Tremendous advances have been made in the treatment of melanoma and other cancers by using immune checkpoint inhibitors targeting the cytotoxic T lymphocyte–associated antigen 4 (CTLA-4) and programmed cell death protein 1 (PD-1); however, responses to these therapies are often heterogeneous and not durable (
1–
3). It has recently emerged that factors beyond tumor genomics influence cancer development and therapeutic responses (
4–
7), including host factors such as the gastrointestinal (gut) microbiome (
8–
10). A number of studies have shown that the gut microbiome may influence antitumor immune responses by means of innate and adaptive immunity (
11,
12) and that therapeutic responses may be improved through its modulation (
13,
14); however, this has not been extensively studied in cancer patients.
To better understand the role of the microbiome in response to immune checkpoint blockade, we prospectively collected microbiome samples from patients with metastatic melanoma starting treatment with anti–PD-1 therapy (
n = 112 patients) (fig. S1 and table S1). Oral (buccal) and gut (fecal) microbiome samples were collected at treatment initiation, and tumor biopsies and blood samples were collected at matched pretreatment time points when possible, to assess for genomic alterations as well as the density and phenotype of tumor-infiltrating and circulating immune cell subsets (
Fig. 1A and fig. S2). Taxonomic profiling using 16
S ribosomal RNA (rRNA) gene sequencing was performed on all available oral and gut samples, with metagenomic whole-genome shotgun (WGS) sequencing performed on a subset (
n = 25). Eligible patients (
n = 89) were classified as responders (R,
n = 54) or nonresponders (NR,
n = 35) on the basis of radiographic assessment using the response evaluation criteria in solid tumors (RECIST 1.1) (
15) at 6 months after treatment initiation. Patients were classified as R if they achieved an objective response (complete or partial response or stable disease lasting at least 6 months) or NR (progressive disease or stable disease lasting less than 6 months). This classification accounts for the subset of patients who may derive long-term disease benefit despite not achieving a bona fide RECIST response and has been used in numerous published studies of patients on checkpoint blockade (
16–
19). Of note, patients in R and NR groups were similar with respect to age, gender, primary type, prior therapy, concurrent systemic therapy, and serum lactate dehydrogenase (table S2). Prior genomic analyses have demonstrated that patients with tumors that have a higher mutational load are more likely to respond to anti–CTLA-4 (
16,
20,
21) or anti–PD-1 therapy (
21–
24); however, a high mutational load alone appears neither sufficient nor essential for response. In this cohort, the total number of mutations and specific melanoma driver mutations were within comparable parameters between R and NR after anti–PD-1 therapy (fig. S3), though the number of tumors available for sequencing (
n = 10, R = 7, NR = 3) was limited and may have reduced our ability to detect a significant association between mutational burden and response.
We first assessed the landscape of the oral and gut microbiome in all available samples in patients (
n = 112) with metastatic melanoma using 16
S sequencing, noting that both communities were relatively diverse with a high abundance of bacteria of the order Lactobacillales in the oral microbiome and Bacteroidales in the fecal microbiome (
Fig. 1B). Bipartite network analysis (
25) demonstrated a clear separation of community structure between the oral and fecal microbiomes in terms of both matched and aggregate samples (fig. S4), suggesting that these communities are distinct in terms of their compositional structure. Loss of microbial diversity (dysbiosis) is associated with chronic health conditions (
26–
28) and cancer (
8–
10) and is also associated with poor outcomes of certain forms of cancer therapy, including allogeneic stem cell transplant (
29). Based on these data, we examined the diversity of the oral and gut microbiomes in eligible patients on anti–PD-1 therapy and found that alpha diversity, or within-sample diversity, of the gut microbiome was significantly higher in R (
n = 30) compared to NR (
n = 13) using several indices (
P < 0.01;
Fig. 1C and fig. S5). No significant differences were observed in the oral microbiome (R = 54, NR = 32,
P = 0.11; fig. S6). We then tested the relationship of diversity and progression-free survival (PFS) in our cohort by stratifying patients based on tertiles of inverse Simpson scores, demonstrating that patients in the highest tertile of fecal alpha diversity had significantly prolonged PFS compared to those with intermediate or low diversity (
P = 0.02 and 0.04, respectively;
Fig. 1, D and E, and fig. S7). No differences in PFS were noted when comparing diversity of the oral microbiome (fig. S8). Importantly, upon visualizing beta diversity, or between-sample diversity, with weighted UniFrac distances (
30) by principal coordinate analysis, we found a notable clustering effect by response status in the gut microbiome of these patients, which was not observed in the oral microbiome (
Fig. 1F and fig. S8E).
Because compositional differences in the microbiome may also influence cancer development and response to therapy (
12,
14,
15,
23), we sought to determine if differences existed in the oral or gut microbiomes of R and NR to anti–PD-1 therapy. To test this, we first compared an enrichment of operational taxonomic units (OTUs) in R versus NR, demonstrating that distinct sets of rare low abundance OTUs were associated with response to anti–PD-1 therapy, with enrichment of orders Clostridiales in R and Bacteroidales in NR in the gut microbiome (
P < 0.01;
Fig. 2, A and B, and fig. S9, A and C). No significant differences in enrichment were noted in the oral microbiome of R versus NR (fig. S9, B and D, and fig. S10). To further explore these findings, we performed high-dimensional class comparisons using linear discriminant analysis of effect size (LEfSe) (
31), which again demonstrated differentially abundant bacteria in the fecal microbiome of R versus NR in response to anti–PD-1 therapy, with Clostridiales order and Ruminococcaceae family enriched in R and Bacteroidales order enriched in NR (
Fig. 2, C and D). No major differences were observed in the oral microbiome between R and NR, with the exception of higher Bacteroidales in NR in response to anti–PD-1 therapy (fig. S11). Pairwise comparisons were then performed for bacterial taxa at all levels by response. In addition to confirming the previous taxonomic differences, these analyses identified the
Faecalibacterium genus as significantly enriched in R (
Fig. 2E and table S3). Metagenomic WGS sequencing further confirmed enrichment of
Faecalibacterium species in addition to others in R, whereas
Bacteroides thetaiotaomicron,
Escherichia coli, and
Anaerotruncus colihominis were enriched in NR (
Fig. 2F and table S4). Importantly, the gut microbiome was shown to be relatively stable over time in a limited number of longitudinal samples tested (fig. S12).
We next asked whether bacterial composition and abundances within the gut and/or oral microbiomes of patients were associated with a specific treatment outcome to anti–PD-1 therapy. We grouped all identified OTUs into clusters of related OTUs (crOTUs) by means of construction of a phylogenetic tree from sequence-alignment data (
32). This technique involves comparison of abundances of different potential groupings of bacteria based on 16
S sequence similarity and helps address the sparse distribution of OTU abundances observed in the absence of this approach (fig. S13). Unsupervised hierarchical clustering of crOTU abundances within the gut and oral microbiomes was then performed without input of response data. We found that patients segregated into two distinct community types. Type 1 was composed entirely of R and was enriched for Clostridiales, whereas type 2 comprised a mixture of R and NR (
P = 0.02) and was enriched for Bacteroidales (
Fig. 3A). To better understand compositional differences between these crOTU community types, we again performed pairwise comparisons of the gut microbiota and identified a pattern very similar to that seen when clustering by response, with Clostridiales and Ruminococcaceae enriched in type 1, and Bacteroidales enriched in type 2 (fig. S14A and table S5). Further, these communities clustered distinctly using principal coordinate analysis of weighted UniFrac distances (fig. S14B). Analysis of crOTUs in the oral microbiome revealed no apparent relationship to treatment response (fig. S15, A and B).
To explore how specific bacterial taxa affect patient treatment response, we compared PFS following anti–PD-1 therapy as it related to the “top hits” that were consistently observed across our analyses. From the Ruminococcaceae family of the Clostridiales order, we focused on the
Faecalibacterium genus in R and Bacteroidales order in NR and stratified patients into high versus low categories on the basis of the median relative abundance of these taxa in the gut microbiome. Patients with high
Faecalibacterium abundance had a significantly prolonged PFS versus those with a low abundance (
P = 0.03). Conversely, patients with a high abundance of Bacteroidales had a shortened PFS compared to that of those with a low abundance (
P = 0.05,
Fig. 3D). This is in line with recently published data on CTLA-4 blockade, where patients with a higher abundance of
Faecalibacterium had a prolonged PFS compared to those with a higher abundance of Bacteroidales in the gut microbiome (
33)
. In addition, univariate Cox proportional hazards analyses demonstrated that the strongest microbial predictors of response to anti–PD-1 therapy were alpha diversity [intermediate hazard ratio (HR) = 3.60, 95% CI = 1.02 to 12.74; low HR = 3.57, 95% CI = 1.02 to 12.52] and abundance of
Faecalibacterium (HR = 2.92, 95% CI = 1.08 to 7.89) and Bacteroidales (HR = 0.39, 95% CI = 0.15 to 1.03) in the fecal microbiome. Our final multivariate model was selected by forward stepwise selection and included
Faecalibacterium abundance (HR = 2.95, 95% CI = 1.31 to 7.29,
P = 0.03) and prior immunotherapy (HR = 2.87, 95% CI = 1.10 to 7.89,
P = 0.03) (table S6). Abundance of
Faecalibacterium and Bacteroidales also outperformed relevant clinical variables in receiver operating characteristic curve (ROC) analysis (fig. S16).
Next, we sought to gain insight into the mechanism through which the gut microbiome may influence response to anti–PD-1 therapy. We first conducted functional genomic profiling of gut microbiome samples using metagenomic WGS sequencing (
n = 25) in R (
n = 14) versus NR (
n = 11). Organism-specific gene hits were assigned to the Kyoto Encyclopedia of Genes and Genomes(KEGG) orthology (KO), and, on the basis of these annotations, metagenomes for each sample were reconstructed into metabolic pathways using the MetaCyc hierarchy of pathway classifications (
34,
35). Unsupervised hierarchical clustering of predicted pathway enrichment identified two groups of patient samples, with response rates of 69.2 and 41.7% (
Fig. 3E). A similar pattern was also noted for KO abundances with 70.6 and 37.5% response rates (fig. S17). Comparisons of pathway enrichment across these groups showed differences in metabolic functions, with anabolic functions predominating in R—including amino acid biosynthesis (
Fig. 3E), which may promote host immunity (
36)—and catabolic functions predominating in NR (
Fig. 3E, fig. S16, and table S7).
There is clear evidence in preclinical models that differential composition of the gut microbiome may influence therapeutic responses to anti–PD-1 therapy at the level of the tumor microenvironment (
12); thus, we next examined the relationship between the gut microbiota and systemic and antitumor immune responses in our cohort of patients on anti–PD-1 therapy. We compared the tumor-associated immune infiltrates using multiparameter immunohistochemistry (IHC) and observed a higher density of CD8
+ T cells in baseline samples of R versus NR (
P = 0.04), consistent with prior reports (
Fig. 4A and fig. S18) (
18,
37). Pairwise comparisons using Spearman rank correlations were then performed between specific bacterial taxa enriched in the gut microbiome of R and NR and immune markers in the tumor microenvironment, demonstrating a statistically significant positive correlation between the CD8
+ T cell infiltrate in the tumor and abundance of the
Faecalibacterium genus, the Ruminococcaceae family, and the Clostridiales order in the gut and a nonsignificant but negative correlation with Bacteroidales (
Fig. 4, B and C, and figs. S19 and S20). No associations were seen between CD8
+ T cell density and diversity or crOTU community type membership (fig. S21). Analysis of systemic immune responses using flow cytometry and cytokine assays revealed that patients with a high abundance of Clostridiales, Ruminococcaceae, or
Faecalibacterium in the gut had higher frequencies of effector CD4
+ and CD8
+ T cells in the systemic circulation with a preserved cytokine response to anti–PD-1 therapy, whereas patients with a higher abundance of Bacteroidales in the gut microbiome had higher frequencies of regulatory T cells (T
regs) and myeloid-derived suppressor cells (MDSCs) in the systemic circulation, with a blunted cytokine response (
Fig. 4D and figs. S22 and S23). To better understand the influence of compositional differences in the gut microbiome on antigen processing and presentation within the tumor microenvironment, we next performed multiplex IHC targeting the myeloid compartment (
38). In these studies, patients with a high abundance of
Faecalibacterium in the gut microbiome had a higher density of immune cells and markers of antigen processing and presentation compared to that of patients with a high abundance of Bacteroidales (
Fig. 4, E and F, and figs. S24 and S25), suggesting a possible mechanism through which the gut microbiome may modulate antitumor immune responses (
12), though this must be validated in a larger cohort.
To investigate a causal link between a “favorable” gut microbiome and response to immune checkpoint blockade, we performed fecal microbiota transplantation (FMT) experiments in germ-free recipient mice (
Fig. 4G). In these studies, mice that were transplanted with stool from responders to anti–PD-1 therapy (R-FMT) had significantly reduced tumor size (
P = 0.04;
Fig. 4H and fig. S26A) by day 14 compared to those transplanted with stool from NR (NR-FMT). Importantly, mice transplanted with R-FMT stool also exhibited improved responses to anti–PD-L1 (PD-1 ligand 1) therapy (
Fig. 4I) in contrast to mice that were transplanted with stool from NR (NR-FMT). Next, we performed 16S sequencing on fecal samples collected from mice treated with FMT, demonstrating that mice transplanted with R-FMT stool also had significantly higher abundance of
Faecalibacterium in their gut microbiome (
P < 0.01) (fig. S27). We also wanted to better understand the mechanism through which the gut microbiome may influence systemic and antitumor immune responses, and so we performed correlative studies on tumors, peripheral blood, and spleens from these mice. These studies demonstrated that tumors of mice receiving R-FMT stool had a higher density of CD8
+ T cells than mice receiving NR-FMT, consistent with human data (
Fig. 4J and fig. S26B, top series). Analysis of CD45
+ myeloid and lymphoid tumor-infiltrating cells by flow cytometry confirmed this result (fig. S26C). Moreover, FMT from R locally increased the number of CD45
+ immune and CD8
+ T cells in the gut compared to FMT from NR (
Fig. 4K and fig. S26B, bottom series). Mass cytometry analysis using t-distributed stochastic neighbor embedding dimension reduction was performed on tumors from mice and demonstrated up-regulation of PD-L1 in the tumor microenvironment of mice receiving R-FMT versus NR-FMT stool (fig. S26D), suggesting the development of a “hot” tumor microenvironment. Further phenotypic studies of tumor immune infiltrates revealed a significant enrichment of innate effector cells (expressing CD45
+CD11b
+Ly6G
+) in mice receiving R-FMT stool (fig. S26E). A lower frequency of suppressive myeloid cells (expressing CD11b
+CD11c
+) was observed in mice receiving R-FMT stool compared to that of mice receiving NR-FMT (fig. S26F). Finally, an increase in the frequency of RORγT
+ T helper 17 cells in the tumor was also detected in mice transplanted with NR-FMT stool (fig. S26G), in line with what we observed in tumors from patients who failed to respond to anti–PD-1 therapy. Mice receiving NR-FMT stool also had higher frequencies of regulatory CD4
+FoxP3
+ T cells (fig. S26H) and CD4
+IL-17
+ T cells (fig. S26I) in the spleen, suggesting impaired host immune responses.
Our results indicate that the gut microbiome may modulate responses to anti–PD-1 immunotherapy in melanoma patients. We propose that patients with a favorable gut microbiome (for example, high diversity and abundance of Ruminococcaceae and Faecalibacterium) have enhanced systemic and antitumor immune responses mediated by increased antigen presentation and improved effector T cell function in the periphery and the tumor microenvironment. By contrast, patients with an unfavorable gut microbiome (for example, low diversity and high relative abundance of Bacteroidales) have impaired systemic and antitumor immune responses mediated by limited intratumoral lymphoid and myeloid infiltration and weakened antigen presentation capacity. These findings highlight the therapeutic potential of modulating the gut microbiome in patients receiving checkpoint blockade immunotherapy and warrant prompt evaluation in cancer patients through clinical trials.
Acknowledgments
The data reported in this paper are tabulated in the main text and supplementary materials. The authors wish to acknowledge all patients and families affected by melanoma. J.A.W. is supported by the Binational Science Foundation, the Melanoma Research Alliance, Stand Up To Cancer, an MD Anderson Cancer Center Multidisciplinary Research Program Grant, and MD Anderson Cancer Center’s Melanoma Moon Shots Program. This project was supported by the generous philanthropic contributions to The University of Texas MD Anderson Cancer Center’s Melanoma Moon Shots program. J.A.W., P.S., and J.P.A. are members of the Parker Institute for Cancer Immunotherapy at MD Anderson Cancer Center. A.R. is supported by the Kimberley Clarke Foundation Award for Scientific Achievement provided by the Odyssey Fellowship program at The University of Texas MD Anderson Cancer Center. K.H. is supported by the National Cancer Institute (NCI) of NIH under award numbers CA016672 (principal investigator R. DePinho) and R25CA057730 (principal investigator S. Chang). J.E.L. is supported by philanthropic contributions to the University of Texas MD Anderson Cancer Center’s Melanoma Moon Shots Program, The University of Texas MD Anderson Cancer Center’s Various Donors Melanoma and Skin Cancers Priority Program Fund, the Miriam and Jim Mulva Research Fund, the McCarthy Skin Cancer Research Fund, and the Marit Peterson Fund for Melanoma Research. The authors acknowledge the Miriam and Sheldon G. Adelson Medical Research Foundation for their support of MD Anderson Cancer Center’s Biospecimen Collection team. L.M.C. acknowledges a Stand Up To Cancer–Lustgarten Foundation Pancreatic Cancer Convergence Dream Team Translational Research Grant, support from the NCI of NIH, and the Brenden-Colson Center for Pancreatic Health. T.T. acknowledges the Oregon Clinical and Translational Research Institute from the National Center for Advancing Translational Sciences at the NIH (NIH #UL1TR000128). J.A.W. acknowledges C. Diaz for administrative support. Fecal, oral and murine 16S, and fecal WGS sequencing data are available from the European Nucelotide Archive under accession numbers PRJEB22894, PRJEB22874, PRJEB22895, and PRJEB22893, respectively. Human WGS sequencing data are available from the European Genome-phenome Archive under accession number EGAS00001002698. J.A.W. and V.G. are inventors on a U.S. patent application (PCT/US17/53717) submitted by The University of Texas MD Anderson Cancer Center that covers methods to enhance checkpoint blockade therapy by the microbiome. T.T. and L.M.C. are inventors on a World Intellectual Property Organization patent (WO 2017/087847) held by Oregon Health and Science University that covers the multiplex technology. M.A.D. is an advisory board member for Bristol-Myers Squibb, Novartis, GlaxoSmithKline, Roche/Genentech, Sanofi-Aventis, and Vaccinex and has received funding from GlaxoSmithKline, Roche/Genentech, Merck, AstraZeneca, and Sanofi-Aventis. A.J.L. is a consultant for MedImmune, Bristol-Myers Squibb, Novartis, and Merck and has received research support from AstraZeneca/MedImmune. Z.A.C. is an employee of MedImmune and owns stock or options in AstraZeneca. J.E.G. is on the advisory board of Merck and receives royalties from Mercator Therapeutics. S.P.P. has honoraria from Speaker’s bureau of Dava Oncology, Merck, and Bristol-Myers Squibb and is an advisory board member for Amgen and Roche/Genentech. P.H. serves on the advisory board of Lion Biotechnologies and Immatics US. R.N.A. has received research support from Merck, Novartis, and Bristol-Myers Squibb. P.S. is a consultant for Bristol-Myers Squibb, Jounce Therapeutics, Helsinn, and GlaxoSmithKline and is also a stockholder from Jounce Therapeutics. J.P.A. is a consultant and stockholder for Jounce Therapeutics, receives royalties from Bristol-Myers Squibb, and has intellectual property with Bristol-Myers Squibb and Merck. J.A.W. has received honoraria from Speakers’ bureau of Dava Oncology, Bristol-Myers Squibb, and Illumina and is an advisory board member for GlaxoSmithKline, Novartis, and Roche/Genentech. The other authors declare no competing interests.