Immune checkpoint blockade (ICB) therapies, targeting proteins like CTLA-4 and PD-1, have revolutionized cancer treatment. Combination ICB (CICB), using both anti-PD-1 and anti-CTLA-4 agents, demonstrates synergistic antitumor activity and is a standard of care for several cancers. However, response rates vary widely (20-60%), and CICB often causes more severe side effects than monotherapy. Predictive biomarkers are needed to guide patient selection and optimize treatment. While tumor mutational burden and mismatch repair deficiency are approved biomarkers, they have limitations and require tumor tissue. The gut microbiome, the community of microorganisms in the gastrointestinal tract, offers a promising alternative. Previous research using 16S rRNA gene sequencing and shotgun metagenomics has linked specific bacterial species (e.g., *Akkermansia muciniphila*, *Faecalibacterium prausnitzii*) to ICB response, but these findings have lacked reproducibility across studies. This may be due to focusing on species-level analysis, which overlooks important strain-level variations in microbial function. This study aimed to investigate whether strain-resolved gut microbiome analysis could improve prediction of ICB response across different cancer types.

Numerous studies have explored the relationship between the gut microbiome and response to ICB. Early work demonstrated associations between specific bacterial species and clinical outcomes in various cancers, such as *A. muciniphila* in lung cancer and *F. prausnitzii* in melanoma. However, meta-analyses have revealed limited reproducibility of these findings across different studies and cohorts, suggesting that species-level analysis may not capture the functional nuances driving the observed associations. The growing recognition of intraspecies (strain) variation within commensal bacteria, with strains exhibiting differing functional potentials and host interactions, has highlighted the potential limitations of higher taxonomic resolution approaches. Thus, a strain-level analysis is necessary to capture the functional heterogeneity within microbial communities and improve the reliability of microbiome-based biomarkers for ICB response prediction.

This study utilized a unique, richly annotated phase 2 clinical trial cohort (CA209-538, n=106) of patients with diverse rare cancers treated with combination ipilimumab (anti-CTLA-4) and nivolumab (anti-PD-1). Baseline fecal samples were collected before treatment and subjected to deep shotgun metagenomic sequencing. A genome-resolved metagenomics approach was employed, involving the assembly of metagenome-assembled genomes (MAGs) and the creation of a study-specific strain reference database. This database included both MAGs and relevant Genome Taxonomy Database (GTDB) species reference genomes (SRGs), allowing for precise strain-level quantification of microbial abundances. Supervised machine learning (ML), using random forest classifiers, was used to predict objective response versus progression (RvsP) and 12-month progression-free survival (PFS12). The predictive performance of strain-level models was compared to models using species, genus, and family-level abundances, as well as models using clinical factors. A meta-analysis of six additional studies (n=364) was conducted to assess the cross-cancer and cross-country generalizability of the identified signatures. The analysis separated cohorts by ICB regimen (CICB vs. anti-PD-1 monotherapy) to investigate regimen-specific effects.

The study found that strain-resolved microbial abundances significantly improved the prediction of RvsP and PFS12 compared to higher taxonomic ranks or clinical factors alone. Clinical factors alone were poorly predictive of RvsP (AUC = 0.56) but were slightly more predictive of PFS12 (AUC = 0.65), suggesting their greater utility as prognostic markers. Strain-level models achieved AUCs of 0.73 for RvsP and 0.70 for PFS12. Notably, the predictive performance of the models improved with increasing taxonomic resolution, from family to strain level. Leave-one-group-out cross-validation within the discovery cohort demonstrated the robustness of the strain-level signatures across different cancer types. SHAP (SHapley Additive exPlanations) analysis identified a small set of key strains driving the predictions, with many belonging to the Firmicutes phylum, particularly within the *Faecalibacterium* genus. The meta-analysis revealed cross-cancer and cross-country validity of the strain-response signatures, but only when training and test cohorts used concordant ICB regimens (CICB or anti-PD-1 monotherapy). This regimen-specificity highlighted the importance of tailoring microbiome-based diagnostics and therapeutics to the specific ICB regimen.

This study demonstrates the significant value of strain-level resolution in gut microbiome analysis for predicting ICB response. The superior performance of strain-level models compared to species-level models or clinical factors underscores the importance of considering intraspecies variations in microbial function. The cross-cancer generalizability of the signatures, especially within concordant ICB regimens, supports the development of pan-cancer microbiome-based diagnostic tools. The striking difference in performance between concordant and discordant ICB regimens highlights the potential for distinct microbiota-host interactions to influence response to different treatment strategies. This suggests that future research should focus on developing regimen-specific microbiome-based biomarkers and therapies.

This large-scale study establishes the importance of strain-level resolution in analyzing the gut microbiome's influence on ICB response. Strain-resolved signatures showed promise in predicting response across cancers, especially when considering the specific ICB regimen. Further research should focus on larger, more diverse cohorts with standardized methodologies, validating these findings and investigating the underlying mechanisms linking specific strains to ICB efficacy. Ultimately, this work paves the way for personalized ICB therapy guided by individual gut microbiome profiles.

Despite the large discovery cohort and meta-analysis, the sample sizes remain relatively small, potentially limiting the statistical power and generalizability of the findings. While advanced bioinformatic techniques were employed, limitations associated with MAG assembly, such as potential errors in assembly and contamination, remain. Further, in vitro and in vivo testing of identified strains is necessary to confirm causality and elucidate the underlying mechanisms driving the observed associations. The study primarily focused on Western populations, limiting its generalizability to other geographical regions. Future studies should use standardized protocols for sample collection and processing to enhance reproducibility.