Longitudinal gut microbiome changes in immune checkpoint blockade-treated advanced melanoma

Abstract

Multiple clinical trials targeting the gut microbiome are being conducted to optimize treatment outcomes for immune checkpoint blockade (ICB). To improve the success of these interventions, understanding gut microbiome changes during ICB is urgently needed. Here through longitudinal microbiome profiling of 175 patients treated with ICB for advanced melanoma, we show that several microbial species-level genome bins (SGBs) and pathways exhibit distinct patterns from baseline in patients achieving progression-free survival (PFS) of 12 months or longer (PFS ≥12) versus patients with PFS shorter than 12 months (PFS <12). Out of 99 SGBs that could discriminate between these two groups, 20 were differentially abundant only at baseline, while 42 were differentially abundant only after treatment initiation. We identify five and four SGBs that had consistently higher abundances in patients with PFS ≥12 and <12 months, respectively. Constructing a log ratio of these SGBs, we find an association with overall survival. Finally, we find different microbial dynamics in different clinical contexts including the type of ICB regimen, development of immune-related adverse events and concomitant medication use. Insights into the longitudinal dynamics of the gut microbiome in association with host factors and treatment regimens will be critical for guiding rational microbiome-targeted therapies aimed at enhancing ICB efficacy.

Fig. 1. High-level view of gut microbiome dynamics in patients with PFS ≥12 and PFS <12 months.

a, For each microbial SGB listed, slopes are shown (that is, whether it is increasing or decreasing over study visits) in patients with PFS ≥12 (n = 83) and PFS <12 months (n = 92), respectively. For increased readability, SGBs differentially abundant in only one study visit have been removed (see Extended Data Fig. 2 for all SGBs). Red and blue colors indicate whether the focal SGB is increasing or decreasing in its abundance over study visits, respectively, with the strength of the colors corresponding to the steepness of the slope, with darker shades indicating steeper increases/decreases. It then shows, in the teal–brown heatmap, the average difference between the two slopes (that is, between patients with PFS ≥12 and PFS <12 months) across the different study visits. Non-gray cells in the heatmap correspond to the focal SGB’s log fold change in abundance between patients with PFS ≥12 and PFS <12 months, respectively. Teal cells correspond to study visits for which the abundance of the focal SGB is higher in patients with PFS ≥12 than with PFS <12 months, and vice versa for brown cells (at 90% BCL). Gray cells denote differences between patients with PFS ≥12 and PFS <12 months whose 90% credible interval cover zero. b, Three example features and how they increase and/or decrease in their expected abundance (represented in centered log ratio coordinates) over the study visits in patients with PFS ≥12 months (yellow slope) and in patients with PFS <12 months (purple slope). For each microbial SGB or pathway, the inset figure then displays the average difference between the two slopes at each study visit, including its 90% credible interval. These averages are the same as depicted in the teal–brown heatmap in a, and significance is deemed by evaluating whether or not the 90% credible interval covers zero. c, Microbial pathways are shown, similar to the format in a. The number (n) represents the number of patient samples at each visit for patients with PFS ≥12 and PFS <12 months.

Fig. 2. A longitudinal balance of microbial taxa (SGBs) predicts OS at baseline.

a, Schematic illustration of a balance between the five SGBs that were consistently higher in patients with PFS ≥12 months (A. butyriciproducens SGB14993 group, I. bartlettii SGB6140, Dorea sp. AF24 7LB SGB4571, L. gasseri SGB7038 group and L. celerecrescens SGB4868) and the four SGBs that were found to be consistently higher in patients with PFS <12 months (R. lactatiformans SGB15271, R. unclassified SGB15265 group, P. copri clade A SGB1626 and an unidentified SGB from the phylum Bacteroidetes SGB1957). In patients with PFS ≥12 and PFS <12 months, the balance is tilted to the left and right side, respectively. b, The balance’s ability to discriminate between patients with PFS ≥12 (n = 83, n₀ = 62, n₁ = 77, n₂ = 38 and n₃ = 30) and PFS <12 months (n = 92, n₀ = 74, n₁ = 69, n₂ = 34 and n₃ = 24) months across study visits (two-sided Wilcoxon test: P_T0 = 0.00085, P_T1 = 0.0007, P_T2 = 0.0005 and P_T3 = 0.1). Boxplots represent minima, Q1, Q2, Q3 and maxima. c, The balance’s predictive ability expressed as the AUC computed from 100 times repeated five-fold cross-validation. Each line shows, for each study visit, the average across the 100 times repeated five-fold cross-validations with the shaded area representing the 95% CI (mean AUC ± s.d.: AUC_T0 0.659 ± 0.092, AUC_T1 0.666 ± 0.091, AUC_T2 = 0.739 ± 0.118 and AUC_T3 0.655 ± 0.129). The dashed diagonal line represents random chance. d, Kaplan–Meier curves and multivariable Cox regression of OS in months for 146 patients at baseline according to high (above median; teal) and low (below median; orange) values of the balance after adjusting for age, sex, BMI, previous therapy, PPI and antibiotics use.

Fig. 3. Different taxon dynamics in patients with PFS ≥12 and PFS <12 months.

a–d, Four different dynamics exemplified by different microbial SGBs with different dynamics in patients with PFS ≥12 (n₀ = 62, n₁ = 77, n₂ = 38 and n₃ = 30) and PFS <12 (n₀ = 74, n₁ = 69, n₂ = 34 and n₃ = 24) months, where the slopes of patients with PFS ≥12 months (yellow slopes) and patients with PFS <12 months (purple slopes) diverge from similar baseline abundances (a, dynamics 2a, Extended Data Fig. 7a), where the slopes of patients with PFS ≥12 and PFS <12 months are crossing (b, generating opposite abundance patterns when comparing baseline to the last study visit, dynamics 3b, Extended Data Fig. 7b), where the slope of the patients with PFS <12 months is relatively unchanged across the study visits compared to the slope of the patients with patients with PFS ≥12 months (c, dynamics 1c, Extended Data Fig. 7c); where the slope of the patients with PFS ≥12 months is relatively unchanged across the study visits compared to the slope of the patients with PFS <12 months (d, dynamics 2c, Extended Data Fig. 7c). The y axis shows the expected abundance (represented in centered log ratio coordinates) for each study visit (x axis). The corresponding inset figures show the average difference between patients with PFS ≥12 and PFS <12 months at each study visit, including its 90% credible interval. The number (n) represents the number of patient samples at each visit for patients with PFS ≥12 and PFS <12 months.

Fig. 4. Divergent signals in monotherapy versus combination therapy.

a–c, Three examples out of the six SGBs that exhibited divergent patterns in monotherapy (PFS ≥12: n₀ = 41, n₁ = 49, n₂ = 29 and n₃ = 25; PFS <12: n₀ = 49, n₁ = 48, n₂ = 25 and n₃ = 18) compared to combination therapy (PFS ≥12: n₀ = 21, n₁ = 28, n₂ = 9 and n₃ = 5; PFS <12: n₀ = 25, n₁ = 21, n₂ = 9 and n₃ = 6): C. eutactus (SGB5121) (a), Butyricicoccus sp. AM29 23AC (SGB14991) (b) and P. merdae (SGB1949) (c). The y axis shows the expected abundance (represented in centered log ratio coordinates) for each study visit (x axis). Left: anti-PD-1 monotherapy. Right: anti-PD-1/anti-CTLA-4 combination therapy. The corresponding inset figures show the average difference between patients with PFS ≥12 and PFS <12 months at each study visit, including its 90% credible interval. The number (n) represents the number of patient samples at each visit for patients with PFS ≥12 and PFS <12 months.

Fig. 5. A balance predictive of ICB-induced colitis at baseline.

A balance between the 10 SGBs associated with the presence of colitis (red; n = 24 patients) and the 12 SGBs associated with the absence of colitis (blue; n = 112 patients) at baseline is predictive of colitis development at baseline. Left: the balance’s discriminatory ability (two-sided Wilcoxon test, P_T0 = 0.00055). Boxplots represent minima, Q1, Q2, Q3 and maxima. Right: the same balance’s predictive ability expressed as the averaged AUC computed from a 100 times repeated five-fold cross-validation (AUC mean ± s.d. of 0.723 ± 0.121). The dashed diagonal line represents random chance.

Extended Data Fig. 1. Study description with sample numbers across study visits.

Samples were collected within 5 sub-cohorts: two prospectively recruited within parallel observational studies (The PRIMM cohorts), and three retrospectively pooled cohorts. Fecal samples were collected at 4 timepoints: at baseline (T0) and at every treatment cycle (T1 to T3) over a period of 12 weeks. The time between two samples was 3 or 4 weeks, depending on the treatment regimen, with Ipilimumab/Nivolumab combination therapy and Pembrolizumab monotherapy administered 3-weekly and Nivolumab monotherapy administered 4-weekly. Treatment continued after the 12 weeks until the patient responded or until the treatment had to pause/stop due to irAEs. Not all subjects provided fecal samples at all study visits. Therefore, gut microbial dynamics were modeled at the level of the population including a random effect for the patient identifier (see Methods). Sample numbers represent patients with complete metadata (that is, no missingness) for all considered covariates/confounders. For the survival analysis, because we adjusted for a smaller number of covariates/confounders, there were n = 147 at baseline (PRIMM-UK = 41; PRIMM-NL = 53; Barcelona = 12; Leeds = 17; Manchester = 24) rather than n = 136 as indicated here. Tumor staging by CT or PET-scans was performed at study entry and at regular intervals during treatment. Tumor response was classified using the Response Evaluation Criteria in Solid Tumors (RECIST) v.1. Endpoints were defined as Progression-free survival at 12 months (PFS12) and overall survival (OS). Immune-related adverse events (irAEs) were assessed using the Common Terminology Criteria for Adverse Events (CTCAE) v.5 (see Table 1). ICB, Immune checkpoint blockade; PRIMM, Predicting Response to Immunotherapy for Melanoma With Gut Microbiome and Metabolomics; NL, Netherland; UK, United Kingdom. The figure was generated in BioRender.com.

Extended Data Fig. 2. Extension of Fig. 1.

This includes all SGBs, that is, also those that were differentially abundant in only one study visit.

Extended Data Fig. 3. Extended longitudinal balance.

The extended balance has 12 SGBs in the numerator (A. butyriciproducens [SGB14993 group], I. bartlettii [SGB6140], D. sp AF24 7LB [SGB4571], L. gasseri [SGB7038 group], L. celerecrescens [SGB4868], R. sp NSJ 71 [SGB4290], GGB9640 [SGB15115], E. rectale [SGB4933 group], E. entriosum [SGB5045], E. sp AM28 29 [SGB6796 group], Clostridium sp AF15 49 [SGB5111], and A. bouchesdurhonensis [SGB17152]) and 9 SGBs in the denominator (R. lactatiformans [SGB15271], R. unclassified [SGB15265 group], P. copri clade A [SGB1626], GGB1420 [SGB1957], Gemmiger [SGB15299], B. obeum [SGB4809], Clostridiales unclassified [SGB15145], P. vulgatus [SGB1814], and B. clarus [SGB1832]). Panel (A) balance’s ability to discriminate between patients with PFS ≥ 12 (n = 83; n₀ = 62; n₁ = 77; n₂ = 38; n₃ = 30) and PFS < 12 (n = 92; n₀ = 74; n₁ = 69; n₂ = 34; n₃ = 24) months across study visits. Boxplots represent minima, Q1, Q2, Q3, and maxima. Panel (B) the balance’s predictive ability expressed as the Area Under the Curve (AUC) computed from 100 times repeated five-fold cross-validation (CV). Each line shows, for each study visit, the average across the 100 repeated five-fold CVs with the shaded area representing the 95% confidence interval. Panel (C) Kaplan–Meier curves and multivariable Cox regression analysis of overall survival in 146 patients at baseline (one patient was removed due to missingness in the included predictor variables) according to high (above median) and low (below median) values of the balance after adjusting for age, sex, BMI, previous therapy, PPI and antibiotic use.

Extended Data Fig. 4. Treating balances as continuous independent variables.

Panel A-C shows a multivariable Cox regression analysis of overall survival (OS) in months for 146 patients at baseline (one patient was removed due to missingness in the included predictor variables) treating (a) the first balance (Fig. 2a), (b) the second balance (that is, the extended longitudinal balance), and (c) the third balance (that is, the ‘baseline only’ balance) as a continuous independent variable. While the histograms show the distribution of each balance (right y-axes), each regression line represent the hazard ratio as a smooth function of each balance (left y-axes). All models are adjusted for age, sex, BMI, PPI and antibiotics use, and previous therapy.

Extended Data Fig. 5. Generalizability of the longitudinal balance (Fig. 2a) across six independent melanoma cohorts.

Panel (A) shows the AUC for each independent baseline cohort, including the current study (in red). Panel (B) shows the AUC for McCullochJA_2022’s post-ICB cohort. Panel (C) shows the average difference in the balance score between patients with PFS < 12 months versus PFS ≥ 12 months from the SpencerCN_2021 cohort. Finally, panel (D) shows Kaplan-Meier curves and multivariable Cox regression analysis of overall survival (OS) in months from 27 patients from McCullochJA_2022’s baseline cohort according to high (≥75 percentile) and low (<75 percentile) values of the balance after adjusting for age, sex, BMI and PPI-use.

Extended Data Fig. 6. Baseline only balance.

A ‘baseline only’ balance containing the 9 species associated with patients with PFS ≥ 12 months at baseline only in the numerator, and the 11 SGBs associated with patients with PFS < 12 months at baseline only in the denominator (see Fig. 1a). Kaplan-Meier curves and multivariable Cox regression analysis of overall survival in 146 patients at baseline (one patient was removed due to missingness in the included predictor variables) according to high (above median) and low (below median) values of the balance after adjusting for age, sex, BMI, previous therapy, PPI and antibiotic use.

Extended Data Fig. 7. Schematic illustration showing the different types of microbial dynamics we observe between patients with PFS ≥ 12 months and PFS < 12 months.

Panels A-E are schematic illustrations (that is, cartoons) showing the breakdown of the different types of taxon dynamics we observe in the overall comparison between patients with PFS ≥ 12 months and patients with PFS < 12 months. Yellow and purple slopes correspond to patients with PFS ≥ 12 and PFS < 12 months, respectively. Panel (A) shows dynamics where patients with PFS ≥ 12 and PFS < 12 months are differentially abundant only after T0 (that is, dynamics 1a and 2a). Dynamics 3a is a particular case of dynamics 1a and 2a where the slopes for patients with PFS ≥ 12 and PFS < 12 months intersect. Panel (B) shows dynamics where patients with PFS ≥ 12 and PFS < 12 months are differentially abundant at early but not at late visits (that is, dynamics 1b and 2b). Dynamics 3b is a special case of dynamics 1b and 2b where the patients with PFS ≥ 12 and PFS < 12 months slopes intersect. Panel (C) shows dynamics where the slope of one of the groups is zero (or close to zero) while the other group is either increasing or decreasing, respectively (that is, dynamics 1c and 2c). In panels (D) and (E), included the inset figures, patients with PFS ≥ 12 and PFS < 12 months exhibit parallel lines (that is, no statistical interactions); Panel 1d and 2d shows dynamics where both patients with PFS ≥ 12 and PFS < 12 months are either increasing or decreasing, respectively, while in panels 1f and 2f, the slopes of patients with PFS ≥ 12 and PFS < 12 months are zero (or close to zero). The number in each plot corresponds to the number of microbial SGBs that follow each type of different dynamics.

Extended Data Fig. 8. Patients with PFS ≥ 12 and PFS < 12 months on monotherapy.

Panel (A) shows, for each microbial SGB listed, its slopes in patients with PFS ≥ 12 months and PFS < 12 months on monotherapy, respectively. Red and blue colors indicate whether the focal SGB is increasing or decreasing in its abundance over study visits, respectively. It then shows the average difference between patients with PFS ≥ 12 and PFS < 12 months across the different study visits. Non-gray cells in the heatmap correspond to the focal SGB’s log-fold change in abundance between patients with PFS ≥ 12 and PFS < 12 months, respectively. Teal cells correspond to study visits for which the abundance of the focal SGB is higher in in patients with PFS ≥ 12 than with PFS < 12 months on monotherapy, and vice versa for brown cells (at 90% BCL). Gray cells denote differences between patients with PFS ≥ 12 and PFS < 12 months on monotherapy whose 90% CI overlapped with zero.

Extended Data Fig. 9. Patients with PFS ≥ 12 and PFS < 12 months on combination therapy.

Panel (A) shows, for each microbial SGB listed, its slopes in patients with PFS ≥ 12 months and PFS < 12 months on combination therapy, respectively. Red and blue colors indicate whether the focal SGB is increasing or decreasing in its abundance over study visits, respectively. It then shows the average difference between patients with PFS ≥ 12 and PFS < 12 months across the different study visits. Non-gray cells in the heatmap correspond to the focal SGB’s log-fold change in abundance between patients with PFS ≥ 12 and PFS < 12 months, respectively. Teal cells correspond to study visits for which the abundance of the focal SGB is higher in in patients with PFS ≥ 12 than with PFS < 12 months on combination therapy, and vice versa for brown cells (at 90% BCL). Gray cells denote differences between patients with PFS ≥ 12 and PFS < 12 months on combination therapy whose 90% CI overlapped with zero.

Extended Data Fig. 10. Patients who developed and not developed colitis.

The figure shows, for each microbial SGB listed, its slopes in patients who developed and not developed colitis, respectively, regardless of response to therapy. Red and blue colors indicate whether the focal SGB is increasing or decreasing in its abundance over study visits, respectively. It then shows the average difference between patients with and without colitis across the different study visits. Non-gray cells in the heatmap correspond to the focal SGB’s log-fold change in abundance between patients with and without colitis, respectively. Teal cells correspond to study visits for which the abundance of the focal SGB is higher in in patients who developed colitis compared to those resistant to colitis, and vice versa for brown cells (at 90% BCL). Gray cells denote differences between patients with and without colitis whose 90% CI overlapped with zero.