Tu211 - Defining Spatial Signature of TNF Inhibition in Ulcerative Colitis with a Deep Learning-Based Algorithm for Case-Control Analysis

Abstract Text: The heterogeneity of autoimmune diseases makes it challenging to characterize general molecular mechanisms or predict optimal personalized treatments. Even predicting the direct molecular effects of treatment has proven difficult, with many single-cell studies failing to stratify patients by treatment. Spatial -omics data can link molecular- to tissue-scale phenotypes, but most methods for analyzing spatial data require tissue annotations that ignore local spatial relationships and/or require error-prone cell segmentation, limiting signal strength. To overcome these limitations, we applied Variational Inference-based Microniche Analysis (VIMA), our new deep learning-based method for case-control analysis of spatial data, to a large tissue atlas of ulcerative colitis.

Through analysis of 52 surface protein markers from the sub-epithelial layers of 42 colonic biopsies, we found significant differences in protein expression and spatial distribution between patients treated with TNF-inhibitors (TNFi) and untreated patients (P=6.5e-4). VIMA identified spatial niches that covaried with TNF treatment duration, stratifying samples from TNFi-naive, through recent TNFi initiation and prior TNFi use, to longtime TNFi treatment. TNFi-associated tissue niches displayed disorganized tissue architecture while TNFi-naive-associated niches contained highly organized, B-cell-dominated lymphoid aggregates. TNFi-associated niches also had lower lymphoid marker levels and higher levels of fibroblast, macrophage, and mesenchymal markers, indicating treatment-associated cell-type-specific tissue infiltration. Intriguingly, TNFi-discontinued samples resembled those from patients currently receiving TNFi, suggesting long-term tissue remodeling induced by TNF inhibition. These results highlight VIMA’s unique ability to isolate subtle spatial signals, shed light on the molecular consequences of TNF inhibition, and expand our understanding of variable responses to TNFi.