Introduction

Advancements in cancer immunotherapy have significantly improved survival rates for melanoma patients, particularly through the use of immune checkpoint inhibitors (ICIs) targeting programmed cell death protein 1 (PD-1) and cytotoxic T-lymphocyte-associated protein 4 (CTLA-4)1,2,3,4. However, clinical responses to ICIs remain inconsistent, suggesting that factors beyond tumor-intrinsic properties and host immune responses influence therapeutic efficacy2,5. Increasingly, research highlights the role of the gut microbiome in modulating immune responses, with specific microbial taxa and functional pathways associated with treatment outcomes6,7. Understanding these interactions is critical for developing microbiome-based strategies to enhance immunotherapy efficacy.

Early studies indicate that specific gut microbial taxa, such as Akkermansia muciniphila, Faecalibacterium prausnitzii, Bifidobacterium longum, and Bacteroides caccae, are enriched in ICI responders5,8,9,10,11. Additionally, fecal microbiota transplantation (FMT) from responders into non-responders can restore PD-1 blockade sensitivity, underscoring the causative role of the microbiome in immunomodulation12,13. Dietary factors have also emerged as a promising avenue, with higher fiber intake correlating with improved progression-free survival in melanoma patients treated with immunotherapy14.

Despite promising findings, research on microbiome-ICI interactions has produced inconsistent results, reflecting real‑world variability in study design, sample collection and handling, dietary and medication exposures, and sequencing techniques. Recognizing that this variability is inherent, our goal is not to eliminate it but to address and measure it: we apply uniform inclusion criteria and one processing pipeline, analyze cohorts separately, and then synthesize effects with random‑effects meta‑analysis while reporting heterogeneity. This framework increases precision where signals replicate across settings and clearly identifies contexts where findings diverge. Meta-analysis serves as a powerful tool to address these challenges, integrating data across multiple studies to enhance statistical power and uncover robust microbial signatures associated with ICI response. However, previous meta-analyses have been limited in scope, largely focusing on microbial taxa while overlooking biosynthetic gene clusters (BGCs) and metabolic pathways that may play critical roles in immunotherapy outcomes11,15,16. Moreover, these studies have analyzed a smaller number of cohorts, restricting their ability to achieve cross-study reproducibility and draw generalizable conclusions. Additionally, prior meta-analyses have not utilized the most up-to-date profiling tools, potentially missing key microbial signatures and functional elements relevant to ICI response.

Recognizing these limitations, we assembled and uniformly reprocessed publicly available whole‑metagenome shotgun datasets from melanoma patients treated with immune checkpoint inhibitors (ICIs). Our meta‑analysis spans 8 studies, 12 independent ICI cohorts (n = 441 patients; 482 stool samples) and 3 trials in which fecal microbiota transplantation (FMT) was combined with ICI (n = 43 patients; 281 longitudinal stool samples). To move beyond taxonomy only analyses, we integrated multiple layers of microbial information, species‑level profiles (MetaPhlAn 417), predicted biosynthetic gene clusters (BGCs), and functional pathways to examine differences between responders (R) and non‑responders (NR). Compared with MetaPhlAn 218, MetaPhlAn 4 leverages a substantially expanded reference genome catalog and improved marker selection, enabling more sensitive detection of low‑abundance and previously uncharacterized species. This allowed us to identify microbial features associated with immunotherapy response with improved taxonomic specificity and to highlight how these results may inform future multimodal therapeutic approaches in melanoma.

Methods

Definition of response to therapy

For each cohort, we harmonized a binary response variable (responder vs non‑responder) using the definitions reported in the original publications (Supplementary Tables S1–S2). In most ICI‑only cohorts, response status was based on best overall response according to RECIST v1.1 or iRECIST19: patients with complete or partial response and, where applicable, durable stable disease were classified as responders, whereas patients with progressive disease or short‑duration stable disease were classified as non‑responders. In Simpson et al. we used the INMC pathological response score at week 6 as the primary endpoint and mapped it to responder versus non‑responder categories. In the ICI + FMT trials, we used each trial’s definition of clinical benefit or response to the combined ICI + FMT regimen. We did not redefine response beyond these study‑specific criteria; all analyses used this harmonized binary variable.

Cohort identification, data collection and sample selection

We systematically searched the literature and public repositories for melanoma studies that combined immune checkpoint blockade with stool shotgun metagenomic sequencing (Fig. 1). Previous meta-analyses have included partially overlapping melanoma cohorts but differed in scope and stratification strategy (Table 1). Inclusion criteria were: (i) unresectable or metastatic melanoma treated with anti‑PD‑1, anti‑CTLA‑4 or their combination, with or without FMT; (ii) at least one stool sample collected before or around treatment initiation; (iii) available response annotations; and (iv) public or controlled‑access raw sequencing data. Eleven primary studies met these criteria, yielding 12 ICI‑only cohorts and 3 ICI + FMT cohorts (15 cohorts total). Sequencing reads were downloaded from the Sequence Read Archive and European Nucleotide Archive, or obtained from authors when required (Supplementary Tables S2–S4). Studies based on 16S rRNA sequencing, non‑stool specimens, or non‑melanoma indications were excluded.

Fig. 1: Overview of Study Design, Omics Analysis, and Workflow.
Fig. 1: Overview of Study Design, Omics Analysis, and Workflow.The alternative text for this image may have been generated using AI.
Full size image

a Clinical context of melanoma and immune checkpoint inhibitor (ICI) therapy. Melanoma arises in the skin and, in advanced disease, spreads via lymphatic dissemination to regional lymph nodes (stage III) and to distant organs (e.g., liver and lung). The table summarizes cohorts analyzed: 8 ICI studies (n = 441 patients; 482 stool samples) and 3 ICI + FMT trials (n = 43 patients contributing 281 longitudinal stool samples). b Multi‑omic features profiled from stool metagenomes: species‑level taxonomic composition (MetaPhlAn 4), functional pathways, and biosynthetic gene clusters (BGCs). c Harmonized pipeline: read QC and host filtering; mapping to curated reference genomes; batch‑effect correction; and multivariable association of microbial features with ICI outcome, comparing responders (R) vs non‑responders (NR). Created with BioRender.com.

Table 1 Melanoma metagenomic samples included in this study compared with recent meta‑analyses
Full size table

Sample selection and repeated measures

Sample selection and timing

In the ICI-only stratum, analyses were generally restricted to a single stool sample per individual collected before the first ICI dose. An exception was McCulloch et al., for which we followed the original study definition of the Pittsburgh early sample cohort and included samples collected before treatment or within 4 months of starting anti-PD-1; the original authors considered this window representative of baseline because the gut microbiome was stable early after anti-PD-1 initiation. Samples collected more than 4 months after anti-PD-1 initiation (the Pittsburgh late sample cohort) were excluded from our ICI-only meta-analysis. In the ICI + FMT stratum, both pre-FMT and post-FMT stool samples were included when available, because response was defined with respect to the combined ICI + FMT regimen; most samples in this stratum were collected after FMT.

Repeated measures

Because participants in the ICI + FMT stratum could contribute multiple samples over time, within-patient correlation was accounted for in the pooled ICI + FMT analysis. Cohort-specific feature tables were merged into a single feature-by-sample matrix paired with a harmonized metadata table, and associations were estimated using MMUPHin lm_meta with a compound Poisson linear model, response as the exposure, age and gender as fixed effects, and study and patient as random effects. This pooled ICI + FMT analysis was conducted separately from the ICI-only pre-treatment meta-analysis.

Glitza cohort handling

In the Glitza et al. cohort, 47 metagenomic samples were available from 6 unique patients, whereas response was annotated at the patient level. To maintain a one-to-one mapping between microbiome features and response labels and to avoid pseudo-replication, species-level relative abundances were averaged within patient across baseline-eligible samples, yielding one patient-level taxonomic profile per individual (n = 6) for meta-analysis.

Quality control, microbiome taxonomics profiling

For quality control purposes, we subjected the whole metagenomic shotgun sequencing reads from these samples to rigorous filtering. This process was executed using KneadData software20, which was applied to the raw reads against the GRCh38 and T2T-CHM1321 human read databases. This step resulted in the generation of trimmed reads, effectively cleaning human-originated contamination. Host-depleted reads were additionally processed with fastp to generate sequencing quality report22. Subsequently, the purified reads were processed through MetaPhlAn 4, utilizing the mpa_vJun23_CHOCOPhlAnSGB_202307 database, to classify and determine their relative abundance17.

Functional pathway profiling

We quantified community‑level functional potential using HUMAnN 3. Host‑depleted reads were mapped to the UniRef90 database to generate gene‑family profiles, which were then regrouped to MetaCyc metabolic pathways. HUMAnN output tables were renormalized to relative abundances. For pathway meta‑analysis, we used the unstratified (community‑level) pathway abundance tables.

Biosynthetic gene cluster profiling

In addition to identifying species and pathways to determine the presence of specific organisms, we aimed to understand their potential functional capabilities, particularly in producing bioactive compounds. Such insights reveal the chemical diversity and potential pharmacological applications within the microbial community. To achieve this, we ran antiSMASH (ANTIbiotics & Secondary Metabolite Analysis SHell) 7.0, a microbial genome mining tool for BGC identification and analysis23 and also employed BGCLens, a computational tool specifically designed for the accurate identification and quantification of the BGC24. BGCLens utilizes a Bayesian reassignment model to enable precise probabilistic inference based on MIBiG database 3.025, accounting for sequencing uncertainty and improving reading assignment to BGC. This robust approach ensures reliable detection and quantification of BGCs, even within the complexity of diverse microbial community structures. This provides high-resolution insights into the functional potential of microbial communities, paving the way for understanding their chemical and pharmacological diversity.

Beta diversity analysis and clustering

Because the included studies used different DNA extraction kits, we evaluated beta diversity to assess between-study similarity. We calculated Bray–Curtis dissimilarity on relative abundance data (non-binary) using the vegdist function in the vegan R package26. Study labels were extracted from sample identifiers to group samples by study. Pairwise mean dissimilarities were computed both within and between studies, with the results stored in a symmetric distance matrix. Hierarchical clustering was then performed on the study-level distance matrix using Ward’s method to minimize variance within clusters. The resulting dendrogram visualized the relationships among studies based on their beta diversity, highlighting ecological similarity in species composition.

Batch effect correction and meta-analysis

To minimize these batch effects, we implemented a standardized correction protocol targeting taxonomy, BGCs, and pathway profiling data. The effectiveness of this correction was evaluated by calculating R2 values, which quantify the proportion of variance in the dependent variable that is explained by the independent variables. In the context of batch effect correction, R² serves as a key metric, reflecting the extent to which unwanted technical variation has been mitigated. By comparing R2 values before and after correction, we assessed improvements in data quality and the reduction of batch-associated biases27,28.

Data from multiple studies were integrated using batch effect removal and meta-analysis. Batch correction was performed with the ‘adjust_batch’ function from the MMUPHin R package29, ensuring consistent interpretation across datasets. Statistical significance and the explanatory power of variables were assessed using permutational multivariate analysis of variance (PERMANOVA), executed via the ‘adonis’ function in the vegan R package.

For further analysis, R2 values were determined to evaluate both statistical significance and the variance explained by the models. To explore associations between response variables and feature abundances while accounting for batch effects, we conducted a meta-analysis using the ‘lm_meta’ function, derived from MaAsLin230. This analysis employed a compound Poisson linear model (CPLM31) with the response variable designated as the exposure and the dataset identifier included as both a batch variable and a random effect.

The input to the CPLM consisted of a batch-adjusted feature abundance matrix, with metadata specifying the response variable and dataset identifiers. The model did not incorporate additional transformations, normalization, or standardization and used “Non-Response” as the reference category for the response variable.

Cross-study predictive modeling and generalization analysis

To evaluate the ability of microbiome profiles to generalize across studies, we performed leave‑one‑dataset‑out (LODO) cross‑study modeling using the seven ICI‑only cohorts with complete taxa, pathway, and BGC profiles (Frankel, Gopalakrishnan, Lee, Matson, McCulloch, Simpson, and Spencer). The Glitza cohort was excluded because only 6 patient-level profiles were available after within-patient aggregation. For each LODO split, we trained a supervised classifier on the remaining cohorts and evaluated its performance (AUC‑ROC) on the held‑out cohort.

We compared two integration strategies. In the concatenated (early‑integration) approach, features from all three modalities were column‑bound into a single design matrix prior to model fitting. In the stacked (late‑integration) approach, we first trained separate modality‑specific models on taxa, pathways and BGCs and then used their predicted probabilities as inputs to a second‑stage model. All models were trained using cross‑validation restricted to the training cohorts, and full implementation details are provided in the accompanying GitHub repository.

To identify features that contributed consistently to prediction across cohorts, we derived a cross‑study feature importance score. For each LODO model we extracted the signed coefficient (for linear models) or feature importance and rescaled it to unit variance. These scores were then averaged across all training cohorts and integration strategies for each feature. The resulting cross‑study importance score reflects both the frequency with which a feature is selected and the consistency of its direction of association. We report the top ten taxa, pathways and BGCs with the highest absolute scores in Fig. 5b.

Ethics statement

This study used de-identified data from previously published studies, including both publicly available datasets and data obtained from authors with permission. Therefore, no ethical approval or informed consent was required or sought.

Statistics and reproducibility

All analyses were performed in R. Unless stated otherwise, statistical tests were two-sided. For species-, pathway-, and BGC-level associations with response (responder vs. non-responder), we fit Tweedie generalized linear models (CPLM) using the MMUPHin lm_meta framework with study specified as the batch variable. In the ICI-only stratum, study was included as a random effect, whereas in the ICI + FMT stratum both study and patient were included as random effects to account for repeated sampling (standardized effect sizes reported as coef/SE). Between-cohort heterogeneity was summarized using I², where applicable. Multiple testing was controlled using the Benjamini–Hochberg false discovery rate (FDR) procedure within each stratum/omics layer; q < 0.25 was used as the significance threshold. Microbiome beta-diversity and metadata variance explained were evaluated using PERMANOVA on Bray–Curtis dissimilarities (vegan/adonis; 999 permutations), with study handled as a blocking factor as specified. Batch effects were mitigated using MMUPHin (adjust_batch). Cross-study prediction used leave-one-dataset-out validation using IntegratedLearner32 with random forest as the base learner; performance was summarized by AUC–ROC on held-out cohorts. Sample sizes were determined by available public datasets (no a priori power calculation). No randomization or blinding was performed because this was a reanalysis of previously generated observational cohorts.

Results

Clarifying variance: moderate cross‑cohort batch effects after harmonized processing

We identified 15 cohorts across 11 studies (Supplementary Table S5) that met the inclusion criteria, which required a focus on melanoma patients undergoing ICI therapy targeting PD-1 and/or CTLA-4 and/or FMT, the use of gut microbiome profiling through shotgun metagenomics sequencing, and the availability of clinical response data stratified by responders and non-responders. Because we uniformly reprocessed the raw reads to derive both taxonomic and functional profiles, author‑provided profiles were not required and, when available, were used only for quality control and sensitivity checks. Despite meeting these criteria, the included studies exhibited variability in DNA extraction toolkits, which can significantly impact downstream analyses due to differences in DNA yield33,34.

To assess inter-study variability, we examined beta diversity (abundance‑based, non‑binary), which measures differences in microbial composition between samples and evaluated the similarity of microbial profiles across studies (Methods). Variability in microbial composition can arise from differences in patient populations, sample processing methods, sequencing techniques, or underlying biological heterogeneity, all of which can introduce batch effects that obscure true biological signals. By analyzing beta diversity metrics, we aimed to determine the extent of divergence between studies and identify clusters of studies with similar microbial communities. We applied this analysis to all 15 cohorts, including both ICI-only and ICI + FMT studies, and performed batch-effect correction for each stratum and for all available omics layers. Using the ICI-only taxonomy profiles as an example, studies such as ‘Spencer’, ‘Gop’, and ‘Matson’ clustered closely together before correction, suggesting high similarity in microbial composition, whereas ‘Glitza_ici’ and ‘Simpson’ formed a separate cluster, indicative of greater dissimilarity (Fig. 2a, left). The hierarchical clustering of the eight studies after batch effect correction demonstrates a reduction in inter-study dissimilarity, as evidenced by lower dissimilarity values (Fig. 2a, right).

Fig. 2: Beta diversity clustering and the effect of batch correction.
Fig. 2: Beta diversity clustering and the effect of batch correction.The alternative text for this image may have been generated using AI.
Full size image

a Hierarchical clustering of eight studies (across 12 cohorts) before(left) and after(right) batch effect correction, respectively. Before batch effect correction, distinct groupings influenced by batch effects were shown. The clustering pattern suggests that study-specific biases contribute significantly to the observed dissimilarities. After batch effect correction, clustering of the same studies after batch effect correction reveals a reduction in batch effects, with studies clustering more closely together. However, some degree of dissimilarity remains, suggesting that while the correction mitigates batch effects, it does not entirely eliminate them. Residual batch effects may still influence clustering patterns, likely due to inherent study-specific differences that persist despite correction. b Principal component analysis (PCA) of species-level profiles colored by study69. Pre-correction (Left), the data points cluster by study batch, showing greater batch-related variation. Post-correction (Right), batch effects are reduced, and the studies align more closely while still retaining meaningful biological signals.

Nevertheless, some study-specific structure persisted, with cohorts such as ‘Lee’ and ‘McCulloch’ still forming distinct clusters. This residual clustering suggests that, while the correction substantially mitigated inter-study variability, it did not eliminate all batch effects. The persistence of these patterns may reflect unmeasured confounders, experimental differences, or inherent biological variability between studies, highlighting the difficulty of fully normalizing batch effects in complex multi-cohort datasets and underscoring the need for cautious interpretation of inter-study comparisons even after correction. The differences in microbial community profiles across studies emphasized the need for subsequent batch (study) effect correction to mitigate potential inter-study variability. To this end, we applied the MMUPHin workflow and quantified the variance explained by batch effects by permutational multivariate analysis of variance (PERMANOVA). Continuing with the ICI-only taxonomy profiling example, PERMANOVA results before batch-effect correction (Fig. 2b, left) showed that dataset grouping explained 12.148% of the variance (R² = 0.12148, p = 0.001), indicating a moderate influence of study-specific effects. After batch-effect correction, the variance explained by dataset grouping decreased to 4.261% (R² = 0.04261, p = 0.001). This reduction demonstrates that batch correction effectively minimized inter-study variability, although a small residual effect remains. The corresponding increase in residual variance (from 87.852% to 95.739%) further supports the conclusion that most of the remaining variability is now attributable to biological or experimental noise rather than systematic batch effects. This reduction in cohort-driven variability highlights the success of the batch correction procedure in minimizing study-specific biases, allowing for a more accurate assessment of the underlying biological signals.

Species-level signatures differ between ICI and ICI + FMT cohorts

For species-level taxonomic profiling, we stratified the studies into ICI-only (k = 8) and ICI + FMT (k = 3) groups and fitted a compound Poisson linear model (CPLM) using the MMUPHin lm_meta framework to associate species abundance with response status (responders (R) vs. non-responders (NR)). In the ICI-only stratum, study was included as a random effect, whereas in the ICI + FMT stratum both study and patient were included as random effects. Species were considered significant at pval < 0.05 with low-to-moderate heterogeneity (I² < 50%, except for Holdemanella porci is 72% and Erysipelatoclostridium ramosum is 55%), and those additionally passing FDR correction (q < 0.25) are marked with asterisks in Fig. 3a. For the ICI-only stratum, all significant species are displayed; for the ICI + FMT stratum, which yielded many associations, we show the eight most responder-enriched and eight most non‑responder–enriched species ranked by standardized effect size (z = coef / s.e.).

Fig. 3: Taxonomic correlates of response across ICI and ICI + FMT cohorts.
Fig. 3: Taxonomic correlates of response across ICI and ICI + FMT cohorts.The alternative text for this image may have been generated using AI.
Full size image

a Standardized coefficients (z = coefficient / s.e.) for species associated with clinical response in ICI‑only cohorts (left, k = 8 studies) and ICI + FMT cohorts (right, k = 3 studies). Bars to the right of zero indicate enrichment in responders and bars to the left enrichment in non‑responders. Blue and salmon bars correspond to ICI‑only species favouring responders and non‑responders, respectively; teal and mustard bars show the same for ICI + FMT. The lower panel highlights species that overlap between the two meta‑analyses. b Longitudinal relative abundance of three overlapping species, including GGB9501 SGB14898, Pseudoflavonifractor capillosus and Clostridiaceae bacterium Marseille Q4143, in the Baruch, Davar and Bertrand ICI + FMT trials. Lines show mean relative abundance (%) in responders (blue) and non‑responders (salmon), with shaded bands indicating 95% confidence intervals. Davar is restricted to days 1-110, and Bertrand is summarized by treatment stages S1-S4. Legend entries (e.g. Responder (2 out of 3)) report the number of patients with detectable abundance of the species out of the total number of responders or non‑responders in each trial.

In the ICI‑only cohorts, several short‑chain-fatty‑acid-producing Clostridiales displayed positive standardized coefficients, indicating an association with response, including Roseburia sp., Dorea formicigenerans, and an unclassified Clostridiales bacterium (Fig. 3a, left). These taxa are broadly consistent with previous reports linking Dorea formicigenerans and other Lachnospiraceae species to improved outcomes or longer progression‑free survival in patients treated with PD‑1–based regimens35. In contrast, Hungatella hathewayi, Clostridium innocuum, and Erysipelatoclostridium ramosum showed negative standardized coefficients, indicating an association with non‑response; these taxa have been linked to dysbiosis, antibiotic exposure, and poorer ICI outcomes in other settings36. Together, these results suggest that, in ICI‑only therapy, gut communities characterized by higher modeled abundance of butyrate‑producing commensals and lower modeled abundance of potentially pathobiontic Firmicutes may favor clinical response.

The ICI + FMT stratum showed a distinct pattern (Fig. 3a, right). Among the top responder-associated taxa were multiple Alistipes and Bacteroides species, Holdemania filiformis, and several unclassified SGBs, many of which also belong to the Clostridiales and Bacteroidales orders. Notably, Holdemania filiformis has repeatedly been reported as enriched in responders to combined CTLA‑4/PD‑1 blockade, consistent with our findings in the FMT‑augmented setting35. Non‑responders in the ICI + FMT cohorts, by contrast, showed higher abundances of Lachnospira eligens, Bifidobacterium pseudocatenulatum, Anaerobutyricum hallii, Coprococcus eutactus, and related Firmicutes. Intriguingly, several of these taxa, including B. pseudocatenulatum and Coprococcus eutactus, have been associated with favorable responses or prolonged progression‑free survival in previous ICI-only studies and reviews6, yet in our ICI + FMT meta-analysis, they were linked to non‑response, highlighting that the same species can behave differently depending on treatment context and ecological background.

We next examined species that were significant in both strata. Three taxa, GGB9501 SGB14898, Clostridiaceae bacterium Marseille Q4143, and Pseudoflavonifractor capillosus, were shared between the ICI-only and ICI + FMT analyses but exhibited discordant directions of association (Fig. 3a, bottom). GGB9501 SGB14898 and P. capillosus were associated with non‑response in ICI-only cohorts yet favored responders in the ICI + FMT stratum, whereas the Clostridiaceae bacterium showed the opposite pattern. Although these species have been only sparsely discussed in the ICI literature, P. capillosus has been implicated in dietary and metabolic interventions that modulate anti‑tumor immunity, including ketogenic diet-based strategies and experimental approaches to enhance immunotherapy37. Such directionally opposite effects across treatment strata mirror recent longitudinal and multi‑cohort observations where species like Coprococcus eutactus show regimen‑dependent associations with progression‑free survival under monotherapy versus combination checkpoint blockade38.

Overall, these species-level meta-analyses reveal that there is no single universal “good” or “bad” bacterium for ICI response. Instead, overlapping sets of SCFA‑producing Clostridiales, Bacteroidales, and Actinobacteria are differentially associated with outcomes depending on whether patients receive ICI alone or ICI combined with FMT. The observation that several taxa previously linked to favorable ICI responses, such as Bifidobacterium pseudocatenulatum and Coprococcus eutactus, are enriched among non‑responders in the ICI + FMT group underscores the context‑dependent nature of microbiome–immunotherapy interactions and suggests that microbiome‑based interventions may reshape not only community composition but also the functional role of individual species in determining treatment outcome.

To further explore these discordant associations, we examined the longitudinal relative abundance of the three overlapping taxa in the ICI + FMT cohorts (k = 3; Baruch, Davar, and Bertrand) (Fig. 3b). In Baruch and Bertrand, GGB9501 SGB14898 and Pseudoflavonifractor capillosus were rare at baseline but showed a clear increase in relative abundance over time or treatment stage in responders, whereas non‑responders remained at consistently low levels. In Davar, these species also displayed intermittent spikes, primarily in responders. By contrast, Clostridiaceae bacterium Marseille Q4143 tended to be more abundant in non‑responders than responders across time in Baruch and at early time points in Davar, with little separation between groups in Bertrand. These temporal patterns support the meta‑analytic finding that the direction of association for these taxa is reversed between the ICI‑only and ICI + FMT strata, and illustrate how FMT-driven ecological shifts can cause the same species to preferentially expand in either responders or non‑responders depending on treatment context.

Distinct functional pathway signatures of ICI response in ICI‑only versus ICI + FMT cohorts

To assess how clinical and technical covariates shape microbiome composition at different functional levels, we performed PERMANOVA with study included as a blocking factor on taxonomic, biosynthetic gene cluster (BGC), and pathway profiles in the ICI‑only and ICI + FMT strata (Fig. 4a). In the ICI‑only cohorts, geographic area explained the largest fraction of between‑sample variation (up to ~5% at the taxonomic level), followed by age and sequencing instrument, all of which showed significant effects across at least one omics layer. Treatment regimen (anti‑PD‑1 monotherapy vs. PD‑1/CTLA‑4 combination) and sex contributed smaller but still detectable fractions of variance. In the ICI + FMT cohorts, age emerged as the dominant covariate, particularly for pathways, with additional contributions from sex and study area. Although these covariates each explained only a modest proportion of the total variance, their consistent effects across taxa, BGCs, and pathways highlight the importance of accounting for host and technical factors when interpreting functional associations with outcome.

Fig. 4: Metadata effects and pathway-level associations with ICI response in ICI-only and ICI + FMT cohorts.
Fig. 4: Metadata effects and pathway-level associations with ICI response in ICI-only and ICI + FMT cohorts.The alternative text for this image may have been generated using AI.
Full size image

a PERMANOVA analysis of clinical and technical covariates on microbiome composition, stratified by treatment group. Bars show the percentage of variance explained (R²) by each metadata variable (x-axis) for species-, BGC-, and pathway-level Bray–Curtis dissimilarities, with study included as a blocking factor. Statistical significance was assessed using permutational multivariate analysis of variance (PERMANOVA; adonis function in the vegan R package) with 999 permutations. All tests were two-sided. Asterisks indicate covariates with p < 0.05. b Pathway-level meta-analysis of associations with response status. For each stratum (left, ICI-only; right, ICI + FMT), volcano plots display standardized coefficients (coef/SE) from compound Poisson linear models (CPLM; MaAsLin2) on the x-axis and −log10(p-value) on the y-axis. Each point represents a MetaCyc pathway; positive coefficients indicate higher modeled pathway abundance in responders and negative coefficients indicate enrichment in non-responders. P-values were derived from two-sided Wald tests. Multiple comparisons were controlled using the Benjamini–Hochberg false discovery rate (FDR) procedure within each stratum, and pathways passing FDR correction (q < 0.25) are highlighted (green, enriched in responders; red, enriched in non-responders) and labeled with pathway names.

We next asked which microbial pathways were directly associated with immunotherapy response after adjusting for study effects. For each stratum, we fitted CPLMs on pathway abundances within each cohort and combined estimates using meta‑analysis, treating the standardized coefficient (coef/SE) as the effect size; positive values indicate pathways with higher modeled abundance in responders, and negative values indicate enrichment in non‑responders. In the ICI‑only group (Fig. 4b, left), several amino‑acid biosynthesis pathways showed positive associations with response, including L‑arginine biosynthesis II (acetyl cycle), L‑methionine biosynthesis III, and the superpathway of L‑lysine, L‑threonine, and L‑methionine biosynthesis II. In contrast, L‑lysine degradation I and (aminomethyl) phosphonate degradation were significantly depleted in responders (qval.fdr < 0.25). These findings suggest that microbiomes of responders are functionally skewed toward anabolic amino‑acid biosynthesis9,35, whereas non‑responders harbor communities with greater potential for amino‑acid and phosphonate catabolism, consistent with the emerging view that microbial amino‑acid metabolism can modulate host immunometabolism39 and responsiveness to checkpoint blockade40,41.

In the ICI + FMT strata (Fig. 4b, right; Supplementary Fig. S5), a different functional pattern emerged. A pyrimidine deoxyribonucleoside salvage pathway was positively associated with response, indicating an enrichment of nucleotide salvage capacity in responders. Nucleotide metabolism, including pyrimidine salvage, is increasingly recognized as a key regulator of immune cell proliferation and effector function, as well as a vulnerability in cancer cells42. In contrast, several carbohydrate and redox‑linked pathways were negatively associated with response, including 1,5‑anhydrofructose degradation, Entner–Doudoroff pathway I, L‑arginine biosynthesis III (via N‑acetyl‑L‑citrulline), and the superpathway of menaquinol‑8 (vitamin K₂) biosynthesis III, together with dTDP‑α‑D‑ravidosamine and dTDP‑4‑acetyl‑α‑D‑ravidosamine biosynthesis. Menaquinone biosynthesis and related vitamin K₂ pathways are known to support the growth and respiratory activity of specific gut bacteria and can shape community composition and host–microbe interactions43. The dTDP‑ravidosamine pathway contributes activated sugar donors used for glycosylation, and more broadly, tumor‑associated glycosylation and elevated UDP‑sugar pools have been implicated in promoting hyaluronan accumulation, immune evasion, and resistance to therapy44.

Together, these pathway‑level analyses indicate that the functional architecture of the gut microbiome associated with ICI response differs markedly between ICI‑only and ICI + FMT settings. In ICI‑only therapy, responders are linked to enrichment of amino‑acid biosynthetic capacity, whereas in the FMT‑augmented setting, response is instead associated with nucleotide salvage and relative depletion of carbohydrate catabolism, menaquinone biosynthesis, and sugar‑nucleotide glycosylation pathways. These context‑dependent functional signatures complement the species‑level findings and underscore that both microbial composition and metabolic potential jointly shape clinical outcome under different immunotherapy regimens.

Multi-omics cross-study profiling uncovers ecological and metabolic divides between ICI responders and non‑responders

To test whether microbiome predictors generalize across cohorts, we trained modality-specific and integrated multimodal models on seven melanoma ICI studies using IntegratedLearner32 with random forest as the base learner (The Glitza cohort was excluded from LODO modeling because only 6 patient-level profiles were available after within-patient aggregation.) and evaluated all pairwise train–test combinations across studies (Fig. 5a). Across most pairs, AUC–ROC values were above random expectation when models were applied to external cohorts, and leave-one-dataset-out (LODO) models performed similarly to within-study models, indicating that a subset of microbial features carries reproducible information about response status.

Fig. 5: Cross-study generalization of microbiome predictors and stable multi-modal features.
Fig. 5: Cross-study generalization of microbiome predictors and stable multi-modal features.The alternative text for this image may have been generated using AI.
Full size image

a Cross-study area-under-the-ROC (AUC-ROC) for predicting response across seven ICI cohorts using IntegratedLearner32 with random forest as the base learner. Each heatmap shows performance when training on the row study and testing on the column study; the boxed diagonal cells indicate within-study performance, and the bottom “LODO” row shows leave-one-dataset-out models trained on all other cohorts. Left, early fusion model using simple concatenation of all microbiome features; right, late fusion model that combines modality-specific predictors using an ensemble. Warmer colors indicate higher AUC-ROC. b Top 10 cross-study stable features per modality (taxa, pathways, and biosynthetic gene clusters). Bars show the cross-study feature importance score, summarizing how consistently each feature contributes to prediction across LODO models.

Using the LODO models, we extracted the top 10 cross‑study stable features per modality (Fig. 5b). In the BGC panel (“bgc”), responder‑associated features included a mix of antimicrobial nonribosomal peptide synthetase (NRPS) and RiPP‑like clusters (for example, a Rothia lanthipeptide, BGC0003096, and the stenothricin NRPS BGC0000431), together with loci involved in exopolysaccharide and capsule‑like glycans, including an EPS locus from Lactobacillus johnsonii (BGC0000767) and Klebsiella‑like capsular clusters (BGC0000730 and BGC0000733). Non‑responder‑associated features comprised an Enterococcus faecalis capsular polysaccharide locus (BGC0000792), the tilivalline NRPS toxin cluster from Klebsiella oxytoca (BGC0000446), a Streptomyces polyketide (BGC0000268), and a salivaricin bacteriocin locus (BGC0000548). NRPS and RiPP clusters often encode narrow‑spectrum antimicrobial peptides that modulate competition within the gut microbiota45, while EPS and capsule loci from organisms such as L. johnsonii, Klebsiella and Enterococcus can influence adhesion, barrier interactions, and immune evasion46. These annotations suggest that stable BGC features on both sides of the classifier are enriched for surface‑associated and competitive functions, rather than fitting a simple “beneficial antimicrobial vs harmful capsule” dichotomy, and they remain hypotheses that will require functional validation.

At the species level (Fig. 5b, “taxa” and Supplementary Fig. S2), responder‑associated features included Veillonella parvula, Fusicatenibacter saccharivorans and Blautia faecis. Veillonella species are obligate anaerobes that ferment host‑derived lactate to short‑chain fatty acids (SCFAs), particularly propionate47. F. saccharivorans and Blautia spp. are common SCFA producers, and B. faecis has been repeatedly identified as a butyrate‑producing commensal decreased in inflammatory disease48. SCFAs such as acetate, propionate and butyrate can support epithelial barrier function and modulate immune responses49, providing a plausible, but still hypothetical-link between these taxa and ICI responsiveness. Several negatively weighted taxa enriched in non‑responders, including Parabacteroides distasonis and recently described Firmicutes clades, have context‑dependent associations in the literature, so their roles in ICI outcome remain uncertain.

In the pathway panel (Fig. 5b, “pwy” and Supplementary Fig. S3), responder‑associated features were primarily central metabolic and biosynthetic pathways involved in carbohydrate fermentation, amino‑acid and cofactor synthesis, and chorismate‑linked metabolism, all functions commonly encoded in commensal anaerobes that contribute to SCFA and vitamin production49. Non‑responder‑associated pathways were enriched for sugar‑nucleotide and related routes that contribute to specialized cell‑surface glycans, including components of capsular and lipopolysaccharide structures in Gram‑negative bacteria, which are important for virulence and immune modulation50.

Overall, the cross‑study stable features (Fig. 5a,b, Supplementary Fig. S4) point to consistent differences between responders and non‑responders in taxa, BGCs, and pathways linked to SCFA production, EPS and capsule biosynthesis, and other surface‑associated functions. These patterns are biologically plausible given existing literature, but should be interpreted as hypotheses about mechanism rather than direct proof of causality.

Discussion

In this study, we integrated multiple melanoma cohorts using a harmonized analytical framework to identify microbiome features associated with response to immune checkpoint inhibitor therapy. By harmonizing inclusion criteria, stratifying analyses by treatment modality, and adjusting for key clinical and technical covariates, we asked which microbial features reproducibly associate with response and to what extent these associations generalize across studies6,11.

At the taxonomic level, we did not identify a single universal “beneficial” or “detrimental” species. Instead, partially overlapping sets of Clostridiales, Bacteroidales, and Actinobacteria showed context‑dependent associations that differed between ICI‑only therapy and ICI combined with FMT6,11,15,16. In ICI‑only cohorts, several short‑chain‑fatty‑acid-producing commensals were enriched in responders8,9,14, whereas taxa linked to dysbiosis, such as Hungatella hathewayi and Clostridium innocuum, were more associated in non‑responders. In contrast, ICI + FMT trials showed a distinct pattern in which Alistipes, Bacteroides and Holdemania filiformis favored response, while species including Bifidobacterium pseudocatenulatum and Coprococcus eutactus were associated with non‑response despite being reported as favorable in some ICI‑only studies6,12,13. Three overlapping species reversed direction between strata, and longitudinal data from FMT trials confirmed that these taxa expand preferentially in either responders or non‑responders depending on the regimen2,12,15.

Functional analyses also differed by treatment context. In ICI‑only cohorts, responders showed higher enrichment associated with amino‑acid biosynthesis pathways, whereas non‑responders were enriched for lysine and phosphonate degradation, suggesting a shift toward anabolic amino‑acid metabolism in communities associated with response9,35. These findings suggest that microbiomes of responders are functionally skewed toward anabolic amino‑acid biosynthesis, whereas non‑responders harbor communities with greater potential for amino‑acid and phosphonate catabolism, consistent with the emerging view that microbial amino‑acid metabolism can modulate host immunometabolism and responsiveness to checkpoint blockade39,40,41. In ICI + FMT cohorts, response was instead linked to a pyrimidine deoxyribonucleoside salvage pathway, while non‑responders exhibited increased carbohydrate degradation, menaquinone biosynthesis and sugar‑nucleotide-linked glycosylation pathways. These findings imply that microbial metabolic potential related to amino‑acid and nucleotide handling, redox balance, and cell‑surface modification may contribute to ICI outcomes, but in a regimen‑specific manner42.

Cross‑study predictive modeling provided an orthogonal view of these patterns. Leave‑one‑dataset‑out classifiers based on combined taxonomic, pathway, and BGC features achieved above‑chance but modest AUC‑ROC values (around 0.6), indicating a reproducible microbiome signal that is not yet sufficient for individual‑level prediction. Nonetheless, this framework highlighted features that were repeatedly selected across cohorts, including commensal taxa such as Veillonella and Blautia species, amino‑acid and cofactor biosynthesis pathways, and several BGCs with putative antimicrobial or exopolysaccharide functions, which represent hypotheses for future mechanistic work45,51.

Although no single BGC achieved study-wide significance in our univariate meta-analysis, the multi-modal cross-study models consistently highlighted several BGCs, particularly RiPP 52 and NRPS-like antimicrobial clusters53, as well as exopolysaccharide and capsule loci, as stable contributors to prediction54,55. This discrepancy likely reflects the high dimensionality and sparsity of the BGC layer, in which many loci are rare, cohort-specific, and only modestly associated with outcome, limiting statistical power after multiple-testing correction56,57. In contrast, the integrated predictive models combine BGC, taxonomic, and pathway information, allowing weak but coherent signals carried by correlated groups of features to accumulate and become detectable. We therefore interpret these BGCs, including RiPP clusters, as composite, hypothesis-generating markers of microbiome state rather than definitive causal determinants of ICI outcome.

Several limitations should be considered. Despite batch correction and adjustment for measured covariates, substantial heterogeneity remains between cohorts, reflecting differences in geography, prior therapies, sample handling, sequencing protocols, and response definitions58,59,60, and some potentially important factors (e.g. antibiotics, diet, prior CTLA‑4 blockade) were incompletely captured. Functional annotations depend on current reference databases (shown in Supplementary Fig. S1) and do not directly measure metabolites or gene expression, and the ICI + FMT strata include relatively small numbers of patients. Finally, the analysis focuses on melanoma and may not generalize to other cancers or immunotherapy regimens.

In summary, this unified re-analysis supports a treatment‑dependent model in which sets of taxa and functions, rather than a single universally “beneficial” organism, reproducibly associate with checkpoint inhibitor response. The partial overlap—and occasional reversal—of response-associated species between ICI‑only and ICI + FMT settings helps reconcile inconsistencies in the literature and argues that “beneficial” microbiome configurations are contingent on the therapeutic regimen and ecological perturbations induced by FMT. Functionally, the recurrent signal linking response to anabolic programs (including amino‑acid 58,59biosynthesis in ICI‑only cohorts) and distinct nucleotide and cell-surface–related pathways in ICI + FMT cohorts suggests that microbiome contributions to immunotherapy may operate through multiple metabolic routes that differ by context. Predictive performance across cohorts remained modest, indicating that microbiome features alone are not yet sufficient for individual-level prediction, but the repeated selection of specific taxa, pathways, and BGC classes in multimodal models provides a refined shortlist of mechanistic hypotheses to prioritize in prospective multi-omic and interventional studies.