Chioso: Segmentation-free Annotation of Spatial Transcriptomics Data at Sub-cellular Resolution via Adversarial Learning

Abstract

Recent advances in spatial transcriptomics technology have produced full-transcriptomic scale dataset with subcellular spatial resolutions. Here we present a new computational algorithm, chioso, that can transfer cell-level labels from a reference dataset (typically a single-cell RNA sequencing dataset) to a target spatial dataset by assigning a label to every spatial location at sub-cellular resolution. Importantly, we do this without requiring single cell segmentation inputs, thereby simplifying the experiments, and allowing for a more streamlined, and potentially more accurate, analysis pipeline. Using a generative neural network as the underlying algorithmic engine, chioso is very fast and scales well to large datasets. We validated the performance of chioso using synthetic data and further demonstrated its scalability by analyzing the complete MOSTA dataset acquired using the Stereo-Seq technology. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=80 SRC="FIGDIR/small/597195v2_ufig1.gif" ALT="Figure 1"> View larger version (29K): org.highwire.dtl.DTLVardef@1521a38org.highwire.dtl.DTLVardef@1376191org.highwire.dtl.DTLVardef@e6b4c3org.highwire.dtl.DTLVardef@65801b_HPS_FORMAT_FIGEXP M_FIG C_FIG