There are no established biomarkers to predict which patients with ulcerative colitis (UC) will relapse after stopping anti-TNF therapy. The authors developed a transcriptomics-based machine learning approach using rectal mucosal biopsies to address this unmet clinical need and to identify molecular features associated with relapse versus sustained remission.
The analysis used samples and clinical data from the prospective, randomized BIOSTOP trial (total n = 136). Participants were randomized to either a withdrawal group (treatment stopped immediately) or a maintenance group (treatment discontinued after 2 years). Mucosal rectal biopsies were collected at baseline from both groups; baseline samples from the withdrawal group were used to train the model, while the maintenance group provided independent test and inference cohorts. The maintenance group contributed baseline samples for the test cohort and 2-year follow-up samples for the inference cohort.
The investigators implemented a semi-supervised machine learning framework applied to RNA-seq data from rectal biopsies. Key components included a shared UMAP manifold representation to capture structure in the gene expression data, combined with LabelSpreading as the semi-supervised classifier. The model operated on a targeted panel of 38 genes selected for the predictive task.
On the test cohort (maintenance-group baseline samples), the combined UMAP + LabelSpreading model applied to the 38-gene panel achieved an area under the receiver operating characteristic curve (AUC) of 0.794, with sensitivity 90.3% and specificity 73.1%.
On the inference cohort (maintenance-group 2-year follow-up samples), performance was more modest: AUC 0.711, sensitivity 82.9%, and specificity 42.3%.
The authors used Shapley Additive exPlanations (SHAP) to rank gene-level contributions to predicted relapse risk. Genes with the highest predictive importance for relapse included IL1B, TYROBP, REG1B, S100A8, and TNF. Genes most predictive of sustained remission included MUC2, HMGCS2, and CDKN1A.
The model’s high sensitivity, particularly on the test cohort, suggests potential utility as a screening tool to reduce the risk of withdrawal-related relapse by identifying patients at high risk. High sensitivity favors minimizing false negatives (missed high-risk patients), which is relevant when the clinical goal is to avoid stopping therapy in those likely to relapse.
However, the specificity was moderate to low—notably reduced in the inference cohort—indicating a risk of false positives. The authors therefore emphasize that the model is not ready for routine clinical deployment and would require further validation.
The study reports model performance within the BIOSTOP trial cohorts but notes the moderate specificity and reduced performance on the inference cohort. The authors state that multi-center validation in larger and more diverse cohorts is required before clinical use. Because this report is a preprint, it has not been peer-reviewed and should not be used to guide clinical practice as presented.
Sequencing data underlying the analysis will be made available upon reasonable request to the corresponding author, subject to ethical and data-protection requirements. Ethical approval for the trial was provided by the Regional Committee for Medical Research Ethics (REK Midt 2016/531). Reported funding includes the South-Eastern Norway Health Authority (2023090). The authors declared no competing interests.
A semi-supervised machine learning model trained on rectal biopsy transcriptomics from the BIOSTOP trial discriminated patients at risk of relapse after anti-TNF withdrawal with high sensitivity and moderate specificity. SHAP analysis highlighted inflammatory and epithelial genes associated with relapse and remission. The approach shows promise as a screening tool but requires multi-center validation in larger, more heterogeneous cohorts before it can be considered for clinical implementation.