smiMAD: Adaptive multi-view graph learning for paired spatial omics integration

Paired spatial omics technologies measure transcriptomic profiles together with epigenomic or proteomic profiles while preserving tissue context. Integration remains challenging because modalities differ in feature distributions, sparsity, and spatial dependencies. Existing methods may not fully model complementary graph views or adapt spatial constraints, limiting preservation of spatial organization and modality-specific information.

We present smiMAD for paired spatial omics integration. smiMAD models spatial-neighborhood and molecular-feature graphs through graph-view-specific and common-information branches. An adaptive gate integrates graph-view and modality information, while an alignment-decorrelation objective promotes common-information embedding consistency and reduces branch redundancy. A dynamic spatial-consistency strategy progressively relaxes spatial constraints during training. Across 17 simulated and 15 real RNA-ATAC and RNA-protein datasets, smiMAD achieved competitive clustering accuracy, spatial coherence, and multimodal integration performance. Molecular marker patterns provided support for the inferred spatial regions.

The source code is available at https://github.com/fdu-wangfeilab/smiMAD, with the version used in this study archived at https://zenodo.org/records/22931624. Data are available at https://zenodo.org/records/17207079.