AMICI: Attention Mechanism Interpretation of Cell-cell Interactions

Spatial transcriptomic data enable study of cell–cell communication, yet current analysis tools often fail to provide dynamic, interpretable estimates of interactions and their spatial range across tissue. We present AMICI, an interpretable attention framework that jointly estimates interaction length scales, adaptively resolves sender–receiver subpopulations, and links communication to downstream gene programs. AMICI recovers ground-truth interactions in semi-synthetic data, uncovers gene programs linked to cell communication in the mouse cortex, and reveals length-scale-dependent tumor–immune signaling that reinforces estrogen receptor (ER) programs in breast cancer.