A gut microbial signature for combination immune checkpoint blockade across cancer types

Immune checkpoint blockade (ICB) with agents targeting CTLA-4 and PD-1/PD-L1 has transformed cancer therapy across multiple tumor types, yet responses are variable and often unpredictable, with higher toxicity for combination regimens. Existing tumor-agnostic biomarkers (for example, tumor mutational burden, mismatch repair deficiency) require tumor tissue and have limitations. The gut microbiome has emerged as a promising tumor-extrinsic biomarker, with prior studies linking specific species such as Akkermansia muciniphila and Faecalibacterium prausnitzii to response to anti-PD-1. However, cross-cohort reproducibility has been poor, potentially because species-level profiling masks functionally relevant strain-level differences. The study aims to test whether strain-resolved gut microbiome features at baseline predict response to combination anti-PD-1 plus anti-CTLA-4 and whether such signatures generalize across cancer types and cohorts, and to determine if microbiome-response signatures are regimen specific (combination versus anti-PD-1 monotherapy).

Previous clinical microbiome studies using 16S rRNA or shotgun metagenomics reported associations between baseline abundances of species like Akkermansia muciniphila (notably in NSCLC) and Faecalibacterium prausnitzii (notably in melanoma) with response to anti-PD-1 therapy. Nonetheless, meta-analyses across studies have shown limited reproducibility of species-level biomarkers, attributable to geographic and methodological heterogeneity and the functional heterogeneity at the strain level within species. There is also growing recognition of substantial intraspecies genomic and functional diversity in key commensals (e.g., Faecalibacterium complex, Akkermansia clade structure), which may explain discordant associations seen at higher taxonomic ranks. Reviews proposing cross-cancer microbiome signatures have largely operated at species resolution and may have conflated different ICB regimens, complicating generalization.

Design and cohort: Prospective, multicenter, single-arm phase 2 clinical trial (CA209-538; NCT02923934) of combination ipilimumab (anti-CTLA-4) plus nivolumab (anti-PD-1) in rare cancers, recruiting 120 adults across five Australian sites. Baseline fecal samples were collected from 106 participants immediately prior to therapy (day −7 to day 0) using OMR-200 kits. Clinical data included demographics, tumor histology (UGB, NEN, GYN), medications, and labs (albumin, NLR, platelets, LDH). Imaging followed RECIST 1.1 at scheduled intervals. Primary outcome for microbiome ML was objective response versus progression (RvsP: CR/PR vs PD/cPD); SD (n=29) excluded for primary endpoint. A sensitivity endpoint was 12-month PFS (PFS12). Sequencing and genome-resolved reference construction: Deep shotgun metagenomic sequencing on NovaSeq 6000 (2×150 bp) yielded median 20.4M clean paired-end reads per sample. Reads were human-decontaminated and quality-controlled (trim-galore, BMTagger). MAG assembly used metaSPAdes and MEGAHIT, followed by binning (MetaBAT2, MaxBin2, CONCOCT), refinement (metaWRAP), and stringent QC (CheckM2, GUNC), producing 2,209 near-complete MAGs. A hybrid custom strain reference database (1,397 genomes) was generated by dereplicating study MAGs with relevant GTDB species reference genomes at 98% ANI (dRep). Taxonomy assigned using GTDB-tk. Read mapping and abundance estimation: Bowtie2 aligned reads to the custom strain database; inStrain performed stringent filtering (min_read_ani 0.95; min_genome_coverage 1), with low-breadth mappings (<50%) removed to avoid spurious mappings. Median mapping rate to reference library was high (≈88%); ≈50% of clean reads were retained for strain quantification after QC. Decontam identified and removed potential contaminants. Microbiome analyses: Alpha diversity (Shannon) and beta diversity (Aitchison distance of CLR-transformed abundances) assessed; PERMANOVA used to test associations with metadata and outcomes. Phylogenetic trees were built via GTDB-tk and FastTree; Faecalibacterium clade structure assessed by FastANI/rapidNJ. Functional genomics: Virulence factor screening with abricate/VFDB; metabolic potential profiling with gapseq. Butyrate-production potential (acetyl-CoA pathway genes thl, bhbd, cro, bcd, but, buk) assessed using HMM profiles; sample-level gene abundances estimated via read mapping. Machine learning: Supervised ML with random forest classifiers (sklearn, imblearn) predicting RvsP (primary) and PFS12 (sensitivity). Feature sets included clinical variables alone (15 features), microbiome abundances at multiple taxonomic ranks (family, genus, species, strain via CLR transformation), and combined features. Extensive hyperparameter tuning via random search with 1,000 combinations and 20× repeated 5-fold CV (100 splits) per feature set; ROC AUC used for selection. Class imbalance addressed by random oversampling in training only. Feature importance interpreted with SHAP TreeExplainer, repeating 1,000 times to stabilize estimates. Leave-one-histology-cohort-out validation tested cross-cancer generalization within CA209-538. Meta-analysis external validation: Six additional shotgun metagenomic cohorts (total n=470; n=383 after excluding SD) from USA, UK, Netherlands, Spain, and Australia were uniformly reprocessed (quality control, mapping to same strain database). Cohorts were stratified by ICB regimen (CICB vs anti-PD-1 monotherapy). External testing evaluated the CA209-538-trained strain-RvsP model on each cohort and performed all pairwise train–test evaluations, comparing regimen-concordant vs -discordant performance. Batch correction beyond CLR was not applied to avoid potential overcorrection and data leakage; technical/geographic contributions to variance were quantified via PERMANOVA.

- Clinical and compositional associations: - Objective response rates in CA209-538 were similar across histologies (24–25%). BOR strongly associated with PFS and OS (log-rank P < 0.0001 for both). - Albumin positively and NLR negatively associated with BOR (Kendall P=0.0056 and P=0.0033, respectively), largely distinguishing rapid clinical progression (cPD). - Alpha diversity (Shannon) positively associated with BOR (τ=0.22, P=0.003). Beta diversity differed by BOR (PERMANOVA P=0.0319, R²=0.043). - Machine learning performance (discovery cohort): - Clinical features alone were weak for RvsP (mean AUC=0.56) but better for PFS12 (AUC≈0.65). - Microbiome features improved with increasing taxonomic resolution; strain-level abundances performed best (AUC=0.73 for RvsP; AUC=0.70 for PFS12), significantly outperforming species-level models. - Adding clinical features did not improve microbiome-only models. - Leave-one-histology-cohort-out validation showed tumor-agnostic performance: mean left-out AUC=0.75 (UGB AUC=0.81; GYN AUC=0.81; NEN AUC=0.64). - RvsP predicted probabilities tracked actual BOR ordering (Kendall P<2.2×10⁻¹⁶) and suggested a trend toward better OS among SD patients with higher predicted response (log-rank P=0.17; underpowered). - Signature taxa and functional insights: - SHAP identified 22 influential strains; most were Firmicutes (Bacillota) and Gram positive. Positive contributors clustered within Ruminococcaceae, notably several Faecalibacterium strains near the F. prausnitzii D clade. - Negative contributors included strains from Lachnospiraceae, Oscillospiraceae, Ruminococcaceae, Bifidobacterium dentium, Akkermansia muciniphila B (distinct from A. muciniphila), and Spyradocola merdavium. - The acetyl-CoA butyrate pathway was complete in all five positive Ruminococcaceae and absent in the two negative Ruminococcaceae strains. However, a strain-agnostic abundance of terminal enzymes (but + buk) was not significantly enriched in responders, underscoring the importance of strain-aware analyses. - External validation and regimen specificity: - Across external cohorts, major drivers of microbial variance included city (9.3%) and DNA extraction kit (8.0%), reflecting technical/geographic heterogeneity. - The CA209-538 strain-RvsP model generalized modestly to CICB cohorts (mean external AUC≈0.65) but not to anti-PD-1 monotherapy cohorts (mean AUC≈0.51). - Pairwise cross-cohort analyses demonstrated significantly higher AUCs when training and testing on regimen-concordant cohorts versus discordant cohorts (Mann–Whitney U P=2.8×10⁻⁷), a pattern observed for both CICB and anti-PD-1 models.

Strain-resolved gut microbial profiling provided superior predictive performance for CICB response compared with species-level profiling and clinical variables, and the learned signatures generalized across distinct cancer histologies within the discovery trial, supporting a tumor-agnostic, host–microbiome interaction. The identification of Faecalibacterium and Ruminococcaceae strains as positive predictors and certain Lachnospiraceae and other taxa as negative predictors aligns with prior observations linking beneficial short-chain fatty acid producers to anti-tumor immunity, while highlighting substantial strain-level heterogeneity even within broadly ‘favorable’ families. The complete acetyl-CoA butyrate pathway selectively present in positive Ruminococcaceae, but not captured by strain-agnostic gene abundance, illustrates how strain-level context is necessary to construct mechanistically plausible hypotheses. Critically, cross-cohort analyses revealed regimen specificity: models trained on CICB cohorts performed best when tested on CICB, and similarly for anti-PD-1 monotherapy. This suggests distinct microbiota–host relationships underpin anti-PD-1 monotherapy versus combination anti-PD-1 plus anti-CTLA-4, consistent with known differences in mechanisms and baseline tumor immune microenvironments, and with CTLA-4 blockade’s effects on gut barrier permeability. These findings argue for developing regimen-specific microbiome biomarkers and therapeutic strategies. The ability of RvsP-trained models to rank BOR categories and to stratify SD patients by outcome indicates potential clinical utility in identifying patients with durable disease control, pending further validation.

This study demonstrates that baseline, strain-resolved gut microbiome signatures predict response to combination ipilimumab plus nivolumab across diverse rare cancers and outperform species-level and clinical-feature models. The signatures generalize across cancer types and geographies but show strong regimen specificity, with substantially better cross-cohort performance when training and testing within the same ICB regimen. Functionally, positive predictor strains cluster within Ruminococcaceae/Faecalibacterium and exhibit complete butyrate-production pathways, suggesting mechanistic links to T cell activation. These results support prioritizing strain-level profiling and regimen-specific biomarker development. Future work should include larger, standardized, geographically diverse cohorts; harmonized sampling and extraction protocols; culture-based strain isolation and genome sequencing; and in vitro/in vivo mechanistic studies to validate causality and enable targeted microbiome therapeutics as adjuncts to ICB.

- Sample sizes, while relatively large for CICB microbiome studies, remain modest overall, limiting power and the stability of feature selection. - Technical and geographic heterogeneity (collection/extraction kits, study sites) contributed substantially to variance and may confound cross-cohort comparisons; batch correction was avoided to prevent artificial inflation and data leakage. - MAG-based references, although stringently curated, can contain assembly artifacts (fragmentation, contamination) and may not capture all strains present in external cohorts. - Lack of culture-derived, patient-specific strain collections precluded functional validation and causal inference. - Generalizability to anti-PD-1 monotherapy cohorts was poor for the CICB-trained signature, reinforcing regimen specificity but limiting immediate clinical utility across regimens.