Response-aware molecular subtyping of ulcerative colitis for patient stratification via interpretable machine...

Ulcerative colitis (UC) is a heterogeneous inflammatory disease with diverse molecular features and variable responses to biologic therapies. Although molecular subtyping has been proposed to characterize disease heterogeneity, most existing approaches do not incorporate treatment response during subtype identification, potentially limiting their clinical relevance. Therefore, we developed an interpretable machine learning framework for response-aware molecular subtyping of UC. Using two cohorts with infliximab (IFX) response information, a predictive model was trained to distinguish responders from non-responders. SHAP values were used to derive response-related representations, which were then used to construct a similarity graph for spectral clustering to identify molecular subtypes. Four UC subtypes with distinct molecular and immune characteristics were identified. Subtypes 1 and 2 showed low IFX response rates, characterized by broad immune activation and an innate immune-dominant profile, respectively. Subtype 3 was characterized by metabolic pathway activation and heterogeneous response patterns, while Subtype 4 demonstrated relatively low immune activation and the highest response rate. The trained framework was subsequently applied to additional cohorts without response information for subtype assignment, where similar subtype-associated molecular patterns were observed, supporting the generalizability of the learned representations. Subtype-associated biomarkers identified by machine learning models showed strong discriminative ability across datasets, including an independent cohort, and were sufficient to recapitulate subtype structure and associated response trends. By incorporating treatment response into subtype identification, this interpretable framework identifies clinically relevant UC subtypes with distinct molecular and immune characteristics, providing a basis for response-aware patient stratification and more informed therapeutic decision-making in UC.