Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 68 of the 70 references it cites were found in a registry or an index; 11 of the 16 sources its claims cite were checked for support.

Cross-cohort gut microbiome associations with immune checkpoint inhibitor response in advanced melanoma

Open access

Abstract

The composition of the gut microbiome has been associated with clinical responses to immune checkpoint inhibitor (ICI) treatment, but there is limited consensus on the specific microbiome characteristics linked to the clinical benefits of ICIs. We performed shotgun metagenomic sequencing of stool samples collected before ICI initiation from five observational cohorts recruiting ICI-naive patients with advanced cutaneous melanoma (n = 165). Integrating the dataset with 147 metagenomic samples from previously published studies, we found that the gut microbiome has a relevant, but cohort-dependent, association with the response to ICIs. A machine learning analysis confirmed the link between the microbiome and overall response rates (ORRs) and progression-free survival (PFS) with ICIs but also revealed limited reproducibility of microbiome-based signatures across cohorts. Accordingly, a panel of species, including Bifidobacterium pseudocatenulatum, Roseburia spp. and Akkermansia muciniphila, associated with responders was identified, but no single species could be regarded as a fully consistent biomarker across studies. Overall, the role of the human gut microbiome in ICI response appears more complex than previously thought, extending beyond differing microbial species simply present or absent in responders and nonresponders. Future studies should adopt larger sample sizes and take into account the complex interplay of clinical factors with the gut microbiome over the treatment course.

Overview

CoChat Summary

TL;DR

In five new observational cohorts of ICI-naive patients with advanced cutaneous melanoma (n = 165 baseline stool metagenomes), integrated with 147 published metagenomes, the gut microbiome was associated with ICI response but in a cohort-dependent manner. Machine learning confirmed some within-cohort predictive signal, yet no single taxonomic or functional biomarker was fully reproducible across cohorts.

Key findings

  • The gut microbiome showed a relevant but cohort-dependent association with ICI response: cohort explained nearly ten times more variance than any other variable, and ORR/PFS12 explained only a limited fraction of total microbiome variation after batch correction.
  • Within-cohort machine learning prediction varied by cohort and endpoint: PRIMM-UK ORR AUC-ROC 0.78 and PRIMM-NL PFS12 AUC-ROC 0.64, whereas PRIMM-NL ORR and PRIMM-UK PFS12 were much lower (0.53 and 0.57, respectively); functional KEGG ortholog profiles exceeded AUC-ROC 0.59 for both datasets and endpoints.
  • Cross-cohort prediction was limited: AUC-ROC reached 0.7 in only two cases and exceeded 0.6 in about one-third of cases (31.4%); leave-one-dataset-out models averaged 0.59 for ORR and 0.60 for PFS12.
  • Response-associated taxa included Roseburia sp. CAG:182 and Roseburia sp. CAG:471 for ORR, Phascolarctobacterium succinatutens and Lactobacillus vaginalis for PFS12, and Akkermansia muciniphila and Dorea formicigenerans for ORR/PFS12; Bacteroides clarus was higher in nonresponders, but no single species was a fully consistent biomarker across all datasets.

Methods

  • Prospective PRIMM-UK and PRIMM-NL cohorts plus Barcelona, Leeds and Manchester cohorts recruited ICI-naive patients with unresectable stage III/IV cutaneous melanoma; baseline stool samples were collected before ICI treatment, with 175 patients recruited and 165 samples passing strict QC, analyzed alongside 147 publicly available metagenomes.
  • Shotgun metagenomic sequencing averaged 7.74 Gb per sample; taxonomic and functional profiling used bioBakery 3 (MetaPhlAn 3.0, HUMAnN 3.0), identifying 608 species, 510 pathways, 6,131 KOs and 2,558 enzyme categories in the 165 samples.
  • Response was classified by RECIST v1.1 as ORR (CR/PR/SD vs PD) and PFS12; analyses included PERMANOVA, an ensemble of eight differential abundance methods with random-effects meta-analysis, and SIAMCAT Lasso logistic regression with nested cross-validation, cross-study validation and leave-one-dataset-out testing.

Limitations

  • Stated: individual cohort sample sizes remain small; collecting very large, properly annotated samples is inherently difficult in this clinical setting; response definition differences (ORR vs PFS12) and many confounding factors are hard to adjust for; cohort microbiome effects are strong; current methods lack characterization of the immunological interface.
  • Stated: the authors call for larger sample sizes, strain-level and mucosal analyses, and extension beyond cutaneous melanoma to uveal/mucosal subtypes and other cancers.
  • Evident from the reported results: cross-cohort reproducibility was limited, with modest AUC-ROC values and no single species consistently associated with response across all datasets.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41591-022-01695-5

Cites

68 of 68 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Association between the gut microbiome and response in the PRIMM-NL and PRIMM-UK cohorts. a, Response evaluated by ORR (Methods) is associated with the overall microbiome structure for PRIMM-UK (P = 0.05), but not for PRIMM-NL or PFS12, as represented visually using principal-component analysis (PCA) of species-level centered log-ratio-transformed relative abundances. P values were calculated using adonis and 999 permutations (Extended Data Figs. 2 and 3 show additional beta diversity analysis). Dim1, dimension 1; Dim2, dimension 2. b, Multivariate analysis showing the amount of inferred variance explained (R2, blue vertical bars) by each identified covariate and their respective P value (orange vertical bars) as determined by PERMANOVA on species-level centered log-ratio-transformed relative abundances. c, Machine learning association analysis between taxonomic (species abundance) and functional profiles (KEGG ortholog abundances) of the microbiome and response showed consistent associations with both response types (ORR or PFS12). The ‘concordant’ label includes only patients who did not progress between 6 and 12 months. AUC-ROC curves are computed using Lasso models trained using 100-repeated fivefold-stratified cross-validations. Shaded areas represent AUC-ROCs from each individual machine learning model. ICI indicates the use of a combination of ipilimumab and nivolumab or single agent. AUC, area under the curve; CV, cross-validation; hPDI, healthy plant-based diet index; PDI, plant-based diet index; uPDI, unhealthy plant-based diet index; mMED, modified Mediterranean diet score.
Fig. 1 Association between the gut microbiome and response in the PRIMM-NL and PRIMM-UK cohorts. a, Response evaluated by ORR (Methods) is associated with the overall microbiome structure for PRIMM-UK (P = 0.05), but not for PRIMM-NL or PFS12, as represented visually using principal-component analysis (PCA) of species-level centered log-ratio-transformed relative abundances. P values were calculated using adonis and 999 permutations (Extended Data Figs. 2 and 3 show additional beta diversity analysis). Dim1, dimension 1; Dim2, dimension 2. b, Multivariate analysis showing the amount of inferred variance explained (R2, blue vertical bars) by each identified covariate and their respective P value (orange vertical bars) as determined by PERMANOVA on species-level centered log-ratio-transformed relative abundances. c, Machine learning association analysis between taxonomic (species abundance) and functional profiles (KEGG ortholog abundances) of the microbiome and response showed consistent associations with both response types (ORR or PFS12). The ‘concordant’ label includes only patients who did not progress between 6 and 12 months. AUC-ROC curves are computed using Lasso models trained using 100-repeated fivefold-stratified cross-validations. Shaded areas represent AUC-ROCs from each individual machine learning model. ICI indicates the use of a combination of ipilimumab and nivolumab or single agent. AUC, area under the curve; CV, cross-validation; hPDI, healthy plant-based diet index; PDI, plant-based diet index; uPDI, unhealthy plant-based diet index; mMED, modified Mediterranean diet score.CC BY · PubMed Central
Fig. 2 Integrated analysis of newly sequenced and publicly available datasets for cross-cohort response–microbiome association. a, Contribution of variables to the overall microbial community composition highlights the heterogeneity of the microbiome structure across cohorts that has a substantially higher effect than both anthropometric and clinical parameters. We either used all available cohorts or newly sequenced cohorts for which additional metadata were available. Batch-correction methods were applied to species-level abundances prior to distance calculations. The plot on the left uses ORR as the outcome variable, whereas the plot on the right adopts PFS12. b, Prediction matrix for microbiome-based prediction of response assessed via ORR (left matrix) and PFS12 (right matrix) within each cohort (values on the diagonal), across pairs of cohorts (one cohort used to train the model and the other for testing) and in the leave-one-cohort-out setting (training the model on all but one cohort and testing on the left-out cohort). We report the AUC-ROC values obtained from Lasso models on species-level relative abundances. Values on the diagonal refer to the median AUC-ROC values of 100-repeated fivefold-stratified cross-validations. Off-diagonal values refer to AUC-ROC values obtained by training the classifier on the cohort of the corresponding row and applying it to the cohort of the corresponding column. The leave-one-out row refers to the performances obtained by training the model using all but the cohort of the corresponding column and applying it to the cohort of the corresponding column. The same prediction matrix using functional microbiome profiles are available in Extended Data Fig. 4. c, ORR (n = 284) cross-validation AUC-ROC values obtained from Lasso models trained using 100-repeated fivefold-stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from Lasso models obtained by training the model using species-level relative abundances and all but the corresponding (circles). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than 1.5× interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles. EC, enzyme category.
Fig. 2 Integrated analysis of newly sequenced and publicly available datasets for cross-cohort response–microbiome association. a, Contribution of variables to the overall microbial community composition highlights the heterogeneity of the microbiome structure across cohorts that has a substantially higher effect than both anthropometric and clinical parameters. We either used all available cohorts or newly sequenced cohorts for which additional metadata were available. Batch-correction methods were applied to species-level abundances prior to distance calculations. The plot on the left uses ORR as the outcome variable, whereas the plot on the right adopts PFS12. b, Prediction matrix for microbiome-based prediction of response assessed via ORR (left matrix) and PFS12 (right matrix) within each cohort (values on the diagonal), across pairs of cohorts (one cohort used to train the model and the other for testing) and in the leave-one-cohort-out setting (training the model on all but one cohort and testing on the left-out cohort). We report the AUC-ROC values obtained from Lasso models on species-level relative abundances. Values on the diagonal refer to the median AUC-ROC values of 100-repeated fivefold-stratified cross-validations. Off-diagonal values refer to AUC-ROC values obtained by training the classifier on the cohort of the corresponding row and applying it to the cohort of the corresponding column. The leave-one-out row refers to the performances obtained by training the model using all but the cohort of the corresponding column and applying it to the cohort of the corresponding column. The same prediction matrix using functional microbiome profiles are available in Extended Data Fig. 4. c, ORR (n = 284) cross-validation AUC-ROC values obtained from Lasso models trained using 100-repeated fivefold-stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from Lasso models obtained by training the model using species-level relative abundances and all but the corresponding (circles). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than 1.5× interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles. EC, enzyme category.CC BY · PubMed Central
Fig. 3 A panel of potential taxonomic and function microbiome biomarkers for response across cohorts. a, Species associated with ORR identified by a meta-analysis using different differential abundance methods. Species shown have random-effects model P values < 0.05 in at least three methods. Values inside the cells refer to unadjusted P values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. The color of the cell was determined by comparing the mean relative abundance in responders to nonresponders; if the mean was higher in responders, then the cells were colored red; if it was higher in nonresponders, then it was colored blue. b, Species associated with PFS12 identified by a meta-analysis using different differential abundance methods. Species shown have random-effects model P values < 0.05 in at least three methods. Values inside the cells refer to unadjusted P values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. c, KOs associated with response status identified by a meta-analysis using different differential abundance methods. The KEGG orthologues shown have random-effects model P values < 0.05 in at least six methods. Values inside the cells refer to unadjusted P values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. d, Species associated with ORR in the two PRIMM cohorts (PRIMM-NL (n = 47) and PRIMM-UK (n = 53)) before and after adjusting for covariates that included PPI, antibiotic and steroid use; gender, performance status; previous therapy; age; and ICIs. Species shown have covariate-adjusted multiple hypothesis testing-corrected q < 0.2 in one of the cohorts identified by ANCOM-BC. Symbols (circles and triangles) show the ANCOM-BC beta coefficient, and error bars represent standard error. Adj, adjusted; NR, nonresponders; R, responders; SMD, standardized mean differences.
Fig. 3 A panel of potential taxonomic and function microbiome biomarkers for response across cohorts. a, Species associated with ORR identified by a meta-analysis using different differential abundance methods. Species shown have random-effects model P values < 0.05 in at least three methods. Values inside the cells refer to unadjusted P values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. The color of the cell was determined by comparing the mean relative abundance in responders to nonresponders; if the mean was higher in responders, then the cells were colored red; if it was higher in nonresponders, then it was colored blue. b, Species associated with PFS12 identified by a meta-analysis using different differential abundance methods. Species shown have random-effects model P values < 0.05 in at least three methods. Values inside the cells refer to unadjusted P values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. c, KOs associated with response status identified by a meta-analysis using different differential abundance methods. The KEGG orthologues shown have random-effects model P values < 0.05 in at least six methods. Values inside the cells refer to unadjusted P values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. d, Species associated with ORR in the two PRIMM cohorts (PRIMM-NL (n = 47) and PRIMM-UK (n = 53)) before and after adjusting for covariates that included PPI, antibiotic and steroid use; gender, performance status; previous therapy; age; and ICIs. Species shown have covariate-adjusted multiple hypothesis testing-corrected q < 0.2 in one of the cohorts identified by ANCOM-BC. Symbols (circles and triangles) show the ANCOM-BC beta coefficient, and error bars represent standard error. Adj, adjusted; NR, nonresponders; R, responders; SMD, standardized mean differences.CC BY · PubMed Central
Fig. 4 Covariate associations with the gut microbiome from the PRIMM cohorts. a, Cross-validation AUC-ROC values obtained from Lasso models trained using 100-repeated fivefold-stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from Lasso models obtained by training the model using species-level relative abundances and all but the corresponding PRIMM cohort (circles). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than 1.5× interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles (PRIMM-NL, n = 55; PRIMM-UK, n = 55). b–d, Species associated with PPI use (<3 months after the start of ICI), toxicity and colitis identified by ANCOM-BC with and without covariate adjustment (PRIMM-NL, n = 47; PRIMM-UK, n = 53). Covariates included in all models were ORR, performance status, previous therapy, age, ICIs (combination of ipilimumab and nivolumab or single agent), gender and antibiotic and steroid use. PPI use was also included as a covariate when analyzing colitis and toxicity. Species shown have covariate-adjusted multiple hypothesis testing-corrected q < 0.2 in one of the cohorts identified by ANCOM-BC. Symbols (circles and triangles) show the ANCOM-BC beta coefficient, and error bars represent standard error.
Fig. 4 Covariate associations with the gut microbiome from the PRIMM cohorts. a, Cross-validation AUC-ROC values obtained from Lasso models trained using 100-repeated fivefold-stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from Lasso models obtained by training the model using species-level relative abundances and all but the corresponding PRIMM cohort (circles). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than 1.5× interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles (PRIMM-NL, n = 55; PRIMM-UK, n = 55). b–d, Species associated with PPI use (<3 months after the start of ICI), toxicity and colitis identified by ANCOM-BC with and without covariate adjustment (PRIMM-NL, n = 47; PRIMM-UK, n = 53). Covariates included in all models were ORR, performance status, previous therapy, age, ICIs (combination of ipilimumab and nivolumab or single agent), gender and antibiotic and steroid use. PPI use was also included as a covariate when analyzing colitis and toxicity. Species shown have covariate-adjusted multiple hypothesis testing-corrected q < 0.2 in one of the cohorts identified by ANCOM-BC. Symbols (circles and triangles) show the ANCOM-BC beta coefficient, and error bars represent standard error.CC BY · PubMed Central
Extended Data Fig. 1 Associations between alpha diversity and response. (a) Alpha diversity measures in the two PRIMM cohorts. # p refers to P values calculated using limma linear models including PPI, antibiotic and steroid use, gender, performance status, previous therapy, age, plant-based diet index, unhealthy plant-based diet index, mediterranean diet score and ICI in the model. $ p refers to p values calculated using the Wilcoxon Rank-Sum Test. The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than ×1.5 interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles. Meta-analysis results obtained using standardized mean differences between responders and nonresponders using (b) Shannon diversity and (c) species richness. The centre of the error bars (grey shaded squares) represent the standardized mean difference for each cohort.
Extended Data Fig. 1 Associations between alpha diversity and response. (a) Alpha diversity measures in the two PRIMM cohorts. # p refers to P values calculated using limma linear models including PPI, antibiotic and steroid use, gender, performance status, previous therapy, age, plant-based diet index, unhealthy plant-based diet index, mediterranean diet score and ICI in the model. $ p refers to p values calculated using the Wilcoxon Rank-Sum Test. The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than ×1.5 interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles. Meta-analysis results obtained using standardized mean differences between responders and nonresponders using (b) Shannon diversity and (c) species richness. The centre of the error bars (grey shaded squares) represent the standardized mean difference for each cohort.CC BY · PubMed Central
Extended Data Fig. 2 Machine learning association analysis of the microbiome and response in PRIMM-NL. (a-e) Machine learning association analysis of the microbiome and response using either metadata alone or in combination with taxonomic (species abundance) and functional profiles (KEGG orthologs’ abundances) in PRIMM-NL. AUC-ROC curves are computed using LASSO models trained using 100-repeated fivefold-stratified cross-validations. Metadata used in the models included: age, gender, performance status, PPI use, antibiotic use, steroid use, ICI and previous therapy. Shaded areas represent AUC-ROCs from each individual machine learning model. (f) Multivariate analysis showing the amount of inferred variance explained (R2, blue vertical bars) by each identified covariate and their respective p value (orange vertical bars) as determined by PERMANOVA on KEGG clr-transformed relative abundances.
Extended Data Fig. 2 Machine learning association analysis of the microbiome and response in PRIMM-NL. (a-e) Machine learning association analysis of the microbiome and response using either metadata alone or in combination with taxonomic (species abundance) and functional profiles (KEGG orthologs’ abundances) in PRIMM-NL. AUC-ROC curves are computed using LASSO models trained using 100-repeated fivefold-stratified cross-validations. Metadata used in the models included: age, gender, performance status, PPI use, antibiotic use, steroid use, ICI and previous therapy. Shaded areas represent AUC-ROCs from each individual machine learning model. (f) Multivariate analysis showing the amount of inferred variance explained (R2, blue vertical bars) by each identified covariate and their respective p value (orange vertical bars) as determined by PERMANOVA on KEGG clr-transformed relative abundances.CC BY · PubMed Central
Extended Data Fig. 3 Machine learning association analysis of the microbiome and response in PRIMM-UK. (a-e) Machine learning association analysis of the microbiome and response using either metadata alone or in combination with taxonomic (species abundance) and functional profiles (KEGG orthologs’ abundances) in PRIMM-UK. AUC-ROC curves are computed using LASSO models trained using 100-repeated fivefold stratified cross-validations. Metadata used in the models included: age, gender, performance status, PPI use, antibiotic use, steroid use, ICI and previous therapy. Shaded areas represent AUC-ROCs from each individual machine learning model. (f) Multivariate analysis showing the amount of inferred variance explained (R2, blue vertical bars) by each identified covariate and their respective p value (orange vertical bars) as determined by PERMANOVA on KEGG clr-transformed relative abundances.
Extended Data Fig. 3 Machine learning association analysis of the microbiome and response in PRIMM-UK. (a-e) Machine learning association analysis of the microbiome and response using either metadata alone or in combination with taxonomic (species abundance) and functional profiles (KEGG orthologs’ abundances) in PRIMM-UK. AUC-ROC curves are computed using LASSO models trained using 100-repeated fivefold stratified cross-validations. Metadata used in the models included: age, gender, performance status, PPI use, antibiotic use, steroid use, ICI and previous therapy. Shaded areas represent AUC-ROCs from each individual machine learning model. (f) Multivariate analysis showing the amount of inferred variance explained (R2, blue vertical bars) by each identified covariate and their respective p value (orange vertical bars) as determined by PERMANOVA on KEGG clr-transformed relative abundances.CC BY · PubMed Central
Extended Data Fig. 4 Cross-cohort response–microbiome associations at the functional level. (a) Contribution of variables to the overall microbial community composition. Batch-correction methods were applied to KEGG abundances prior to distance calculations. The plot on the left uses ORR as the outcome variable, whereas the plot on the right adopts PFS12. (b) Prediction matrix for microbiome-based prediction of response assessed via ORR (left matrix) and PFS12 (right matrix) within each single cohort (values on the diagonal), across pairs of cohorts (one cohort used to train the model and the other for testing), and in leave-one-cohort-out setting (training the model on all but one cohort and testing on the left-out cohort). We report the AUC-ROC values obtained from LASSO models on KEGG relative abundances (top) and level 4 enzyme categories (bottom). Values on the diagonal refer to the median AUC-ROC values of 100-repeated fivefold stratified cross-validations. Off-diagonal values refer to AUC-ROC values obtained by training the classifier on the cohort of the corresponding row and applying it to the cohort of the corresponding column. The leave-one-out row refers to the performances obtained by training the model using all but the cohort of the corresponding column and applying it to the cohort of the corresponding column.
Extended Data Fig. 4 Cross-cohort response–microbiome associations at the functional level. (a) Contribution of variables to the overall microbial community composition. Batch-correction methods were applied to KEGG abundances prior to distance calculations. The plot on the left uses ORR as the outcome variable, whereas the plot on the right adopts PFS12. (b) Prediction matrix for microbiome-based prediction of response assessed via ORR (left matrix) and PFS12 (right matrix) within each single cohort (values on the diagonal), across pairs of cohorts (one cohort used to train the model and the other for testing), and in leave-one-cohort-out setting (training the model on all but one cohort and testing on the left-out cohort). We report the AUC-ROC values obtained from LASSO models on KEGG relative abundances (top) and level 4 enzyme categories (bottom). Values on the diagonal refer to the median AUC-ROC values of 100-repeated fivefold stratified cross-validations. Off-diagonal values refer to AUC-ROC values obtained by training the classifier on the cohort of the corresponding row and applying it to the cohort of the corresponding column. The leave-one-out row refers to the performances obtained by training the model using all but the cohort of the corresponding column and applying it to the cohort of the corresponding column.CC BY · PubMed Central
Extended Data Fig. 5 Machine learning association analysis using random forest. (a) Prediction matrix for microbiome-based prediction of response assessed via ORR (left matrix) and PFS12 (right matrix) within each single cohort (values on the diagonal), across pairs of cohorts (one cohort used to train the model and the other for testing), and in leave-one-cohort-out setting (training the model on all but one cohort and testing on the left-out cohort). We report the AUC-ROC values obtained from Random Forest models on species-level relative abundances. Values on the diagonal refer to the median AUC-ROC values of 100-repeated fivefold stratified cross-validations. Off-diagonal values refer to AUC-ROC values obtained by training the classifier on the cohort of the corresponding row and applying it to the cohort of the corresponding column. The leave-one-out row refers to the performances obtained by training the model using all but the cohort of the corresponding column and applying it to the cohort of the corresponding column. (b) Cross-validation AUC-ROC values obtained from Random Forest models trained using 100-repeated fivefold stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from Random Forest models obtained by training the model using species-level relative abundances and all but the corresponding PRIMM cohort (circles). PRIMM-NL (n = 55) and PRIMM-UK (n = 55). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than ×1.5 interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles.
Extended Data Fig. 5 Machine learning association analysis using random forest. (a) Prediction matrix for microbiome-based prediction of response assessed via ORR (left matrix) and PFS12 (right matrix) within each single cohort (values on the diagonal), across pairs of cohorts (one cohort used to train the model and the other for testing), and in leave-one-cohort-out setting (training the model on all but one cohort and testing on the left-out cohort). We report the AUC-ROC values obtained from Random Forest models on species-level relative abundances. Values on the diagonal refer to the median AUC-ROC values of 100-repeated fivefold stratified cross-validations. Off-diagonal values refer to AUC-ROC values obtained by training the classifier on the cohort of the corresponding row and applying it to the cohort of the corresponding column. The leave-one-out row refers to the performances obtained by training the model using all but the cohort of the corresponding column and applying it to the cohort of the corresponding column. (b) Cross-validation AUC-ROC values obtained from Random Forest models trained using 100-repeated fivefold stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from Random Forest models obtained by training the model using species-level relative abundances and all but the corresponding PRIMM cohort (circles). PRIMM-NL (n = 55) and PRIMM-UK (n = 55). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than ×1.5 interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles.CC BY · PubMed Central
Extended Data Fig. 6 Taxonomic overview of species associations with response. Cladogram showing species associated with Responders (red) and nonresponders (blue) using ORR and identified by a minimum of 2 meta-analysis methods. The height of the outer bar plots reflects the number of meta-analysis methods supporting the association.
Extended Data Fig. 6 Taxonomic overview of species associations with response. Cladogram showing species associated with Responders (red) and nonresponders (blue) using ORR and identified by a minimum of 2 meta-analysis methods. The height of the outer bar plots reflects the number of meta-analysis methods supporting the association.CC BY · PubMed Central
Extended Data Fig. 7 Reproducible biomarkers for colorectal cancer across cohorts. Species associated with CRC identified by a meta-analysis using different differential abundance methods. Species shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of patients with CRC and controls.
Extended Data Fig. 7 Reproducible biomarkers for colorectal cancer across cohorts. Species associated with CRC identified by a meta-analysis using different differential abundance methods. Species shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of patients with CRC and controls.CC BY · PubMed Central
Extended Data Fig. 8 Microbiome biomarkers of response across cohorts. (a) KEGG orthologues associated with PFS12 identified by a meta-analysis using different differential abundance methods. KEGGs shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. (b) Level 4 enzyme categories associated with PFS12 identified by a meta-analysis using different differential abundance methods. ECs shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. (c) Species associated with PFS12 in the two PRIMM cohorts before and after adjusting for confounders that included PPI, antibiotic and steroid use, gender, performance status, previous therapy, age and ICI. PRIMM-NL (n = 47) and PRIMM-UK (n = 52). Species shown have covariate-adjusted multiple hypothesis testing-corrected q < 0.2 in one of the cohorts identified by ANCOM-BC. Symbols (circles and triangles) show the ANCOM-BC beta coefficient and error lines represent the standard error. (d) Level 4 enzyme categories associated with ORR identified by a meta-analysis using different differential abundance methods. ECs shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders.
Extended Data Fig. 8 Microbiome biomarkers of response across cohorts. (a) KEGG orthologues associated with PFS12 identified by a meta-analysis using different differential abundance methods. KEGGs shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. (b) Level 4 enzyme categories associated with PFS12 identified by a meta-analysis using different differential abundance methods. ECs shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders. (c) Species associated with PFS12 in the two PRIMM cohorts before and after adjusting for confounders that included PPI, antibiotic and steroid use, gender, performance status, previous therapy, age and ICI. PRIMM-NL (n = 47) and PRIMM-UK (n = 52). Species shown have covariate-adjusted multiple hypothesis testing-corrected q < 0.2 in one of the cohorts identified by ANCOM-BC. Symbols (circles and triangles) show the ANCOM-BC beta coefficient and error lines represent the standard error. (d) Level 4 enzyme categories associated with ORR identified by a meta-analysis using different differential abundance methods. ECs shown have random-effects model p values < 0.05 in at least 6 methods out of 8 methods. Values inside the cells refer to unadjusted p values < 0.05 obtained by two-tailed Wilcoxon tests on differences in the relative abundance of responders and nonresponders.CC BY · PubMed Central
Extended Data Fig. 9 Clinical parameters associated with response and the microbiome. (a) Forest plot showing Cox logistic regression multivariate analysis of progression-free survival. Error lines represent the 95% confidence interval of the hazard ratio. (b) Cross-validation AUC-ROC values obtained from LASSO models trained using 100-repeated fivefold stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from LASSO models obtained by training the model using KEGG relative abundances and all but the corresponding PRIMM cohort (circles). PRIMM-NL (n = 55) and PRIMM-UK (n = 55). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than ×1.5 interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles.
Extended Data Fig. 9 Clinical parameters associated with response and the microbiome. (a) Forest plot showing Cox logistic regression multivariate analysis of progression-free survival. Error lines represent the 95% confidence interval of the hazard ratio. (b) Cross-validation AUC-ROC values obtained from LASSO models trained using 100-repeated fivefold stratified cross-validations (boxplots) and leave-one-dataset-out AUC-ROC values from LASSO models obtained by training the model using KEGG relative abundances and all but the corresponding PRIMM cohort (circles). PRIMM-NL (n = 55) and PRIMM-UK (n = 55). The lower and upper hinges of boxplots correspond to the 25th and 75th percentiles, respectively. The midline is the median. The upper and lower whiskers extend from the hinges to the largest (or smallest) value no further than ×1.5 interquartile range from the hinge, defined as the distance between the 25th and 75th percentiles.CC BY · PubMed Central

Verification

Claims the paper makes

6 Valid · 2 Not valid · 1 Not checked · 70 Own results and interpretations

23 more claims

Create a free account to see all 79 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 24
  • It may state two separate claims as one. 34

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

22 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Citations

Cites

Show all 70 references

Cited by

See all 464

Metadata

Authors
Lee, K. A., Thomas, A. M., Bolte, L. A., Björk, J. R., de Ruijter, L. K., Armanini, F., Asnicar, F., Blanco-Miguez, A., Board, R., Calbet-Llopart, N., Derosa, L., Dhomen, N., Brooks, K., Harland, M., Harries, M., Leeming, E. R., Lorigan, P., Manghi, P., Marais, R., Newton-Bishop, J., Nezi, L., Pinto, F., Potrony, M., Puig, S., Serra-Bellver, P., Shaw, H. M., Tamburini, S., Valpione, S., Vijay, A., Waldron, L., Zitvogel, L., Zolfo, M., de Vries, E. G. E., Nathan, P., Fehrmann, R. S. N., Bataille, V., Hospers, G. A. P., Spector, T. D., Weersma, R. K., & Segata, N.
Affiliations
  • King's College London (GB)
  • University of Trento (IT)
  • University Medical Center Groningen (NL)
  • Lancashire Teaching Hospitals NHS Foundation Trust (GB)
  • Inserm (FR)
  • University of Manchester (GB)
  • University of Leeds (GB)
  • Guy's Hospital (GB)
  • European Institute of Oncology (IT)
  • Instituto de Salud Carlos III (ES)
  • The Christie NHS Foundation Trust (GB)
  • Mount Vernon Cancer Centre (GB)
  • University of Nottingham (GB)
  • City University of New York (US)
  • Hospital Clínic de Barcelona (ES)
Corresponding author
  • Spector, Tim D.
  • Weersma, Rinse K.
  • Segata, Nicola
Venue
Nature Medicine
Published
2022-02-28
Volume
28
Issue
3
Pages
535-544
ISSN
  • 1078-8956 (ISSN-L)
  • 1546-170X
OpenAlex
W4214673864
Cited by
464
References
70
Citations by year
2026: 102 · 2025: 113 · 2024: 104 · 2023: 99 · 2022: 46
FWCI
38.06
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Research Explorer (The University of Manchester) (publishedVersion, other-oa)
  • Repository@Nottingham (University of Nottingham) (submittedVersion, cc-by)
  • Iris (University of Trento) (submittedVersion, cc-by)
  • Research Portal (King's College London) (submittedVersion, cc-by)
  • HAL (submittedVersion, cc-by)
  • ARCA (Università Ca' Foscari Venezia) (submittedVersion, other-oa)
  • University of Groningen research database (University of Groningen / Centre for Information Technology) (submittedVersion, cc-by)
Funders
  • Wellcome Trust (100282, 100282/Z/12/Z)
  • Seerave Foundation (825410)
  • Cancer Research UK (A27412 and A22902, A22902, A27412, 10589, 19167, 11963, 22902)
  • Ministero dell’Istruzione, dell’Università e della Ricerca (H2020, RBFR13EWWI_001)
  • Università degli Studi di Trento
  • KWF Kankerbestrijding (10034 POINTING, 10034)
  • Institut Gustave-Roussy
  • National Institutes of Health (1u01ca230551-01a1, H2020, 1U01CA230551)
  • Horizon 2020 Framework Programme (MASTER-818368, OncoBiome 825410, 818368, ONCOBIOME-825410 project, MASTER-818368 project, 825410)
  • Medical Research Council (MR/MO19012/1, MR/M019012/1)
  • Instituto de Salud Carlos III (PI15/00716, PI15/00956, H2020)
  • European Regional Development Fund (H2020)
  • National Cancer Institute (1U01CA230551, U01 CA230551, R01CA230551)
Topics
  • Health Sciences › Medicine › Oncology › Cancer Immunotherapy and Biomarkers
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gut microbiota and health
  • Health Sciences › Medicine › Oncology › Cutaneous Melanoma Detection and Management
Keywords
gut microbiota, immune checkpoint inhibitors, advanced melanoma, treatment response prediction, Bifidobacterium pseudocatenulatum, Roseburia, Akkermansia muciniphila, microbiome biomarkers, overall response rate, progression-free survival, immunotherapy response
MeSH terms
Gastrointestinal Microbiome, Immune Checkpoint Inhibitors, Humans, Melanoma, Skin Neoplasms, Reproducibility of Results

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Cross-cohort gut microbiome associations with immune checkpoint inhibitor response in advanced melanoma. https://app.cochat.ai/papers/10.1038/s41591-022-01695-5
The paper
Lee, K. A., Thomas, A. M., Bolte, L. A., Björk, J. R., de Ruijter, L. K., Armanini, F., Asnicar, F., Blanco-Miguez, A., Board, R., Calbet-Llopart, N., Derosa, L., Dhomen, N., Brooks, K., Harland, M., Harries, M., Leeming, E. R., Lorigan, P., Manghi, P., Marais, R., . . . Segata, N. (2022). Cross-cohort gut microbiome associations with immune checkpoint inhibitor response in advanced melanoma. Nature Medicine. https://doi.org/10.1038/s41591-022-01695-5

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.