Randomized Placebo-Controlled, Biomarker-Stratified Phase Ib Microbiome Modulation in Melanoma: Impact of Antibiotic Preconditioning on Microbiome and Immunity

Although the advent of immune-checkpoint blockade (ICB) in the treatment of both metastatic and locally advanced melanoma has dramatically improved outcomes (1–4), the challenges of patient selection and overcoming resistance have become more relevant than ever. Although recent combination immunotherapy trials continue to show incremental improvements, with complete response (CR) rates of 22% and 27% from nivolumab plus ipilimumab (5) and nivolumab plus relatlimab (6), respectively, but there remains a strong need for modulation strategies of host factors to enhance response or overcome resistance. The importance of the gut microbiome in solid tumor therapy, particularly as it pertains to ICB response, has been most well demonstrated in the setting of advanced melanoma. Early seminal work in both preclinical models and clinical cohorts demonstrated differences in the taxonomic composition of the gut microbiome and dietary fiber intake that correlate with both response to ICB and immune-related toxicity (7–11). Several reports specifically identified certain spore-forming taxa, such as those in the Firmicutes phyla and the Ruminococcaceae family, to be associated with improved response (8, 12).

Stemming from these observational studies validated in preclinical models, recent work has also focused on microbiome modulation as a possible modifiable mechanism to overcome ICB resistance, most notably with fecal microbiota transplantation (FMT); FMT with stool from donors who had responded to immunotherapy demonstrated the ability to promote response and overcome preexisting resistance to ICB therapy in melanoma (13, 14).

SER-401 (Seres Therapeutics) is a proprietary formulation of bacterial Firmicutes spores that are fractionated and purified from the human stool of a healthy donor, whose fecal microbiome embodied the taxonomic communities characterized as favorable for response to anti–PD-1 therapy (8). SER-401 contained a high proportion of spore products from the Ruminococcaceae family, with the hypothesis that treatment with SER-401 prior to ICB initiation may increase Ruminococcaceae abundance and potentially lead to improved response. Similar formulations of donor-derived products have been rigorously evaluated in human randomized clinical trials, most notably SER-109 for the prevention of recurrent Clostridioides difficile infections (15). Utilizing a similar Firmicutes-spore–based product, treatment with SER-109 after standard antibiotic therapy for C. difficile infection (majority with oral vancomycin) was well tolerated with a similar adverse event (AE) rate compared with the placebo arm; importantly, patients in the SER-109 arm had a reduced incidence of recurrent C. difficile infection (12% compared with 40%).

Although such interventions have had some success in the treatment of infectious or inflammatory conditions, there have not yet been any published prospective, randomized, placebo-controlled trials utilizing microbiome modulation prior to ICB (particularly in treatment-naïve patients with advanced melanoma). In this context, this multicenter phase Ib trial across five tertiary centers in the United States was designed to evaluate the safety and tolerability of treatment with oral microbiome intervention (SER-401) versus matched placebo in combination with nivolumab, in participants with metastatic or unresectable melanoma without previous ICB treatment. Given the previously observed correlation of a higher abundance of Ruminococcaceae family taxa in ICB responders in the setting of melanoma (8), there was prospective patient stratification of high versus low baseline Ruminococcaceae abundance utilizing quantitative PCR (qPCR) of patient stool within each arm (using a threshold of 16% relative abundance when comparing the qPCR signal of Ruminococcaceae-specific genetic elements against the total bacterial content). The primary endpoints were incidence and severity of AEs associated with the trial regimen. Although the study design was not powered to detect differences in efficacy between the treatment and placebo arms, secondary endpoints included objective response rates (ORR), progression-free survival (PFS), overall survival (OS), and change in tumor CD8 T-cell infiltration, as well as correlative longitudinal exploratory analyses related to the microbiome and systemic inflammation.

The initial study design contained four intervention arms; the two SER-401 arms were to receive vancomycin preconditioning followed by SER-401 and then nivolumab with SER-401 maintenance dosing, compared with placebo preconditioning followed by placebo microbiome intervention and nivolumab with placebo maintenance. The FMT arms were to receive vancomycin preconditioning followed by FMT and subsequent nivolumab, compared with placebo preconditioning followed by nivolumab (refer to Supplementary Fig. S1A for intended schema).

Given the constraints of limited accrual due to the COVID-19 pandemic during the study enrollment time, the trial was discontinued prior to assigning any participants to the FMT arms, leading to a two-arm placebo-controlled design (Fig. 1A).

From February 18, 2019, until the primary study completion date of March 4, 2022, 14 patients in total were accrued to the microbiome intervention arm (Supplementary Fig. S1B, CONSORT diagram). Baseline patient characteristics are listed in Supplementary Table S1. The median age was 64 (range, 26–84 years), and the majority were male (10, 71%) with cutaneous subtype (12, 86%). Three patients (21%) were stage IIIc at the time of enrollment, whereas the majority (11, 79%) were stage IV. Of the 14 patients accrued, 8 (57%) were randomized to the vancomycin + SER-401 + nivolumab arm [5 (63%) with low and 3 (38%) with high baseline Ruminococcaceae abundance], whereas 6 (43%) were randomized to the placebo + nivolumab arm [4 (67%) with low and 2 (33%) with high baseline Ruminococcaceae abundance]. There were no clear clinicopathologic differences in patients classified as Ruminococcaceae high versus low in the baseline gut microbiota, with the exception of higher median age in the group of patients with high Ruminococcaceae (P = 0.02 by Wilcoxon rank sum test; Supplementary Table S2). In the SER-401 arm, 4 patients received prior adjuvant immunotherapy (more than 12 weeks prior to enrollment); in the placebo arm, 1 patient received prior adjuvant immunotherapy.

Safety analysis (Table 1A) was based on all participants who received at least 1 dose of any study intervention (vancomycin/placebo, SER-401/placebo, and nivolumab). Serious AEs (SAE) were reported in 2 patients (14%); in the SER-401 arm, 1 patient experienced grade 4 primary adrenal insufficiency (patient 102-S0005, who experienced primary adrenal insufficiency only after the trial period had ended following initiation of combination immunotherapy as a standard of care, see Supplementary Table S3). Both SAEs were designated unrelated to the study drug. In the placebo arm, 1 patient experienced grade 3 dehydration, duodenitis, and gastritis related to nivolumab initiation, leading to their withdrawal from the trial. Treatment-related adverse events (TRAE) of any grade were reported in 4 of 8 patients in the SER-401 arm (50%) and 5 of 6 patients in the placebo arm (83%). There were no grade 3 or 4 TRAEs reported in the SER-401 arm; in the placebo arm, 1 patient (17%) reported grade 3 dehydration, and none had grade 4 TRAEs. The full list of all AEs is detailed in Supplementary Table S3.

Among the 14 patients enrolled, the median time on treatment was 5.2 months (range, 1.8–10.6 months) in the SER-401 arm and 10.8 months (range, 6–11.2 months) in the placebo arm. The median cumulative dose of vancomycin or matching placebo was 2 g. Participants received a median of 62.5 doses of SER-401 or a matching placebo (of an expected 63 doses) in both arms; the median cumulative dose was 312.5 × 10⁶ spore colony-forming units (SCFU) in the SER-401 arm. Participants received a median of 6.0 doses of nivolumab in the SER-401 arm and 11.5 doses in the placebo arm (of a maximum of 12 doses).

Although the study was not powered for comparison of efficacy between arms, response and survival were analyzed in the modified intent-to-treat (mITT) population (all 14 patients, Table 1B). A summary of the best-confirmed responses is demonstrated in Fig. 1B. The objective response rate (ORR) was 25.0% [95% confidence interval l (CI), 3.2–65.1] for the SER-401 arm and 66.7% (95% CI, 22.3–95.7) for the placebo arm. The disease control rate (DCR) was 37.5% (95% CI, 8.5–75.6) for the SER-401 arm and 83.3% (95% CI, 35.9–99.6) for the placebo arm (Supplementary Fig. S2). The median duration of response was not reached by data cutoff for both arms (range, 13.6–15.2 months for the SER-401 arm and 5.7–18.3 months for the placebo arm; see Fig. 1C). Of the responders, 3 patients (50%) in the placebo arm and 2 patients (25%) in the SER-401 arm had an ongoing response at the point of trial completion (Fig. 1C; Supplementary Table S4).

The median OS was 21.1 months (95% CI, 9.7–not reached) in the SER-401 arm, with 1 year OS rate of 75% (95% CI, 31.5–93.1); median OS was not reached in the placebo arm, with 1 year OS rate of 100% (95% CI, 100-100; see Fig. 1D). Median PFS was 5.2 months (95% CI, 2.2–not reached) in the SER-401 arm and 15.0 months (95% 3.3–not reached) in the placebo arm (Fig. 1E). There was no statistically significant difference in OS or PFS between the two arms.

To assess the associations of baseline abundance of Ruminococcaceae with clinical outcome, we next assessed response and survival—stratifying patients into high versus low abundance of Ruminococcaceae [based on our prespecified cutoff of 16%, with >16% classified as high (n = 5) and <16% classified as low (n = 9) by qPCR]. In this analysis, we did not observe any statistically significant differences in response or survival based on high versus low Ruminococcaceae at baseline in the gut microbiota (Supplementary Table S5; Supplementary Fig. S3A–S3C); however, this analysis was limited by a small sample size.

Recognizing the limitation of assessing the predictive capability of baseline Ruminococcaceae abundance in this small cohort, we next performed an analysis of response and survival as it related to the relative abundance of Ruminococcaceae at baseline (using the same cutoff of 16% in a previously published cohort of patients with metastatic melanoma treated with ICB (11). In the 114 patients with 16S microbiome sequencing and clinical response data available, we observed a higher response rate and prolonged survival in patients with high (≥16%) versus low (<16%) abundance of Ruminococcaceae (median PFS 991 days in patients with high Ruminococcaceae vs. 342 days in patients with low Ruminococcaceae, P = 0.077 by log-rank test; Supplementary Fig. S4).

Longitudinal stool samples were collected at multiple time points and analyzed with whole metagenome sequencing (WMS); these time points included baseline prior to any intervention, after vancomycin preconditioning but before any SER-401 (D-7), after SER-401 but before any nivolumab (C1D1), and on treatment time points (Fig. 1A). We first verified that the qPCR-derived Ruminococcaceae relative abundance correlated strongly with the baseline Ruminococcaceae abundance calculated by WGS (R² = 0.83, P < 0.001 by Spearman correlation; Supplementary Fig. S5). In the active arm, at D-7 there was a significant shift in the microbiome taxonomic composition after vancomycin preconditioning, with depletion of target families such as Ruminococcaceae (P = 0.008, Wilcoxon signed-rank test), as well as other spore formers (e.g., Bacteroidaceae, Lachnospiraceae, Clostridiaceae; Fig. 2A and B; Supplementary Fig. S6). Although there was recovery of these families at C1D1 after administration of SER-401 when compared with baseline, there was no further increase in abundance of these taxa after C1D1, suggesting suboptimal engraftment of SER-401 species (Fig. 2B; Supplementary Fig. S7A and S7B). The placebo arm did not have any significant changes in composition even after initiation of nivolumab.

When stratifying by baseline Ruminococcaceae abundance, in the low Ruminococcaceae group of the SER-401 arm, there was a trend toward a higher abundance of Ruminococcaceae family species after SER-401 administration at C2D1 compared with baseline (P = 0.06); this trend was not seen in the high Ruminococcaceae group. There were no appreciable trends in Ruminococcaceae abundance in the placebo arm in either high or low Ruminococcaceae groups (Fig. 2C; Supplementary Fig. S8). Of note, the only complete responder (patient 113-S0001, SER-401 arm) had the lowest confirmed baseline Ruminococcaceae abundance (<0.5% at baseline); after SER-401 administration, the Ruminococcaceae abundance increased to 12.5% at C1D1, with subsequent decrease and stabilization at 2.7% at last time point (Fig. 2C; Supplementary Fig. S8).

With respect to overall microbiome diversity, there was a significant decrease in alpha diversity indices compared with baseline at D-7 in the SER-401 arm (after vancomycin preconditioning alone); in contrast, patients who received placebo antibiotic in the placebo arm did not have any changes in alpha diversity (Supplementary Fig. S9A and S9B). The overall structural differences between baseline stool samples and longitudinal trial samples, as determined by the Bray-Curtis dissimilarity, demonstrated a near complete shift at D-7 in the SER-401 arm, followed by a drop in dissimilarity after C1D1 and beyond; in the placebo arm, there was expected low-level variation of composition over time (Fig. 2B; Supplementary Fig. S10A and S10B). There were no statistically significant differences in alpha or beta diversity, or baseline Ruminococcaceae abundance, between the treatment arms at baseline (Supplementary Fig. S11A–S11C). There were shifts in Ruminococcaceae family species after vancomycin preconditioning, and some patients, including the patient with CR, developed numerous new species (going from 2 at baseline to 6 at C7D1 (Supplementary Fig. S12A); there were no such differences noted in the placebo arm (Supplementary Fig. S12B).

Next, the co-occurrence patterns of bacterial taxa were analyzed across patients at each time point in each arm to generate commensal bacterial network maps, which showed detectable disruption of the network after vancomycin preconditioning at D-7 (Supplementary Fig. S13A). Although the taxonomic diversity may have recovered by C1D1 in the SER-401 arm, the degree of commensal network connectedness did not return to baseline until C1D8; there was no obvious change in the networks of the placebo arm across time points (Supplementary Fig. S13B).

Utilizing metagenomic analysis from stool WMS, changes in the imputed functional capacity of the gut microbiota were compared within each arm with respect to baseline stool, focusing on MetaCyc pathways previously found to be associated with ICB response or resistance in melanoma patients (see Supplementary Table S6 for specific pathways; ref. 16). Although there were minimal differences in either response or resistance-associated pathways (16) at baseline between arms, at D-7 (after vancomycin preconditioning) there was a significant increase of multiple ICB resistance–associated pathways and a concurrent decrease in response-associated pathways (such as butyrate biosynthesis; adjusted P = 0.0009). Importantly, at C1D1 (after SER-401 or placebo microbiome intervention but before the first dose of nivolumab), none of the significant differences between arms in both ICB response and resistance-associated pathways persisted (Fig. 2D and E; Supplementary Fig. S14A and S14B).

Longitudinal stool metabolomic analysis demonstrated no differences between arms at baseline via principal component analysis (PCA); however, at D-7, there was a significant separation of the two arms (R² = 0.26, P = 0.001 by PERMANOVA). But similar to the metagenomic analysis, by C1D1 and beyond, these significant differences did not persist (Supplementary Fig. S15). Although the exact relationship between shifts in taxa and metabolic function remains difficult to prove without functional experiments outside of the scope of this project, these data suggest that in this small data set, there were significant gut metabolomic changes after vancomycin preconditioning, which did not persist after SER-401 and before nivolumab treatment.

As part of the translational analysis, we sought to evaluate detectable changes from the gut microbiome intervention related to systemic and antitumor immunity. A secondary endpoint was CD8 infiltration percentage; baseline and on-treatment biopsies (taken at C2D1) were analyzed for CD8 tumor-infiltrating lymphocyte (TIL) infiltration using IHC. There was an increase in CD8% (17.6%–37.8%, P = 0.06; Fig. 3A) when considering all patients. Where paired biopsies were available for the same patient (n = 3), there was a quantifiable increase after treatment in both the placebo arm (63% increase) and the SER-401 arm (27% increase; range, 18–37%, Fig. 3B). The degree of increase was not associated with ORR (Supplementary Fig. S16).

Next, to determine whether microbiome intervention caused a shift in the phenotype of circulating lymphocytes, peripheral blood mononuclear cells (PBMC) across time points were analyzed using high-parameter flow cytometry. This revealed subpopulations of circulating regulatory T-cell (Treg) types that were significantly increased in the SER-401 arm at D-7 from baseline (P < 0.05, vancomycin preconditioning only window), including CD25^hi Tregs expressing PD-1 and TIGIT (Fig. 3C, left). Longitudinally, as compared with baseline by mean fold change, there was a numerical (though not statistically significant) increase in regulatory T cells that persisted into C1D1 across the identified cell populations (Fig. 3C, right), suggesting that antibiotic and SER-401 administration contributed to the increased proportion of PBMC-derived Tregs. Differences were also noted when comparing the peripheral immune population by Ruminococcaceae status and response (Supplementary Fig. S17A–S17D).

To determine whether these increased circulating cellular phenotypes translated to systemic immune functional changes, longitudinal serum proteomics using Olink was performed. At D-7 (after vancomycin but before SER-401 or nivolumab), there was evidence of increased systemic inflammation coupled with greater immune activation in the SER-401 arm compared with the placebo arm. These were driven by pathways in cellular stress/cell death, T-cell receptor (TCR) signaling, and Toll-like receptor signaling [increase in proteins such as CASP8 (17), GRAP2 (18), and IRAK4 (19), respectively]. This change from baseline was no longer detectable after SER-401 at C1D1, though certain peptides such as those noted above did trend higher longitudinally during the treatment in the SER-401 arm (Fig. 3D; Supplementary Fig. S18A AND S18B). Additionally, we observed no significant differences in the frequency of regulatory T cells between patients with high versus low abundance of Ruminococcaceae at baseline (Supplementary Fig. S19).

Although these data remain exploratory in nature without the power to draw true intergroup comparisons, they suggest parallels between microbiome, metabolomic, and immune features in this small group of patients. They also demonstrate the importance of a deeper and nuanced correlative analysis beyond taxonomic diversity or TIL infiltration, to understand the modulators of microbiome-mediated ICB response.

This is the first-of-its-kind placebo-controlled randomized controlled study with a prospective biomarker stratification (baseline Ruminococcaceae abundance in stool microbiome), testing microbiome intervention prior to initiation of ICB in advanced melanoma patients. Although the trial was severely limited by the challenges in accrual during the COVID pandemic, vancomycin preconditioning followed by microbiome intervention with SER-401 and subsequent nivolumab treatment was well tolerated, and only one grade 3 or higher serious adverse event (adrenal insufficiency, deemed by the site investigator to be unrelated to the initiation of treatment) was identified. In comparison, 1 patient in the placebo arm developed grade 3 gastritis after the initiation of nivolumab which led to withdrawal from the study. These toxicities resolved in both patients before the end of the trial data collection period.

There were difficulties with accrual from the COVID pandemic, which affected our ability to open the originally planned FMT arm; however, even among the four additional patients who were eventually able to be enrolled during the pandemic in 2020, no COVID-related events were reported through the study period. Difficulties with accrual from the COVID pandemic notwithstanding, our study was still able to enroll and randomize patients from five different tertiary melanoma treatment centers across the United States. Furthermore, prospective stratification by qPCR of baseline stool into high versus low Ruminococcaceae abundance groups prior to ICB initiation was feasible and accurate, as evidenced by the raw Ruminococcaceae abundance by WMS (Supplementary Fig. S8). Although it was not powered to test the efficacy of response and definitive conclusions about efficacy and the impact of vancomycin preconditioning and SER-401 are challenging to draw, the trial demonstrates the feasibility of such a biomarker-driven investigation—something that will be ever in more need as microbiome interventions seek the transition from translational studies to the clinical realm.

Prior to this study, there has not been a randomized, in-depth prospective effort into the impact on the gut microbiome of not only the microbiome intervention but also the preconditioning antibiotic regimen in patients with cancer, either in the form of FMT or other consortia-based approach. In addition, although others studies have focused on anti–PD-1 refractory melanoma patients, our study only enrolled patients who were anti–PD-1 naïve, undergoing treatment in the first-line setting. In our study, the preconditioning with oral vancomycin starting 1 week prior to initiation of ICB had a detectable impact on not only the gut microbiome but also the metabolomic function from the gut prior to initiating ICB. Via metagenomic analysis of the stool microbiome, imputed pathways previously published by Simpson and colleagues (16) were utilized to test the abundance of ICB response–associated pathways and ICB resistance–associated pathways. At D-7 in the SER-401 arm (after vancomycin preconditioning), there was a significant decrease in alpha diversity (Supplementary Fig. S9B); concurrently, there were decreases in certain response-associated pathways (most notably butyrate biosynthesis), with a significant increase in multiple resistance-associated pathways (Fig. 2D and E). This shift in the gut functional landscape at D-7 was also seen in the stool metabolomic profiling (Supplementary Fig. S15). These changes did not persist at C1D1 after SER-401 treatment and before nivolumab initiation. Importantly, we verified that the baseline fundamental characterizations of diversity metrics between patients in each arm were not statistically different, in both alpha and beta diversity (Supplementary Fig. S11A–S11C).

The timing of the gut functional change also coincided with detectable increases in measures of systemic inflammation, namely, in serum proteomics and circulating flow cytometry. At D-7, serum proteins involved in cellular stress and death (such as CASP8 and GRAP2) were differentially increased in the SER-401 compared with the placebo samples. There was also a concurrent increase in circulating regulatory T-cell types compared with baseline at D-7 (vancomycin preconditioning only window) in the SER-401 arm, but not in the placebo arm. Given the timing, these findings of systemic immune inflammation seen at D-7, after vancomycin preconditioning but before SER-401 or nivolumab, could be related to the gut microbiome changes given the very limited bioavailability of systemic vancomycin after oral administration. Although these gut metabolomic or systemic immune changes did not persist at C1D1 after SER-401 treatment before nivolumab initiation, they highlight the importance of the preconditioning regimen choice, duration, and interval prior to ICB initiation (beyond the microbiome intervention itself) in optimizing response.

Although the study is not powered to compare response rates between the two arms, it is possible that the impact of vancomycin preconditioning prior to nivolumab initiation may have contributed to the lower ORR (25% in the SER-401 arm, compared with 67% in the placebo arm). However, there may have been other confounding factors in the randomization of the small number of enrolled patients that led to the lower ORR rate. For example, there was a higher proportion of stage III patients in the placebo group (33% vs. 12.5%), and stage IV M1c-d patients comprised only 17% of the placebo arm, compared with 50% of the SER-401 arm. Furthermore, the lack of detailed dietary information in patients enrolled in our study represents a limitation, particularly considering the significant impact of diet on the response of the gut microbiome to immunotherapy, as emphasized in our previous work (11). Future clinical trials should incorporate an assessment of dietary patterns in enrolled patients and should also consider dietary guidelines for patients treated with gut microbiome modulation and immunotherapy.

A component of the trial design was to ensure that the placebo and SER-401 arms were similar in their baseline makeup in terms of the relative abundance of Ruminococcaceae, a family of spore-forming bacteria shown to favorably modulate response to immunotherapy (8, 12, 20). There was no difference in the proportion of patients with high versus low Ruminococcaceae abundance between arms (Supplementary Table S1), and the baseline abundance category did not appear to impact ORR (Supplementary Fig. S20). Notably, there was only a trend toward increased Ruminococcaceae abundance compared with baseline in the low baseline Ruminococcaceae group treated with SER-401 after vancomycin preconditioning. In fact, the only patient with CR had the lowest baseline Ruminococcaceae abundance and saw a detectable and persistent increase in Ruminococcaceae abundance after C1D1. Although direct correlations between the degree of Ruminococcaceae increase and ORR were difficult to measure given the low sample size, this suggests the potential importance of personalizing and tailoring microbiome interventions in such prospective biomarker-driven precision trials (which requires the development and optimization of rapid companion diagnostics).

There were many limitations to the study; at the initial study design, 60 patients were planned to be enrolled into the doublet randomized arms of the SER-401 and also the FMT groups. Due to the COVID pandemic, which started after study initiation, the FMT arm never started enrollment. As the current “gold standard” for microbiome intervention, FMT from melanoma patient donors who have previously responded to immunotherapy would have served as an interesting and valuable control for microbiome modulation compared with SER-401. It would also have served to validate work previously published on responder FMT increasing sensitivity to immunotherapy (13, 14), but in the treatmentnaïve and placebo-controlled setting.

The other control that was lacking in the study was the preconditioning regimen and duration. Recently, there have been studies, published after this trial was initiated, demonstrating the negative impact of antibiotic use prior to ICB initiation on overall response (10, 21). Although the choice of vancomycin was driven by known sensitivity associations relating to the presence of spore-forming gut microbes (such as those in the Ruminococcaceae family; ref. 22), it is possible that the disruption of the microbiome and its potential lasting metabolic effects would have been ameliorated by the SER-401 microbiome intervention better if there was a slightly narrower antibiotic regimen utilized (or longer interval from ICB treatment).

It should be noted that there have been other more recently published microbiome intervention trials (within the context of enhancing immunotherapy response) that did not utilize any antibiotics in their preparative regimen prior to intervention. A seminal study that demonstrated the potential of FMT for promoting response in ICB-refractory patients was performed without the use of a preparative regimen prior to FMT and ICB (14). A more recent nonrandomized trial of responder donor FMT enrolled 20 patients with advanced melanoma who were ICB-naïve (NCT03772899); they underwent nonantibiotic bowel preparation with 4 liters of polyethylene glycol without any antibiotic the night before FMT capsule administration (1 week prior to ICB initiation; ref. 23). They demonstrated ORR of 65%, with 10% achieving CR; ICB therapy was discontinued in 10% of patients due to toxicity. Administration of single bacterial strains rather than whole FMT has also been associated with improved outcomes in certain patient cohorts (24–26). This includes a retrospective analysis of patients with non–small cell lung cancer who were treated with ICB and demonstrated clinical benefit (with improved response and prolonged PFS) in patients who received CBM588 (a strain used in a commonly prescribed probiotic in Japan). This was particularly evident in patients who received treatment with antibiotics (26). In contrast, a randomized (but not placebo-controlled) phase I trial testing CBM588 in the setting of advanced renal cell carcinoma patients undergoing nivolumab/ipilimumab treatment also did not utilize preparative antibiotics; this demonstrated improved PFS in patients receiving CBM588 concurrently with ICB (NCT03829111; ref. 27). These data suggest that a preparative antibiotic regimen to clear space in the gut microbiota may not be as necessary within the context of microbiome modulation for cancer immunotherapy.

Based on the available data from this study, the vancomycin preconditioning regimen and subsequent washout period given prior to administration of SER-401 appears to have had a deleterious impact on the gut microbiota and markers of systemic immunity and inflammation. This could be potentially due to disruption of the microbiome following antibiotic preconditioning (before administration of SER-401) or due to the incomplete recovery and suboptimal restructuring of the microbiome in a subset of patients prior to the administration of ICB. Notably, SER-401 was administered after a 2- to 3-day delay to enable washout of vancomycin accumulated in the GI tract. This washout delay has not been used in other studies of spore-forming Firmicute products and may have contributed to the suboptimal results. These findings have important implications for the design of future clinical trials and strategies to modulate gut microbes in the setting of treatment with immunotherapy for cancer. Antibiotic and other preconditioning regimens should be carefully considered (and ideally tested prospectively) alongside testing of optimal strategies to modulate gut microbes to improve immunotherapy response.

Taken together, this study demonstrates the feasibility of a multicenter placebo-controlled randomized trial of microbiome intervention with prospective biomarker stratification in the setting of a preparative regimen for ICB. It highlights the importance of controlling for the baseline microbiome of each patient, as well as the impact of both the preparative preconditioning regimen and microbiome intervention itself, in future clinical designs aimed at harnessing the true potential of microbiome modulation in enhancing immunotherapy response.

The study (NCT03817125) was approved by the US Food and Drug Administration and according to the Institutional Review Board at each of the 5 tertiary treatment centers. The study protocol is provided in the Supplementary Appendix. This study enrolled participants ages 18 years or older with histologically confirmed stage III cutaneous, acral, or mucosal melanoma that was judged inoperable or stage IV cutaneous melanoma. Additional eligibility criteria included the following: participants were allowed to participate in this study who had received prior radiotherapy at least 2 weeks prior to study intervention administration; participants must have been naïve to anti–programmed cell death ligand 1 (PD-L1) or anti–PD-1 agents except for BRAF-targeted therapy in the metastatic setting or anti–cytotoxic T-lymphocyte associated protein-4 (CTLA-4) therapy in the adjuvant setting; and measurable disease as defined by Response Evaluation Criteria in Solid Tumors (RECIST) v1.1.

Written informed consent was obtained from enrolled patients. This study was conducted in accordance with the protocol and with the consensus ethical principles derived from international guidelines, including the Declaration of Helsinki and the Council for International Organizations of Medical Sciences International Ethical Guidelines.

Study investigators, site personnel, and participants were blinded to the assignment of antibiotics and microbiome study intervention throughout the course of the study. Nivolumab was administered open label to all study intervention arms. Of note, patients who reported the use of antibiotics within 30 days prior to randomization or indicated a planned or required need for antibiotic prophylaxis for more than 24 consecutive hours during the study were not eligible. The use of probiotics was not allowed and was listed under prohibited concomitant therapy.

A 2-week antibiotic/microbiome study intervention lead-in phase consisted of 4 days of pretreatment with oral vancomycin (125 mg) or placebo administered 4 times daily, starting on day −14, followed by a 2 (+1) day washout. The washout may have been extended by an additional day if needed to accommodate scheduling. Participants unable to tolerate the antibiotic pretreatment were discontinued from the study and replaced prior to the administration of the microbiome/ nivolumab combination.

A 1-week loading regimen of the microbiome study intervention was performed 2 to 3 days following antibiotic/placebo administration. Participants in the SER-401/placebo arm received 7 doses of SER-401 investigational intervention or placebo administered 1 dose every day (q.d.) for 1 week. The first dose of SER-401 or matching placebo was administered in the clinic and was followed by a 1-hour observation period after dosing. SER-401 was to be taken at approximately the same time on each day and the administration date was recorded on the diary card. The participant was to observe an overnight or at least a 6-hour fast (no food or drink except for small amounts of water) prior to SER-401 administration. Subsequent doses of SER-401 or matching placebo were dispensed to the participant for self-administration q.d. at home. During the antibiotic/microbiome lead-in phase, participants used a participant-reported diary to track compliance to the oral antibiotic/placebo and SER-401/placebo, as well as to record, on a daily basis during administration, the occurrence of solicited symptoms commonly associated with microbiome study intervention.

Nivolumab 480 mg every 4 weeks (Q4W) therapy for all participants began on cycle 1 day 1 (C1D1), which was approximately one day following the 7 days of lead-in with the microbiome intervention/placebo. A nivolumab study intervention cycle was defined as 4 calendar weeks, starting on the day of each nivolumab infusion. The protocol active treatment phase with nivolumab continued for up to 12 cycles, unless the participant experienced confirmed disease progression or unacceptable toxicity in the judgment of the treating physician. During cycles 1 and 2, participants received one daily dose of microbiome study intervention (SER-401 or placebo) in addition to nivolumab. Beginning at cycle 3, all participants only received nivolumab. Participants were to be followed for approximately 2 years from the time of the initiation of the study intervention.

Blood and stool biomarkers were collected throughout the study. Participants were supplied with home fecal collection kits to obtain stool samples for metagenomic and metabolomic analyses. Following randomization, stool was collected with a window of up to 2 days prior to the following time points, when feasible: start of microbiome study intervention (i.e., collected after completion of antibiotic/ placebo and prior to the start of microbiome study intervention) and C1D1 (i.e., 7 days after start of microbiome study intervention). Thereafter, stool was collected with a window of ± 3 days of the following time points, when feasible: cycle 1 day 8 (C1D8; i.e., 7 days after the first dose of nivolumab); cycle 2 day 1 (C2D1); cycle 3 day 1 (C3D1); cycle 4 day 1 (C4D1; i.e., 4 weeks after last dose of microbiome study intervention); cycle 7 day 1 (C7D1); and at the end of treatment visit.

The efficacy analysis was performed on the mITT population, which comprised all randomized participants who received at least 1 dose of any study intervention (vancomycin/placebo, SER-401/ placebo, and nivolumab). Participants were analyzed according to the study interventions to which they were allocated.

There were no primary efficacy objectives or endpoints for this study. Response rates for ORR and DCR were estimated within each study intervention arm and 95% CIs were estimated using the Clopper–Pearson method. PFS and OS were estimated using Kaplan–Meier techniques, and the median survival time and 95% CIs were estimated within each study intervention arm. In addition, the 1-year PFS and OS rates were estimated within each study intervention arm and 95% CIs were estimated using the Clopper–Pearson method. For participants who experienced an objective response [CR or partial response (PR)] during the study, duration of response was defined as the time from the first tumor assessment that documented response (CR or PR, whichever was recorded first) to the first documentation of radiographic progressive disease per RECIST v. 1.1 or death due to any cause, whichever occurred first. The same clinical outcomes analysis was stratified for subset analysis using the prospective qPCRderived baseline Ruminococcaceae family relative abundance using the previously defined cutoff of 16%.

Although this study was not powered to assess differences in groups, baseline patient characteristics were compared by arm and baseline Ruminococcaceae status (qPCR) to aid in the interpretation of study findings. Statistical differences between groups were assessed by the Fisher exact test (categorical variables) and Wilcoxon rank sum test (continuous variables) using a threshold of P < 0.05.

Safety was based on the safety population, which consisted of all participants who received at least 1 dose of any study intervention (vancomycin/placebo, SER-401/placebo, and nivolumab). Participants were analyzed according to the study interventions received. Safety was assessed through summaries of AEs, clinical laboratory test results, vital signs, electrocardiograms (ECG), and Eastern Cooperative Oncology Group performance status.

In the retrospective cohort analysis, we included an initial total of 217 advanced melanoma patients (with baseline 16S gut microbiome sequencing) from a cohort previously published (11). These data sets were pulled into a phyloseq object, and the data set was rarified to 5,000 reads per sample. After rarefaction, 214 samples remained; these data were agglomerated at the bacterial family level, and relative abundance was calculated. After filtering for receipt of ICB and available PFS data, 114 patient samples remained. This cohort was split by a cutoff of 16% relative abundance of Ruminococcaceae family bacteria, into Ruminococcaceae high (>16%) and Ruminococcaceae low (less than or equal to 16%) groups. Survival analysis was performed using the survfit function from the survival R package. The log-rank test was performed to compare PFS.

Fecal samples were collected from participants at 8 time points for assessment of SER-401 engraftment using WMS: baseline, D-7, D1, C1D1, C1D8, C2D1, C3D1, C4D1, and C7D1.

WMS was performed on all available samples as a measure of drug species engraftment. We quantify engraftment, which is a measure of SER-401 pharmacokinetics, based on WMS of both participant stool samples and SER-401 investigational drug product. Engraftment of SER-401 was assessed as a secondary endpoint. Taxonomic profiling of WMS data for both SER-401 and participant samples was performed using the MetaPhlAn4 software (28), which additionally includes taxonomic markers of custom-curated species of interest (involving isolation and functional screening of strains of interest) not available in the public version of the database. The MetaPhlAn4 method involves mapping filtered sequencing reads to a set of species-specific genomic markers, followed by an estimation of the relative abundance of each species based on the distribution of mapped reads. Raw WMS sequence reads were preprocessed to remove human sequences following methods used in the HMP (The Human Microbiome Project Consortium, 2012) and to filter out technical artifacts before mapping.

Metagenomic imputed functional profiling was completed using the HUMAnN 3.6 in order to profile the presence/absence and abundance of microbial pathways, mapped to genes known to be associated with the pathways, from a community of metagenomic sequencing (http://huttenhower.sph.harvard.edu/humann; RRID: SCR_014620). Microbiome data were imported into R programming environment version 4.2.3 using Phyloseq R package version 1.42.0 to analyze the taxonomic composition data. Alpha diversity was calculated using the inverse Simpson metric. Bray distance was used to calculate the distance between the time points for all the groups and the beta diversity metric.

Intrataxonomic relationships among groups were computed using Analysis of Compositions of Microbiomes with Bias Correction (ANCOM-BC) R package version 1.6.2. Secom_linear function from ANCOM-BC was run with a prevalence cut of 20%, maximum P = of 0.05, and Spearman rank correlation with a minimum rho value of 0.70. The resultant output was then visualized as a network using visNetwork R package version 2.1.2 with co-occurrence of taxa at species in more than 60% of the samples at all time points and for both groups. Inferred networks were further quantified using the pagerank algorithm for centrality and centrality_degree function for degree calculation using tidygraph package version 1.2.2. The analysis was performed using stratified taxonomic data from MetaPhlAn4 and compared between time points (with respect to baseline) using the maaslin2 tool for multivariable association determination. For the taxonomic heat map generation, Maaslin2 version 1.16.0 was used to perform differential analysis using the negative binomial method, trimmed mean of M-values normalization, and visit variable set as a fixed effect.

For prospective stratification of the baseline patient stools prior to randomization, quantitative PCR was performed with methodologies previously published (29). Using these prespecified methods, patients were grouped into having either high or low baseline abundance of the Ruminococcaceae family in their gut microbiota using a threshold of 16%. This threshold was determined using data from prior publications (8) in addition to unpublished expanded patient data at the time of trial design. The primers utilized for the quantification of Ruminococcaceae were the following:

• –

  Ruminococcaceae forward primer (Clep886F): 5′-TTAACACAAT AAGTWATCCACCTGG-3′

• –

  Ruminococcaceae reverse primer (Clep1240R): 5″-ACCTTCCTCC GTTTTGTCAAC-3″

For a list of available biospecimens utilized for each time point and each patient, please refer to Supplementary Fig. S21.

Tumor samples were formalin-fixed and paraffin-embedded, including pretreatment and on-treatment samples. From each tissue block, a hematoxylin and eosin–stained slide was examined to evaluate tumor cellularity. Immunohistochemical (IHC) studies were performed in a CLIA-certified laboratory using an automated slide stainer (Leica Bond Max, Leica Biosystems) and an antihuman CD8 primary antibody (Lab Vision, Thermo-Fisher Scientific; clone: C8/144B; dilution 1:20) with 3,3′-diaminobenzidine (DAB) chromogen, and counterstained with hematoxylin. All slides were stained using previously optimized conditions with positive and negative control tissue placed on the same slide adjacent to the test tissue. IHC and hematoxylin and eosin–stained slides were converted into high-resolution digital images at 20× magnification using an Aperio slide scanner. Image analysis software (Aperio ImageScope) was applied to quantify the number and percentage of IHC-positive lymphocytes within designated areas marked by a pathologist manually (i.e., hand drawn around the tumor foci for needle core biopsies and other uniquely shaped fragments of tissue). The Aperio image analysis software quantified (i) the percentage of CD8-positive lymphocytes and (ii) the number of CD8-positive lymphocytes in a given area (the latter reported as CD8⁺ cells/mm²).

Untargeted global metabolomic profiling of stool was performed by Metabolon Inc. Briefly, stool was thawed from −80°C storage, and proteins were precipitated with methanol addition and centrifugation. The organic solvent was removed using a TurboVap and the resulting extract was analyzed using a combination of four approaches including reverse-phase UPLC MS/MS on positive and negative mode as well as HILIC/UPLC MS/MS. This resulted in the relative quantification of 1,396 separate chemicals within each sample, which were then normalized across batches using a standard composed of a mixture of all samples.

To compare global metabolite profiles across time points, PCA was performed. Chemicals with invariant quantification across samples were removed resulting in a total of 1,380 metabolites included in the analysis. Batch-normalized chemical quantities were log-transformed and scaled to a mean of 0 and standard deviation of 1. The resulting matrix was used to perform PCA using the prcomp function from the base stats package in R version 3.5.2. For each time point, a permutational ANOVA (PERMANOVA) was performed to test for the influence of the treatment arm on sample metabolomic distances as quantified by the Euclidean distance of samples after PCA projection. PERMANOVA R² and P values were calculated using the adonis function in the vegan R package (version 2.5-5) with 10,000 permutations.

Serum proteins were quantified using Olink multiplex proximity extension assay (PEA) panels (Olink Proteomics; www.olink.com) according to the manufacturer’s instructions (30). The assay was performed at the Olink Analysis Service Center. The basis of PEA is a dualrecognition immunoassay, where two matched antibodies labeled with unique DNA oligonucleotides simultaneously bind to a target protein in solution. This brings the two antibodies into proximity, allowing their DNA oligonucleotides to hybridize, serving as a template for a DNA polymerase-dependent extension step. This creates a doublestranded DNA “barcode,” which is unique for the specific antigen and quantitatively proportional to the initial concentration of the target protein. The hybridization and extension are immediately followed by PCR amplification and the amplicon is then finally quantified by microfluidic qPCR using Fluidigm BioMark HD system (Fluidigm Corporation). Data were normalized using internal controls in every single-sample, interplate control, negative controls, and correction factor, and expressed as log₂-scale, which is proportional to the protein concentration. The final assay readout is reported as normalized protein expression (NPX) values, which is an arbitrary unit on a log₂ scale where a higher value corresponds to a higher protein expression. One NPX difference equals to the doubling of the protein concentration. In this study, two Olink panels (Target96 Immuno-Oncology and Target96 Immune Response) were used, which consist of 172 unique analytes. Additional details about the analytes, detection range, data normalization, and standardization are available at https://www.olink.com/resources-support/document-download-center/.

Cryopreserved PBMC samples for fluorescent flow cytometry were analyzed in the Translational Cytometry Laboratory of the Penn Cytomics and Cell Sorting Shared Resource (University of Pennsylvania, Philadelphia, PA) on an extensively prequalified 28-color BD Symphony A5 cytometer (BD Biosciences). Staff were blinded to treatment arm and clinical outcome. At the time of analysis, cryopreserved PBMC samples were thawed in 37°C prewarmed RPMI-1640 medium (Gibco) containing 10% FBS and 100 U/mL of penicillin–streptomycin (Gibco). Samples were washed, counted, and resuspended in a medium containing 1 mg/mL DNase I (Roche) and 5 mmol/L magnesium chloride, and incubated at 37°C for 1 hour. After resting, cells were washed with PBS without additives (Corning) and transferred to staining tubes. PBMCs were incubated with 1 μL (0.2 μg) of 0.2 mg/mL nivolumab antibody (Selleck Chemicals) for 5 minutes at RT, followed by the addition of a Fixable Viability Stain 510 for 10 minutes at RT in the dark. Cells were then washed twice with FACS wash buffer (PBS, 1% BSA, 2 mmol/L EDTA). A surface antibody cocktail (T-cell phenotyping antibody panel; Supplementary Table S7) was prepared daily and used to stain up to 1 × 10⁷ cells per tube. Cells were incubated for 20 minutes at RT followed by washing twice with FACS staining buffer. The cells were resuspended in FoxP3 Transcription Factor Staining Buffer Fix/Perm solution (eBioscience) and incubated for 1 hour at RT to prepare the cells for intracellular staining. Post fixation, the samples were washed with Foxp3 permeabilization buffer. A freshly prepared cytoplasmic/intracellular staining cocktail master mix was added to the samples and incubated overnight at 4°C. The following day, the samples were washed with permeabilization buffer and resuspended in FACS wash buffer. Cells were stored at 4°C in the dark and acquired within 2 hours. Following daily QC, the instrument was standardized by setting hard dyed beads [BD Biosciences, Cytometer Setup and Tracking Beads (CS&T)] to predetermined target channels. Compensation controls (Invitrogen UltraComp eBeads or cells for Live/Dead stain) were prepared daily along with a frozen PBMC process control. The compensation matrix was calculated in Diva software (BD Biosciences) and used only for that day’s run. Data were analyzed using CellEngine cloud-based flow cytometry analysis software (CellCarta). Highlevel gates were tailored per patient across all time points by at least two investigators blinded to patient outcome. Single marker gates were drawn uniformly for analysis across patients and time points, with a representative gating strategy provided in Supplementary Fig. S22).

After gating for live cells and the CD3⁺ population, T-cell populations were defined as following, as shown in Supplementary Fig. S22: a combination of CD45RA, CD27, and CCR7 expression on CD4⁺ and CD8⁺ T cells was used to define naïve (CD45RA⁺CD27⁺CCR7⁺), T central memory (CM; CD45RA⁻CD27⁺CCR7⁺), T effector memory 1 (EM1; CD45RA⁻CD27⁺CCR7⁻), T effector memory 2 (EM2; CD45RA⁻CD27⁻CCR7⁺), T effector memory 3 (EM3; CD45RA⁻CD27⁻CCR7⁻), and terminally differentiated effector memory (EMRA; CD45RA⁺ CD27⁻CCR7⁻) subpopulations. CD4⁺ regulatory T cells were defined as Foxp3⁺CD25^hiCD127^−/low. The non-naïve CD4⁺ and CD8⁺ T-cell populations used in time-series and survival analyses included the defined effector memory, central memory, and TEMRA populations defined above. Expression of additional differentiation, activation, and inhibitory markers was evaluated within each of these compartments.

Translational analysis related to immune features was performed using Parker Institute in-house software CANDEL (Kitch and Kamphaus; ref. 31), and statistical analysis of these data was done using R version 4.2.2. For ×50 analysis, immune cell populations were quantified by percent-of-parent and analyzed in multiple ways. First, percentages of each gated population at day −7 were normalized to baseline and then compared between SER-401–active arm versus placebo arm. After t tests were performed for each population, uncorrected P values and effect sizes (by fold change) were reported (see Supplementary Table S8). Cell population data were also normalized to baseline levels and then plotted over time as a function of mean fold change, both individually and by mean of each arm to evaluate longitudinal trends. Single time-point analysis of mean fold-change differences was performed using Mann–Whitney tests, again using an uncorrected P value threshold = 0.05. For the subset analysis based on baseline Ruminococcaceae status, percentages of each gated population at day −7 were normalized to baseline and then compared between SER-401–active arm and placebo arm, and also by both baseline Ruminococcaceae status and response. After t tests were performed for each population, uncorrected P values and effect sizes (by fold change) were reported (see example Supplementary Table S8). Cell population data were also normalized to baseline levels and then plotted over time as a function of mean fold change, both individually and by means of each arm/group to evaluate longitudinal trends. Single time-point analysis of mean fold-change differences was performed using Mann–Whitney tests, using an uncorrected P value threshold = 0.05. Subset analysis of Treg populations was performed based on baseline Ruminococcaceae abundance characteristics, with the same statistical analysis performed as described for the high-parameter flow cytometry experiments.

Olink data were reported using values from the Normalized Protein eXpression (NPX) values. Per Olink, NPX is Olink’s arbitrary unit which is in log₂ scale. It is calculated from Ct values and data preprocessing (normalization) is performed to minimize both intraand interassay variation. NPX data allow users to identify changes in individual protein levels across their sample set, and then use these data to establish protein signatures. The NPX scale is inverted compared with that of Ct. This means that a high NPX value equals a high protein concentration. Because NPX is in a log₂ scale, a 1 NPX difference means a doubling of protein concentration. If needed NPX values can be converted into linear scale: 2^NPX = linear NPX. For instances where markers were duplicated (due to slight overlap in panels), the average of all values for that patient’s measurement at each time point was used. The Olink data were analyzed in a similar fashion to ×50. First, percentages of each marker at day −7 and then cycle 1, day 1 of anti–PD-1 were normalized to baseline and then compared between the SER-401–active arm and placebo arm. After t tests were performed for each population, uncorrected P values and effect sizes (by fold change) were reported (see Supplementary Table S9). Olink data were also normalized to baseline levels and then plotted over time as a function of mean fold change, both individually and by means of each arm to evaluate longitudinal trends.

Statistical analysis for deduced functional data with MetaCyc annotation was conducted with Maaslin2 adjusting for age, visit, sex, batch, and stage with a minimum abundance of 10 reads. P values were adjusted using false discovery rate correction.

Global differences in stool metabolomic analysis as well as overall microbial taxonomic communities were performed using PCA. Differences between study arms at each time point were evaluated using permutation multivariate analysis of variance (PERMANOVA) using the R “vegan” package (http://cran.r-project.org/web/packages/vegan/index.html; RRID:SCR_011950).

All of the whole-genome sequencing data, in addition to associated clinicopathologic deidentified metadata, are available through the NCBI Sequence Read Archive (SRA) under BioProject ID PRJNA978931. All of the 16S data utilized for the retrospective cohort analysis can be found under the SRA BioProject ID PRJNA770295. Other raw data pertaining to correlative analyses can be found in the supplementary tables. Further details and data are available upon reasonable request from the authors.

All codes and associated metadata utilized for the microbiome analysis can be accessed at https://github.com/mda-primetr/SER-401_Trial.

An overview of the trial patients’ representativeness can be found in Supplementary Table S10.

I.C. Glitza reports grants from BMS during the conduct of the study; grants, personal fees, and other support from Pfizer, personal fees from Novartis, grants from Merck, and personal fees from Midatech outside the submitted work. C.N. Spencer reports other support from the Parker Institute for Cancer Immunotherapy during the conduct of the study; personal fees from the Rare Cancer Research Foundation outside the submitted work; in addition, C.N. Spencer has a patent for improving responses to ICB therapy by enhancing the diversity and composition of the gut microbiome in patients with cancer licensed. J.R. Wortman reports other support from Seres Therapeutics outside the submitted work; in addition, J.R. Wortman has a patent for US20210361721A1 pending. J. Fairchild reports personal fees from the Parker Institute for Cancer Immunotherapy during the conduct of the study. C.B. Peterson reports grants from NIH/NCI CCSG 5P30CA016672 (Biostatistics Resource Group) during the conduct of the study. B. Weiner reports other support from Immuneering Corporation outside the submitted work. N. Hicks reports other support from Seres Therapeutics outside the submitted work. J. Aunins reports other support from the Parker Institute for Cancer Immunotherapy during the conduct of the study; other support from Seres Therapeutics, Inc. outside the submitted work. C. McChalicher reports employment and shareholder relationship with Seres Therapeutics. E. Walsh reports other support from Seres Therapeutics outside the submitted work. O. Hamid reports personal fees from Alkermes, Amgen, Beigene, Bioatla, Eisai, Roche Genentech, Georgiamune, GigaGen, GSK, GSK, Idera, Incyte, Instilbio, IO Bio, Iovance, Janssen, KSQ, Merck, Moderna, NGM, Obsidian, Sanofi, Seattle Genetics, Tempus, Vial Health, Zelluna, Bactonix, BMS, Immunocore, and Novartis, Pfizer, Regeneron during the conduct of the study; and contracted research for institution: Arcus, Aduro, Akeso, Amgen, Bioatla, BMS, Cytomx, Exelixis, Roche Genentech, GSK, Immunocore, Idera, Incyte, Iovance, Merck, Moderna, Merck Serono, Nextcure, Novartis, Pfizer, Regeneron, Seattle Genetics, Torque, Zelluna. P.A. Ott reports grants from Parker Institute during the conduct of the study; grants and personal fees from Bristol-Myers Squibb, Genentech, Merck, grants from Celldex, personal fees from Evaxion, Servier, Phio, and Pharmajet outside the submitted work. G.M. Boland reports grants from Olink Proteomics, Teiko Bio, Palleon Pharmaceuticals, personal fees from Iovance, Merck, Novartis, and Ankyra Therapeutics, and grants and personal fees from InterVenn Biosciences outside the submitted work. R.J. Sullivan reports grants and personal fees from Merck, personal fees from Novartis, Pfizer, Marengo, and Replimune outside the submitted work. N.J. Ajami reports other support from Seres Therapeutics and other support from Parker Institute for Cancer Immunotherapy during the conduct of the study. T. LaVallee reports grants and other support from PICI during the conduct of the study; personal fees from PICI outside the submitted work. M.R. Henn reports other support from the Parker Institute for Cancer Immunotherapy during the conduct of the study; personal fees and other support from Seres Therapeutics outside the submitted work. H.A. Tawbi reports grants and personal fees from Bristol-Myers Squibb, Merck, Novartis, grants from Genentech, personal fees from Pfizer, Iovance, Eisai, grants from Dragonfly, RAPT Therapeutics, personal fees from Karypharm, Jazz Pharmaceuticals, and Medicenna outside the submitted work. J.A. Wargo reports other support from Parker Institute for Cancer Immunotherapy and Seres Therapeutics during the conduct of the study; other support from Gustave Roussy Cancer Center, EverImmune, OSE Immunotherapeutics, Micronoma, and Daiichi Sankyo outside the submitted work; in addition, J.A. Wargo has a patent for PCT/US17/53.717 issued. No disclosures were reported by the other authors.

I.C. Glitza: Conceptualization, resources, supervision, investigation, writing–original draft, project administration, writing–review and editing. Y.D. Seo: Formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. C.N. Spencer: Conceptualization, resources, data curation, formal analysis, visualization, methodology, writing–original draft. J.R. Wortman: Conceptualization, resources, data curation, formal analysis, supervision, methodology. E.M. Burton: Conceptualization, resources. F.A. Alayli: Conceptualization, resources, data curation, formal analysis, visualization. C.P. Loo: Conceptualization, resources, formal analysis, visualization. S. Gautam: Conceptualization, resources, software, formal analysis. A. Damania: Data curation, software, formal analysis, investigation, visualization, methodology. J. Densmore: Conceptualization, resources, supervision. J. Fairchild: Conceptualization, resources, supervision. C.R. Cabanski: Conceptualization, resources, supervision. M.C. Wong: Data curation, software, formal analysis. C.B. Peterson: Formal analysis, methodology. B. Weiner: Conceptualization, resources, formal analysis. N. Hicks: Conceptualization, resources, formal analysis. J. Aunins: Conceptualization, resources, data curation, formal analysis. C. McChalicher: Conceptualization, resources, data curation, supervision. E. Walsh: Data curation, formal analysis. M.T. Tetzlaff: Conceptualization, resources, investigation, project administration. O. Hamid: Conceptualization, resources, investigation, project administration. P.A. Ott: Conceptualization, resources, investigation, project administration. G.M. Boland: Conceptualization, resources, investigation, project administration. R.J. Sullivan: Conceptualization, resources, investigation, project administration. K.F. Grossmann: Conceptualization, supervision, investigation. N.J. Ajami: Conceptualization, resources, data curation, software, formal analysis, supervision, investigation, methodology, writing–original draft, project administration, writing–review and editing. T. LaVallee: Conceptualization, resources, data curation, supervision. M.R. Henn: Conceptualization, resources, data curation, formal analysis, supervision, validation, writing–review and editing. H.A. Tawbi: Conceptualization, resources, investigation, project administration. J.A. Wargo: Conceptualization, resources, formal analysis, supervision, investigation, writing–original draft, project administration, writing–review and editing.

I.C. Glitza, Y.D. Seo, N.J. Ajami, and J.A. Wargo acknowledge Rebecca Soto, Ph.D., for her invaluable contribution to the review and revision of the manuscript. We also wish to acknowledge MD Anderson’s Platform for Innovative Microbiome and Translational Research (PRIME-TR) for supporting the analysis and interpretation of the microbiome results presented herein (J.A. Wargo and N.J. Ajami are the program director and executive scientific director for PRIME-TR, respectively). The work was in part supported by The University of Texas MD Anderson Cancer Center SPORE in Melanoma (5P50CA221703-03), as well as Training of Academic Surgical Oncologists grant (2T32CA009599-35). The clinical trial was funded in part by Seres Therapeutics (Cambridge, MA). Most importantly, the study team wishes to thank all patients and their families who contributed their time, samples, and data to this research. This work was supported by the Parker Institute for Cancer Immunotherapy and the University of Texas MD Anderson Cancer Center.