CellPie: a fast spatial transcriptomics topic discovery method via joint factorization of gene expression and imaging data

Abstract

MotivationSpatially resolved transcriptomics has enabled the study of expression of genes within tissues while retaining their spatial identity. The lack of single-cell resolution for most of the current high-throughput spatial transcriptomics technologies led to the development ofin-silicomethods, to disentangle the spatial profiles of individual cell-types. However, most current approaches ignore useful information from associated imaging data that can help to better resolve cell-types or spatial domains.ResultsWe presentCellPie, a fast, reference-free topic modelling method, based on joint non-negative matrix factorisation between spatial RNA transcripts and histological or molecular imaging features. This synergy of the two modalities can lead to improved single-cell deconvolution and spatial clustering. We assessedCellPiein two different tissues and imaging settings, showing an improved accuracy against published deconvolution and clustering methods. In addition, in terms of computational efficiency,CellPieoutperforms all tested deconvolution methods by at least two orders of magnitude, without the use of GPUs. Availability:https://github.com/ManchesterBioinference/CellPieContact:sokratia.georgaka@manchester.ac.uk