Fig. 1: General TACIT Workflow.

a Multiplex imaging employs both spatial proteomics (top) and spatial transcriptomics (bottom). After segmentation (b top), a CELLxFEATURE matrix is generated (c). Hierarchical cell type structures (b bottom) are formulated based on panel design, expert knowledge, and scRNA-seq marker matching, resulting in a CELLTYPExMARKER matrix (c). Cells are organized into microclusters (MCs) by a community-based Louvain algorithm, averaging 0.1–0.5% of the population (d top). These matrices are then used to compute Cell Type Relevance (CTR) scores for all cell types across cells (d bottom). Optimal thresholds are established to classify cells as clean if they meet one threshold or mixed if multiple (e). Threshold derivation extends to segmental regression on ordered median CTR scores across all MCs to identify breakpoints (f, g), defining “low relevance group (LRG)” and “high relevance group (HRG)” (h). The determined CTR threshold minimizes classification error within 478 MCs, distinguishing between LRG and HRG (i, j). Cells above the threshold are highlighted in red on the UMAP, while those below are in gray (k). After identifying thresholds for all CTRs, cells that meet only one threshold are classified as “clean cells,” while cells that meet multiple thresholds are classified as “mixed cells” (l). The UMAP with all features shows no clear separation between two distinct cell types (m – top left); however, clear segregation appears when only relevant features are used in the UMAP embedding (m – top right). Mixed identities are resolved by analyzing the mode of cell types within their k-nearest neighbors (m – bottom). Validation is performed via heatmaps comparing mean marker and cell type values with the CELLTYPExMARKER matrix (n), and by calculating enrichment scores for each cell type (n). The UMAP plot illustrates spatial distributions with cell type annotations (o top-right) and connections of cell type clusters (o bottom-left), combining cell type and state analyses (o bottom-right).