Response-aware molecular subtyping of ulcerative colitis for patient stratification via interpretable machine
Tianyi Shi1, Xiucai Ye2, Yuta Nakazawa3
1Department of Computer Science, University of Tsukuba, Tsukuba, Japan.
Computational Biology and Chemistry
|July 18, 2026
Summary
This study introduces a novel machine learning approach for molecular subtyping in ulcerative colitis (UC) that considers treatment response. It identifies four distinct UC subtypes with varying infliximab (IFX) response rates and unique molecular profiles.
Area of Science:
- Immunology
- Computational Biology
- Gastroenterology
Background:
- Ulcerative colitis (UC) is a complex inflammatory bowel disease with significant heterogeneity.
- Current molecular subtyping methods often overlook treatment response, limiting clinical applicability.
- Personalized medicine requires patient stratification based on molecular profiles and predicted treatment outcomes.
Purpose of the Study:
- To develop an interpretable machine learning framework for response-aware molecular subtyping of UC.
- To identify distinct UC subtypes based on molecular features and infliximab (IFX) response.
- To establish a foundation for improved patient stratification and therapeutic decision-making in UC.
Main Methods:
- Developed an interpretable machine learning framework integrating treatment response data.
- Trained a predictive model using two cohorts with IFX response information to differentiate responders and non-responders.
- Utilized SHAP values for response-related representations, followed by spectral clustering to identify molecular subtypes.
- Validated the framework on independent cohorts to assess generalizability.
Main Results:
- Identified four distinct UC molecular subtypes with unique immune and molecular characteristics.
- Subtypes 1 and 2 exhibited low IFX response rates, linked to immune activation profiles.
- Subtype 3 showed metabolic pathway activation, while Subtype 4 had low immune activation and the highest IFX response rate.
- Machine learning-derived biomarkers effectively discriminated subtypes across datasets and predicted response trends.
Conclusions:
- The developed framework successfully integrates treatment response into UC molecular subtyping.
- Identified clinically relevant UC subtypes with distinct molecular and immune signatures.
- This approach facilitates response-aware patient stratification and informs personalized therapeutic strategies for UC management.