CCIDeconv: Hierarchical model for deconvolution of subcellular cell-cell interactions in single-cell data

CCIDeconv deconvolutes whole-cell cell-cell interactions into cytoplasmic and nuclear subcellular CCI, achieving median composite score 0.75 across nine human sST datasets in leave-one-dataset-out validation.

• Why it matters: Subcellular spatial transcriptomics can reveal compartment-specific interaction signals, but predicting these patterns from non-spatial scRNA-seq requires a method that preserves biological localization fidelity.
• What they did: The approach estimates modified CellChat-derived communication scores separately for cytoplasmic and nuclear compartments, then uses a hierarchical classification and regression framework with Human Protein Atlas protein localization data to deconvolve whole-cell scores.
• The result: In nine publicly available human sST datasets, leave-one-dataset-out validation yields mean R2 of 0.87 (cytoplasmic) and 0.80 (nuclear), and performance without spatial features approaches spatial models as training datasets increase.

AI summary of the abstract · may contain errors · not medical advice · read the paper.