NOLAN: CONSTRUCTING GRAPH REPRESENTATION OF TISSUE STRUCTURE WITH SELF-SUPERVISED LEARNING

Artemy Bakulin, Nathan Levy, Can Ergen, Jonas Maaskola, Nir Yosef

Track: Full Paper Track

Keywords: Tissue structure, spatial transcriptomic, self-supervised learning

TL;DR: We propose a method for learning niche representation from spatial transcriptomic data and using it to reconstruct topological structure of the tissue.

Abstract: Understanding tissue architecture is fundamental to deciphering the complex interplay of cells in both health and disease, yet many current approaches to spatial transcriptomic analysis focus on discrete cellular niches while overlooking the transitional regions that connect them. Here, we present NOLAN, a self-supervised framework that learns neighborhood-informed cell representations, capturing additionally the variation within niches and the dynamics at their interfaces. Leveraging these representations, NOLAN constructs a graph-based abstraction of tissue, modeling it as a network of interconnected regions bridged by transitional zones. In a spatial transcriptomics atlas covering eight different cancer tissues, NOLAN reveals a landscape of tumor microenvironments characterized by both tissue-specific niches and shared niches. NOLAN enables multi-sample comparative analysis by providing a unified coordinate system of spatial niches.

Attendance: Artemii Bakulin

Submission Number: 100