Learned Morphological Features Guide Cell Type Assignment of Deconvolved Spatial Transcriptomics

Abstract

Spatial transcriptomics enables to study the relationship between gene expression and tissue organization. Despite many recent advancements, existing sequencing-based methods have a spatial resolution that limits identification of individual cells. To address this, several cell type deconvolution methods have been proposed to integrate spatial gene expression with single-cell and single-nucleus RNA sequencing, producing per spot cell typing. However, these methods often overlook the contribution of morphology, which means cell identities are randomly assigned to the nuclei within a spot. In this paper, we introduce MHAST, a morphology-guided hierarchical permutation-based framework which efficiently reassigns cell types in spatial transcriptomics. We validate our method on simulated data, synthetic data, and a use case on the broadly used Tangram cell type deconvolution method with Visium data. We show that deconvolution-based cell typing using morphological tissue features from self-supervised deep learning lead to a more accurate annotation of the cells.