Evaluating Spatial Encoding Strategies for Cell Type Annotation with Spatial Omics Data

Merel Kuijs, Alma Andersson, Ehsan Hajiramezanali, Tommaso Biancalani, Aicha BenTaieb

Keywords: spatial omics, cell type labelling, graph neural networks

TL;DR: This study critically evaluates the importance of spatial information in cell type annotation tasks, questioning the importance of encoding spatial context for this prediction task.

Abstract: Recent spatial omics research leverages the assumption that spatial information enhances model performance on the cell type annotation task. This study investigates and challenges that assumption by conducting benchmark experiments comparing the performance of spatial and non-spatial models. We show that graph-based spatial models do not consistently outperform non-spatial models, provide theories to explain our findings, and make recommendations for future work on spatial encoding strategies.

Submission Number: 49