Enhancing patient stratification and interpretability through class-contrastive and feature attribution techniques

A crucial component of treating genetic disorders is identifying the genes and gene modules that drive disease processes. While Next-Generation Sequencing (NGS) provides rich data for this task, current machine learning approaches often lack explainability and fail to account for gene correlations. We develop a comprehensive framework of machine learning techniques for explainable patient stratification in inflammatory bowel disease, focusing on Crohn’s disease (CD) subtypes: CD with deep ulcer, CD without deep ulcer and IBD-controls. Our approach combines Gaussian Mixture Modelling for patient stratification, a modified kernelSHAP algorithm accounting for gene co-expression, systematic identification of gene modules, and class-contrastive analysis for explaining individual patient phenotypes. This framework confirms known disease-associated genes while unveiling novel genetic factors potentially underlying CD heterogeneity. Gene Ontology enrichment analysis validates the biological relevance of identified gene modules and associated pathways. Our methods offer a versatile toolkit for analysing high-dimensional, correlated biological data across diverse disease contexts.