Longitudinal multiomics analysis of endocrine therapy effects and gut microbiota in breast cancer recurrence.

Introduction
Breast cancer was the second most commonly diagnosed cancer worldwide in 2022, accounting for nearly 2.3 million new cases (11.6% of all cancer cases), and was the leading cause of cancer-associated death in women and fourth leading cause of cancer-related deaths globally with 666,000 cases (6.9% of all cancer deaths)1. Endocrine therapy (ET), including selective estrogen receptor modulators (SERMs) such as tamoxifen and aromatase inhibitors (AIs) such as letrozole, has been used for decades as the main treatment for estrogen receptor-positive (ER+) breast cancer. The adjuvant ET with tamoxifen or letrozole is used as the standard therapy for at least five years after surgery and reduces the mortality rate of breast cancer by 30%2. However, most breast cancer-related death is caused by recurrence, which occurs in ~30% of patients3,4.
The relationship between the gut microbiota and breast cancer has been widely examined in recent years5,6. Gut microbiota can affect breast cancer development and progression through several mechanisms7–10. By far, one of the well-known mechanisms involves the “estrobolome,” which refers to the gut bacterial genes (gut microbiota) whose products can metabolize estrogens11. The estrobolome is enriched with some gut microbiota with GUS/BG genes encoding β-glucuronidase and β-galactosidase/glucosidase, which are involved in estrogen metabolism in humans12,13. Conjugated estrogens can be deconjugated by estrobolomes, resulting in the reabsorption of estrogens into circulation14,15, leading to the accumulation of endogenous estrogens that increase the risk of ER+ breast cancer.
Nevertheless, not only does the gut microbiota affect estrogen metabolism, but also estrogen can affect the gut microbial composition. For example, an isoflavone-rich diet or soy isoflavone consumption can affect the microbial composition, suppress inflammation in, and decrease the incidence of, breast cancer, and modulate the gut microbiota to benefit metabolism16–19. Long-term administration of estrogen also directly affects the diversity and composition of the gut microbiota and further reduces β-glucuronidase activity20–22.
However, studies of the effects of isoflavone and estrogen supplementation on the gut microbiota have mainly focused on animals rather than on humans. Moreover, the effect of oral medication with ET on the gut microbiota, particularly long-term administration, has not been widely examined in humans. In addition, although the clinical benefits of ET, ~30% of patients experience recurrence after long-term treatment4,23. The gut microbiota is considered a potential biomarker of recurrence24,25. However, the profiles of gut microbiota on breast cancer recurrence remain unclear. Thus, in this study, we aimed to investigate the long-term effects of oral medication with SERMs and AIs on the profiles of the human gut microbiota, especially in patients between non-recurrence and recurrence. Moreover, the longitudinal analysis of gut microbiota and multi-omics studies of non-invasive blood cell-free DNA (cfDNA) and tissue EndoPredict score (EPclin risk score) for three to five years of follow-up to provide a multi-omics view of the effects of long-term treatment with ET and the human gut microbiota on breast cancer recurrence.
Here, we demonstrate that long-term endocrine therapy does not significantly alter the overall gut microbial composition. Patients with recurrence exhibit reduced α-diversity and increased abundances of Sutterella and Ruminococcus compared with non-recurrent patients, whereas cfDNA profiles show no significant differences between groups. High EPclin scores are associated with predicted chemotherapy benefit, yet recurrence still occurs in a subset of patients. In these cases, gut microbial markers effectively discriminate recurrence and are associated with poorer progression-free survival, particularly in patients with larger tumors. Collectively, these findings indicate that, in addition to genetic profiles, the gut microbiota represents a critical factor influencing and predicting breast cancer recurrence during long-term endocrine therapy.

Methods

Patient recruitment
The study was approved by the Internal Review Board of Kaohsiung Medical University Hospital, and informed consent was obtained from all participants [KMUHIRB-G(II)−20180018]. Forty-eight female participants were included in this study and were recruited from the Division of Breast Oncology and Surgery, Department of Surgery, Kaohsiung Medical University Chung-Ho Memorial Hospital. All participants were diagnosed with naive ER-positive/HER2-negative breast cancer via pathological examination. Patients were excluded if they were diagnosed with malignancies other than breast cancer, had a history of polyps, or were administered antibiotics within 4 weeks before the screening visit. According to the ET regimen and the purpose of this study, the 48 patients were divided into four groups: Tam-T, patients who were only administered tamoxifen without other therapeutic regimen and achieved progression-free status (follow-up duration for an average of 44 months); Let-L, patients who were only administered letrozole without other therapeutic regimen and achieved progression-free status (follow-up duration for an average of 50 months); CLet-CL, patients who received chemotherapy (CTx) with an anthracycline–taxane sequence plus letrozole and achieved progression-free status (follow-up duration for an average of 57 months); Recu-R, patients who were received CTx plus letrozole and experienced a recurrence during the ET period (follow-up duration for an average of 54 months;). Fecal samples were serially collected from the patients at two different time points: before the patients underwent ET, CTx, or surgery (the first time point: Tam-T1, Let-L1, CLet-CL1, and Recu-R1) and after taking tamoxifen or letrozole for ~1 year (the second time point: Tam-T2 or Let-L2, respectively) and after taking CTx plus letrozole [the second time point: total of ~1 year; CLet-CL2 and Recu-R2 (The second sample collection was conducted ~12 months before the recurrence event)]. In addition, blood and tissue samples were collected before therapy. Thus, we serially collected 96 fecal samples from the 48 patients: the first time point of before therapy/surgery (Tam-T1, Let-L1, CLet-CL1, and Recu-R1) and the second time point of one year after the first time point (Tam-T2, Let-L2, CLet-CL2, and Recu-R2). We also collected 48 blood samples and 34 tissue EPclin risk scores at the first time point (Tam-T1, Let-L1, CLet-CL1, and Recu-R1). Patients with breast cancer were followed up for three to five years. Moreover, all patients enrolled in this study were receiving standard care for their breast cancer in the medical center. All specimens were collected under a standard therapeutic regimen, without any interruptions or intentionally administered treatments.

Fecal DNA extraction and 16S sequencing
DNA was extracted according to the manufacturer’s instructions using a QIAamp Fast DNA Stool Mini Kit (Qiagen, Hilden, Germany). Measurement with a NanoDrop 2000 spectrophotometer (Wilmington, DE, USA) showed that the samples had an OD 260/280 ratio of 1.8–2.0. The V3–V4 hypervariable region of the 16S rDNA was amplified using bacterial-specific forward (5ʹ TCGTCGGCAGCGTCAGATGTGTATAAGAGACAGCCTACGGGNGGCWGCAG 3ʹ) and reverse (5ʹ GTCTCGTGGGCTCGGAGATG TGTATAAGAGACAGGACTACHVGGGTATCTAATCC 3ʹ) primer set. The PCR product was purified using AMPure XP magnetic beads (Beckman Coulter, Brea, CA, USA), and indexed adapters were added to the amplicons using a Nextera XT Index Kit (Illumina, San Diego, CA, USA) according to the manufacturer’s instructions. Nuclease-free water was included as a no-template control. The size of the amplified DNA was verified using a 4200 TapeStation System (Agilent Technologies, Santa Clara, CA, USA). After library construction, samples were mixed with the MiSeq Reagent Kit v3 (600-cycle) and loaded onto a MiSeq cartridge. A 2× 300 bp paired-end sequencing run was performed using a MiSeq platform (Illumina)26.

Taxonomic composition, 16S sequencing analysis, and the ratio of gut microbiota
Raw sequencing reads were analyzed using the DADA2 pipeline to identify amplicon sequence variants (ASVs). For initial quality control, R1 and R2 reads were trimmed for primer sequences, synchronized in orientation, truncated low-quality ends, and filtered with an expected error (EE) over 8 or those containing any ambiguous bases. The retained R1 and R2 reads were subjected to learning of the error model, and denosing was performed using pseudo-pooling. After merging the paired reads and removing chimeras, the taxonomy of ASVs was identified using the Greengenes 13.8 database. Raw paired-end reads were also analyzed using the BaseSpace Ribosomal Database Project (RDP) classifier. Relative abundance and the α-diversity (Shannon entropy) were determined using the CLC Genomics Workbench with the Microbial Genomics Module (Qiagen), BaseSpace (Illumina), and GraphPad Prism version 9 software (GraphPad, Inc., La Jolla, CA, USA). The β-diversity (PCoA-weighted UniFrac) was analyzed by using the R package ade4. The linear discriminant analysis (LDA) effect size (LEfSe) was conducted using MicrobiomeAnalyst 2.027 to identify specific microbial markers between groups with a p-value cutoff = 0.05 and an LDA score cut-off of 2.0. The heat tree analysis was conducted using MicrobiomeAnalyst 2.027 with the parameters of Reingold–Tilford layout and Wilcoxon p-value cutoff = 0.05. The machine learning of the random forest algorithm was also conducted using MicrobiomeAnalyst 2.027. Pairwise correlation analysis was performed among taxa at the genus level. Correlations were evaluated using Pearson’s correlation and visualized using the R package ‘corrplot’. The metabolic functional profile of the microbiota was predicted using the PICRUSt2 algorithm based on ASV abundance. The predicted abundance of functional groups in KEGG Orthology (KO) was normalized to the total read count per sample for downstream analysis. Differential enrichment analysis was performed using the ANOVA and Kruskal–Wallis test. The resulting functional profiles of the top ten pathways were visualized using a boxplot to identify those showing significant differences26. The ratio of gut microbiota was obtained by the following: divided the percentage of Sutterella by that of Ruminococcus at the first time point of before therapy (CLet-CL1 and Recu-R1) to obtain the S1 score. Then, divided the percentage at the second time point of after therapy (CLet-CL2, Recu-R2) to obtain the S2 score. Finally, divided S1 by S2 to obtain the ratio between the CLet and Recu groups.

Cell-free DNA extraction, sequencing, and analysis
Whole blood samples were collected into Cell-free DNA BCT Streck Tubes (Streck, La Vista, NE, USA) and the plasma fraction was separated from the blood. Cell-free DNA was extracted from a 5 ml plasma sample using a MagMAX™ Cell‑Free Total Nucleic Acid Isolation Kit (Thermo Fisher Scientific, Waltham, MA, USA) according to the manufacturer’s instructions. Quantitative and qualitative analysis of cfDNA were performed using a Qubit 2.0 Fluorometer (Thermo Fisher Scientific) and TapeStation system (Agilent Technologies). Libraries were constructed using Oncomine™ Breast cfDNA Research Assay v2 according to the manufacturer’s instructions. Briefly, two-cycle multiplex touchdown PCR was conducted, and the amplicon product was purified using Agencourt AMPure XP magnetic beads (Beckman Coulter, Brea, IN, USA). Sample-specific barcodes were added to the amplicons using Ion Torrent™ Tag-Sequencing adapters according to the manufacturer’s instructions and then cleaned up using Agencourt AMPure XP magnetic beads. The libraries were diluted and quantified by qPCR using an Ion Universal Quantitation Kit (Thermo Fisher Scientific). The Ion ChefTM System was used for template preparation and chip loading. The template-loaded 540™ chip was loaded on an Ion S5 XL sequencer system for sequencing. Raw data were processed automatically using Torrent Suite™ software and aligned to the reference hg19 genome to create BAM files. The BAM files were uploaded to the Ion Reporter software with an Oncomine™ plug-in for variant calling and annotation. All charts were created using R 4.2.3 software (R Core Team, 2021).

FFPE tissue RNA extraction
FFPE tumor blocks were obtained from patients with hormone receptor-positive, HER2-negative primary breast cancer. Sections (10 µm) were prepared, and adjacent hematoxylin–eosin (H&E) slides were reviewed by a pathologist to confirm ≥30% invasive carcinoma. Total RNA was extracted from paraffin-embedded tissues using the Tissue Preparation Reagents kit (Myriad International GmbH, Germany) following the manufacturer’s instructions. The yield of total RNA and the potential for DNA contamination were evaluated using Q²-Control Strips on the Mx3005P PCR Instrument. This method utilizes quantitative PCR (qPCR) to measure the amount of specific target genes in a sample. RPL37A served as the RNA input control, while HBB was used to monitor for DNA contamination.

EndoPredict® RUO assay
The EndoPredict® RUO Assay was performed using EndoPredict® RUO PCR plates (Myriad International GmbH), which contained pre-loaded primers and probes for the target genes. These plates were designed to analyze a total of 12 genes, comprising 8 signature genes, 3 normalization genes, and 1 DNA contamination control gene. For each reaction, RNA input was normalized based on the RPL37A Ct values to account for variations in starting material. The assay included EndoPredict PICO® qPCR-H₂O as a negative control and EndoPredict® PICO qREF as a positive control to ensure assay integrity and performance. All PCR amplifications were conducted on the Mx3005P PCR Instrument following the default EndoPredict CE thermal profile. Raw qPCR data were processed with the cloud-based EndoPredict® Report Generator, which integrates tumor size and lymph node status to generate the EPclin score and estimate the 10-year risk of distant metastasis. The EPclin score was calculated by combining the 12-gene molecular score (EP Score) with the patient’s tumor size and nodal status to obtain the combined EndoPredict score (EPclin score). Patients with EPclin scores of less than 3.3 were considered to be at low risk for recurrence, and those with scores greater than or equal to 3.3 were considered to be at high risk of recurrence28,29.

Statistical analysis
Different groups were compared using the Mann–Whitney test, Wilcoxon rank sum test, chi-squared test, and Fisher’s exact test. Values of p less than p <0.05, p <0.01, and p <0.001 were considered to indicate significant differences. Error bars represent the standard error of the mean (SEM) from independent experiments. Additionally, error bars represent mean ± SEM from biological replicates. The specificity and sensitivity of the microbial markers were determined using the receiver operating characteristic curve (ROC curve) and the area under the curve (AUC) value. Correlations were calculated using the Pearson’s correlation coefficients. All statistical analyses were performed using GraphPad Prism version 9 software30. Kaplan–Meier curves were generated to estimate PFS in all patients, as well as within the CLet and Recu subcohorts stratified by microbiome abundance, S1/S2, and diversity dichotomized at the median cutoff. Additional Kaplan–Meier analyses were performed according to tumor size and gut microbiota [including abundance, S1/S2, and diversity (median cutoff)] within distinct tumor size strata. P-values were calculated using the log-rank test to compare survival distributions. All analyses were performed using R software, version 4.4.1.

Results

Patients with recurrence showed lower α-diversity before therapy
In this study, we recruited 48 participants and divided into four arms: the first and second arms included 14 and 17 patients who used single tamoxifen (Tam) and letrozole (Let), respectively, with a progression-free status; the third arm included 12 patients who used CTx plus letrozole with a progression-free status (CLet); and the fourth arm consisted of 5 patients who received CTx plus letrozole and experienced recurrence (one for local recurrence and four for distant recurrence; the recurrence occurred on average about 21 months after the completion of chemotherapy and the initiation of hormonal therapy.) during ET (Recu). We serially collected fecal samples from the same patients at two time points of before therapy (Tam-T1, Let-L1, CLet-CL1, and Recu-R1) and at one year after therapy initiation (Tam-T2, Let-L2, CLet-CL2, and Recu-R2). Blood and tissue EPclin scores were determined before therapy. The clinical characteristics and EPclin score of the individuals in the four groups are summarized in Table 1. Patients were followed-up for three to five years. A representative diagram of the conception and results of this study is illustrated in Fig. 1 (this figure was created with BioRender.com). A total of 31 patients receiving long-term ET and 17 patients treated with chemotherapy plus letrozole. Fecal samples were serially collected from the same patients at two time points: before therapy and after ~1 year of tamoxifen or letrozole treatment and blood/tissue EPclin scores were collected at one-time point of before therapy. A brief summarize of results were also shown in the figure.
Initially, we analyzed the α-diversity among the four arms. In comparisons with the same regimen of third arms of the CLet-CL1 group, α-diversity was significantly reduced in patients with recurrence of the Recu-R1 group (p = 0.03, Fig. 2A). There was no significant difference in α-diversity following a single long-term administration of ET (Tam-T1 vs. Tam-T2; Let-L1 vs. Let-L2). To evaluate β-diversity, the PCoA was performed to determine the total microbial composition. The PERMANOVA analysis revealed no significant difference between before and after therapy (Tam-T1 vs. Tam-T2, Let-L1 vs. Let-L2, CLet-CL1 vs. CLet-CL2, and Recu-R1 vs. Recu-R2; Fig. 2B). In addition, the relative abundance of major taxa at the phylum and species levels did not significantly differ among the Tam-T1, Tam-T2, Let-L1, Let-L2, CLet-CL1, and CLet-CL2 groups, whereas Fusobacteria was abundant in the Recu-R2 group (Fig. 2C). These results indicate that long-term administration of ET did not significantly impact the total microbial composition. However, patients who may experience recurrence during future ET are expected to show lower α-diversity than are patients without recurrence.

Patients with recurrence and those without recurrence displayed different microbial markers
We previously found that the α-diversity and microbial composition differed in patients with recurrence than those without recurrence. Thus, we aimed to identify crucial microbial markers in patients with recurrence and focused on the same regimens in the CLet-CL and Recu-R groups. Microbial markers were verified using a heat tree analysis, machine learning of random forest algorithm, and LEfSe analysis. In heat tree analysis, the hierarchical taxonomic structure was classified from low abundance (blue) to high abundance (red). The labeled nodes represent microbiomes of significance, and the layers from the inside to the outside indicate different taxonomic levels. Fig. 3A shows that similar to the results of β-diversity, there was no apparent difference and few trees for the hierarchical taxonomic abundance in the Tam-T2 vs. Tam-T1 and Let-L2 vs. Let-L1 groups. In contrast, Sutterella was observed in the CLet and Recu heat trees (Fig. 3A). In the random forest algorithm, a higher mean decrease in accuracy (MDA) indicates that the variable or feature is more important for successful classification. In Fig. 3B, the abundances of several critical features at the genus level are displayed in different colors (low abundance: blue; high abundance: red) for the Tam-T1 to Recu-R2 groups, particularly for Sutterella and Ruminococcus for the CLet-CL1 to Recu-R2 groups (Fig. 3B). These critical features at the genus level were also observed in LEfSe analysis of the CLet-CL1 vs. Recu-R1 and CLet-CL2 vs. Recu-R2 groups (Fig. 3C).

Metabolic functional pathway in patients with recurrence
To verify the microbial markers in patients with recurrence, we utilized BaseSpace with the Ribosomal Database Project (RDP) classifier to confirm critical features at the genus and species levels. Using the RDP classifier, we screened all potential microbial markers from heat tree, random forest, and LEfSe analyses. We eliminated microbiomes with low percentages or those that were not significant in the Tam, Let, CLet, and Recu groups. Like the results of β-diversity and heat tree analysis, there was no statistically significant difference in the microbiome in the Tam-T2 vs. Tam-T1 and Let-L2 vs. Let-L1. In contrast, the abundance of Sutterella was significantly increased in Recu-R1 compared to CLet-CL1 and the second time point of Recu-R2 (Fig. 4A). On the other hands, the abundance of Ruminococcus was significantly higher in Recu-R2 than in the CLet-CL2 (Fig. 4A). We further found the species level of Ruminococcus was classified as Ruminococcus gnavus and exhibited the same trend in abundance as Ruminococcus (Fig. 4A).
Next, because the abundances of Sutterella and Ruminococcus showed opposite trends at two different time points of before and after treatment for Recu-R1 and Recu-R2, we further investigated the relationship between these genera in each group from Tam-T1 to Recu-R2. Interestingly, only the Recu-R1 group showed a significant difference. There was a significant positive correlation between the abundances of Sutterella and Ruminococcus, with a very high correlation coefficient of 0.962 (Fig. 4B). A negative correlation was observed for Recu-R2 but was not significant. To further understand the function of the microbiome in recurrence, we employed PICRUSt2 to predict the abundance of different microbiomes based on KEGG Orthology (KO). We focused on the recurrence groups from Let-L1 to Recu-R2 to identify functional pathways of the gut microbiota. Interestingly, compared with the non-recurrent group of Let and CLet-CL, Recu-R1 was enriched in pathways contributing to the abundance of the microbiome against the β-glucosidase and long-chain acyl-CoA synthetase. These results were observed in the CLet-CL1 vs. Recu-R1 groups, with β-glucosidase and long-chain acyl-CoA synthetase as the top two pathways (Fig. 4C).

Mutation profiles of cfDNA in different regimens of breast cancer
Before the patients received therapy, we also collected matched blood specimens for non-invasive cfDNA analysis to examine the circulating genomic profiles besides the microbiome. Figure 5A shows the alteration types of each target in the different groups, we constructed a donut chart of the proportion of matched cfDNA targets in the inner rings and proportion of alteration types (pathogenic and likely pathogenic) in the outer rings. The parameter was the patient number annotated only using Clinvar_clnsig, and the percentage was calculated using the variant number. The donut chart shows that the distribution of pathogenic/likely pathogenic variants in Recu-R1 was similar to that in CLet-CL1. We further visualized genomic alterations in pathogenic/likely pathogenic genes and single nucleotide variant (SNV) using OncoPrint. In Fig. 5B, the y-axis shows a summary of the alteration frequencies and alteration types of the 10 genes in each patient. The x-axis shows a summary of the alteration rates and alteration frequencies in each gene. The distribution and proportion of pathogenic/likely pathogenic variants in the Tam-T1 group were significantly lower than those in the other three groups. In contrast, the distribution and proportion of pathogenic/likely pathogenic variants was similar in the Recu-R1, Let-L1, and CLet-CL1 groups. As shown in Fig. 5C, the Tam-T1 group differed significantly from the other three groups with regard to the distribution and proportion of pathogenic/likely pathogenic variants, but this phenomenon was not observed between the CLet-CL1 and Recu-R1 groups.

Prediction value of the gut microbiota in breast cancer recurrence
Besides stool and blood, we also collected matched tissue EPclin scores for long-term observation of the benefits of chemotherapy and to evaluate the proportion of recurrence. Among all patients with long-term follow-up for progression-free survival (PFS), the PFS in the CLet/Recu group was significantly lower than that in the Tam/Let groups (p = 0.002; Fig. 6A). In our EPclin score results, 71% of patients (N = 34) had an EPclin score of up to five years of follow-up and 98% survival rates (Table 1). In the Tam group, 100% of patients (N = 11) were defined as having a low risk of EPclin. In the Let group, 83% of patients (N = 10) were defined as having a low EPclin risk, and two patients were at high risk without adjuvant chemotherapy. All patients in the Tam and Let groups were still alive with a disease-free status. In the CLet (N = 9) and Recu (N = 2) groups, 100% of patients were defined as EPclin high risk and all patients received adjuvant chemotherapy. To date, patients in the CLet group are still alive and non-recurrent, indicating the benefit of chemotherapy in high-risk patients. In long-term follow-up with PFS, patients with high-risk EPclin scores showed a significantly shorter PFS compared with those with low-risk EPclin scores (p = 0.056; Fig. 6A).
However, two high-risk patients who received adjuvant chemotherapy remain experienced disease recurrence during ET therapy. Because the cfDNA and EPclin showed identical patterns and scores for these patients but different gut microbiota profiles between the CLet and Recu groups, we further evaluated the prediction value of the gut microbiota. We divided the percentage of Sutterella by that of Ruminococcus at the first time point of before therapy (CLet-CL1 and Recu-R1) to obtain the S1 score. Then, we divided the percentage at the second time point of after therapy (CLet-CL2, Recu-R2) to obtain the S2 score. Finally, we divided S1 by S2 to obtain the ratio between the CLet and Recu groups (CLet: 2.11 versus Recu: 27.71, p = 0.001). The average value was evaluated using a receiver operating characteristic curve (ROC curve). The area under the curve (AUC) was 0.967 (p = 0.003), indicating outstanding discrimination for distinguishing between the CLet and Recu groups (Fig. 6B).
In addition to evaluating the predictive value of the gut microbiota between the CLet and Recu groups, we also assessed patient outcomes within this subcohort based on microbiome abundance, the S1/S2 ratio, and Shannon diversity. Progression-free survival (PFS) in the CLet and Recu subcohort was stratified by the median cutoff into low and high values for abundance, S1/S2, and diversity. We found that patients with high Sutterella and low Ruminococcus abundance exhibited significantly poorer PFS (p = 0.006). Furthermore, patients with a high S1/S2 ratio (p = 0.007) and low diversity (p = 0.06) also showed significantly worse PFS (Fig. 6C). We next evaluated several clinical characteristics and found that larger tumor size was associated with poor PFS across all four patient groups (Tam, Let, CLet, and Recu). In contrast, when analyzing only the CLet/Recu subcohort, tumor size alone was not statistically significant (Fig. 7A). Therefore, we further stratified this subcohort by combining tumor size with gut microbiota features. As shown in Fig. 7B, in patients with tumor size ≤20 mm, neither microbiome abundance nor the S1/S2 ratio reached statistical significance. However, in patients with tumor size ≥20 mm, those with high Sutterella, high S1/S2 ratio, and low Ruminococcus abundance exhibited significantly poorer PFS (Fig. 7C). Together, these findings suggest that in the CLet and Recu groups, the gut microbiota not only serves as a potential predictive marker but is also associated with poor prognosis, particularly in relation to tumor size.

Discussion
In breast cancer, the gut microbiota with an important role in estrogen metabolism is named as the “estrobolome.” However, not only does the gut microbiota affect estrogen metabolism, but estrogen can also affect the gut microbial composition. Isoflavone consumption can affect microbiome composition and suppress inflammation in, and decrease the incidence of, breast cancer16–19. The composition and diversity of the gut microbiota were influenced by 17β-estradiol in mouse models of colorectal cancer22. Long-term estrogen and bazedoxifene supplementation impact the composition of the gut microbiota and reduce β-glucuronidase activity, indicating that estrogen inhibition affects the gut microbiota and provides the first evidence of the effects of ET on the gut microbiota21. Then, two studies showed that administration of tamoxifen alters the gut microbiota diversity and composition and increases inflammation in breast cancer31,32. Recently, one study showed that administration of tamoxifen does not affect overall microbiota composition and abundance33. However, the effects of bazedoxifene and tamoxifen on the gut microbiota were all mainly focused on animal studies. To date, no clear research has emerged to suggest how the effect of tamoxifen and letrozole on the profiles of the gut microbiota, particularly with long-term administration of human study and in patients between non-recurrence and recurrence.
In this study, we provide the first evidence that concerns the composition of the human gut microbiota under the long-term administration of tamoxifen and letrozole. From the β-diversity of total microbial composition did not significantly differ between before and after ET, particularly in patients who were only administered tamoxifen (Tam-T1 vs. Tam-T2) and letrozole (Let-L1 vs. Let-L2). In addition, there was no apparent difference and few trees in the abundance of heat tree analysis in Tam-T2 vs. Tam-T1 and Let-L2 vs. Let-L1. In the random forest algorithm, less feature was observed in the difference of color within Tam-T1 to Tam-T2 and Let-L1 to Let-L2. All critical features from the above analysis were verified using the RDP classifier, which revealed no significant differences between Tam-T2 vs. Tam-T1 and Let-L2 vs. Let-L1. These results indicate that one-year single oral medication with tamoxifen or letrozole did not impact the human gut microbial composition. Besides, these results indicated that there were some discrepancies between animal and human studies, which should be considered in further analysis of the gut microbiota.
In addition to single ET administration, another critical issue in this study was CTx plus letrozole with non-recurrence (CLet) and recurrence (Recu). The gut microbiota is considered a potential biomarker for predicting the risk of recurrence in colorectal, lung, and liver cancers24,25,34. However, little is known about the effects of the gut microbiota profile on breast cancer recurrence. In 2020, Okubo et al. reported the first study of the association between gut microbiota and breast cancer recurrence35, but the results were based on chemotherapy and analyzed the fear of cancer recurrence (FCR) grade using the Concerns About Recurrence Scale (CARS). In this study, we compared the same regimen as the CLet group with the Recu group and found that α-diversity was significantly reduced in the Recu group before therapy (Recu-R1 vs. CLet-CL1). The α-diversity is a general indicator used to evaluate the condition of the gut environment, with lower α-diversity correlated with an unhealthy gut environment with disease status36,37. Lower α-diversity was also observed in recurrent Clostridium difficile-associated diarrhea and linked to a higher FCR grade35,38. Thus, under the regimen of CTx plus letrozole, a lower α-diversity before therapy may indicate the potential for breast cancer recurrence.
In addition to the lower α-diversity at the first-time point of before therapy, we also found that the microbial abundance of Sutterella was significantly increased at the first-time point of Recu-R1 compared with that for CLet-CL1. Sutterella is a typical pathogenic bacterium associated with autism, metabolic syndrome, and therapeutic failure in ulcerative colitis39–41. Sutterella also plays an important role in the malignancy of colorectal and pancreatic cancers and, most importantly, is a universal breast cancer biomarker26,42,43. Moreover, Sutterella was enriched in patients who did not respond to treatment with immune checkpoint inhibitors (ICIs) and involved in chemoresistence44,45. Until now, Sutterella has been reported to possess the ability to degrade IgA and induce local inflammation. Most importantly, it can promote M2 macrophage polarization, thereby altering the tumor microenvironment (TME) and affecting prognosis. Therefore, it is hypothesized that Sutterella may influence therapeutic response and prognosis by modulating the mucosal barrier and tumor microenvironment41,46,47. In contrast, the abundance of Ruminococcus and Ruminococcus gnavus were increased at the second time point in Recu-R2. Ruminococcus or Ruminococcus gnavus are involved in several diseases, including irritable bowel syndrome, colorectal cancer, and neurological disorders48,49. In addition, in colorectal cancer, a higher intratumoral abundance of microbial Cluster 1 (Ruminococcus gnavus) was significantly associated with poorer disease-free survival (DFS) after surgery (HR = 1.26, p = 0.0009)50. Consumption of a Western or high-fat diet increases the abundance of Ruminococcus, which leads to breast tumorigenesis51,52. Most importantly, Ruminococcus is enriched in patients with castration-resistant prostate cancer (CRPC) and promotes endocrine resistance via androgen biosynthesis. Ruminococcus gnavus has the ability to convert steroid precursors (pregnenolone/hydroxypregnenolone) into active androgens (DHEA/testosterone). During endocrine therapy (such as androgen deprivation therapy for prostate cancer), these gut-derived androgens can partially compensate for the loss of host endogenous androgens, thereby contributing to tumor resistance to endocrine therapy and promoting disease progression53. Thus, the study of CRPC and our results indicate that Ruminococcus may also plays a role in endocrine resistance and breast cancer recurrence.
Based on the above findings showing that Sutterella and Ruminococcus are associated with poorer cancer prognosis, our results indirectly support this observation — different gut microbial features showed a more significant association with prognosis (PFS; p = 0.006–0.06), particularly in relation to tumor size. In our previous study, we provided the first comprehensive analysis of gut microbiome profiles in breast cancer patients across different menopausal statuses26. Indeed, in the current cohort, we also observed some differences in the gut microbiome between the Tam and Let groups, which is consistent with their significant differences in age and menopausal status (and consequently, in ET regimens; Table 1). However, in the most critical comparison between the CLet and Recu groups, patients shared similar age distributions, treatment regimens, and comparable menopausal statuses. Therefore, the observed differences in gut microbiome composition between these two groups can be interpreted as independent of menopausal status, age, or treatment effects.
Interestingly, based on the opposite abundances of Sutterella and Ruminococcus in Recu-R1 and Recu-R2, we found a significant positive correlation between these genera, which was only observed in the Recu-R1 group. The more likely explanation for this is that a high abundance of Sutterella at Recu-R1 may positively correlate with and promote the abundance of Ruminococcus over a small range. Next, the abundance of Ruminococcus gradually increased after ~1 year of Recu-R2 treatment. The synergistic interaction of Sutterella and Ruminococcus is similar to that of Bacteroides fragilis and Escherichia coli in colorectal cancer54; however, the abundance of Sutterella and Ruminococcus varied at the two different time points, suggesting that these genera are crucial microbial markers of breast cancer recurrence. In functional pathway analysis, compared with CLet-CL1, the microbes in Recu-R1 were highly involved in β-glucosidase and long-chain acyl-CoA synthetase. The β-glucosidase is an enzyme component of the estrobolome; it increases the risk of estrogen receptor-positive breast cancer8. Long-chain acyl-CoA synthetase is involved in drug resistance and aggressive phenotypes of breast cancer55,56.
In addition to evaluating the major role of the gut microbiota, we analyzed another non-invasive specimen of blood cfDNA from the same patient collected at the first time point. The cfDNA is a small nucleic acid fragment released from cells via necrosis, apoptosis, and circulation in the blood. The cfDNA shows clinical utility, as it contains direct genetic information and has a short half-life, providing real-time information57–59. In total variants, the Tam-T1 group had a lower proportion of total variants than did the Let-L1, CLet-CL1, and Recu-R1 groups. This may be partly because more than 90% of patients in Tam and Let groups are in stage I. Moreover, patients taking tamoxifen are younger than those taking letrozole. Similarly, the high proportion of total variants in the Recu-R1 group may be explained by the fact that patients in Recu are stage II and III. These results are similar to those of most previous cfDNA studies demonstrating that the mutation rates of cfDNA in the early stage were significantly lower than those in later stages, which is an inherent limitation of cfDNA analysis60–63. On the other hand, aging is associated with an increased proportion of mutations in cfDNA64. Moreover, not all of our patients with breast cancer harbored cfDNA mutations because of the high proportion of stage I patients. In the ClinVar of pathogenic/likely pathogenic variants, we found that the distribution and proportion of ClinVar variants in Recu-R1 were similar to those in CLet-CL1, indicating that the cfDNA mutation profiles did not differ between non-recurrent and recurrent patients at the first time point of before therapy in our study. In fact, because of its short half-life and ability to provide real-time information, cfDNA is widely used to monitor treatment responses and recurrence65–67. However, we did not collect blood specimens at the second time point of after therapy or perform serial collection, which is a limitation of this study.
Moreover, we also collect tissue EPclin risk scores with follow-up for three to five years. EndoPredict is a well-established and clinically used multigene prognostic test (MPT) for breast cancer. In clinical practice, the use of EndoPredict can predict an individual’s risk of early or late breast cancer recurrence from 0 to 15 years to guide early treatment decisions (prognostic value). Most importantly, EndoPredict can predict an individual’s benefit from chemotherapy to guide who will benefit from added chemotherapy and who will only require ET alone and safely forgo adjuvant chemotherapy29,68,69. In our study, 71% of patients had EPclin risk scores for up to five years of follow-up and a 98% survival rate. In the Tam and Let groups treated with ET alone, 91% of patients were defined as EPclin low risk, and only two patients were defined as EPclin high risk without adjuvant chemotherapy in those who required longer follow-up. All patients in Tam and Let groups were still alive and disease-free, indicating that patients with a low EPclin risk score could safely forgo adjuvant chemotherapy. In the CLet group, 100% of patients were defined as EPclin high risk. All patients in CLet group had already received adjuvant chemotherapy and were still alive and non-recurrent, indicating that patients with a high EPclin risk could benefit from chemotherapy70,71.
However, two patients (EPclin high risk) still experienced recurrence after receiving adjuvant chemotherapy (Recu group). The gut microbiota profile of the Recu group was significantly different from that of the CLet group. The difference in the gut microbiota differed from the cfDNA profiles and EPclin risk scores, which were similar between the CLet and Recu groups. Moreover, in the CLet/Recu group, stratification based on different gut microbial features showed a more significant association with prognosis (PFS; p = 0.006–0.06), particular in association with tumor size. These results indicate that, besides the well-recognized genetic and pathological factors associated with recurrence, the gut microbiome may also represent one of the potential factors worthy of future attention. In recent years, the influence of the gut microbiota on treatment responses has been widely evaluated. The gut microbiota influences therapeutic efficacy; toxicity; and the side effects of chemotherapy, endocrine therapy, and immune checkpoint inhibitors, and subsequently affects recurrence72–75. Moreover, under the same chemotherapy plus ET regimen in patients defined as EPclin high risk, the gut microbiota ratio could provide outstanding discrimination for further distinguishing between CLet and Recu. Thus, the different profiles and outstanding AUC value of the gut microbiota between CLet and Recu provide us a hint that besides genetic factors, the gut microbiota is another critical factor that we should consider in the influence and prediction of breast cancer recurrence.
The present study had some limitations. Because we used a multi-omics approach long-term follow-up, and high response/survival rates of breast cancer, the number of subjects in this study was quite small, particularly in the Recu group. In addition, blood specimens were not collected at the second time point to monitor the potential treatment response using cfDNA mutation profiles. In the future, the gut microbial markers will be verified in a larger number of clinical participants for a longer follow-up period for EPclin score observation. However, the present study still provides valuable results and represents the most complete study to date using multi-omics approaches in long-term ET follow-up of the human gut microbiota.
In conclusion, we provide valuable and the first evidence of human study with long-term ET administration to elucidate the profiles of the gut microbiota. Additionally, we provide a comprehensive viewpoint using long-term follow-ups of stool, blood, and tissue specimens to evaluate the role and prediction utility of the gut microbiota in breast cancer recurrence. Long-term administration of ET did not significantly affect the total microbial composition. However, patients who may experience recurrence following future ET showed lower α-diversity than did patients without recurrence. Moreover, the abundances of the critical microbial markers Sutterella and Ruminococcus increased significantly before treatment and at one year after treatment in patients with recurrence. The abundances of Sutterella and Ruminococcus were positively correlated with each other only at the first time point in patients with recurrence who were enriched in pathways contributing to the abundance of the microbiome against the estrobolome and long-chain acyl-CoA synthetase. The cfDNA profiles at the first time point did not significantly differ between patients taking the same regimens who showed recurrent and non-recurrent disease. Additionally, patients with EPclin high risk and chemotherapy were indeed benefit from adjuvant chemotherapy. However, some patients with a high EPclin risk score show recurrence, even with chemotherapy. Under these circumstances, gut microbial markers may be useful for distinguishing between patients who are non-recurrence and those with recurrence, providing us a critical hint that, besides genetic profiles, the gut microbiota is another critical factor that we should consider in the influence and prediction of breast cancer recurrence and prognosis in the future.