Limited predictive value of the gut microbiome and metabolome for response to biological therapy in inflammatory bowel disease

Abstract Introduction The complexity of Inflammatory Bowel Diseases (IBD) presents challenges for the management of these diseases. Predicting treatment outcomes remains difficult, leading to suboptimal outcomes and high costs. Emerging evidence suggests the potential of the gut microbiome in predicting response to biologic treatments. In this prospective study we aimed to predict treatment response to vedolizumab and ustekinumab in 79 IBD patients by integrating clinical data, gut microbiome profiles and fecal metabolites and validating these findings in a replication cohort of 47 IBD patients. Methods Treatment response was defined as either continuation or discontinuation of the biologic at six months. We performed whole genome metagenomic shotgun sequencing on the baseline fecal samples to detect microbial and functional profiles. Additionally, over 1000 metabolites were captured through untargeted metabolomic profiling. Baseline diversity, dissimilarity and differential abundance analyses compared responder and non-responder groups. The prediction tool CoDaCoRe was used to identify predictive log-ratio biomarkers. We tested our identified ratios in an external cohort and attempted to replicate the microbiome-based signals of previous studies. Finally, we used a neural-network framework to model the relationship between metabolites and microbes, comparing these clusters in responders and non-responders and tested our approach with different definitions of response. Results We identified seven metabolites to be differentially abundant between responders and non-responders (FDR < 0.05). However, no significant differences in bacterial species and pathways were detected at baseline between responders and non-responders. Our prediction analysis indicated only marginal predictive utility of the gut microbiome and fecal metabolites for treatment response, when compared to a clinical model using fecal calprotectin, disease duration and disease activity, among other factors (AUC only clinical features: 0.71±0.13, AUC microbial and clinical features: 0.73±0.12). The main predictive features of these models were the disease activity and previous anti-TNF use combined with high abundance of Phocaeicola vulgatus, Bacteroides uniformis and Alistipes onderdonkii, and the low abundance of Ruminococcus gnavus and Faecalibacterium prausnitzii. Testing our identified ratio of 10 species in an external cohort of 47 IBD patients reinforced the lack of predictive power of the microbiome. No replication of previously published predictive signals of the microbiome was observed. Additionally, we identified 2 metabolite clusters and 1 microbiome cluster associated with response, and observed that predictors were highly dependent on the definition of response. Conclusion While previous studies of similar size have shown that microbial features can predict response to either vedolizumab (VedoNet), or vedolizumab, ustekinumab and anti-TNF, our comprehensive study found no significant differences in the gut microbiome at baseline between responders and non-responders among IBD patients treated with ustekinumab or vedolizumab. Microbial features added no predictive power to drug response in our IBD patient cohort or an independent replication cohort. Addition of metabolite features did not improve predictive power. These findings suggest minimal impact of the pre-treatment gut microbiome on treatment outcomes with these medications in IBD patients with long term chronic disease. Generalizability beyond initial study cohorts is limited, leaving predictors for individualized IBD-medication selection unidentified, but based on this work, likely do not lie in the fecal microbiome.