Longitudinal gut microbiome changes in immune checkpoint blockade-treated advanced melanoma

Immune checkpoint blockade (ICB) improves survival across multiple advanced cancers, but only a subset of patients respond and immune-related adverse events (irAEs), including colitis, are common. Many cross-sectional (baseline) microbiome studies report inconsistent microbial biomarkers of response, likely due to methodological, biological and clinical confounders and high intra- and inter-individual variability. With multiple microbiome-directed trials (including FMT) underway, there is an urgent need to understand longitudinal gut microbiome dynamics during ICB. This study longitudinally profiled the gut microbiome via shotgun metagenomics at baseline and three early on-treatment visits over 12 weeks in 175 ICB-treated advanced melanoma patients from five cohorts. The authors hypothesized that: (1) microbial abundances change over the treatment period as cumulative ICB effects accrue, and (2) responders (PFS ≥12 months) and non-responders (PFS <12 months) exhibit distinct baseline and on-treatment microbiome patterns.

The article situates its work within prior baseline-focused studies linking specific taxa to ICB response across melanoma and other tumors, but notes a lack of consensus and reproducibility. SCFA-producing taxa have been repeatedly implicated in favorable outcomes, whereas taxa associated with chronic inflammatory and metabolic diseases often relate to resistance. The field has also begun testing microbiome interventions, including FMT from responders or healthy donors in ICB-refractory or naive patients, with encouraging response rates. However, without longitudinal context, interpreting such interventions is challenging. The authors reference recent MetaPhlAn4-based meta-analyses and trials, and evidence of diet and medication (for example, PPIs) influencing the gut microbiome and potentially ICB outcomes or irAEs, underscoring the need for time-resolved and context-aware analyses.

Study design and cohort: 175 patients with unresectable stage III/IV cutaneous melanoma were recruited across five cohorts (PRIMM-UK n=54; PRIMM-NL n=74; Manchester n=17; Leeds n=19; Barcelona n=11). Treatments: anti-PD-1 monotherapy (nivolumab or pembrolizumab; 67%) or anti-PD-1 plus anti-CTLA-4 combination (ipilimumab; 33%). Clinical endpoints: PFS at 12 months (PFS12) and overall survival (OS). PFS defined from therapy start to progression or death; OS defined from start to death from any cause (subset n=147). irAEs assessed per CTCAE v5; colitis defined excluding non-immune etiologies. Follow-up up to 7.3 years (median 4.3). Sampling: Fecal samples at baseline (T0) and prior to cycles at ~3–4-week intervals for T1, T2, T3 over 12 weeks. Sample handling: Home collection kits with freezing/shipping; storage at −80 °C; DNA extraction at King’s College London (MagMax Core protocols, mechanical lysis). Sequencing: Libraries via Illumina Nextera DNA Flex; 300 bp paired-end on NovaSeq6000. Total 1,283 Gb; average ~53.9 million reads/sample pre-QC. QC and preprocessing: Centralized pipeline (SegataLab/preprocessing). Inclusion required >1 Gb data and passing QC, yielding 447 samples from 195 patients; after strict QC and metadata completeness, 408 samples from 175 patients used. Taxonomic and functional profiling: MetaPhlAn4 (vJan21 SGB database) for species-level genome bins (SGBs); HUMAnN3 for MetaCyc pathway profiling. Prevalence filtering retained features present in ≥20% baseline samples and ≥10% longitudinal samples: 434 SGBs and 395 pathways across 408 samples. Longitudinal statistical modeling: Compositional data handled via log-ratio framework. Applied Bayesian multinomial logistic normal regression (Pibble model in R package fido v1.0.4), modeling log-ratio-transformed proportions with multivariate normal errors, allowing positive/negative covariation among features. Modeled visit-varying intercepts and slopes to capture linear changes over visits, with higher-order interactions to test differences by response group (PFS≥12 vs <12) and moderator variables: therapy regimen (mono vs combo), colitis (no/yes), and baseline PPI use (no/yes). Adjusted for prior antibiotics use, prior BRAF/MEK therapy, time since first injection, cancer center, other irAEs (non-colitis), age, sex, BMI, and included patient random effects. Posterior inference at 90% Bayesian confidence level (BCL) primarily, with other BCLs provided in supplements. Post hoc contrasts computed using reference grids and marginal means to compare groups at each visit, averaging over covariates. Balance (log-ratio) analyses: Constructed balances as log ratios of geometric means of predefined SGB subsets (from longitudinal model hits). Evaluated discrimination between PFS groups across visits via Wilcoxon tests and predictive performance by AUC from 100× repeated five-fold cross-validation. Survival analyses: Multivariable Cox models for OS and PFS (time-to-event), adjusting for sex, age, BMI, PPI, antibiotics, previous therapy, colitis, other irAEs; checked proportional hazards via Schoenfeld residuals. External validation: Reprocessed six independent melanoma cohorts (baseline, and one post-ICB set) with MetaPhlAn4 under same settings; computed balances and assessed predictive ability (logistic regression AUCs) and association with outcomes where available. Subgroup/context analyses: Stratified analyses for monotherapy vs combination therapy (excluding colitis and PPI users), colitis vs no colitis (averaging over response and regimen), and PPI users vs non-users in monotherapy without colitis. Software: MetaPhlAn4, HUMAnN3, fido, phyloseq, tidyverse, caret, PROC, survival, ggsurvfit. Data and code availability provided via ENA accessions and GitHub.

- Cohort and outcomes: 175 patients; 117 (67%) monotherapy; 58 (33%) combination. PFS≥12 achieved by 83 (47%). Median OS 34.1 months. - Longitudinal SGB dynamics and discrimination: - At 90% BCL, 62 (14.3%) SGBs had nonzero slopes in PFS≥12; 41 (9.4%) in PFS<12 (directional changes over visits). - 99 (22.8%) SGBs discriminated PFS≥12 vs <12 in at least one visit (range 342 at 50% BCL to 3 at 100% BCL). - Of these 99 SGBs, 20 were baseline-only, 42 were post-initiation-only, and 5 and 4 remained consistently higher across all visits in PFS≥12 and PFS<12, respectively. - Consistently higher in PFS≥12 across visits (5 SGBs): Agathobaculum butyriciproducens (SGB14993 group), Intestinibacter bartlettii (SGB6140), Dorea sp. AF24 7LB (SGB4571), Lactobacillus gasseri (SGB7038 group), Lacrimispora celerecrescens (SGB4868). Several showed increasing on-treatment abundances. These fiber-degrading, SCFA-capable taxa align with plant-based diets. - Consistently higher in PFS<12 across visits (4 SGBs): Ruthenibacterium lactatiformans (SGB15271), Prevotella copri clade A (SGB1626), Ruminococcaceae unclassified (SGB15265 group), and an unidentified Bacteroidetes SGB (SGB1957). - Pathways: - Responders (PFS≥12): higher abundances of SCFA or precursor pathways across multiple visits (PWY-6396 superpathway of 2,3-butanediol biosynthesis; PWY-P124 Bifidobacterium shunt; PWY-6435 4-hydroxybenzoate biosynthesis V; PWY-5088 L-glutamate degradation VIII to propanoate). - Non-responders (PFS<12): enriched menaquinol (vitamin K) synthesis pathways at baseline and early treatment; these are linked to chronic inflammatory and cardiometabolic diseases and correlate with Prevotella/Bacteroides; diet-manipulable. - Responders showed higher post-baseline polyamine biosynthesis (POLYAMINSYN3-PWY) across T1–T3 (not baseline), suggesting benefits of polyamines (e.g., spermidine) and autophagy induction. - Predictive balances: - Primary balance (5 PFS≥12 SGBs vs 4 PFS<12 SGBs) discriminated PFS groups at T0–T2 (Wilcoxon P_T0=0.00085, P_T1=0.0007, P_T2=0.0005) but not T3 (P=0.1). Cross-validated AUCs: 0.659±0.092 (T0), 0.666±0.091 (T1), 0.739±0.118 (T2), 0.655±0.129 (T3). - OS association at baseline: Above-median balance associated with longer OS (35.4 vs 28.4 months; HR 1.669, P=0.035). As continuous predictor: HR_OS=0.828, P=0.001; HR_PFS=0.829, P=0.0005. - Extended balance (adding SGBs present in all but last visit) improved AUCs: 0.771±0.088 (T0), 0.706±0.094 (T1), 0.783±0.118 (T2), 0.765±0.138 (T3). OS association: 37.0 vs 26.9 months; HR 1.792, P=0.014. As continuous: HR_OS=0.752, P=0.0002; HR_PFS=0.727, P=8.93×10−6. - Baseline-only balance (9 responder- and 11 non-responder-associated baseline SGBs) predicted OS at baseline (35.5 vs 28.4 months; HR 1.639, P=0.034). - External validation: Comparable AUCs across several independent cohorts; significant discrimination in a cohort with N=112 (P=0.04). Predicted OS in a small cohort (N=27; P=0.024). - Baseline-only SGBs: Responders enriched at baseline with Romboutsia timonensis (SGB6148), Limosilactobacillus fermentum (SGB7106), Blautia schinkii (SGB4825); non-responders enriched with Eubacterium siraeum (SGB4198 group), Oscillibacter sp. ER4 (SGB15254), Dysosmobacter sp. NSJ 60 (SGB15124) at baseline only. - Post-initiation discriminators: Responders increased several SCFA producers (Lachnospiraceae: Coprococcus comes (SGB4577 group), C. catus (SGB4670), Gemmiger (SGB15295 group), Anaerobutyricum hallii (SGB4532)). Non-responders increased Clostridium spiroforme (SGB6747); Blautia hydrogenotrophica (SGB4677), Blautia wexlerae (SGB4837 group), Ruminococcus torques (SGB4608), Sellimonas intestinalis (SGB4617), Eisenbergiella tayi (SGB4988); Erysipelotrichaceae members Turicibacter sanguinis (SGB6847) and Faecalibacillus faecis (SGB6750). - Pattern dynamics: - 22 SGBs showed intersecting slopes (reversals) between groups (e.g., Streptococcus thermophilus, T. sanguinis, Blautia schinkii). - Divergence after ICB from similar baselines: increases in non-responders and decreases in responders for Christensenellaceae bacterium NSJ 53, E. tayi, Mediterraneibacter massiliensis, S. intestinalis, Hydrogeniiclostridium mannosilyticum. - 16 SGBs changed mainly in responders (e.g., increases in Lachnospiraceae bacterium OF09 6, E. siraeum; decreases in F. faecis, Fusicatenibacter saccharivorans). - 14 SGBs changed mainly in non-responders (e.g., Bilophila wadsworthia; several Clostridium spp.). - Therapy context effects: - Monotherapy vs combination therapy showed shared and divergent associations; six SGBs had opposite patterns by regimen, including Coprococcus eutactus (SGB5121), Butyricicoccus sp. AM29 23AC (SGB14991), Parabacteroides merdae (SGB1949). - Under monotherapy, non-responders had increasing abundances of several Bacteroides spp. across visits, not observed with combination therapy, supporting agent-dependent biphasic Bacteroides effects. - SGBs higher in responders regardless of regimen included Lacticaseibacillus rhamnosus, an unknown Firmicutes (SGB47850), Dorea sp. AF24 7LB, Dorea formicigenerans, Coprococcus comes and several unclassified Ruminococcaceae. - Colitis associations and prediction: - Patients without colitis had higher abundances of butyrate producers (Roseburia inulinivorans, R. hominis, A. butyriciproducens, Eubacterium rectale, Bacteroides thetaiotaomicron) and Faecalibacterium prausnitzii subspecies; decreases were larger in those who developed colitis. - A. muciniphila was higher at baseline in those who developed colitis but declined thereafter; the no-colitis group had lower but increasing abundances. - Baseline balance (10 colitis-associated vs 12 colitis-resistant SGBs) predicted colitis development (Wilcoxon P=0.00055; AUC 0.723±0.121). - Colitis incidence was higher with combination vs monotherapy (0.310 vs 0.128; Δ=0.182; 95% CI 0.036–0.329; χ²=7.259; P=0.007). Microbial signatures with colitis under combination resembled IBD (enrichment of Streptococcus, Veillonella, Bacteroides, Eggerthella). - PPI use: - Among monotherapy patients without colitis, PPI users shared 33 associations with non-users, but some taxa showed regimen-specific slope differences by PPI status (e.g., S. thermophilus increased in non-users with PFS≥12 but decreased in users; Christensenellaceae bacterium NSJ 53 and Bacteroides caccae increased in non-users but decreased in users with PFS<12).

This study demonstrates that the gut microbiome in ICB-treated melanoma patients exhibits dynamic, response-specific trajectories that differ from baseline associations and vary by clinical context (therapy regimen, colitis, PPI use). The findings support the concept that SCFA-producing, fiber-degrading taxa are beneficial and remain elevated or increase during treatment in responders, while taxa linked to inflammatory states tend to rise in non-responders. Crucially, some taxa reverse or diverge after ICB initiation, meaning that therapeutic targets derived from baseline-only data may produce opposite or unexpected effects on-treatment. The constructed longitudinal balances combine consistent responder- and non-responder-enriched SGBs into a practical composite that can discriminate PFS status at early timepoints and associate with OS, showing moderate predictive performance and partial external generalizability. Pathway-level signals (SCFA, polyamine, menaquinone) suggest mechanistic links to immunomodulation and diet-addressable metabolic functions. Context-dependent differences (e.g., Bacteroides patterns by regimen; colitis-associated signatures; PPI interactions) further indicate that microbiome-based interventions and monitoring should account for treatment and host factors. The results align with effects observed in FMT trials, suggesting potential synergy between microbiome modulation and ICB. Overall, longitudinal, fine-resolution microbiome profiling, integrated with clinical covariates, is essential to inform rational microbiome-targeted strategies for enhancing ICB efficacy and mitigating irAEs.

The study profiles longitudinal gut microbiome changes in 175 advanced melanoma patients undergoing ICB and identifies distinct, context-dependent microbial taxa and pathway dynamics associated with favorable (PFS≥12) versus unfavorable outcomes. Five SGBs consistently enriched in responders and four in non-responders form balances that predict PFS class across early visits and associate with OS. SCFA- and polyamine-linked pathways characterize responders, whereas menaquinol synthesis is elevated in non-responders. Treatment regimen, colitis, and PPI use modulate microbial trajectories, underscoring the need for context-aware, time-resolved biomarkers and interventions. These insights provide a roadmap for designing rational microbiome-targeted therapies and interpreting microbiome-focused trials, including FMT. Future work should validate these longitudinal signatures at larger scales, integrate metagenomics with metabolomics and immunomics, refine causal mechanisms, and test tailored interventions (dietary, probiotic/consortia, FMT) stratified by clinical context.

- Modeling assumed linear changes across visits, potentially oversimplifying complex dynamics. - Differences in taxonomic databases across studies can limit comparability with prior literature. - Some post hoc subgroup comparisons involved small sample sizes, reducing generalizability and statistical power. - Observational design precludes causal inference. - Heterogeneity in DNA isolation and sequencing platforms across validation cohorts and partial clinical metadata availability constrained adjustment and validation analyses.