Crohn’s disease (CD) is an immune-mediated chronic inflammatory condition of the gastrointestinal tract (Andoh and Nishida, 2023). It has a multifactorial disease aetiology that is a result of an aberrant immune response towards the microbiome in a genetically susceptible host (Turpin et al., 2018). This notion is supported by clinical and experimental evidence: in CD patients, diversion of the faecal stream through surgical interventions can reduce inflammation in affected bowel segments (Burke, 2019), while in experimental animal models, intestinal inflammation is markedly more difficult to generate in germ-free conditions (Hernández-Chirlaque et al., 2016).

The faecal microbiome of CD patients is associated with an overall reduced microbial α-diversity and species richness compared to the microbiome of healthy individuals (Becker et al., 2015). The observed dysbiosis is characterized by increased prevalence of a low cell count Bacteroides 2 enterotype-like composition, with the bacterial load associating inversely with systemic and intestinal inflammation (Vieira-Silva et al., 2019).

CD patients are treated with mainstream immunosuppressive medications and biological agents such as the antibodies infliximab (IFX) and adalimumab (ADA), which both recognize TNFα (anti-TNFα), vedolizumab (VDZ), targeting the α₄β₇ integrin and ustekinumab (USTE) that binds the p40 subunit of interleukin (IL)-12 and IL-23 (Torres et al., 2020). Unfortunately, on average 40% of patients treated with biological agents will fail to respond or lose their response over time (Sands et al., 2004; Papamichael et al., 2015; Peyrin-Biroulet et al., 2019). Hence, clinicians have to decide upon treatments with different modes of action without a diagnostic test that can predict which therapy is suited for the individual patient. Understanding mechanisms behind treatment resistance and finding biomarkers that predict response to biological treatment would advance the current practice for CD patients (Joustra et al., 2025).

Faecal microbial signatures have been investigated for their predictive potential regarding patient response to anti-TNFα (Busquets et al., 2015; Aden et al., 2019; Ribaldone et al., 2019; Sanchis-Artero et al., 2021; Ventin-Holmberg et al., 2021; Chen et al., 2022; Park et al., 2022), VDZ (Ananthakrishnan et al., 2017) and USTE therapy (Doherty et al., 2018). However, their performance has been limited, which may be attributable to two well-established confounding factors of the faecal microbiome: diet and stool consistency (Vandeputte et al., 2016; Dinsmoor et al., 2021). It is suggested that the adherent microbiome, consisting of a bacterial community that directly adheres to the inner lining of the gut and is therefore better representative of interactions with intestinal epithelial cells and the underlying immune cells (Shi et al., 2017; Yilmaz et al., 2019; Mavragani et al., 2020; Sakurai et al., 2020; Juge, 2022), may potentially provide a more stable signature for biomarker purposes (Dovrolis et al., 2020). Likewise, the aim of this study was to assess the potential of the intestinal tissue-adherent gut microbiome to predict response of patients with Crohn’s disease to treatment with biological agents.

We collected ileal and colonic biopsies from CD patients prior to starting treatment with anti-TNFα, VDZ or USTE. Treatment response was recorded at 26–52 weeks after start of therapy, and 16S rRNA gene sequencing of biopsies was performed to assess the baseline tissue-adherent microbiome in ileum and colon. Using supervised machine learning, we identified a therapy response model based on microbial composition. The anti-TNFα colonic microbiome model outperformed randomization and showed good predictive performance. Feature importance analysis identified Mediterraneibacter gnavus (formerly Ruminococcus gnavus) ASVs associated with non-response, and Blautia ASVs with response. Building on our previous finding that anti-TNFα therapy response depends on the induction of CD206⁺ expressing regulatory type macrophages (Vos et al., 2011; Wildenberg et al., 2017), we performed proof-of-concept mixed lymphocyte reactions (MLRs) to explore whether M. gnavus and Blautia luti may influence macrophage polarization and, potentially, anti-TNFα responsiveness.

In this study, we demonstrate that the intestinal tissue-adherent microbiome obtained from diagnostic colonic biopsies prior to treatment initiation contains informative signatures associated with response to anti-TNFα therapy in Crohn’s disease. By profiling pre-therapy intestinal biopsies using 16S rRNA gene sequencing, we characterized baseline mucosa-associated microbial communities and leveraged these data to identify microbial features linked to treatment outcome. Using machine learning techniques incorporating stability selection and extra trees classifier, we identified a robust panel of ASVs with high discriminatory performance for classifying responders and non-responders to anti-TNFα therapy based solely on baseline colonic microbiota profiles. Importantly, these associations were observed prior to treatment initiation, highlighting the potential clinical relevance of tissue-adherent microbial signatures for stratifying patients before exposure to biological agents. To explore potential biological mechanisms underlying these associations, we performed a proof-of-concept anti-TNFα macrophage polarization experiment. In this setting, the presence of M. gnavus was associated with reduced numbers of M2-like macrophages, supporting a functional link between specific microbial taxa and host immune responses relevant to anti-TNFα efficacy. Collectively, our findings indicate that the colonic mucosa-associated microbiota represents a promising source of baseline biomarkers for anti-TNFα therapy in CD and suggest that M. gnavus may contribute to microbial-immune interactions associated with treatment non-response.

Most microbiome studies to date have relied on faecal samples because they are easily accessible. While some of these investigations have incorporated machine learning approaches, predictive performance has generally been moderate and often lacks validation across independent training-test splits (Ventin-Holmberg et al., 2021). When appropriately processed, intestinal biopsies have been shown to comprise both loosely-adherent and strictly-adherent microbial communities (Mukhopadhya et al., 2022), which more closely reflect the physiological environment of the gut mucosa, including mucus secretion and anti-microbial peptide secretion, and show less diet induced variability (Zoetendal et al., 2002; Vaga et al., 2020). Moreover, analysis of tissue-adherent microbiota circumvents well-established confounders associated with faecal sampling, such as stool consistency (Vandeputte et al., 2016). Based on these considerations, we hypothesized that tissue-adherent bacteria may represent more stable biomarkers than faecal microbiota and therefore provide improved prediction of medication response.

Indeed, for anti-TNFα our approach demonstrated good predictive performance for colonic biopsies. However, prediction using ileal biopsies did not exceed performance expected by chance, that is, permuted labels. This discrepancy may relate to α-diversity, which was significantly higher in colonic biopsies from responders than non-responders, a pattern not observed in the ileum. We speculate that a relatively rich adherent microbiota in the colon may support treatment response in CD. Nevertheless, the primary aim of this study was to use machine learning to identify predictive microbiome features rather than to compare microbial composition using conventional metrics. Accordingly, interpretations based on α- and β-diversity should be approached with caution.

Our approach was unable to provide accurate prediction of response to VDZ or USTE, which we attribute to prior exposure to biologics. Nearly 50% of patients receiving anti-TNFα were biologic-naïve, whereas most patients treated with VDZ and USTE had undergone multiple rounds of biologic therapy. We found that prior exposure to VDZ or USTE shifted the adherent microbial composition, driving it toward a more “dysbiotic” state and thereby limiting the predictive value of post-treatment microbiome profiles. Overall, these results suggest that microbiome-based prediction may be less feasible in heavily pretreated patient populations.

We focused on ASVs profiled before treatment initiation. ASVs linked to M. gnavus and Blautia emerged as the strongest predictors of treatment outcome, associated with non-response and response to anti-TNFα, respectively. These findings align with previous studies showing that M. gnavus is enriched in baseline samples of non-responders and decreases following successful ADA therapy in Crohn’s disease patients (Ribaldone et al., 2019). Conversely, Blautia has been consistently associated with response to IFX (Zhuang et al., 2020), and is more abundant in pre-treatment samples of responders (Park et al., 2022).

M. gnavus is a key indicator of a low-diversity, dysbiotic “Bacteroides2-like” enterotype composition (Bresser et al., 2022), which is often observed in patients with CD (Vieira-Silva et al., 2019). Several studies have elucidated mechanisms relevant to M. gnavus’ s role in human health and disease (Crost et al., 2023). For one, M. gnavus demonstrates the ability to adapt and adhere to the gut lining; it is known to produce antimicrobial peptides, including bacteriocins to kill other bacterial taxa (Crost et al., 2023). Furthermore, it can use complex carbohydrates as nutrients and degrade intestinal mucus for its own energy source (Hall et al., 2017; Franzin et al., 2021; Crost et al., 2023). In IBD, it is hypothesized that aberrant mucus degradation might lead to increased gut permeability, possibly enhancing the inflammatory response (Hall et al., 2017). M. gnavus also harbours glucorhamnan on the cell surface, a polysaccharide that potently induces TNFα secretion by dendritic cells, in a toll-like receptor 4-dependent manner (Henke et al., 2019). Thus, in CD patients with high relative abundance of M. gnavus, i.e. anti-TNFα non-responders, glucorhamnan may play a role in uncontrolled secretion of TNFα. This could partially explain why anti-TNFα is inefficacious in these patients and warrants future investigations into this polysaccharide and its possible relation to treatment failure. Besides glucorhamnan, some M. gnavus strains harbour another, recently described, capsular polysaccharide that promotes a tolerogenic instead of a proinflammatory immune response. Henke et al. showed that the absence of this protective polysaccharide, as observed in some IBD-derived isolates, elicited a robust inflammatory immune response (Henke et al., 2021). Lack of this capsular M. gnavus polysaccharide could also be important in response’ s failure. The possible relevance of the two different M. gnavus expressed polysaccharides in therapy response can be addressed by in vitro assays, for which tissue adherent strains will have to be isolated from an anti TNFα treatment cohort.

To investigate whether ASVs annotated to M. gnavus and Blautia may influence anti-TNFα responsiveness, we performed a series of in vitro proof-of-concept experiments. We focused on anti-TNFα-induced polarization of macrophages toward the M2 phenotype, which is known for its anti-inflammatory properties and role in tissue repair. The balance between (pro-inflammatory)-M1 and (anti-inflammatory)-M2 macrophages is critical for the development and progression of CD (Zhou et al., 2019; Zhang et al., 2023). In previous mouse transfer colitis studies, we observed that anti-TNFα therapy induces a shift toward CD14⁺CD206⁺ M2 regulatory macrophages (Wildenberg et al., 2017). This shift was confirmed in vitro using both mouse and human MLR assays and corroborated in a small in situ study of anti-TNFα treated patients (Vos et al., 2011; Wildenberg et al., 2017), where induction of regulatory macrophages occurred only in successfully treated patients.

Here, we assessed the effect of M. gnavus and B. luti on anti-TNFα induced M2 macrophage polarization in human MLRs. Because fresh biopsies were not available, isolation of bacterial strains from patient biopsies was not feasible; therefore, these experiments were restricted to the strains M. gnavus RJX1124 (Henke et al., 2019) and B. luti DSM14534. The M. gnavus RJX1124 strain was selected because it was originally isolated from an IBD biopsy. However, this strain is characterized by the absence of a tolerogenic polysaccharide, expression of pro-inflammatory glucorhamnan, and a strong capacity to induce TNFα production in murine bone-marrow-derived dendritic cells (Henke et al., 2021). These functional properties cannot be inferred from our ASV-level resolution, and the strains driving the predictive associations in our models may therefore differ, potentially leading to different outcomes in the MLR assay. The B. luti DSM14534 strain was selected in an even more arbitrary manner. In contrast to M. gnavus, for which BLAST-based querying allowed species-level annotation of associated ASVs, ASVs assigned to Blautia could not be resolved to the species level using either the SILVA database or NCBI BLAST searches. Consequently, strain selection could not be guided by sequence-based annotation. We therefore selected B. luti based on species-level relevance, as it represents one of the most abundant Blautia species in the human gut microbiome (Liu et al., 2021), providing a biologically plausible proof-of-concept model. However, this approach inherently implies that alternative Blautia species or strains could exert different immunomodulatory effects, and that the outcome of the MLR assays may vary accordingly.

The successful induction of CD14⁺CD206⁺ regulatory macrophages by IFX was completely reversed by the addition of heat killed M. gnavus. This result may explain negative treatment outcome in patients with high pre-therapy level of adherend M. gnavus. Yet, it has to be acknowledged that B. luti also diminished, albeit to a lesser extent, the polarizing effect of IFX. As mentioned above, this may relate to the difficulty of identifying which Blautia spp. are relevant in our predictions for therapy response. Possibly other species of the Blautia genus would have given a different, i.e. less inhibiting, response. Better identification on species level could become relevant for future therapy as well. One can envision, that identification of the exact species contributing to therapy success may eventually lead to a probiotic-like add on treatment for anti-TNFα therapy.

Our study has several strengths. CD patients were rigorously classified as responders or non-responders based on stringent clinical, endoscopic, and biochemical criteria. We focused on adherent microbial signatures from intestinal biopsies rather than faecal samples, which are often suboptimal for capturing mucosa-associated microbial dynamics. While intestinal biopsies are inherently low in biomass and therefore more susceptible to contamination during sampling, storage, DNA extraction, and amplification, we implemented a comprehensive decontamination screening protocol to ensure high-quality input for both conventional and machine learning analyses. Additionally, working at the ASV level, rather than the broader genus level, enabled the detection of more specific microbial signals, such as M. gnavus, which appeared in four of the 15 most informative ASVs within the anti-TNFα colonic cohort. Importantly, our predictive ASVs were identified in colonic samples from anti-TNFα–treated patients regardless of disease location (ileal, colonic, or ileocolonic CD), indicating that the predictive model is robust across different CD subtypes.

One primary limitation of our approach is that the anti-TNFα patient cohort included individuals treated with two different biologicals, ADA and IFX. This choice was guided by several considerations: (a) both ADA and IFX target TNFα; (b) sample sizes for ADA and IFX separately were too small to support machine learning analyses, which were the primary focus of this study; and (c) α-diversity analyses within the ADA and IFX subgroups showed that responders to both therapies tended to have higher α-diversity, although the difference was not significant for IFX, likely due to limited sample size. Despite structural differences (IFX is a humanized mouse monoclonal antibody, whereas ADA is fully human (Mpofu et al., 2005)), we have previously demonstrated that both antibodies similarly induce macrophage polarization and inhibit T-cell proliferation (Vos et al., 2011). In the current study, the most informative ASVs predicting non-response to anti-TNFα showed higher relative abundance in non-responders compared with responders in both the ADA and IFX subgroups, further supporting the reliability of our findings. Taken together, pooling ADA and IFX is unlikely to have affected microbe-drug interactions; nonetheless, future validation in independent cohorts will be important to replicate these findings.

In this study we present a machine-learning based model to predict success or failure of anti-TNFα therapy in patients with CD. Future studies on cohorts with larger sample size are needed to validate our findings, and possibly reveal microbial signatures during treatment to assess the effect of therapy on the microbiome. M. gnavus ASVs were most predictive for non-successful therapy response whereas Blautia spp. related to treatment success. In vitro assays suggested that M. gnavus may interfere with the M2 polarizing effect of anti-TNFα. Eventually, designing predictive tests to target tissue-adherent microbial biomarkers could improve the current clinical practice and inform on personalized treatment strategies for CD patients.