Gene panel design for spatial transcriptomics with prioritized gene sets

Abstract

A fundamental limitation of the emerging single-cell spatial transcriptomics (sc-ST) technologies is their panel size. Being based on fluorescence in situ hybridization, an sc-ST dataset can profile only a pre-determined panel of a few hundred genes. This often forces biologists to build panels from only the marker genes of different cell types and forgo other genes of interest, e.g., genes encoding ligand-receptor complexes or genes in specific pathways. We propose scGIST- a deep neural network that designs sc-ST panels through constrained feature selection. On four datasets, scGIST outperformed alternative methods in terms of cell type detection accuracy. Moreover, unlike other methods, scGIST allows genes of interest to be prioritized for inclusion in the panel while staying within the its size constraint. We demonstrate through diverse use cases that scGIST includes large fractions of prioritized genes without compromising cell type prediction efficacy making it a valuable addition to sc-STs algorithmic toolbox.