STORM: Unified spatial, temporal, and multimodal data integration into biologically interpretable embeddings

Spatiotemporal multi-omic atlases are rapidly emerging and shifting tissue biology from static, single-modality snapshots to dynamic, multi-layer atlases across development and disease. Yet integrating space, time, and modality while preserving tissue geometry, transient niches, and real biological change remains difficult. Here we introduce STORM (Spatial Temporal multi-Omics Representation Model), among the first end-to-end frameworks to jointly model space–time–modality while grounding the representation in gene programs. STORM learns a unified cell embedding using a multimodal graph neural network variational autoencoder, jointly aligning timepoints and modalities while correcting time-dependent technical variation. Leveraging curated gene programs, STORM decomposes complex biology into interpretable, program-level components across modalities and reveals time-evolving program dynamics. STORM extends to multiple timepoints, sequencing platforms, tissues, and biological conditions, as demonstrated in three paired spatial RNA–ATAC datasets spanning postnatal (P0–P22) and embryonic (E11.0–E18.5) mouse brain development and a lysolecithin-induced demyelination model. Across datasets, STORM preserves stage-dependent anatomy, resolves small anatomically meaningful populations (including the island of Calleja), tracks coordinated spatiotemporal program trajectories, and enables program-level RNA–ATAC contrasts that expose regulatory priming not apparent from expression alone. Our results demonstrate that STORM provides a scalable, interpretable route to dissect spatiotemporal molecular programs from time-course spatial multi-omic data.