including dendritic cells (DCs), macrophages, T cells, and neutrophils, through type I interferons such as IFN-α and β, leading to antitumor immunity. C Schematic representation of the live cell imaging process. PBMCs activated by NA media or supernatant from KCTC 3007 were applied to HEC-1B cancer cells, with imaging conducted over 4 h. D Plot showing the differential attachment of activated PBMCs from NA and 3007 groups to a single HEC-1B cancer cell after 4 h of imaging. A paired t-test was used for analysis. E Representative images illustrating the interaction between HEC-1B cancer cells and live bacteria KCTC 3007 over 4 h, noting that no definite direct effects of the bacteria on cancer cells were observed (Supplementary Video 1). F Representative images displaying the interaction between HEC-1B cancer cells and activated PBMCs. Contrary to the interaction with KCTC 3007, PBMCs directly engage with cancer cells. Furthermore, immune cells in the 3007 group exhibit a higher count compared to those in the NA group, as observed in Supplementary Videos 2–7 TMA-producing microbe KCTC 3007 acts through CutC gene expression to promote immune cell proliferation and enhance antitumor immunity. A Diagram and results of metabolomics analysis showing the conversion of choline into TMA and TMAO, analyzed using UPLC-MS/MS. The relative abundance of choline in each group was compared using paired t-test ( N  = 3 for each group). B Summary of findings from the study by Mirji et al., 2022 [ 50 ]. TMAO activates various immune cells, including dendritic cells (DCs), macrophages, T cells, and neutrophils, through type I interferons such as IFN-α and β, leading to antitumor immunity. C Schematic representation of the live cell imaging process. PBMCs activated by NA media or supernatant from KCTC 3007 were applied to HEC-1B cancer cells, with imaging conducted over 4 h. D Plot showing the differential attachment of activated PBMCs from NA and 3007 groups to a single HEC-1B cancer cell after 4 h of imaging. A paired t-test was used for analysis. E Representative images illustrating the interaction between HEC-1B cancer cells and live bacteria KCTC 3007 over 4 h, noting that no definite direct effects of the bacteria on cancer cells were observed (Supplementary Video 1). F Representative images displaying the interaction between HEC-1B cancer cells and activated PBMCs. Contrary to the interaction with KCTC 3007, PBMCs directly engage with cancer cells. Furthermore, immune cells in the 3007 group exhibit a higher count compared to those in the NA group, as observed in Supplementary Videos 2–7 Subsequently, we sought to understand the connection between TMAO and antitumor immunity. Importantly, a study by Mirji et al. [ 50 ] provided a crucial reference for our investigation. They demonstrated that TMAO enhances immune activation in pancreatic cancer, as summarized in Fig. 5 B. Specifically, TMAO was shown to potentiate the type I interferon pathway, exerting antitumor effects in a type I interferon-dependent on this pathway. Single-cell data from their study indicated the involvement of CD8 + T cells, DCs, macrophages, and neutrophils, driven by several regulators identified in their research. IL1B, TNF, IFNα/β, IFNAR1, STING1, and others were identified as upstream regulators. Notably, IFNAR1 was highlighted as a core regulator, a finding that aligns with our observations (Fig. 4 F). Moreover, the presence of bacteria carrying the cutC gene was correlated with extended survival in pancreatic cancer patients, consistent with our results. In addition to this study, Wang et al. [ 51 ] demonstrated a similar mechanism of action for TMAO-mediated antitumor immunity in triple-negative breast cancer. These insights strongly support the notion that TMAO serves as a key mediator in the antitumor immunity associated with BM KCTC 3007, substantiating its potential therapeutic role. Finally, we conducted live cell imaging using Tomocube technology. We began by seeding HEC-1B cells in a 6-well plate one day before imaging. These were then co-cultured with live KCTC 3007 and activated PBMCs, and we observed them over a four-hour period (Fig. 5 C). Interestingly, we found that KCTC 3007 did not directly interact with or attack the cancer cells (Fig. 5 E; Supplementary Video 1). Notably, consistent with our RNA sequencing results, we observed a higher number of immune cells per cancer cell in the 3007 group compared to the NA group, indicating enhanced immune cell proliferation in PBMCs treated with KCTC 3007 ( p  = 0.0054; Fig. 5 D, F; Supplementary Videos 2–7). To further evaluate immune cell recruitment in the presence of KCTC 3007, we performed triple co-culture experiments with cancer cells, live KCTC 3007, and PBMCs using two ECa cell lines. Quantification of PBMC attachment per cancer cell revealed an increasing trend in the KCTC 3007-treated HEC-1B group compared to controls ( p  = 0.0784; Figure S27A; Supplementary Videos 8–13), and a significant increase in the AN3CA system ( p  = 0.0486; Figure S27B; Supplementary Videos 14–23), suggesting enhanced immune cell-tumor cell interactions with this bacterial strain. To validate the role of TMAO in mediating antitumor immunity, we conducted experiments using two ECa cell lines: HEC-1B and AN3CA. We first evaluated whether TMAO exhibits direct cytotoxic effects on cancer cells. WST assay results showed that 300 μM TMAO treatment did not produce significant cytotoxic effects on either cell line (Fig. 6 A and B), suggesting that the antitumor activity of TMAO is mediated through immune modulation rather than direct cancer cell killing. Fig. 6 Enhanced antitumor immunity via type I interferon response in mouse and human immune cells treated with TMAO. A WST assay results showing relative HEC-1B cell viability across treatment groups (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. B WST assay results showing relative AN3CA cell viability across treatment groups (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. C IFNA1 gene expression levels measured by qPCR in human immune cells (300 μM). Analyzed using Friedman test (left) and Wilcoxon test (right), N  = 10 per group. D IFNB1 gene expression levels measured by qPCR in human immune cells (300 μM). Analyzed using Friedman test (left) and Wilcoxon test (right), N  = 10 per group. E CD8 + T cell population (normalized to CD3 +) measured by FACS in mouse splenocytes (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. F CD8. + memory T cell population (normalized to CD3 +) measured by FACS in mouse splenocytes (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. G WST assay results showing relative HEC-1B cell viability with TMAO-activated PBMCs (600 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 9 per group. H WST assay results showing relative AN3CA cell viability with TMAO-activated PBMCs (600 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 9 per group. I PBMC attachment per residual HEC-1B cancer cell across treatment groups (600 μM). Paired t-test was employed ( N  = 5). J Representative time-lapse images showing interaction between HEC-1B cancer cells and activated PBMCs. TMAO-treated PBMCs directly engage with cancer cells with higher cell counts compared to MTM control, as shown in Supplementary Videos 24–33. K Pyroptotic events per HEC-1B cancer cell across treatment groups (600 μM) with representative pyroptosis image. Paired t-test was employed ( N  = 4) Enhanced antitumor immunity via type I interferon response in mouse and human immune cells treated with TMAO. A WST assay results showing relative HEC-1B cell viability across treatment groups (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. B WST assay results showing relative AN3CA cell viability across treatment groups (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. C IFNA1 gene expression levels measured by qPCR in human immune cells (300 μM). Analyzed using Friedman test (left) and Wilcoxon test (right), N  = 10 per group. D IFNB1 gene expression levels measured by qPCR in human immune cells (300 μM). Analyzed using Friedman test (left) and Wilcoxon test (right), N  = 10 per group. E CD8 + T cell population (normalized to CD3 +) measured by FACS in mouse splenocytes (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. F CD8. + memory T cell population (normalized to CD3 +) measured by FACS in mouse splenocytes (300 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 10 per group. G WST assay results showing relative HEC-1B cell viability with TMAO-activated PBMCs (600 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 9 per group. H WST assay results showing relative AN3CA cell viability with TMAO-activated PBMCs (600 μM). Analyzed using Tukey's multiple comparisons test (left) and paired t-test (right), N  = 9 per group. I PBMC attachment per residual HEC-1B cancer cell across treatment groups (600 μM). Paired t-test was employed ( N  = 5). J Representative time-lapse images showing interaction between HEC-1B cancer cells and activated PBMCs. TMAO-treated PBMCs directly engage with cancer cells with higher cell counts compared to MTM control, as shown in Supplementary Videos 24–33. K Pyroptotic events per HEC-1B cancer cell across treatment groups (600 μM) with representative pyroptosis image. Paired t-test was employed ( N  = 4) We next examined whether TMAO activates immune cells to enhance antitumor responses. PBMCs treated with 300 μM TMAO showed increased expression of type I interferon genes, with IFNA1 showing a trend toward significance ( p  = 0.0645; Fig. 6 C) and IFNB1 demonstrating significant upregulation ( p  = 0.0020; Fig. 6 D). To assess immune cell population changes, we performed FACS analysis using mouse splenocytes due to limited availability of antibodies for human PBMC analysis. TMAO treatment resulted in a significant increase in the CD8 + T cell population ( p  = 0.0195; Fig. 6 E and F), indicating enhanced cytotoxic T cell responses. To further evaluate the functional antitumor activity of TMAO-activated immune cells, we pre-treated PBMCs with 600 μM TMAO for 4 days and subsequently co-cultured them with cancer cells. WST assay revealed that TMAO-activated PBMCs showed a trend toward enhanced cytotoxic activity against AN3CA cells ( p  = 0.0628; Fig. 6 H), while no significant effect was observed in HEC-1B cells ( p  = 0.6398; Fig. 6 G). Although qPCR analysis of these co-cultured PBMCs did not show significant changes in IFNA1 , IFNB1 , and IFNG expression ( p  = 0.4961, p  = 0.8203, and p  = 0.8258, respectively; Figures S28C-S28E), real-time imaging using Tomocube demonstrated increased PBMC attachment to cancer cells in the TMAO-treated group for HEC-1B ( p  = 0.0806; F ig. 6I; Supplementary Videos 24–33), indicating enhanced immune cell-tumor cell interactions. To further characterize the mode of cancer cell death, we quantified representative features of pyroptosis and apoptosis in the co-culture system. The TMAO-treated group showed a significant increase in pyroptosis and a significant decrease in apoptosis compared to controls ( p  = 0.0063 and 0.0252, respectively; Figs. 6 K and S28F), suggesting that TMAO-activated PBMCs preferentially induce pyroptotic rather than apoptotic cell death in cancer cells. Collectively, these results demonstrate that TMAO enhances antitumor immune responses in ECa through immune cell activation and improved immune cell-tumor cell engagement, rather than through direct cytotoxic effects on cancer cells. To investigate whether TMA affects angiogenesis and cell proliferation in ECa, we treated HEC-1B and AN3CA cell lines with TMA at concentrations of 300 μM and 600 μM. Angiogenesis potential was assessed by qPCR analysis of VEGF gene expression, and cell proliferation was evaluated using WST assays. TMA treatment suppressed cell proliferation in HEC-1B cells at both 300 μM ( p  = 0.0408; Figure S29A) and 600 μM ( p  = 0.0413; Figure S29E) concentrations. However, VEGF gene expression showed no significant changes at either 300 μM ( p  = 0.2059; Figure S29C) or 600 μM ( p  = 0.2912; Figure S29G) in HEC-1B cells. In AN3CA cells, TMA treatment did not significantly affect cell proliferation at 300 μM ( p  = 0.2135; Figure S29B) or 600 μM ( p  = 0.2220; Figure S29F) concentrations, nor did it alter VEGF expression at 300 μM ( p  = 0.3284; Figure S29D) or 600 μM ( p  = 0.4739; Figure S29H) concentrations. Despite a preliminary study suggesting the potential of TMA to promote angiogenesis or cell proliferation in colon cancer [ 52 ], our findings indicate that TMA does not enhance angiogenesis or cell proliferation in ECa cell lines, suggesting that in ECa, TMA does not contribute to ECa progression through these mechanisms.

Our study represents substantial progress in endometrial microbiome research. We have successfully identified functionally relevant microbes, elucidated their mechanisms of action, and characterized their associated metabolites in the endometrium—a tissue historically considered nearly sterile. Moreover, we have rigorously validated these findings through multi-layered experimental approaches. This work provides significant advances in the field and establishes a methodological framework applicable to future multi-omics microbiome studies. Our study distinguishes itself through comprehensive integration of three multi-omics datasets: metatranscriptomic, transcriptomic, and metabolomic analyses from the same patient cohort. This approach overcomes limitations of previous research, which has been largely restricted to transcriptomic analyses of public databases [ 34 , 35 ] or has failed to integrate metabolomic and microbiome data [ 52 – 54 ]. By analyzing three distinct omics datasets from an in-house cohort, we traced a complete mechanistic pathway: linking a specific microbial strain to its key metabolite production and subsequently connecting this metabolite to changes in host gene expression. This end-to-end analysis provides deeper insights into the microbe-metabolite-host interplay and represents a significant advancement over prior studies. Applying this integrated framework, we identified BM KCTC 3007 as a beneficial microbe and established its association with improved RFS in ECa. Although the genus Bacillus is recognized for its TMA-producing capability through the presence of the cutC gene and CutD protein, its antitumor efficacy mediated by TMAO has not been previously established. Through convergent evidence from in vitro experiments, in silico immune profiling, live cell imaging, and survival data, we provide the first demonstration that BM KCTC 3007 promotes antitumor immunity in ECa through TMAO production and subsequent activation of type I interferon-mediated immune responses. Previous studies have demonstrated the antitumor potential of BM-derived components through multiple mechanisms. The lipopeptide Iturin A effectively re-sensitized docetaxel-resistant breast cancer cells by inhibiting the Akt signaling pathway, reducing cell proliferation and enhancing apoptosis [ 53 ]. Additionally, BM-derived peptidoglycans showed significant tumor resistance enhancement in mouse fibrosarcoma models, with efficacy dependent on peptide structure and administration route [ 54 ]. Furthermore, co-culture fermentation products of BM and Flammulina velutipes exhibited potent growth inhibitory effects against gastric adenocarcinoma cells, achieving 77.75% inhibition in vitro and demonstrating superior antitumor activity compared to single-strain cultures. Despite our findings and referenced preliminary studies highlighting its antitumor efficacy (Fig. 6 ), TMAO has a protumorigenic effect as described in the Introduction section. Though our cohort did not show a statistically significant difference in cardiovascular events according to serum TMAO levels, TMAO is also known for its deleterious effects in increasing the risk of cardiovascular diseases [ 43 , 44 ]. TMAO has emerged as a significant cardiovascular risk factor, with elevated plasma levels being strongly associated with increased cardiovascular disease incidence and mortality [ 55 , 56 ]. Mechanistically, TMAO promotes cardiovascular pathology through multiple pathways, including acceleration of atherosclerosis, amplification of inflammatory responses, and enhancement of thrombotic potential [ 57 , 58 ]. Given this dual nature of TMAO, the therapeutic administration of TMAO itself warrants extensive validation and careful consideration. In this context, the use of probiotic strains such as KCTC 3007, which produce a complex mixture of metabolites rather than TMAO alone, may offer a more balanced and potentially safer therapeutic approach. BM KCTC 3007 holds therapeutic promise. A randomized, double-blind, placebo-controlled study on a related BM strain (MIT411) demonstrated safety and tolerability over 45 days in healthy adults with no adverse effects [ 59 ], supporting the safety profile for short-term use. Future work should leverage these findings to accelerate clinical translation of KCTC 3007 as a novel microbial therapeutic for ECa. Additionally, bioengineering can offer a powerful avenue to address any limitations identified through future research and to develop an even more potent therapeutic strain. However, clinical translation requires careful consideration of TMAO's dual nature—antitumor effects versus potential protumor or cardiovascular risks. Promising approaches include local administration (intravaginal or intrauterine) to maximize endometrial concentrations while limiting systemic exposure, dose optimization in preclinical models, cardiovascular safety monitoring, and patient stratification based on cardiovascular risk profiles. Comprehensive preclinical safety studies assessing both efficacy and toxicity are essential prerequisites for responsible clinical advancement. We acknowledge the lack of significant BMI differences between age groups despite age-related discrepancies in our ECa cohort (Table S2). This can be explained by two factors: ethnic baseline and disease heterogeneity. First, Korean women exhibit a substantially different BMI profile than Western populations. NHANES data show Western women aged in their 40 s-60 s with BMI averages of 29–35 kg/m 2 , while KNHANES data show Korean women in the same age range maintain a narrow BMI distribution of approximately 23–24 kg/m 2 with limited age-related variation [ 60 ]. This inherently stable BMI profile minimizes observable age-related differences. Second, our cohort appears to include a relatively higher proportion of Type 2 ECa patients, who are weakly associated with BMI, naturally attenuating the expected BMI-ECa association [ 61 ]. Combined with our limited sample size, these factors resulted in statistically nonsignificant BMI differences. These limitations underscore the need for validation in multiethnic, larger cohort studies. Our observation of decreased alpha diversity in ECa patients (Fig. 1 C) compared to benign controls is supported by previous findings. Lu et al. (2020) reported significant reductions in all alpha diversity metrics in endometrial tissue samples from 25 ECa patients compared to 25 patients with benign uterine lesions [ 62 ]. Notably, both their study and ours identified microbial dysbiosis characterized by the predominance of specific bacterial taxa— Micrococcus in their cohort and Pseudomonas and Klebsiella in ours—suggesting that ECa is associated with disrupted microbial balance favoring opportunistic pathogens. Building upon these findings, our RNA sequencing approach enabled transcriptomic analysis that provides mechanistic insights into this microbiome-disease relationship. GSEA of our benign versus ECa groups revealed significant enrichment of inflammatory response-related immune pathways in the ECa group (Figure S1C), corroborating Lu et al.'s observation of elevated inflammatory cytokine levels associated with endometrial dysbiosis. Collectively, these findings support the hypothesis that alterations in the endometrial microbiome—characterized by reduced diversity and pathogen enrichment—may contribute to the creation of a pro-inflammatory microenvironment that facilitates ECa pathogenesis. Lactobacillus species are widely recognized as beneficial bacteria with probiotic properties. However, in our study, we observed a higher abundance of LF in the endometrial tissue of the ECa group compared to controls (Fig. 1 F). The biological significance of this finding remains unclear given the contrasting roles of LF reported across different cancer types. In colorectal cancer, LF exhibits anticancer properties by inducing apoptosis, suppressing cell proliferation via PI3K/AKT and Wnt/β-catenin pathway inhibition, and reducing inflammatory cytokines through NF-κB modulation [ 63 , 64 ]. Animal studies further demonstrate that LF improves gut microbiome composition, thereby suppressing tumor development [ 64 , 65 ]. These findings indicate that LF exerts anticancer effects through both direct cytotoxic and immunomodulatory mechanisms in colorectal cancer. Conversely, studies in pancreatic ductal adenocarcinoma reported that Lactobacillus presence (without species-specific distinction) was associated with antitumor immune suppression, decreased intratumoral T cell infiltration, and poor prognosis [ 66 ]. These contradictory findings suggest that the role of LF is highly cancer type-specific. Given the increased abundance of LF in ECa observed in our study and the conflicting data from previous research, further investigation is warranted to elucidate whether this bacterium exerts protumor or antitumor effects specifically in the context of ECa. In our study, tissue-resident CT and MP appeared to be associated with poor prognosis in ECa patients (Figure S7). For CT, this preliminary observation aligns with emerging evidence from other cancer types that suggests possible adverse prognostic effects of Cupriavidus species. In breast cancer, large-scale TCGA data analysis revealed that higher intratumoral abundance of CT correlated with worse prognosis and reduced survival rates, particularly in Black women [ 67 ]. Similarly, in colorectal cancer, elevated Cupriavidus abundance was associated with increased neural invasion, advanced clinical staging, shorter overall survival, elevated systemic inflammation response index (SIRI), and reduced disease-free survival [ 68 , 69 ]. These consistent findings across endometrial, breast, colorectal cancers suggest that the adverse prognostic impact of tissue-resident CT may represent a common pattern across different female and gastrointestinal malignancies. Notably, CT tended to show survival differences survival differences in the ECa and ECa_aged groups but not in the Endometrioid group. While this may suggest an adverse prognostic impact preferentially in the Serous subtype, this finding should be interpreted cautiously given the very limited number of Serous cases ( N  = 5). In contrast, MP demonstrated significant survival differences across all subgroups, suggesting its association with poor prognosis regardless of histological subtype. No previous studies have reported associations between MP and cancer prognosis, making our observation a novel finding. The mechanisms underlying these associations remain unclear for both bacteria, and species-specific functional studies are needed to elucidate how CT and MP influence tumor progression and patient outcomes in a subtype-specific manner. The GSEA results suggest that high BM abundance is associated with activation of antitumor immune responses centered on type I interferon (IFN-α/β) signaling. GSEA showed enrichment of both interferon alpha and gamma response pathways (Fig. 2 G). Type I interferons enhance CD8 + T cell function through multiple mechanisms: upregulating MHC class I expression on tumor cells to increase their visibility to CD8 + T cells [ 70 , 71 ]; directly promoting CD8 + T cell activation, proliferation, and differentiation into effector and memory cells [ 72 , 73 ]; and enhancing DC antigen cross-priming and T cell activation [ 71 , 74 ]. Therefore, the strong type I interferon signature in the high BM group provides a mechanistic link to effective T cell-mediated antitumor immunity. Notably, complement gene sets also showed enrichment in GSEA. Type I interferons and complement systems work synergistically: interferons promote adaptive immunity by activating dendritic cells and CD8 + T cells, while complement mediates innate immunity through tumor cell lysis and generation of anaphylatoxins (C3a/C5a) that recruit macrophages and neutrophils [ 75 – 78 ]. The STING pathway connects these systems by inducing type I IFN secretion and immune cell activation [ 79 , 80 ]. This concurrent activation suggests coordinated immune responses that amplify antitumor immunity [ 77 , 78 , 81 ]. Regarding neutrophil enrichment in TIMER (Figs. 2 H, S11A-C), we acknowledge the dual role of tumor-associated neutrophils (TANs), which differentiate into antitumor (N1) or protumor (N2) phenotypes depending on microenvironmental signals [ 82 , 83 ]. While TANs shift from N1 to N2 as tumors progress [ 83 , 84 ], type I interferons promote N1 polarization by enhancing neutrophil cytotoxic capacity [ 82 , 85 ]. Given the IFN-α/β-dominant microenvironment associated with high BM abundance, we suggest that neutrophil enrichment likely reflects beneficial antitumor N1 polarization. We acknowledge that the correlation between the relative abundance of BM and serum TMAO levels in the entire ECa cohort showed a statistical trend ( p  = 0.0629; Fig. 2 L) but did not reach the conventional threshold for significance ( p  < 0.05). However, more detailed subgroup analysis revealed statistically significant correlations (Figures S16E-S16G). We believe this overall biological trend indicates a meaningful association between BM and host TMAO levels. Therefore, despite the limitation in the primary cohort analysis, the consistent and significant correlations observed across multiple subgroups strongly support our hypothesis that BM plays a biologically relevant role in the host's TMAO metabolism. The age-based stratification, considering biological factors such as menopause and associated hormonal changes, revealed that the BM-TMAO relationship is age-dependent rather than universal (Figure S17). The presence of a significant correlation only in the aged group suggests that postmenopausal metabolic and hormonal environments may be necessary for BM to effectively influence systemic TMAO levels. These findings indicate that the contradictory results regarding the role of TMAO in previous studies may partially stem from heterogeneous age distributions and failure to account for age-related metabolic differences in study populations. Regarding the different immune response patterns observed between days 3 and 4 (Figs. 3 , S20-S24), our findings align with established temporal dynamics where PBMCs stimulated with microbial products show rapid interleukin upregulation ( IL-1β, IL-6, IL-12 ) within 1–3 days, followed by delayed interferon expression ( IFNA, IFNB, IFNG ) at 3–4 days [ 86 – 88 ]. This sequential pattern reflects stepwise immune activation from initial inflammation to antiviral and antitumor responses [ 86 – 88 ]. However, gene expression patterns can vary depending on experimental conditions, and given that no prior literature matches our specific experimental setup, some divergence from the reported patterns may be expected in our results. The significant upregulation of IFNA1, a key Type I interferon, suggests activation of an anti-tumor immune response mediated through Interferon-Stimulated Genes (ISGs) [ 89 , 90 ]. This 'ISG signature' is a hallmark of an inflamed, or 'hot,' tumor microenvironment, as observed in ovarian, colorectal, and lung cancers [ 91 – 93 ]. These ISGs typically include T-cell recruiting chemokines (CXCL9, CXCL10), antigen presentation components (TAP1, PSMB9), and anti-proliferative effectors (OAS1, MX1) [ 94 , 95 ]. These findings indicate that the IFNA1-ISG axis is a critical component of the anti-tumor response in our cohort. The upstream regulator analysis predicts activation of both IFNAR1 and RNY3 (Fig. 4 F). Type I interferons activate IFNAR1/2 receptors, triggering JAK-STAT signaling that induces interferon-stimulated genes (ISGs) including CIITA , ISG15 , and IFITM3 [ 96 , 97 ]. CIITA enhances MHC class II-mediated antigen presentation, ISG15 amplifies immune responses, and IFITM3 provides antiviral effects [ 97 , 98 ]. This IFNAR1-mediated cascade coordinates innate and adaptive immunity to promote antitumor responses. RNY3 is a non-coding Y-RNA that regulates immune responses and cancer progression. Decreased RNY3 expression is associated with tumor progression and poor prognosis in bladder and prostate cancers [ 99 , 100 ], while it modulates cytokine expression, particularly IL-13 in T lymphocytes [ 101 ]. Furthermore, increased actin filaments are consistent with IFN-driven immune activation. Actin remodeling enhances type I IFN production through RIG-I-like receptor activation [ 102 , 103 ] and serves as a scaffold for immune signaling complexes [ 102 , 104 ]. IFN stimulation induces further actin remodeling, creating a positive feedback loop that sustains immune activation [ 104 , 105 ] and supports cell migration, antigen presentation, and cytotoxic functions in activated immune cells [ 106 – 108 ], linking cytoskeletal dynamics to antitumor immunity. While our results demonstrate that TMAO plays a significant role in the antitumor immunity associated with KCTC 3007, the underlying mechanisms may be more complex than TMAO-mediated immunity alone. Although PBMC activation by TMAO showed similar response patterns to those induced by microbial supernatant, we could not identify identical responses across all conditions. This discrepancy may be attributed to differences in immune stimulation kinetics between microbial supernatant and purified TMAO, resulting in temporal misalignment, as well as the potential contribution of other bacterial metabolites acting in concert with TMAO. Despite extensive metabolomic analysis, TMAO was the only metabolite we successfully identified and validated from KCTC 3007. Given that KCTC 3007 produces multiple metabolites simultaneously, we acknowledge that other immune pathways may work in combination with TMAO-mediated responses. Therefore, while our data support an important role for TMAO in the antitumor effects of KCTC 3007, further investigation is warranted to fully elucidate the complete mechanisms and identify other potential bioactive metabolites. The association between BM and favorable outcomes should be interpreted within the host's hormonal and immunological context. In ECa, the 'estrobolome' establishes a baseline hormonal environment through bacterial species like LF, which produces β-glucuronidase to modulate local estrogen levels [ 109 ]. This hormonal milieu interacts with the BM-TMAO-IFN axis in an age-dependent manner. The BM-TMAO correlation strengthens in postmenopausal patients (Figure S18), suggesting that the BM-TMAO-IFN pathway functions as a compensatory immune mechanism in low-estrogen environments. This interplay between estrogen-regulating and immune-activating microbes helps explain heterogeneity in prior TMAO studies and suggests that microbiota-targeted therapies should consider patients' hormonal status. We acknowledge several limitations in our study. First, the absence of medication use and dietary habit data represents an important limitation, as these factors can significantly influence microbial communities and serve as confounding factors in differential abundance analysis [ 110 , 111 ]. Second, our use of a modified risk stratification system, while necessary for adequate statistical power, may reduce clinical prognostic granularity compared to the original ESGO/ESTRO/ESP classification. By consolidating risk groups due to limited sample sizes, we may have obscured specific microbiome differences between high-intermediate and low-intermediate risk categories. Third, due to the relatively small sample size and the low number of events, we were unable to perform multivariable survival analysis to adjust for potential confounders. Therefore, our findings based on univariate analysis should be interpreted with caution and should be validated in larger prospective cohorts. In silico immune profiling has inherent limitations including incomplete capture of biological complexity, dependence on input data quality [ 112 , 113 ], poor correlation with clinical immunogenicity, and challenges in cross-dataset generalization [ 113 , 114 ]. To address these constraints, we performed rigorous experimental validation through in vitro and in vivo studies, coupled with iterative model refinement based on comparative analysis with experimental data [ 115 , 116 ]. This multi-layered validation approach substantiated our computational predictions and strengthened the reliability of our findings. We acknowledge multiple technical constraints. First, although bioinformatics analysis identified specific microbial species, the absence of remnant samples precluded experimental validations such as culturing or fluorescence in situ hybridization (FISH) to confirm their presence. Direct culturing from tissue might have yielded more therapeutically effective strains. To address this gap in future prospective studies, we recommend implementing an enhanced sample processing protocol: upon collection, samples should be aliquoted with one portion cryopreserved for biobanking to enable subsequent validation experiments, while a separate aliquot is immediately treated with stabilizing agents (e.g., ethanol) to prevent metabolite degradation and improve metabolomics data quality. Second, we faced technical limitations in directly quantifying TMA and TMAO in culture supernatants due to extremely low metabolite concentrations, sample quenching during processing, and sample depletion from repeated analyses. However, the BM-TMAO correlation is strongly supported by multiple converging lines of evidence: robust statistical associations in our ECa cohort, biological plausibility established by BM's known enzymatic capacity ( cutC/cutD genes), and genome database confirmation. Future studies should employ supplementary methodologies such as NMR spectroscopy and isotope tracing for absolute structural and functional confirmation. Third, due to limited PBMC donor availability and antibody constraints for human FACS analysis, initial TMAO validations used mouse splenocytes, resulting in non-paired experimental designs (300 μM for mouse splenocytes; 600 μM for human PBMCs). While we recognize this methodological inconsistency, our key findings remained remarkably consistent: TMAO robustly enhanced antitumor immune responses, including promotion of CD8 + T cell populations, increased immune cell-tumor cell interactions, and preferential induction of pyroptotic cell death, across both human ECa cell lines tested. These convergent results across different experimental systems and TMAO concentrations strengthen the overall conclusion that TMAO acts as a potent immune modulator in the context of ECa. Finally, our findings are based on in vitro models and a single-center Korean patient cohort, which have inherent limitations. In vitro systems cannot fully recapitulate the complex three-dimensional tumor microenvironment and systemic immune interactions present in vivo, while our cohort's ethnic homogeneity may limit generalizability to other populations with different genetic backgrounds and microbiome compositions. To address these limitations and advance translational impact, future validation should employ patient-derived organoid models that better simulate the tumor microenvironment while maintaining patient-specific characteristics, in vivo animal models to evaluate systemic immune responses and therapeutic efficacy, and multiethnic, multicenter cohort studies to assess cross-population applicability.

Endometrial cancer (ECa) is one of the most common gynecologic malignancies representing a global burden, with an estimated 420,242 new ECa cases worldwide in 2022 [ 1 ]. In the United States, ECa is projected to be the fourth most common cancer and the fifth most common cause of cancer deaths among women in 2024 [ 2 ]. Considering its prevalence and disease burden, prevention, early detection, and precise prognostication of ECa are of paramount importance. Most patients with ECa are diagnosed at an early stage, and surgery is the mainstay of treatment [ 3 ]. Based on pathologic evaluation, patients are stratified into risk groups, and those who are at risk of recurrence are recommended to receive adjuvant treatment, consisting of radiation, chemotherapy, or both [ 4 , 5 ]. Despite refined, risk-based adjuvant therapy, patients with metastatic or recurrent ECa have limited responses to standard chemotherapy and poor survival outcomes, suggesting the necessity to increase antitumor responses [ 3 , 6 ]. Beyond the addition of novel therapeutic agents, there is a need to explore ways to enhance antitumor responses by approaching ECa from the perspective of the bacterial microbiota. The bacterial microbiota has emerged as a key player in carcinogenesis with cancer-modulating effects. Researchers are elucidating links between the bacterial microbiota and cancer, focusing in particular on immune responses, dysbiosis, genotoxicity, and metabolism [ 7 ]. Traditionally, the uterus was regarded as a sterile organ, with an understanding that microbial cultures do not thrive within it. However, increasing evidence indicates that a microbiome does exist within the uterine environment [ 8 ]. The endometrium is now known to host a microbiome characterized by low abundance and high diversity [ 9 – 12 ]. Lactobacillus has been shown to be the dominant genus in the endometrium of healthy women [ 9 – 11 ]. Additionally, genera such as Streptococcus , Bifidobacterium , Gardnerella , and Prevotella have also been identified [ 11 ]. Studies suggest that a dominance of Lactobacillus is associated with a higher rate of pregnancy or implantation success [ 13 ]. Dysbiosis and/or specific bacteria in the endometrium may have an active role in the development, progression, and metastasis of ECa through direct and indirect mechanisms [ 14 ]. ECa patients exhibit altered compositions of uterine, vaginal, and gut microbiota compared to healthy controls, characterized by increased microbial diversity and changes in specific bacterial abundance [ 15 , 16 ]. Notably, anaerobic bacteria such as Anaerococcus , Atopobium , Porphyromonas , Peptoniphilus , and Prevotella are more frequently found in ECa tissues, and these strains demonstrate carcinogenesis-related mechanisms including oxidative stress induction and increased inflammatory cytokine production [ 15 , 17 ]. The reduction of beneficial bacteria such as Lactobacillus and increased microbial diversity are associated with cancer development and progression, with microbial dysbiosis potentially increasing cancer risk through chronic inflammation, hormonal metabolism disruption, and immune response alterations [ 18 , 19 ]. Gut microbiota can also act as a risk factor for ECa through estrogen metabolism, immune regulation, and metabolic syndrome associations [ 19 , 20 ]. In a comprehensive analysis of endometrial samples from 19 non-pregnant women, Verstraelen et al. (2016) identified 183 unique 16S rRNA gene amplicons, yet only 15 phylotypes were commonly shared across all samples, underscoring the limited and variable microbial presence in the uterus [ 21 ]. They also observed highly diverse microbial communities across individual samples, with some dominated by Bacteroides , while others contained varying proportions of Lactobacillus , Prevotella , Atopobium , and Mobiluncus [ 21 ]. Quantitative diversity metrics support these observations, as multiple studies have reported higher alpha diversity indices (Shannon, Chao1) in the uterine microbiome compared to the vaginal microbiome, indicating the presence of multiple species at relatively low abundances [ 21 – 23 ]. While DNA-based sequencing methods such as 16S rRNA gene sequencing or shotgun metagenomic sequencing have been widely used, studies based on RNA sequencing have the advantage of extracting gene expression profiles while simultaneously estimating microbial abundance. Analyzing the microbiome using metatranscriptome data offers the benefit of more accurately reflecting the presence of live, active bacteria compared to metagenome data, as it captures RNA transcripts [ 24 ]. Traditional metagenomic approaches for microbial detection in tissue samples are limited by their inability to distinguish between metabolically active and dormant microorganisms, necessitating more sophisticated analytical methods. Metatranscriptomics addresses this limitation by analyzing mRNA expression profiles to identify functionally active microbial strains within tissue environments, providing crucial insights into which microorganisms are genuinely contributing to local biological processes rather than being merely present as passive residents [ 25 , 26 ]. This approach extends beyond simple taxonomic identification to reveal the functional roles and physiological states of tissue-resident microbiota by characterizing active gene expression patterns related to metabolism, pathogenicity, and antimicrobial resistance [ 25 , 27 ]. Furthermore, metatranscriptomics demonstrates particular utility in analyzing tissue samples with low microbial biomass, where optimized methodological approaches enable accurate detection and characterization of metabolically active microorganisms that might be overlooked by conventional sequencing methods [ 27 , 28 ]. The intricate relationship between tissue-resident microbiota and host immunity is largely mediated through bacterial metabolites that serve as critical signaling molecules in immune regulation. Short-chain fatty acids (SCFAs) from dietary fiber fermentation promote regulatory T cells, suppress inflammation, and maintain barrier integrity [ 29 , 30 ]. Tryptophan metabolites activate AhR signaling for mucosal homeostasis [ 30 , 31 ], while bile acids regulate immune differentiation through FXR and TGR5 pathways [ 29 , 30 ]. These metabolites exert local and systemic effects, influencing both tissue-resident and distant immune populations [ 29 , 32 ]. Disruption of this metabolic signaling through dysbiosis is associated with inflammatory bowel disease, autoimmune disorders, and other immune pathologies [ 31 , 33 ], highlighting the therapeutic potential of targeting microbiome-immune interactions. Among bacterial metabolites, TMAO has been reported to be involved in carcinogenesis-related mechanisms including tumor cell proliferation, angiogenesis, and metastasis promotion in colorectal and liver cancers, with studies showing that higher blood TMAO concentrations are associated with increased cancer risk and poor prognosis [ 34 , 35 ]. In colorectal cancer specifically, TMAO can promote cell proliferation and angiogenesis, and induce cancer progression through FXR signaling inhibition [ 19 , 35 ]. In liver cancer, TMAO also promotes cell proliferation, migration, and invasion, and is associated with POSTN and MAPK signaling pathway activation [ 36 , 37 ]. However, contradictory results have been reported in some cancers such as pancreatic cancer and triple-negative breast cancer, where TMAO may promote antitumor immunity and enhance immunotherapy responses [ 38 , 39 ]. Meta-analysis results showed that the association between TMAO and cancer incidence demonstrated only partial positive correlations in blood TMAO levels, failing to reach consistent conclusions [ 40 , 41 ]. To date, no studies have directly examined the association between TMAO and ECa. Based on these findings, our research team aimed to identify significant microbes and metabolites associated with ECa through multi-omics analyses, and to experimentally verify their mechanisms.

Supplementary Material 1. Supplementary Data. Supplementary Material 2. Supplementary Figure 1-29. Supplementary Material 3. Supplementary Table 1-4. Supplementary Material 4. Supplementary Video 1-33. Supplementary Material 5. Supporting Data 1-3. Supplementary Material 1. Supplementary Data. Supplementary Material 2. Supplementary Figure 1-29. Supplementary Material 3. Supplementary Table 1-4. Supplementary Material 4. Supplementary Video 1-33. Supplementary Material 5. Supporting Data 1-3.