Liver- and Microbiome-derived Bile Acids Accumulate in Human Breast Tumors and Inhibit Growth and Improve Patient Survival Free

Breast cancer is a heterogeneous disease and its biology is closely related to the expression of hormone receptors. Studies of the transcriptome discovered that breast tumors can be classified into subtypes with distinct biology (1, 2). Of all subtypes, basal-like and HER2-positive tumors tend to produce the most aggressive disease (3, 4). Estrogen receptor–positive tumors mainly fall into the luminal A subtype category while tumors lacking the expression of the estrogen, progesterone, and HER2 receptors have been described as triple-negative tumors and largely overlap with the basal-like subtype. Factors besides molecular subtypes that associate with the prognosis of patients with breast cancer include disease stage, the tumor proliferation index, the differentiation status and vascularization of a tumor, and a patient's body mass (5–8).

The relationship of tumor metabolism with breast cancer subtypes and disease prognosis remains incompletely understood (9). Large differences in metabolite abundance have been observed between tumor and adjacent noncancerous tissues and by estrogen receptor status (10, 11). It has also been shown that TP53 and Myc signaling influence breast cancer metabolism (9, 11, 12). Preclinical and clinical studies have linked tumor metabolism to therapeutic response and patient survival (9, 13). Thus, current research suggests that metabolic profiling of tumors may provide new opportunities to improve outcomes in breast cancer. Here, we interrogated the metabolome of human breast tumors to describe metabolites whose accumulation affects tumor biology and disease prognosis. We applied large-scale metabolomics and machine learning-based feature selection to identify metabolites that show a robust association with tumor biology and disease outcome. Key observations were corroborated using absolute quantification of selected metabolites and analysis of an independent dataset, and cell culture experiments. Our approach led to the identification of bile acids as metabolites that accumulate in breast tumors. Bile acids are important signaling molecules (14). Accordingly, tumors with increased bile acid content exhibited a distinct disease biology and an association with patient survival.

Patients with breast cancer were recruited between 1993 and 2003, as described previously (11, 15, 16). Patients completed a questionnaire and provided biospecimens. Samples of fresh-frozen tumor tissue and adjacent noncancerous tissue were processed by a pathologist immediately after surgery at the Department of Pathology, University of Maryland (Baltimore, MD). Clinical and pathologic information was obtained from medical records and pathology reports. The collection of biospecimens and the clinical and pathologic information was approved by the University of Maryland Institutional Review Board (IRB) for the participating institutions (UMD protocol #0298229). IRB approval of this protocol was then obtained at all institutions (Veterans Affairs Medical Center, Union Memorial Hospital, Mercy Medical Center, and Sinai Hospital, Baltimore, MD). The research was also reviewed and approved by the NIH Office of Human Subjects Research Protections (OHSRP #2248). Informed written consent was obtained from all patients and the research followed the ethical guidelines set by the Declaration of Helsinki.

Human breast cancer cell lines (MCF7, MDA-MB-175-VII, T47D, ZR-75-30) were acquired from the ATCC and were cultured in RPMI (Life Technologies) with 10% FBS (Life Technologies). Cells were regularly authenticated using either a short tandem repeat analysis with GenePrint10 or STR profile creation (17 loci plus Amelogenin) from ATCC and tested for Mycoplasma contamination. Four bile acids, deoxycholate (≥99%), chenodeoxycholate (≥97%), glycodeoxycholate (≥97%), and glycochenodeoxycholate (≥97%), were purchased from Sigma-Aldrich and resolved in DMSO to a stock concentration of 100 mmol/L and then added to culture medium at indicated concentrations. The TGR5 and OATP1B1/3 antagonists, NF449 and rifampicin, were also obtained from Sigma-Aldrich.

Cells were preseeded in 96-well cell culture plates (Corning Life Sciences) overnight and then cultured with added bile acids. Untreated and bile acid–treated cells were cultured for 24 hours and then assessed for cell proliferation using the BrdU colorimetric ELISA assay (Roche Applied Science) according to the provided protocol. Six to 12 biological replicates were analyzed per treatment condition and reported as mean ± SEM. To study the ability of TGR5 and OATP1B1/3 antagonists to inhibit the antiproliferative effects of deoxycholate, T47D was pretreated with 10 μmol/L of antagonist for 24 hours before 20 μmol/L deoxycholate was added for an additional 48 hours.

Metabolomic profiling of human breast tissues was performed using both an untargeted discovery approach and a targeted approach for validation and absolute quantification. Untargeted metabolic profiling of known and unknown metabolites in the discovery set included 67 human breast tumors and 65 tumor-adjacent noncancerous tissues and was performed by Metabolon Inc, as described previously (11, 17). Additional metabolic profiling with absolute quantification of the four bile acids, deoxycholate, chenodeoxycholate, glycodeoxycholate, and glycochenodeoxycholate was performed at the Alkek Center for Molecular Discovery of Baylor College of Medicine (Houston, TX) using selective reaction monitoring and the Agilent 6490 Triple Quadrupole Mass Spectrometer system. For absolute quantification of bile acids in tissue extracts from 20 tumor-adjacent noncancerous tissue pairs, a random subset of the 67 tumors described above, we prepared serial dilutions of bile acid standards (all Sigma-Aldrich) to generate the calibration curve.

Gene expression data existed for 61 breast tumors with metabolome data in the discovery set, as described previously (11). We performed RNA sequencing to obtain additional gene expression data for bile acid-treated cell lines (T47D, MDA-MB-175-VII). Here, cells were treated with deoxycholate for 24 hours prior to RNA isolation. RNA sequencing was performed using 3–4 biological replicates. The data were deposited in the NCBI database (http://www.ncbi.nlm.nih.gov). The NCBI BioProject ID is PRJNA544091. See Supplementary Methods for more details about the gene expression analysis.

For pathway enrichment analysis, genes were ranked by t-statistic and imported into the Gene Set Enrichment Analysis Preranked module (https://software.broadinstitute.org/gsea/index.jsp; ref. 18). Hallmark and KEGG gene sets were selected within MSigDB as references for pathway analysis. Visualization of particular pathways was performed using pathview, a tool set for pathway-based data integration and visualization in Bioconductor.

Large-scale proteome and metabolome data existed for 58 breast tumors. The mass spectrometry–based analysis of the proteome was performed as described previously (19). More details are provided in Supplementary Methods.

We selected the array-based gene expression profiles of 11 cell-cycle genes (BIRC5, CCNB1, CDC20, CEP55, MKI67, NDC80, NUF2, PTTG1, RRM2, TYMS, UBE2C) and summed them into a meta-gene score as a marker for tissue proliferation, as described previously (19–21). This proliferation signature also contains MKI67, the transcript that encodes Ki67, a commonly used proliferation marker using IHC.

We built predictive models for patient survival from a training set using Cox Proportional-Hazards Regression modeling with L1 penalized log partial likelihood estimation (22). The predictive power of metabolomics data, when integrated with clinical variables, was then assessed by the concordance index (C-index; ref. 23). The C-index is a nonparametric value to quantify the power of a predictive model, wherein a C-index of 1 indicates perfect prediction accuracy while a C-index of 0.5 indicates a model not better than random chance. More details are provided in Supplementary Methods.

Analyses were conducted using R version 3.5 (R Foundation for Statistical Computing; http://www.r-project.org/). Within R, the following packages were used: pheatmap_1.0.12, pathview_1.24.0, org.Hs.eg.db_3.8.2, beeswarm_0.2.3, survival_2.44-1.1, ggplot2_3.1.1, DESeq2_1.24.0, Biobase_2.44.0, GenomicRanges_1.36.0, and S4Vectors_0.22.0. All statistical tests were two-sided. P < 0.05 was considered statistically significant. The nonparametric Wilcoxon rank test was used for group comparisons with continuous data. Box and whisker plots were used to display data in graphs. We used the Spearman rank correlation test and calculated ρ to assess correlations between tumor markers. Survival analysis was performed for the 67 patients with breast cancer in the discovery set with existing metabolome data. These patients had long-term follow up for breast cancer–specific survival. The Cox Proportional-Hazards Regression model was applied to estimate hazard ratios (HRs) and a Wald test was used to evaluate the significance of outcome differences between risk groups. In the analysis, tumor metabolite levels were generally median-dichotomized to define high-abundance and low-abundance groups, except for glycochenodeoxycholate abundance measurements in the Metabolon dataset where the cutoff was set at the detection limit (above vs. at/below) because this metabolite was at/below the detection limit in more than 50% of the tumors in this dataset. P_trend for the Cox regression analysis was calculated using continuous data for tumor metabolite levels.

In a previous study of 67 human breast tumors and 65 tumor-adjacent noncancerous tissues, our group described relative abundance levels of 536 metabolites in these tissues (11). Here, we further investigated the association of metabolites with patient survival and disease biology using this cohort. The characteristics of the 67 patients with breast cancer are described in Supplementary Table S1. We restricted our analysis to those metabolites that were measurable in at least 25% of the tumors (n = 398) and explored their association with breast cancer survival in the context of other patient data using a cross-validation approach and LASSO-based feature selection for Cox regression modeling (Fig. 1A). The predictive power of metabolomics data, when integrated with clinical variables, was assessed by the C-index, as described under Methods. Using only metabolomics data, top models achieved substantial predictive power with C-indexes being significantly higher than 0.5 (mean C-index = 0.71; Fig. 1B), indicating the usefulness of metabolome data for survival prediction. When we applied an integrated model of metabolomics and clinical variables, we significantly improved the predictive power (mean C-index = 0.80, Wilcoxon signed rank test P < 0.001 comparing top metabolomics-based models vs. integrated models of metabolomics and clinical variables). The features that were most commonly selected by LASSO for survival prediction in the integrated models included the variables patient's age, body mass index (BMI), menopausal and node status, and receipt of chemotherapy, as well as the five metabolites, deoxycarnitine, butyrylcarnitine, 3-dehydrocarnitine, glycochenodeoxycholate (GCDC), and 2-hydroxypalmitate (Fig. 1C). Among these metabolites, GCDC is a liver metabolism–derived primary bile acid with an unexpected accumulation in a subset of breast tumors and adjacent noncancerous tissue (Supplementary Fig. S1). The other four metabolites are surrogates for changes in fatty acid metabolism in mitochondria and peroxisomes. The inclusion of these five metabolites into the integrated model for breast cancer survival prediction was substantiated with Kaplan–Meier plots and Cox proportional hazards analyses for GCDC (Fig. 1D), deoxycarnitine, butyrylcarnitine, 3-dehydrocarnitine, and 2-hydroxypalmitate (Supplementary Fig. S2). Of these metabolites, only the increase in tumor GCDC levels associated with improved patient survival [HR = 0.29; 95% confidence interval (CI): 0.1–0.83; P_trend = 0.032 when continuous data were used in the Cox regression model]. GCDC was associated with survival independent of patient's age, BMI, menopausal and node status, and receipt of chemotherapy, or when the multivariable Cox regression analysis was additionally adjusted for tumor subtypes (HR = 0.08; 95% CI: 0.02–0.65). Noteworthy, GCDC levels in breast tumors did not associate with patients' BMI, arguing against an obesity effect that led to GCDC accumulation.

To further examine the occurrence of GCDC and other bile acids in breast tumors, we selected four bile acids, deoxycholate (DC), chenodeoxycholate (CDC), glycodeoxycholate (GDC), and GCDC, for the analysis. DC is a gut microbiome–derived bile acid, while the other three originate from liver metabolism. Forty breast tissues (20 tumor-adjacent normal pairs) were analyzed using selective reaction monitoring and the Agilent 6490 Triple Quadrupole Mass Spectrometer system for absolute measurements. With this approach, we could detect these bile acids in all 40 tissues at mean concentrations of 5.7 μmol/L for DC, 19.3 μmol/L for CDC, 1.3 μmol/L for GDC, and 1.8 μmol/L for GCDC (Fig. 2A). A robust correlation (r = 0.83) was observed when we compared between GCDC measurements using the Metabolon platform and our absolute quantification method (Fig. 2B). GDC, GCDC, and DC levels showed high correlations with each other (r > 0.5) in the breast tissues, whereas CDC followed a dissimilar accumulation pattern. The bile acids tended to be slightly elevated in tumor tissue when compared with adjacent noncancerous tissues (Fig. 2C), but no clear difference emerged between estrogen receptor (ER)-positive and ER-negative tumors (Fig. 2D). GCDC levels were higher in luminal A tumors than other tumors (Fig. 2E), which we confirmed with the analysis of the only publicly available dataset (“Duke cohort”; Supplementary Fig. S3A and S3B) that contains bile acid measurements for 25 breast tumors (12). Finally, in an exploratory survival analysis using the absolute measurements of GCDC in the 20 breast tumors, we could corroborate our previous finding that increased GCDC in breast tumors is associated with improved patient survival (Fig. 2F).

To further characterize the tumor biology associated with increased GCDC, we examined relationships of GCDC with the transcriptome and proteome. Having existing gene expression and proteome data from 61 and 58 breast tumors, respectively, in our NCI-Maryland breast cancer cohort (11, 19), we initially analyzed the transcriptome differences between the 15 GCDC-high versus the 46 GCDC-low tumors. The investigation identified 73 transcripts that were significantly differentially expressed using a false discovery rate (FDR) < 5% and fold difference > 2 as stringent cutoffs (Supplementary Table S2). A list of 15 genes including BUB1, NUF2, AURKA, CDC20, CENPE/F/I, CCNB2, CDK1, FOXM1, PLK1, and TOP2A, all key regulators of the cell cycle and cell proliferation, showed significantly decreased expression in GCDC-high tumors (all were decreased between 2.1- and 2.5-fold), while genes such as PROL1, PIP, WIF1, HBA1/2, IGF2, and COL14A1 that have a relationship with extracellular matrix and exosome secretion were upregulated in these tumors (Fig. 3A). A gene set enrichment analysis (GSEA) including all differently expressed genes further corroborated these findings. The Cell Cycle KEGG pathway was the top downregulated pathway (enrichment score 0.65; FDR = 0) in GCDC-high tumors, with the downregulation of G₂–M checkpoint genes being the most significant change (Fig. 3B and C). We also assessed the proliferation score of each tumor, as described under Methods, and then correlated this score with GCDC levels in tumors. This analysis showed that GCDC-high tumors tend to have a low proliferation score, while the score is prominently higher in the GCDC-low tumors (Fig. 3D). Our finding was validated with an analysis of the “Duke cohort” (12), the independent dataset with both metabolome and TCGA transcriptome data, showing that increased GCDC levels are associated with a significantly decreased proliferation rate in breast tumors (Supplementary Fig. S3C). Similar inverse associations between tumor bile acid levels and the proliferation score were observed for GCDC, DC, and GDC based on the analysis of their absolute measurements in 20 breast tumors (Supplementary Fig. S4). However, we did not find that GCDC levels affected the tissue proliferation score in tumor-adjacent noncancerous tissues (Supplementary Fig. S5), indicating a distinct effect of GCDC on tumor biology.

As proteome data existed for 58 breast tumors in our study, we analyzed the relationship of tumor GCDC levels with proteome-defined tumor characteristics by applying GSEA with proteome-annotated genes. We made two key observations. GSEA identified the complement and coagulation cascade gene set as one of the top upregulated proteome-defined gene sets in GCDC-high tumors and corroborated the downregulation of G₂–M checkpoint genes in these tumors (Supplementary Figs. S6A–S6C and S7). The proteome data also suggested a downregulation of fatty acid metabolism genes in GCDC-high tumors.

To gain an understanding of how bile acids are taken up into breast tumors, we examined the relationship of tumor bile acid content with the following known bile acid transporters: ASBT (SLC10A2), NTCP (SLC10A1), and OATP1B1 and B3 (SLCO1B1 and SLCO1B3). We used the tumor content data for GCDC and DC from the absolute measurements in 20 tumors (Fig. 2A) for this analysis. We found that GCDC tumor content significantly correlated with the transcript expression of SLCO1B1 (ρ = 0.56; P = 0.02), whereas DC correlated with the expression of both SLCO1B1 (ρ = 0.37; P = 0.03) and SLCO1B3 (ρ = 0.34; P = 0.04). To further assess the contribution of the organic anion transporter OATP1B1 to bile acid uptake into breast cancer cells, we pretreated T47D cells, which express SLCO1B1, but not SLCO1B3, with an antagonist of the OATP1B1/3 transport system, rifampicin (24), and then challenged the cells with DC (Supplementary Fig. S8). This experiment showed rifampicin attenuates the growth-inhibitory effect of DC, indicating the importance of OATP1B1 for bile acid uptake into T47D cells.

To further gain an understanding of metabolites that are correlated with GCDC tissue levels, we conducted a correlation analysis between GCDC and the other 397 metabolites across all 132 tissue samples in the study. The analysis identified 51 metabolites that significantly correlated with GCDC (Fig. 4). Of those, 21 were positively and 30 were negatively correlated with GCDC (all P < 0.05). Further analysis showed that these metabolites were enriched for members in the sterol/steroid metabolism subpathway. Enrichment in this pathway positively correlated with GCDC levels (hypergeometric test P = 2.1 × 10⁻⁴). Androsterone sulfate and dehydroepiandrosterone sulfate (DHEA-S) were among these metabolites and had the highest correlation of all metabolites with GCDC (r = 0.72 and r = 0.69, respectively; P_adj < 0.001 each, adjusted for multiple comparison analysis). When we restricted the analysis to tumor samples only, this positive correlation with metabolites in the sterol/steroid pathway remained (r = 0.63 each; P_adj < 0.001). We then validated this finding with the analysis of our independent dataset (“Duke cohort”), verifying the positive correlation between tumor GCDC levels and various sterol/steroid pathway metabolites, including androsterone sulfate, DHEA-S, epiandrosterone sulfate, and 4-androsten-3beta, 17beta-diol disulfate 1 (r = 0.38–0.61; all P < 0.05).

Multiple studies previously examined effects of bile acids on cancer cell growth and reported both pro- and antiproliferative effects of these bile acids in different cancer cell models including breast cancer (25–29). To revisit this question and to obtain further clarification whether bile acids are growth-inhibitory in breast cancer cells, four luminal A breast cancer cell lines, MCF7, MDA-MB-175-VII, T47D, and ZR-75-30 (30), were treated with either 10 to 100 μmol/L of DC or CDC, or with 10 to 500 μmol/L of GDC or GCDC, and proliferation was assessed with a BrdU incorporation assay (Fig. 5). Although all bile acids had a growth-inhibitory effect in the luminal A breast cancer cells, both DC and CDC were more potent in inhibiting cell growth than their glycine conjugates, GDC and GCDC, which is likely explained by the limited uptake of GDC and GCDC into cultured cells. DC showed significant growth-inhibitory effects in cell culture for three cell lines, T47D, MDA-MB-175-VII, and ZR-75-30, at concentrations that are found in breast tumors (see Fig. 2A) while GCDC did not.

To further characterize the changes that are induced by bile acids in breast cancer cells, T47D and MDA-MB-175-VII cells were treated with 20 and 50 μmol/L DC, which are physiologically relevant concentrations and induce growth arrest in these cells according to our BrdU assay. Following treatment for 24 hours, we performed RNA sequencing to examine the gene expression changes induced by DC. A principal component analysis pointed to significant differences in gene expression between treated and untreated cells (Fig. 6A and E), accompanied by significantly decreased cell proliferation scores in treated cells (Fig. 6B and F), consistent with our data from human breast tumors. The DC-induced expression changes were dose-dependent (Fig. 6C and G) and affected a core set of pathways, as shown by GSEA (Fig. 6D and H). Notable, cell-cycle- and DNA replication–related genes were downregulated, and so were many DNA repair–related genes. In contrast, DC treatment led to the induction of steroid biosynthesis genes (Supplementary Fig. S9; Supplementary Table S3), consistent with human breast tumor data that pointed to an increase in sterol/steroid metabolism–related metabolites in GCDC- and DC-high breast tumors. DC also altered expression of genes in the steroid hormone biosynthesis pathway (Supplementary Fig. S10; Supplementary Table S4). We followed up on this observation and examined the expression of key enzymes in the estrogen metabolism pathway and their correlation with tumor GCDC content. Within the group of luminal A tumors that tend to have the highest GCDC content and express estrogen receptor α, we found a positive correlation with GCDC for catechol-O-methyltransferase (COMT; ρ = 0.45; P = 0.03), 17β-hydroxysteroid dehydrogenase 2 (HSD17B2; ρ = 0.49; P = 0.02), and sulfotransferase 1E1 (SULT1E1; ρ = 0.37; P = 0.08), with an average increase in expression of these genes in GCDC-high tumors by 15%–30%, based on transcript levels. Notably, all three enzyme activities are predicted to decrease estradiol availability in the tumor.

Here, we report a distinctive tumor biology and less aggressive disease in patients with breast cancer that accumulate bile acids in their tumors. Most notable, these tumors exhibit a low proliferation index based on gene expression data, which is consistent with our cell culture data showing that bile acids tend to inhibit cell-cycle progression. Increased tumor proliferation is one of the most important adverse markers for clinical outcome in breast cancer (7), specifically in luminal A tumors (8). Because luminal A tumors may contain the highest bile acid concentrations, as our data for GCDC indicate, GCDC and other bile acids may determine the tumor biology of this subtype more so than of other breast cancer subtypes. The reason of why bile acids may accumulate in these tumors remains to be determined, but our data imply that the process is not related to body mass or obesity.

Bile acids are produced by the liver and are commonly modified by the gut microbiome after their secretion into the bile and the lumen of the small intestine (31). It remains unclear whether they are modified by the microbiome of organs other than the gut. Recent publications indicate the presence of bacteria in breast tumors (32, 33). Similarly, bacteria can be found in normal human breast tissue and breast milk (34–36). Thus, it is possible that this microflora has an influence on the composition of bile acids in human breast tissue. Few previous studies investigated the relationships of bile acids with breast disease. It was shown that DC, CDC, lithocholate, and cholate can be detected in breast cyst fluid from women with fibrocystic disease (37). In a follow-up study, the same authors showed that deuterated CDC quickly appears in cyst fluid after being ingested by volunteers (38). Thus, there is uptake of bile acids into breast tissue. Others reported slightly elevated plasma bile acid concentrations in postmenopausal women with breast cancer comparing 20 patients with matched controls (39). We observed that the bile acid contents of tumor and adjacent noncancerous tissues are not significantly different, indicating that bile acids are not selectively taken up into tumors. However, we observed an inverse relationship between GCDC tissue level and the proliferation score only in tumor tissue. The data argue for a distinct effect of bile acids on tumor biology.

Several studies have investigated the effects of bile acids in breast cancer cell lines. The findings are rather heterogenous suggesting pro- and antitumor effects. Two studies reported growth-stimulatory effects of bile acids at up to 100 μmol/L concentrations in MCF-7 cells (25, 40). Others reported that DC promotes survival of MDA-MB-231 and 4T1 breast cancer cells (41, 42). In contrast, antitumor effects, including inhibition of growth, were observed by different investigators (27–29). One of these studies reported decreased long-term survival of CDC- and DC-treated MCF-7 cells and destabilization of HIF1α (28). In our investigations, DC but not GCDC, led to an inhibition of proliferation in human luminal A breast cancer cell lines at concentrations that we observed in breast tumors. Bile acids may directly interfere with the cell cycle and synergize in their effects on tumor biology because GCDC, GDC, and DC levels show high correlations with each other in breast tissues, as shown by data in this study. Yet, the inverse relationship between their tumor levels and the tissue proliferation rate may involve other mechanisms, such as the interference of bile acids with the hypoxia response, as postulated by Phelan and colleagues (28). Additional mechanisms may include an increase to drug sensitivity in presence of bile acids, as was reported recently (43), or an effect on steroid hormone biosynthesis toward a decreased estrogen availability in tumors, as suggested from our gene expression data in GCDC-high luminal A breast tumors. In addition, bile acids may also accumulate in tumor-associated fibroblasts and adipocytes and interfere with estrogen metabolism in these cells.

Bile acids signal through the farnesoid X receptor (FXR) and G-protein–coupled bile acid receptor 1 (TGR5; ref. 14). Our assessment of available gene expression data for these two receptors in breast tumors and breast cancer cell lines indicated a rather low expression. Nevertheless, treatment of T47D cells with NF449, a TRG5 antagonist, attenuated the antiproliferative effects of DC (Supplementary Fig. S8). Other studies have shown FXR protein expression in these cell lines and tumors using Western blot analysis and IHC (27, 44–46). As a general finding, FXR signaling in these cell lines led to oncosuppressive effects (27, 44, 45). Consistent with these observations, high FXR expression in human breast tumors was associated with favorable patient survival in one analysis (46). These data further substantiate our findings that the tumor bile acid content associates with a less aggressive disease and improved patient survival in breast cancer. Moreover, the upregulation of genes in the sterol/steroid metabolism in GCDC-high breast tumors is consistent with FXR and TGR5 activation and signaling, although most of our understanding about these receptors comes from studies of the liver and gut (47, 48). Finally, our observation that the complement and coagulation cascade is upregulated in GCDC-high tumors based on proteome data for these tumors could be related to the presence of FXR and TGR5 in immune cells and adipose tissue, or an activation of FXR signaling in the liver (49).

Bile acids interfere with glucose, fatty acid, and lipid metabolism through FXR signaling (14). Our proteome data suggest a downregulation of fatty acid metabolism genes in GCDC-high tumors. In addition, bile acids can inhibit carnitine acetyltransferase activity, as shown previously (50). Carnitines facilitate the transport of long-chain fatty acids into mitochondria for β-oxidation. Interestingly, the three carnitines that were found to be associated with reduced breast cancer survival in our study, butyrylcarnitine, 3-dehydrocarnitine, and deoxycarnitine, all show an inverse association in tissue levels with GCDC (ρ = −0.39, P < 0.001; ρ = −0.35, P = 0.003; ρ = −0.43, P < 0.001, respectively). Thus, the accumulation of bile acids in breast tumors may also lead to a disturbance in carnitine metabolism and inhibition of fatty acid metabolism, which should be investigated in future studies.

In summary, we report the first evidence that liver- and microbiome-derived bile acids accumulate in a subset of human breast tumors and inhibit their growth and improve patient survival. Future research is needed to gain more insight into the bile acid–induced signaling cascade in breast tumors and the involvement of cancer-associated fibroblasts and adipocytes in this biology.

No potential conflicts of interest were disclosed.

Conception and design: W. Tang, N. Putluri, S. Ambs

Development of methodology: W. Tang, V. Putluri, N. Putluri

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): W. Tang, V. Putluri, C.R. Ambati, N. Putluri

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): W. Tang, V. Putluri, C.R. Ambati, N. Putluri, S. Ambs

Writing, review, and/or revision of the manuscript: W. Tang, N. Putluri, S. Ambs

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): W. Tang, T.H. Dorsey

Study supervision: S. Ambs

We would like to thank personnel at the University of Maryland and the Baltimore Veterans Administration Hospital for their contributions with the recruitment of subjects. The research was supported by the Intramural Research Program of the Center for Cancer Research (ZIA BC 010887, to S. Ambs), American Cancer Society grant 127430-RSG-15-105-01-CNE and NIH grants R01CA220297 and R01CA216426 (all to N. Putluri), and support grants for Metabolomics Shared Resources (P30 CA125123) and CPRIT Proteomics and Metabolomics Core Facility at Dan L. Duncan Cancer Center (RP170005).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.