The Promise and the Problem Early studies suggested that specific gut bacteria were associated with better responses to checkpoint blockade in cancer patients, raising hope that microbiome profiling could predict who would benefit from treatment. However, different studies identified different bacterial species as beneficial, and no consistent signature had emerged across patient cohorts.
Study Design and Scale This study assembled shotgun metagenomic sequencing data from stool samples collected before treatment from 165 patients across five observational cohorts in the UK, Netherlands, and Spain - including two prospectively recruited PRIMM cohorts with over 50 patients each, both larger than all previously published datasets. These were combined with 147 samples from five existing published datasets for cross-cohort analysis.
Machine Learning Approach A Lasso-based machine learning framework was used to estimate whether combinations of microbial species abundances and functional gene families could predict checkpoint blockade response (overall response rate) and progression-free survival. This approach went beyond simple presence/absence comparisons to model the predictive value of the entire microbiome profile simultaneously.
Key Finding: Cohort Dependence The microbiome does have a statistically detectable association with immunotherapy response in melanoma, but the strength and specific features of this association are substantially cohort-dependent. No single species serves as a fully consistent biomarker across all study populations, pointing to a more complex relationship than previously thought.
Shotgun Metagenomic Sequencing Stool DNA was extracted and subjected to shotgun metagenomic sequencing at an average depth of 7.74 gigabases per sample - deeper than typical microbiome studies. This approach sequences all DNA in the sample rather than targeting only bacterial 16S rRNA genes, enabling species-level identification and functional gene profiling using the bioBakery 3 analysis pipeline.
PRIMM Cohorts: Prospective Design The two PRIMM cohorts (UK and Netherlands) were the backbone of the analysis. Both enrolled previously treatment-naive advanced melanoma patients and collected stool, serum, and blood samples before and during ICI treatment alongside detailed clinical and dietary data. Response was assessed using standardized RECIST v1.1 criteria at 6 and 12 months.
Confounding Factor Control A notable strength was the collection of detailed metadata to control for confounding factors: body mass index, age, sex, prior treatments, antibiotic use, proton pump inhibitor use, steroid use, and dietary patterns from food frequency questionnaires converted into Mediterranean diet and plant-based diet indices. Multivariate analyses accounting for these factors confirmed that microbiome-response associations were not simply artifacts of clinical differences between responders and non-responders.
Cross-Cohort Validation Machine learning models trained on one cohort were tested on others to assess whether microbiome signatures could generalize. This approach is more stringent than internal cross-validation and better reflects how a clinical biomarker would perform in practice when deployed in a new patient population.
PRIMM-UK Shows Stronger Associations than PRIMM-NL Overall microbiome composition differed between responders and non-responders in PRIMM-UK (borderline significance, PERMANOVA P = 0.05) but not in PRIMM-NL (P = 0.61). Multivariate analysis confirmed that in PRIMM-UK, response was the dominant variable explaining microbiome variance, while in PRIMM-NL, clinical factors such as proton pump inhibitor use, sex, and prior therapy explained more variance - potentially masking microbiome-response associations.
Machine Learning Prediction Performance Lasso-based models using species-level features achieved AUC-ROC values of 0.78 for overall response in PRIMM-UK and 0.64 for 12-month progression-free survival in PRIMM-NL, but only 0.53 and 0.57 for the swapped endpoints, demonstrating that the same microbiome features do not equally predict both endpoints or generalize across cohorts.
Functional Gene Families Add Predictive Power Models using predicted microbial functional gene abundances (KEGG orthologs) showed more consistent prediction across cohorts and endpoints, with AUC values exceeding 0.59 across all comparisons. This suggests that microbiome functional capacity may be a more reproducible predictor than taxonomic composition alone.
Alpha Diversity Not Consistently Predictive Overall microbiome diversity (alpha diversity as measured by Shannon index) was generally not associated with immunotherapy response after accounting for confounding factors, with only one cohort-endpoint combination reaching statistical significance. This challenges earlier reports that simply higher microbial diversity predicts better response.
Partially Consistent Species Associations Across the integrated multi-cohort analysis, several species were more commonly associated with responders: Bifidobacterium pseudocatenulatum, Roseburia intestinalis, Roseburia inulinivorans, and Akkermansia muciniphila appeared in multiple cohorts as enriched in responders, consistent with some prior reports. However, none of these associations was statistically significant across all cohorts after correcting for multiple testing.
Limited Cross-Cohort Reproducibility of Signatures When machine learning models trained on one dataset were applied to others, predictive performance dropped substantially compared to within-cohort cross-validation. This finding - that microbiome signatures identified in one population do not reliably transfer to another - is a central cautionary conclusion of the study.
Dietary and Clinical Confounders Are Important Significant differences in dietary patterns between UK and Dutch cohorts, as well as differences in treatment regimens (combination vs. single-agent ICI) and BRAF mutation rates, likely contribute to cohort-specific microbiome compositions. The microbiome signature of a responder in the UK may reflect different dietary and demographic background than a responder in the Netherlands, making a single universal signature unrealistic.
Comparison with Published Cohorts Analysis of the five publicly available published datasets alongside the new cohorts confirmed inconsistency in which specific species were enriched in responders. Species highlighted in prior publications - including Faecalibacterium prausnitzii, Bacteroides caccae, and Bifidobacterium longum - did not consistently replicate across all datasets in this analysis.
Biomarker Development Requires Larger Studies The finding that microbiome-response associations are cohort-dependent argues against clinical deployment of microbiome biomarkers based on small single-cohort studies. Robust biomarker development will require prospective multi-national cohorts with standardized sample collection, processing, and microbiome analysis protocols to identify features that are truly biology-driven rather than artifacts of population or methodological differences.
Fecal Microbiota Transplant Implications Randomized trials of fecal microbiota transplantation (FMT) from immunotherapy responders to non-responders have been initiated partly based on earlier microbiome-response associations. The current study does not contradict the biological plausibility of this approach, but suggests that identifying the right donor microbiome composition is more complex than simply selecting for high abundance of one or two species.
Microbiome as Part of a Multi-Factor Prediction Model The data support the view that the gut microbiome is one of several factors influencing immunotherapy response, alongside tumor mutational burden, PD-L1 expression, CD8 T cell infiltration, and clinical features. Future prediction models should incorporate microbiome data alongside these established factors rather than treating the microbiome as a standalone biomarker.
Diet and Medications as Confounders and Targets Because diet strongly shapes microbiome composition, and because dietary patterns differed between cohorts, dietary intervention may be a modifiable factor that influences immunotherapy response through the microbiome. Prospective studies pairing dietary standardization with ICI treatment and microbiome monitoring could disentangle this relationship.
Standardized Multi-Center Studies The authors recommend future studies adopt standardized collection protocols, unified bioinformatics pipelines, and larger sample sizes to enable more reliable cross-cohort comparisons. International consortia modeled on the PRIMM study design - prospective, multi-site, with harmonized clinical data collection - represent the most promising path forward.
Longitudinal Sampling The current study used pre-treatment baseline samples. Microbiome composition changes dynamically with ICI treatment, dietary changes, and clinical status. Longitudinal sampling before, during, and after treatment could reveal whether changes in specific species or functional pathways during treatment predict durable response better than baseline composition alone.
Mechanistic Studies To move from correlation to causation, mouse experiments with controlled microbiome compositions, germ-free models, and FMT experiments are needed to determine which specific microbial species or metabolites causally influence CD8 T cell activation, tumor infiltration, and checkpoint inhibitor efficacy.
Integration with Other Omics Combining metagenomics with metabolomics of circulating microbial metabolites, transcriptomics of peripheral blood immune cells, and proteomics of stool samples could identify the molecular pathways through which specific microbiome features modulate systemic anti-tumor immunity - providing mechanistic targets for microbiome-based therapeutic interventions.