2025/03/16 by Elyas Heidari, Andrew Moorman, Dániel Unyi +9 · 1 voice · 2 citations
Biochemistry, Genetics and Molecular Biology · #Single-cell and spatial transcriptomics #Cell Image Analysis Techniques #Molecular Biology Techniques and Applications
paper · pdf · doi:10.1101/2025.03.14.643160
The accurate assignment of transcripts to their cells of origin remains the Achilles heel of imaging-based spatial transcriptomics, despite being critical for nearly all downstream analyses. Current cell segmentation methods are prone to over- and under-segmentation, misassign transcripts to cells, require manual intervention, and suffer from low sensitivity and scalability. We introduce segger, a versatile graph neural network based on a heterogeneous graph representation of individual transcripts and cells, that frames cell segmentation as a transcript-to-cell link prediction task and can leverage single-cell RNA-seq information to improve transcript assignments. On multiple Xenium dataset benchmarks, segger exhibits superior sensitivity and specificity, while requiring orders of magnitude less compute time than existing methods. The user-friendly open-source software implementation has extensive documentation (https://elihei2.github.io/seggerdev/), requires little manual intervention, integrates seamlessly into existing workflows, and enables atlas-scale applications.