Fig. 2: TACCO compositional annotation, cell segmentation and analysis of spatial expression data.

a, Analysis of in silico mixtures. UMAP embedding of scRNA-seq profiles of mouse colon (1)²⁹ and of in silico mixed scRNA-seq data before (2) and after (4) applying TACCOs splitting procedure into pure contributions. (3) L2 error (y axis) of cell-type annotations for each simulated bead size (x axis). Dashed lines: categorical annotation methods. b, A Slide-seq puck of normal mouse colon²⁹ colored by TACCO cell-type annotations (1, colors are weighted and summed per bead) or by TACCO-defined regions (2) (Methods; Extended Data Fig. 3); (3) short-range (up to 20 µm) neighborship enrichment z scores over randomly permuted annotation assignments; (4) long-range (up to 500 µm) dependence of the composition of cell types around beads (log₂(p(annotations|center)); y axis) at different distances from region 2 (muscularis; x axis) for each cell-type annotation (color). c, Comparison of TACCO and Baysor for single-molecule cell-type-of-origin annotation performance based on the single-molecule FISH: (1–3) Entire section profiled by osmFISH²⁷ (left) and zoom-in on a small region (right) colored by (1) the published annotation of cells from watershed-based segmentation of the poly(A) signal or (2,3) the segmentation-free single-molecule annotation for cell-type-of-origin by TACCO (2) or Baysor (3) (Extended Data Fig. 6a); (4) measured astrocyte marker gene expression on the zoom-in as ground truth for the astrocyte annotation in (1–3). Gray molecules are unannotated. d, Effective cell segmentation of osmFISH data by TACCO. tSNE embeddings of RNA profiles in cell-like objects from the published watershed-based segmentation colored by the published cell-type annotation (1), from TACCOs annotation-based segmentation (using the annotations in c (2)) (2, 3), colored by either TACCOs single-molecule annotations summed per cell (2) or by the published segmented cell annotation pulled back to single molecules (as in c(1)) and summed per cell (3) (Methods; Extended Data Fig. 5). e, Recovery of layered tissue structure in osmFISH data. Long-range (up to 2,000 µm) dependence of the composition of cell types (p(annotations|center = hippocampus), y axis) recovered by TACCO for TACCO-segmented objects at different distances from the hippocampus region annotation for each cell-type annotation (Extended Data Fig. 6c).