Integration and annotation of spatial multi-omic data with DIRAC highlights spatial organization of lymphoid organs

Integration across spatial omics modalities followed by multi-omic spatial domain and cell-type annotation are two fundamental tasks for downstream analysis. We present Domain Invariant Representation through Adversarial Calibration (DIRAC), a geometric deep learning model that unifies both tasks by treating horizontal integration (different cells/spots, same omic modality) and vertical integration (same cells/spots, different omics modalities) under a generalized domain adaptation framework.

DIRAC uses an adversarial domain discriminator to integrate multiple spatial omics modalities into a unified domain-invariant embedding space and to automate cell-type annotation by transferring labels from reference multi-omic data. DIRAC delineated more biologically meaningful spatial domains and improved clustering and cell-type annotation performance across omics modalities (histone marks, chromatin accessibility, RNA, and protein) and technology platforms (sequencing and imaging-based).