Machine learning predicts relapse after anti-TNF withdrawal in ulcerative colitis using rectal biopsy transcriptomics

Background There are currently no available biomarkers that predict the clinical outcome of ulcerative colitis (UC) patients after anti-TNF withdrawal. Here we developed a semi-supervised machine learning framework on RNA-seq data derived from rectal biopsies to predict relapse following anti-TNF discontinuation. Methods We included UC patients (n=136) enrolled in the prospective, randomized controlled BIOSTOP trial. Patients were assigned to a withdrawal group (stopped treatment immediately) or a maintenance group (stopped treatment after 2 years). A machine learning framework was trained on gene expression data from mucosal biopsies collected at baseline from the withdrawal group and validated in the maintenance group using baseline samples (test cohort) and 2-year follow-up samples (interference cohort). Results A semi-supervised machine learning model, combining a shared UMAP manifold representation with LabelSpreading on a targeted 38-gene panel achieved an AUC of 0.794, a sensitivity of 90.3%, and a specificity of 73.1% on the test cohort, and an AUC of 0.711, a sensitivity of 82.9%, and a specificity of 42.3% on the inference cohort. Applying Shapley Additive exPlanations (SHAP) we found that IL1B, TYROBP, REG1B, S100A8, and TNF showed the highest predictive importance for relapse risk, whereas MUC2, HMGCS2, and CDKN1A displayed the highest predictive importance for sustained remission. Conclusion The high sensitivity makes the model a promising screening tool to prevent withdrawal-related relapse, though its moderate specificity underscores the need for multi-center validation in larger, more diverse cohorts before clinical deployment.