Integration and annotation of spatial multi-omic data with DIRAC

Spatial multi-omic technologies measure multiple molecular layers within intact tissues. Two central challenges are vertical integration, which combines different molecular modalities, and horizontal integration, which aligns cell types across studies. Here we show domain invariant representation through adaptive calibration (DIRAC), a geometric deep learning model that unifies these tasks in a generalized domain adaptation framework. DIRAC learns task-discriminative features that are insensitive to domain shifts caused by batch effects, omic modalities or study sites, enabling a shared embedding space and automated cell-type annotation through label transfer from reference datasets. Across sequencing-based and imaging-based datasets, DIRAC identified a new microvascular proliferation domain in glioblastoma, corrected cortical-layer annotations in severe Alzheimer’s disease, resolved subtle structures in mouse embryos and remained robust to missing data and unseen cell types. We generated mouse spleen and thymus Stereo-CITE-seq data and used DIRAC to reveal thymic T cell migration and splenic immune cell interactions. DIRAC is fast and scales to millions of cells. DIRAC is a geometric deep learning model that enables both vertical and horizontal integration and annotation of spatial multi-omic data.