The workflow supports clinical researchers in comparing two cohorts of spatially-resolved single-cell omics data (e.g., tumour vs. healthy, metastatic vs. non-metastatic) to identify differentiating biomarkers at multiple levels of detail — from cell type abundance through pairwise co-localisation patterns to complex multi-cellular microenvironments — while continuously relating abstract findings to spatial tissue context and detecting within-cohort outliers.
The input consists of two cohorts, each comprising sets of tissue samples. Each sample contains segmented cells with assigned type labels and spatial positions. A user-defined distance parameter specifies the neighbourhood radius for microenvironment computation. The system computes: (a) per-sample abundance of each cell type (absolute and relative counts), (b) per-sample pairwise co-localisation frequencies (how often each cell type appears in each other cell type's microenvironment), and (c) separability metrics (Silhouette, Dunn's index) ranking cell types and pairwise combinations by their power to differentiate the two cohorts.
The system generates raincloud plots (superposed kernel density estimates with one-dimensional jitter plots) for each cell type, using complementary cohort colours (blue/orange) that blend to neutral grey where distributions overlap. Plots are sorted by separability metric, placing the most differentiating cell types at the top. The analyst inspects these to identify cell types whose abundance distributions clearly separate the cohorts. They can search by keyword, combine related cell types via drag-and-drop (e.g., merging CD4 and CD8 T-cell plots into a unified T-cell plot), and detect outlier samples visible as isolated lines in the jitter portion. Selected samples are linked to the tissue view for spatial validation.
The system displays a difference heatmap of pairwise co-localisation patterns: rows represent centre cell types, columns represent microenvironment cell types, and colour encodes which cohort has higher co-localisation frequency (using the same complementary colour scheme). The analyst identifies strongly-coloured cells indicating pairwise combinations that differentiate the cohorts. Clicking a heatmap cell pre-populates the detail microenvironment explorer with that combination.
Starting from a pairwise combination identified in the heatmap, the analyst iteratively builds complex microenvironment patterns. The detail view shows: a "Selected" raincloud plot for the current microenvironment definition (centre cell type + required neighbourhood cell types), and "Remaining" raincloud plots showing what happens when each additional cell type is added to the neighbourhood. Remaining plots are ranked by separability. The analyst assesses whether the current microenvironment sufficiently differentiates the cohorts. If not, they drag the most promising remaining cell type into the specification, extending the microenvironment pattern, and the system recomputes. This cycle repeats — extending the pattern, observing separation improvement, and assessing — until a sufficiently differentiating microenvironment is found. Throughout, outlier samples are identified and validated against spatial tissue images.
At any point, the analyst selects samples from raincloud plots to view in the tissue panel, where identified cell types or microenvironments are highlighted in their spatial context (with non-selected structures faded). This enables validation of abstract findings against biological plausibility (e.g., confirming that co-localised cells form tertiary lymphoid structures). The analyst synthesises findings: confirmed biomarkers (cell types or microenvironment patterns differentiating cohorts), identified outliers warranting further investigation, and spatial validation of the biological relevance of discovered patterns.