Toward a Dual-Axis Model of Microbiome Modulation in Cancer Immunotherapy: Pathobiont Elimination and Functional Ecosystem Restoration

Abstract

Purpose

The gut microbiome is increasingly recognized as a modulator of cancer immunotherapy efficacy, including responses to immune checkpoint inhibitors (ICIs) and chimeric antigen receptor T-cell (CAR-T) therapy. Recent clinical trials of microbiome-targeted interventions such as fecal microbiome transplantation (FMT) and live biotherapeutic products (LBPs) suggest the potential to enhance antitumor immunity and improve clinical outcomes. Yet responses remain heterogeneous and are not fully explained by engraftment of donor taxa alone

Methods

We integrate evidence from interventional trials, observational cohort studies, and principles from gut microbial ecology to develop a model hypothesis on how microbiome-targeted therapies may shape response to immunotherapy, with potential to inform future trial design, analyses, and interpretation.

Results

Drawing on the available evidence, we propose that therapeutic perturbation of the gut microbiome may augment immunotherapy efficacy through two parallel axes: (1) elimination of immunosuppressive pathobionts that restrain CD8+ T-cell activation and promote myeloid-mediated immunosuppression, and (2) functional restoration of the gut ecosystem through engraftment of taxa that provide metabolites, structural cues, and immunoregulatory signals required for effective antitumor immunity. The success of both axes appears to depend on ecological processes governed by predator-prey dynamics, including colonization resistance, resilience of the resident microbiota, and the ability of administered organisms to displace entrenched dysbiotic communities. This ecological lens may help to explain discrepancies across trial designs, donor types, and intervention modalities, and suggests that complete donor engraftment is neither necessary nor sufficient for clinical benefit.

Conclusions

A dual-mechanism model of pathobiont elimination and functional microbial restoration may help explain microbiome-mediated enhancement of cancer immunotherapy, highlighting a balanced immune permissive gut ecosystem as a key determinant of therapeutic success.

INTRODUCTION

Foundational studies in murine models showed that the gut microbiota can strongly influence the efficacy of cancer chemotherapy and immunotherapy^1–4. Mechanistically, these effects were linked to microbial regulation of innate inflammatory responses and support of adaptive immune functions required for antitumor activity^1,2,4. Clinical translation followed in 2018, when three landmark studies, later reinforced by a growing body of evidence, showed that distinct gut bacterial taxa were associated with favorable or unfavorable responses to PD-1 immune checkpoint inhibitor (ICI) therapy in melanoma and other epithelial cancers, including non-small cell lung cancer (NSCLC) and renal cell carcinoma (RCC)^3,5–8. Collectively, these observations established the gut microbiome as a major modifier of cancer immunotherapy response. Parallel preclinical work demonstrated causality through inbred laboratory mice, germ-free mouse models and FMT experiments, showing that gut microbiome from responding patients or specific commensals that were associated to favorable response in patients (Bifidobacterium longum, Akkermansia muciniphila) could enhance ICI efficacy and that the microbiome shaped antitumor responses^3,5–7. Subsequent studies extended these findings to patients treated with anti-cytotoxic T-lymphocyte-associated protein 4 (CTLA-4) and other single or combination immunotherapies^9–11, as well as to chimeric antigen receptor T cell (CAR-T) treated hematologic malignancy patients¹² - linking higher microbial diversity and certain microbial taxa to ICI response, while linking Bacteroides intestinalis to ICI toxicity via mucosal IL-1β upregulation, and showed that antibiotic-induced dysbiosis was associated with markedly inferior survival^13,14.

Subsequent multi-cohort studies revealed that the relationship between gut microbiota composition and ICI (or CAR-T) efficacy was substantially more complex than early reports suggested. No single bacterial species consistently predicted response across independent cohorts - with “cohort effects” (driven by geography, diet, prior therapies, methodological variation or combinations thereof) being responsible for more microbiome variance than response status alone^15–19. What emerged with reasonable consistency was a phylum/family-level pattern: Actinobacteria and the Lachnospiraceae/Ruminococcaceae families of Firmicutes broadly favored response, in part through production of immunomodulatory metabolites, including short-chain fatty acids (SCFA) and others, while Gram-negative members of the order Bacteroidales were enriched in non-responders and associated with a systemic pro-inflammatory myeloid program^15–19. Bjork et al. added a critical dimension by demonstrating that microbiome changes during treatment - not just at baseline - were associated with outcomes, and that the same taxa could exhibit opposite trajectories depending on the treatment regimen (anti-PD-1 monotherapy vs. combination blockade), which cautioned against drawing conclusions regarding microbiome signatures based solely on baseline cross-sectional data¹⁹. Concordantly, strain-resolved microbial analyses that compared patients treated with anti-PD-1 monotherapy to various combination regimens (anti-PD-1 + anti-CTLA-4, or anti-PD-1 + TLR9 agonist) revealed that microbial response signatures were highly context-dependent, varying with the specific immunotherapy platform used^10,11. Quantitative modeling approaches have begun to formalize the relationship between microbiome composition and immunotherapy efficacy in terms that move beyond correlative association. Hadjigeorgiou et al., working in the laboratory of Rakesh Jain - whose foundational contributions to tumor vascular biology and the physical determinants of immune access to solid tumors have long established the value of rigorous mathematical frameworks in cancer biology - integrated agent-based and differential equation modeling with association analyses across human cohort data to show that the gut microbiome influences immunotherapy outcomes by modulating two mechanistically central parameters: the rate of immune-cell activation and the efficiency of tumor-cell killing²⁰. Critically, distinct bacterial taxa were found to exert quantitatively synergistic or antagonistic effects on these parameters, with the magnitude of microbiome-associated modulation comparable to that of established tumor-intrinsic determinants of immunotherapy response. This framework offers a formal quantitative basis for the dual-axis model proposed in this review: bacterial taxa that antagonize immune-cell activation or tumor killing correspond to the pathobiont axis, actively restraining antitumor immunity, while those that synergize with these parameters represent the functional restoration axis, amplifying the conditions required for effective checkpoint inhibitor response. The incorporation of such modeling approaches into trial design and biomarker development represents a natural next frontier at the intersection of mathematical oncology and microbiome science.

Bacterial metabolites, bile acids (BA) and SCFA synthesis, mucosal barrier integrity, and the modulation of dendritic cell and T cell priming represent interconnected outputs of microbial community ecology, rather than the isolated contributions of individual taxa. Certain bacteria produce molecules that activate innate immune receptors, thereby boosting immune responses against tumors. For instance, components from Enterococcus and Bifidobacterium species can stimulate receptors such as NOD2 and TLR2, enhancing cancer immunity^21–23. In addition, normal microbial communities help maintain baseline type I interferon (IFN) signaling, which is important for immune readiness^24,25. When the microbiome is disrupted - such as in germ-free or antibiotic-treated animals - this signaling is diminished, leading to weaker immune responses.

Early animal studies demonstrated that microbial products, such as lipopolysaccharides (LPS), can enhance the effectiveness of T cell–based therapies by promoting immune activation²⁶. Similarly, shifts in gut bacterial populations, including expansion of certain Proteobacteria, can improve therapeutic outcomes by increasing IL-12 production²⁷. In clinical settings, however, broad-spectrum antibiotics that reduce microbial diversity are linked to poorer responses to therapies like CAR-T cells, whereas bacteria that produce short-chain fatty acids (SCFAs) are associated with better outcomes. SCFAs can enhance immune function partly by influencing gene regulation and cytokine production^12,28–30. Microbial metabolites are also key contributors to anti-cancer immunity. Compounds such as inosine can enhance T cell activity both by signaling through specific receptors and by serving as an energy source under nutrient-limited conditions³¹. Other metabolites, including SCFAs, trimethylamine N-oxide (TMAO), and tryptophan derivatives, have been linked to improved immunotherapy responses through various mechanisms, such as activating immune cells, promoting T cell persistence, or increasing tumor cell susceptibility to immune attack^32–37. Some bacterial species can even modulate immune checkpoint pathways directly, suggesting new therapeutic targets³⁸. Overall, these findings highlight the complex and multifaceted ways in which the microbiome influences cancer immunotherapy outcomes. Artificial intelligence (AI) and machine learning (ML) approaches are increasingly leveraged to integrate these multidimensional datasets and to identify emergent functional patterns that are not apparent from compositional analyses alone.

In this review, we integrate evidence from interventional trials, observational cohort studies, and principles of gut microbial ecology to develop and defend a dual-axis interpretive framework for microbiome-mediated modulation of cancer immunotherapy. The first axis centers on the elimination of immunosuppressive pathobionts - Gram-negative and Gram-positive taxa that drive tonic myeloid activation, IL-8-mediated immune exclusion, and disruption of epithelial barrier integrity - whose depletion appears necessary to relieve a dominant brake on CD8⁺ T cell-mediated antitumor immunity. The second axis concerns the functional restoration of an immune-permissive gut ecosystem: the engraftment or expansion of microbial communities capable of producing the metabolites, structural signals, and mucosal immune cues - including short-chain fatty acids, secondary bile acids, and tryptophan catabolites - that are required for effective dendritic cell conditioning and cytotoxic T cell priming. We argue that both axes operate through a combination of direct immunomodulatory mechanisms and ecological community dynamics, and that the relative contribution of each axis varies with the treatment backbone, disease setting, and baseline microbiome composition of the host. This framework, grounded in but extending beyond the available clinical evidence, may offer a basis for reconciling discrepant findings across FMT, live biotherapeutic product, and dietary intervention trials, and suggests specific principles for the rational design of next-generation microbiome-targeted therapies that we develop in the sections that follow.

MICROBIOME INTERVENTIONAL TRIALS IN CANCER: FECAL MICROBIOME TRANSPLANTATION (FMT), LIVE BACTERIAL SUPPLEMENTATION (LBP), AND DIETARY INTERVENTIONS

Therapeutic perturbation of the gut microbiome by FMT, whether delivered by colonoscopy or capsules, is better understood not as the administration of specific bacteria but as the transfer of an entire biological ecosystem, that includes bacteria, archaea, fungi, viruses, bacteriophages, and their collective metabolic outputs. That breadth is both FMT’s central strength and its primary limitation. Its strength lies in the ability, under favorable ecological conditions, to comprehensively remodel the recipient gut microbiome. Such remodeling can displace pathobionts through multiple parallel mechanisms, including direct microbial competition, bacteriophage-mediated predation, SCFA–driven pH shifts, and restoration of mucosal immune function that selectively eliminates taxa incompatible with the new equilibrium. At the same time, efficacy is constrained by host-dependent colonization resistance, donor variability, and limited mechanistic resolution, making durable benefit neither uniformly reproducible nor easily optimized.

Initial proof-of-concept studies demonstrated that FMT could rescue some patients with PD-1 refractory melanoma patients by restoring sensitivity to anti-PD-1 ICI. These trials by Davar et al. and Baruch et al. evaluated responder-derived FMT combined with PD-1 rechallenge in patients with PD-1–refractory melanoma^39,40. Despite small sample sizes and moderate objective response rates (ORR), both studies demonstrated that a subset of patients (≈40%) exhibited durable disease control, exceeding the <10% expected with PD-1 rechallenge alone. Clinical benefit correlated with durable donor microbiome engraftment, increased CD8⁺ T-cell infiltration, enhanced interferon-γ signaling, and suppression of IL-8–driven myeloid inflammation^39,40. Together, these findings provided the first direct human evidence that therapeutically microbiome state could causally reverse resistance to ICI in cancer patients.

Subsequent front-line studies demonstrated substantially greater efficacy, suggesting that microbiome modulation is most effective before immune escape mechanisms and T-cell dysfunction become entrenched. A phase I trial in immunotherapy-naïve cutaneous melanoma (MIMIC) that combined healthy donor FMT (LND-101) with PD-1 blockade and reported 65% ORR including 20% complete responses⁴¹; exceeding historical outcomes with anti-PD-1 monotherapy in cutaneous melanoma^42,43. Notably, sustained donor–recipient microbiome convergence occurred only in responders, directly linking clinical benefit to durable ecological engraftment of LND-101. The FMT-LUMINate trial extended this strategy by evaluating LND-101 in first-line cutaneous melanoma in combination with dual anti-PD-1 and anti-CTLA-4 therapy, as well as in PD-L1–high NSCLC in combination with anti-PD-1 therapy⁴⁴. High ORRs were reported in both settings (ORR 75% in melanoma, 80% in PD-L1–high NSCLC); and while responders developed a distinct post-FMT gut microbiome, this was driven by neither donor–recipient similarity nor strain-level engraftment. However, responsiveness after FMT was associated with establishment of a microbial compositions consistent with those previous reported response to either anti-PD-1 or anti-PD-1/CTLA-4 combination^10,17,44. In addition, response to LND-101 in FMT-LUMINate was associated with substantial loss of baseline bacterial species, particularly depletion of Enterocloster citroniae, E. lavalensis, and Clostridium innocuum; an observation buttressed by a re-analysis of the combined data of MIMIC trial in PD-1 naïve melanoma, and the Baruch et al. and Davar et al. trials in PD-1 refractory melanoma^39–41. Donor-specific effects were also evident with respect to toxicity. Engraftment of donor-derived Prevotella species was associated with the development of immunotherapy-related adverse events (irAEs) in the context of anti-CTLA-4 treated melanoma. Donor 5, a Prevotella-rich donor from the earlier MIMIC trial, was associated with grade ≥3 adverse events in 25% of melanoma patients treated with FMT plus anti-PD-1 and anti-CTLA-4. However, this same donor did not induce toxicity in patients receiving anti-PD-1 alone, nor was Prevotella-rich FMT associated with severe adverse events in NSCLC patients treated with anti-PD-1 monotherapy, indicating that toxicity risk depends on the interaction between donor microbiome composition and the immunotherapy backbone rather than donor characteristics alone. These data underscore that the second axis - functional ecosystem modulation - is not uniformly beneficial; the immune-activating capacity of certain donor taxa (herein Prevotella spp.) can exceed the therapeutic window in the context of combination checkpoint blockade targeting CTLA-4, where mucosal inflammatory thresholds are already lowered. Donor selection must therefore account for immunotherapy backbone and patient immune context, not donor microbiome composition alone.

In RCC, the PERFORM trial reported 50% ORR when LND-101 was combined with standard first-line ICI regimens, with signals suggesting reduced immune-related toxicity among responders⁴⁵. More conclusively, the randomized, placebo-controlled TACITO trial demonstrated superior median progression-free survival (24 vs. 9 months) and higher objective response rates (52% vs. 32%) with responder-derived FMT⁴⁶. Although the prespecified 12-month PFS endpoint did not reach statistical significance - likely reflecting limited statistical power in this phase 2 trial - the magnitude of clinical benefit was substantial, with median PFS of 24 versus 9 months and ORR of 52% versus 32%, providing compelling evidence that microbiome manipulation can meaningfully alter immunotherapy outcomes in RCC.

The principal limitations of FMT are its unpredictability and lack of scalability, particularly with respect to donor selection and mechanistic interpretability. Although the FMT-LUMINate trial demonstrated impressive ORRs in immunotherapy-naïve cutaneous melanoma using a healthy donor FMT product (LND-101), the specific elements of the transplanted ecosystem responsible for clinical benefit remain incompletely defined. The observation that responders exhibited greater loss of baseline microbial species, along with selective depletion of Enterocloster spp. and Clostridium innocuum, suggests that efficacy may be driven in part by targeted displacement of pathobionts - potentially mediated in part by donor-derived bacteriophages with specificity for these taxa. Responder-derived FMT, as implemented in the Baruch, Davar, and TACITO trials, represents a rational refinement of this approach. In this model, donor ecosystems are selected a priori for pro-immunogenic properties, theoretically enriching both functional microbial content and pathobiont-displacing capacity. The strain-level associations and progression-free survival benefit observed in TACITO lend support to this strategy.

Despite its biological appeal, a responder-based FMT paradigm presents substantial practical challenges, particularly in donor identification, standardization, and large-scale implementation. Trials conducted to date have begun to clarify some of the ecological and molecular mechanisms involved, but the number of patients studied remains too limited to support robust interpretation of the results. Moreover, although the evidence that the intestinal microbiome influences responses to different types of immune checkpoint inhibitors is strong, the precise mechanisms, the bacterial taxa involved, and their ecological synergies or antagonisms remain incompletely characterized. As discussed later, machine learning approaches may help advance personalized immunotherapy by predicting which immune checkpoint blockade strategy is most likely to benefit a given patient and by estimating whether fecal microbiota transplantation could improve treatment responsiveness. Recent models have shown that post-FMT microbiome composition can be predicted from donor and recipient microbiome profiles, providing a useful foundation for more systematic donor selection. Incorporating these tools into clinical decision-making could enable more effective donor–recipient matching, improve FMT efficacy, and reduce the likelihood of adverse or suboptimal outcomes.

Live biotherapeutic products (LBPs) - such as the probiotic Clostridium butyricum strain CBM-588^47–51, or the defined bacterial consortia SER-401 and VE800⁵² - represent a more reductionist alternative. Rather than transferring an entire ecosystem, these approaches introduce one or a small number of well-characterized strains designed to deliver a specific function - most notably, in the case of CBM-588, stimulation of endogenous Bifidobacterium populations through bifidogenic metabolite production. This strategy offers advantages in manufacturing consistency, safety characterization, and mechanistic clarity, but it sacrifices the ecological complexity that may be necessary for durable community restructuring.

The CBM-588 trials highlight the fundamental trade-off of reductionist microbiome strategies. In two independent randomized studies in renal cell carcinoma, CBM-588 was added to an immunotherapy backbone with Bifidobacterium enrichment as the primary endpoint^53,54. Clinical benefit was reproducible across treatment backbones, with the magnitude of benefit in the nivolumab–cabozantinib trial (74% versus 20% ORR) ranking among the largest effect sizes reported in microbiome-oncology studies. However, neither trial met the primary endpoint of Bifidobacterium enrichment, suggesting that CBM-588 alone was insufficient to overcome colonization resistance at the doses and durations evaluated. This apparent disconnect between limited microbiome remodeling and robust clinical efficacy raises several questions. From a trial design perspective, surrogate microbiome endpoints - such as enrichment of a target taxon - may be entirely decoupled from the immunological mechanisms driving clinical benefit, and their use as primary endpoints risks incorrectly classifying active microbiome interventions as failures. However, it is equally possible that CBM-588 exerts immunomodulatory effects independent of durable engraftment. These effects may include transient metabolite production during intestinal transit, trophic support of existing commensal populations, and direct interactions with mucosal immune cells. As a butyrate-producing organism, C. butyricum could transiently enhance epithelial barrier integrity, reduce systemic lipopolysaccharide translocation, mitigate inflammation, and attenuate myeloid-mediated immunosuppression - effects sufficient to augment checkpoint inhibitor efficacy without requiring long-term colonization. Another explanation is that the immunotherapy combinations themselves, particularly dual checkpointinhibition, partially compensate for incomplete microbiome restructuring. Regimens such as nivolumab plus ipilimumab activate broader immune programs that may overlap with or substitute for microbiome-mediated immune priming. In contrast, single-agent PD-1 blockade may be more dependent on microbiome-derived signals to achieve effective T-cell activation thresholds. Conversely, reductionist products that failed to engraft or meaningfully restructure the microbiome, such as SER-401, VE800 and EDP1503, showed no improvement in ORR or PFS^55–57. Across trials, success consistently correlated with either durable ecosystem-level restructuring or sufficient functional activity to alter immune tone. Products lacking both properties reliably underperformed, clarifying the mechanistic reasons for these failures.

Dietary composition is a major determinant of gut microbiota structure, with measurable alterations occurring within hours of changes in food intake⁵⁸. Large-scale epidemiologic evidence underscores the relevance of diet to cancer risk: the Umbrella study, encompassing 860 meta-analytic comparisons across cancer types, identified numerous significant associations between dietary patterns and cancer incidence⁵⁹. Within the specific context of ICI-treated solid tumors, consistent epidemiologic signals have emerged linking higher fiber intake, adherence to a Mediterranean dietary pattern, and lower consumption of non-nutritive sweeteners to improved clinical outcomes^60–62.

Against this backdrop, dietary intervention trials that prioritize fiber intake have emerged as a distinct and increasingly credible subset of nutritional studies in oncology. Unlike earlier diet trials focused on weight loss or quality-of-life endpoints, these interventions are explicitly designed to modulate the gut microbiome and downstream immune function. The strongest prospective evidence comes from the DIET and NUTRIVENTION trials. In DIET, melanoma patients pending ICI initiation in the neoadjuvant or metastatic settings were randomized to either a controlled high-fiber diet (30–50 g/day) or a healthy control diet (~20 g/day)⁶³. The high-fiber diet was relatively well-tolerated, alongside favorable microbiome remodeling and resulted in improved event-free survival (EFS) in the neoadjuvant ICI setting, although no changes in pathologic response (neoadjuvant setting) or ORR (advanced setting) were seen⁶³. Fiber-based dietary interventions have also shown promise beyond advanced solid tumors, particularly in premalignant disease settings. In the NUTRIVENTION trial, a high-fiber, plant-based diet administered to patients with myeloma precursor states (MGUS or smoldering myeloma) led to coordinated improvements in metabolic parameters, systemic inflammation, and gut microbiome composition, accompanied by an absence of progression to multiple myeloma at one year⁶⁴. Supported by concordant preclinical models, these findings suggest that fiber-driven microbiome modulation may be most effective when immune and metabolic systems remain relatively intact, positioning dietary intervention as a potential disease-modifying strategy in early or pre-cancer states rather than solely as supportive care.

Taken together, these data highlight that the critical next step in microbiome modulation for cancer immunotherapy is not choosing between FMT, LBPs, or dietary intervention in isolation, but rather matching the intervention to the biological and clinical context. FMT remains the most powerful modality for large-scale ecosystem remodeling, yet its intrinsic regulatory, safety, and scalability challenges constrain widespread use outside of rationally designed, standardized allogeneic products. One such example is MaaT Pharma’s MaaT033, which has demonstrated safety, robust microbiota reconstruction, and favorable pharmacodynamic signals - including donor-associated microbial engraftment, increased gut microbiota richness, reduced systemic inflammatory markers, and restoration of SCFA production - in heavily pretreated AML patients receiving intensive chemotherapy and broad-spectrum antibiotics⁶⁵. These data suggest that controlled, industrialized ecosystem replacement may be feasible in select settings where profound dysbiosis is anticipated and tolerated. A central unresolved question across FMT-based strategies is whether donor engraftment, donor–recipient convergence, or selective loss of recipient taxa is required for therapeutic success. While initial trials of FMT in PD-1 naive or PD-1 refractory melanoma demonstrated that response was associated with engraftment^40,41, metagenomic strain-profiling suggested that engraftment was critical to FMT success in recurrent Clostridioides difficile infection but may be less critical in immuno-oncologic settings, where functional reorganization or selective pathobiont displacement may suffice⁶⁶. The More recent studies linking clinical benefit to loss of specific baseline species rather than donor strain acquisition further support this latter model⁴⁴. This mechanistic controversy - engraftment versus convergence versus subtraction - will be directly addressed in ongoing randomized trials of FMT-based interventions in cancer, including CanBiome2 (NCT06623461) and PICASSO (NCT04988841), and is likely to shape the future design of microbiome therapeutics. LBPs (such as CBM-588, SER-401 and VE800) represent a more reductionist alternative. Rather than transferring an entire ecosystem, these approaches introduce one or a limited number of well-characterized strains intended to deliver specific immunomodulatory functions. This strategy offers advantages in manufacturing consistency, safety characterization, and regulatory tractability, but may lack the ecological breadth required for durable community restructuring. However, more recently developed products such as MB097 suggest that rationally designed multi-strain consortia can deliver complementary immune-stimulatory activities - such as DC IL-12 induction and metabolite-mediated relief of myeloid suppression - without requiring durable ecosystem-level engraftment⁶⁷, and prospective data from the MELODY-1 study (NCT06540391) that tested MB097 in ICI-refractory and ICI-naïve melanoma is awaited.

Dietary interventions occupy a complementary niche. While carbohydrate-restricted diets that simultaneously reduce fermentable fiber - most notably ketogenic diets - have produced heterogeneous and sometimes conflicting results⁶⁸, fiber-preserving or fiber-enriching dietary strategies show more consistent biological alignment with microbiome-mediated immune modulation. High-fiber dietary interventions reliably improve microbial diversity, metabolic output, and mucosal immune tone, and appear most effective in early-stage or premalignant settings where immune and metabolic systems remain relatively intact⁶⁹. Collectively, these findings point to a unifying principle: microbiome-targeted strategies are most likely to influence cancer outcomes when they preserve or restore fermentable substrate availability and engage the microbiome–immune axis, rather than focusing narrowly on macronutrient restriction or systemic metabolic endpoints alone.

FUNCTION OVER TAXONOMY: MICROBIOME-ENCODED IMMUNE STATES GOVERN RESPONSE AND TOXICITY IN IMMUNOTHERAPY TREATED CANCER

Across ICI, CAR-T therapy, and allogeneic hematopoietic stem cell transplantation (alloSCT), observational and interventional studies converge on a central principle: the functional state of the gut microbiome, rather than the presence or absence of specific taxa, governs therapeutic efficacy and immune-mediated toxicity. The inconsistent performance of taxa-level biomarkers across cohorts reflects several biological realities. First, the same immunologically relevant biochemical mechanisms may be encoded by distinct species in different individuals, while conversely, closely related strains within a single species may diverge markedly in functional capacity - a phenomenon supported by strain-resolved analyses^17,66. Second, different enterotypes and geographic contexts may harbor distinct organisms that nonetheless mediate equivalent immune-modulatory functions⁷⁰. While gene- and pathway-level analyses therefore offer a more principled framework than taxonomy alone, their predictive performance to date has been modest, despite encouraging signals reported using machine-learning approaches by multiple groups^16,17,71. Importantly, microbiome-mediated enhancement of antitumor immunity is likely multifactorial - potentially involving 10–20 distinct mechanisms described to date - and these pathways are unlikely to be simultaneously active in all patients. As a result, individual “beneficial” taxa may not emerge as strong or consistent signals across studies, even when they are mechanistically relevant. In contrast, deleterious microbiome states may be more conserved across patients, suggesting that enrichment of a microbiome-mediated inflammatory program is often more predictive of poor response than enrichment of any single beneficial organism^17,72.

Within this functional framework, recurring bacterial lineages should be viewed primarily as markers of underlying immune-relevant programs, including inflammatory signaling, epithelial barrier integrity, and microbial metabolite production, rather than as direct caUnited Statesl agents. The concept of the pathobiont provides a unifying mechanistic explanation for microbiome-associated resistance and toxicity across diverse therapeutic modalities, including antibiotics, chemotherapy, and immunotherapy. In ICI-treated solid tumors, Gram-negative pathobionts can promote tonic innate immune activation - most prominently via TLR4-dependent myeloid skewing and IL-8–driven immune exclusion - thereby impairing effective CD8⁺ T-cell responses¹⁷. In this context, recurrent enrichment of genera such as Gram-negative Bacteroides or Gram-positive Enterocloster in non-responders reflects a shared inflammatory function albeit through distinct mechanisms - TLR4-driven myeloid activation for Bacteroides and loss of SCFA-mediated barrier support for Enterocloster species, rather than species-specific causality. Conversely, responder microbiomes in melanoma and other ICI-treated tumors are enriched for taxa within Actinobacteria and Firmicutes, particularly Lachnospiraceae and Ruminococcaceae, not because these organisms are intrinsically beneficial, but because they participate in functional programs - such as SCFA, secondary bile acid metabolism, and tryptophan catabolism - that support epithelial integrity and cytotoxic T-cell function^{5,7,11,15,17,18,32,40,41,60,73–76}.

Strikingly similar functional patterns emerge in CAR-T therapy. Multi-center studies of CD19 CAR-T recipients demonstrate that higher baseline microbial diversity and preservation of anaerobic, Clostridia-associated metabolic functions correlate with complete response and improved survival, whereas disruption by broad-spectrum, anaerobe-depleting antibiotics is associated with inferior outcomes and increased immune effector cell–associated neurotoxicity syndrome (ICANS)^{12,15,29,30,77}. Here again, the clinical signal is not driven by a single organism but by loss of microbiome functions that restrain inflammatory tone and regulate cytokine amplification, predisposing patients to cytokine release syndrome and ICANS. In alloSCT, these same principles are amplified. Large, multi-institutional cohorts have shown that loss of gut microbiome diversity and domination by single taxa - particularly Enterococcus and other facultative pathobionts - predict increased transplant-related mortality, graft-versus-host disease (GVHD), and infection-related death^78–81. Conversely, preservation of diverse, anaerobe-rich communities dominated by Clostridiales and SCFA-producing taxa is consistently associated with improved survival across geographic regions and age groups. Taken together, these data indicate that microbiome composition is not therapy-specific but instead represents a shared tumor-extrinsic determinant of T-cell mediated immune performance.

Meta-analyses and toxicity-focused studies further reinforce that efficacy and safety are governed by overlapping functional microbiome programs. In melanoma, Lachnospiraceae-associated communities correlate with improved ICI response but also with an increased propensity for irAEs, consistent with heightened immune activation¹⁷. In contrast, Streptococcaceae-enriched microbiomes are associated with colitis and systemic toxicity while conferring poor antitumor efficacy, highlighting a dysregulated inflammatory state that favors toxicity without effective tumor control¹⁷. Similarly, in alloSCT, expansion of Enterococcus, Enterobacteriaceae, and antibiotic-resistant organisms predicts GVHD severity, bloodstream infections, and non-relapse mortality independently of conditioning intensity or donor characteristics^81,82. Across platforms, these observations support a shared principle: pathobiont expansion amplifies inflammatory programs that destabilize T-cell–based therapies, either by suppressing antitumor immunity or by exacerbating immune toxicity.

Taken together, these data argue against a taxonomy-centric interpretation of microbiome effects. The core-microbiome and functional guild framework helps explain why taxa-level associations with immunotherapy outcomes are often inconsistent across cohorts^83,84. In this model, host-relevant effects are mediated by functionally coherent microbial guilds - such as SCFA-producing, barrier-supportive, or inflammatory communities - whose taxonomic composition may vary across enterotypes, geographies, and diets while their functional outputs remain conserved. This explains why negative, pathobiont-driven inflammatory states are often more reproducible predictors of resistance and toxicity than enrichment of any single beneficial organism. Diet acts as a dominant upstream regulator of these guilds by shaping substrate availability and ecological competition, and recent studies show that dietary interventions can rapidly reprogram microbiome functional states even without durable taxonomic replacement, reinforcing that microbiome effects on ICI, CAR-T, and alloSCT are best understood at the level of function rather than taxonomy⁸⁵. Specific taxa matter insofar as they represent functional states - pathobiont-driven inflammation versus metabolically supportive, immune-permissive ecosystems. The consistent recurrence of these functional poles across ICI, CAR-T, and alloSCT indicates that the gut microbiome acts as a shared, tumor-extrinsic regulator of T-cell immunity, with outcomes determined less by “who is there” than by what the microbial community is doing.

ECOLOGICAL DETERMINANTS OF THERAPEUTIC SUCCESS: AI AND ML APPROACHES TO DRIVERS OF RESPONSE

The gut microbiome is not a collection of independently regulated organisms, but a complex ecological system governed by the same principles that shape macroscopic ecosystems, including competitive exclusion, trophic interactions, spatial structure, and stability landscapes^86,87. Within this system, established microbial communities exert colonization resistance against incoming organisms through multiple mechanisms: competition for shared nutrients, production of bacteriocins and other antimicrobials, and activation of host immune responses that selectively eliminate non-resident strains⁸⁶. These ecological forces strongly constrain which microbes can establish durable populations following microbiome-targeted interventions, helping to explain the heterogeneity of clinical responses observed across trials⁸⁸.

Schluter et al. provided quantitative evidence that these ecological interactions translate directly into immune regulation in humans. In a large cohort of cancer patients undergoing alloSCT⁸² - a clinical setting in which both the immune system and the microbiome are profoundly disrupted and subsequently reconstituted. Using Bayesian inference, they demonstrated reproducible, day-to-day associations between specific gut bacterial genera and circulating neutrophil, lymphocyte, and monocyte counts⁸². Notably, the magnitude of these microbiome-associated effects on immune cell dynamics was comparable to that of immunomodulatory medications. Genera such as Faecalibacterium and Ruminococcus, which also tend to be associated with favorable responses in ICI-treated cancer¹⁷, showed among the strongest positive associations with white blood cell reconstitution, demonstrating that the microbiome is not merely a passive correlate of immune recovery, but instead plays an active regulatory role.

Given this complexity, microbiome management in cancer should be framed as an ecological control problem rather than a taxonomic optimization exercise. This framing naturally motivates the application of artificial intelligence (AI) and machine learning (ML) approaches, which are uniquely suited to integrating high-dimensional, longitudinal microbiome data with host immune, clinical, and treatment variables. Rather than focusing on static taxa-level associations, AI- and ML-based models can infer latent ecological states, identify functionally relevant microbial guilds, and capture non-linear dynamics that govern transitions between immune-permissive and immune-suppressive microbiome configurations^83,89,90.

ML models using gut microbiome data have shown potential for predicting responses to ICI, though early models achieved only moderate accuracy^71,74. Geographic variation, with location biased prevalence of different taxa or guilds/gut microbiotypes and differences across cancer types and treatments further complicate model generalization, though these factors can be partially addressed with proper data correction^10,17,91. As expected, somewhat better prediction was observed when gene abundance rather than taxonomy was used for training ML models^11,16,71. Performance improves when functional microbiome data or multi-omics inputs - such as fungal profiles and microbial metabolites - are included. In some cases, these models outperform traditional cancer biomarkers^92–95. Evidence suggests that microbiome features associated with poor outcomes are more consistent than those linked to positive responses further supporting an ecological equilibrium with negative species suppressing cancer immunity or preventing the dominance of immunity permissive species^17,72.

Despite strong associations between microbiome composition and immunotherapy outcomes, establishing causality remains challenging. Importantly, some microbiome signatures reflect an overall disease associated phenotype rather than treatment response, highlighting the need to distinguish between these outcomes^83,96. New approaches, including genetic analyses and network-based methods, are beginning to clarify causal relationships^24,97–99. Greater precision will likely require focusing on strain-level differences and functional gene variation as well as on the balance within strains in the different ecological niches of the gastrointestinal tract^10,17,100. An ecological perspective also clarifies the mixed effects of antibiotic pre-conditioning prior to FMT. While antibiotics can transiently reduce colonization resistance and disrupt entrenched pathobiont dominance, they may also create a highly permissive but unstable ecosystem in which both beneficial and harmful organisms can colonize indiscriminately^55,101. The therapeutic optimum therefore lies between insufficient perturbation and wholesale ecosystem collapse: enough disruption to enable functional remodeling without inducing ecological chaos. Achieving this balance will likely require ecologically informed, AI-guided strategies that tailor microbiome interventions to the host’s baseline ecosystem, dynamically monitor response, and adaptively steer the system toward stable, immune-permissive states

CONCLUSIONS

Over the past 13 years, preclinical and clinical studies have generated compelling evidence that the gut microbiome is both a determinant of, and a therapeutic target for, response to cancer immunotherapy. Although fecal microbiome transplantation (FMT) has provided important proof of concept that therapeutic modulation of the microbiome can overcome resistance in a subset of patients, its biological complexity and interindividual variability underscore the need for more precise, mechanism-driven strategies. Both the depletion of pathobionts, which may contribute to an immunosuppressive state through disruption of mucosal integrity and induction of chronic inflammation, and the acquisition or expansion of beneficial taxa, which may promote antitumor immunity through the production of immunostimulatory signals or metabolites, appear to be critical determinants of immunotherapy efficacy. Artificial intelligence and machine learning approaches capable of predicting the ecological equilibrium of the recipient gut microbiome on the basis of donor fecal composition or live biotherapeutic products, together with the ecological features of the recipient microbiome, may prove valuable in anticipating the effects of therapeutic microbiome modulation⁶⁶. Progress in clinical application has been slow due to biological differences between model systems and humans, as well as challenges in microbial colonization. Future strategies may involve designing stable bacterial consortia to improve therapeutic effectiveness, an approach that has shown high efficacy in mouse models but still failed to achieve clinical success⁵². A precise definition of the microbial, immunologic, and ecological features that underline durable clinical benefit will therefore be essential for translating advances in microbiome science into reproducible and effective cancer therapies.

ABBREVIATIONS

AI

artificial intellingence

alloSCT

allogeneic hematopoietic stem cell transplantation

BA

bile acid

CAR-T

chimeric antigen receptor T-cell

CTLA-4

cytotoxic t-lymphocyte associated protein 4

FMT

fecal microbiome transplantation

GVHD

graft versus host disease

ICI

immune checkpoint inhibitor

irAE

immunotherapy-related adverse event

ML

machine learning

LBP

live biotherapeutic product

NSCLC

non-small cell lung cancer

ORR

objective response rate

PD-1

programmed death 1

PFS

progression-free survival

RCC

renal cell carcinoma

SCFA

short-chain fatty acid