INTERLACE: Learning Cross-Modal Molecular Programs from Multilayer Spatial Omics Graphs

Spatial multi omics technologies measure diverse molecular signals within intact tissues, but different modalities often lack a shared common feature space or an appropriate similarity measure. Here, we present INTERLACE, a flexible multilayer graph framework for identifying multilayer spatial programs, defined by groups of molecular features with coordinated spatial activity within individual modalities and shared spatial organization across modalities. We refer to these modality-specific groups of molecular features as modules. Each modality is represented by a molecular graph whose node features are representations of their spatial profiles, such as gene co-expression, immune receptor sequence similarity, and chemical-structure similarity. Within each modality, INTERLACE learns soft assignments of molecular features to modules. These assignments determine module-level spatial profiles by aggregating their spatial profiles. INTERLACE then alternates between estimating cross-modal correspondences from the current module-level spatial profiles using unbalanced optimal transport with KL marginal relaxation and updating the within-modality module assignments using both modality-specific graph information and these cross-modal correspondences. In simulations, INTERLACE improved the recovery of within-layer modules and cross-layer correspondences, with its largest advantages under sparsity and noise. In human tonsil, INTERLACE distinguished two follicle-specific associations linking B cell programs to immunoglobulin clonotypes. In breast cancer, INTERLACE identified a spatially coordinated B-cell - T-cell activation niche and a purine-associated myoepithelial-like niche.