Engineering Spatial and Molecular Features from Cellular Niches to Inform Predictions of Inflammatory Bowel Disease

Area of Science:

• Computational biology
• Genomics
• Immunology

Background:

• Differentiating between Crohn's disease (CD) and ulcerative colitis (UC), the two main subtypes of Inflammatory Bowel Disease (IBD), presents a significant clinical challenge due to overlapping symptoms.
• Existing diagnostic methods often struggle to precisely distinguish between IBD subtypes, necessitating novel approaches.

Purpose of the Study:

• To develop and validate a novel computational framework utilizing spatial transcriptomics (ST) for accurate and explainable classification of IBD subtypes.
• To investigate the distinct biological mechanisms underlying CD and UC through spatial analysis of colonic mucosa.

Main Methods:

• Analysis of ST data from colonic mucosa of healthy controls (HC), UC, and CD patients.
• Identification of cellular niches using Non-negative Matrix Factorization (NMF).
• Engineering of 44 features capturing niche composition, neighborhood enrichment, and niche-gene signals for training a multilayer perceptron (MLP) classifier.

Main Results:

• The MLP classifier achieved an accuracy of 0.774 ± 0.161 for three-class classification (HC, UC, CD) and 0.916 ± 0.118 for two-class classification (IBD vs. HC).
• Explainability analysis revealed that spatial organization of niches predicts general inflammation, while specific niche-gene signatures differentiate UC from CD.
• The framework successfully transformed descriptive spatial data into an accurate and explainable predictive tool.

Conclusions:

• The developed computational framework offers a potential new diagnostic paradigm for IBD subtypes.
• The study provides deeper insights into the distinct biological mechanisms driving CD and UC.
• This approach demonstrates the power of integrating spatial transcriptomics with explainable machine learning for complex disease classification.