SpaTRACE: Spatiotemporal recurrent auto-encoder for reconstructing signaling and regulatory networks from...

Cell-cell communication and gene regulatory programs jointly coordinate cellular behaviors during development, regeneration, and disease. Recent advances in spatial transcriptomics enable measurement of gene expression with spatial context across developmental trajectories, providing new opportunities to study dynamic signaling and regulatory processes. However, most existing methods analyze either ligand-receptor (LR) signaling or gene regulatory networks (GRNs) separately, rely on curated interaction databases, and often assume steady-state gene expression, limiting their ability to capture temporal regulatory dynamics and discover novel interactions. We present SpaTRACE, a spatiotemporal recurrent autoencoder framework for joint inference of intercellular signaling and gene regulatory networks from spatial transcriptomics data. SpaTRACE models time-lagged dependencies along pseudotime-sampled cellular trajectories using an attention-based encoder-decoder architecture that predicts future target gene expression from upstream intra- and intercellular signals. The learned attention structure enables simultaneous reconstruction of transcription factor-target gene (TF-TG) regulatory interactions, ligand-receptor-target gene (LR-TG) signaling pathways, and ligand-receptor binding relationships without requiring predefined LR databases. Across synthetic benchmarks, SpaTRACE accurately recovers both GRN and signaling interactions and outperforms existing GRN and cell-cell communication inference methods. Applications to mouse midbrain development reveal transcriptional regulators and signaling programs associated with neuronal differentiation, while analysis of axolotl brain regeneration identifies stage-specific signaling dynamics and candidate interactions involved in tissue repair. Availability: Source code and documentation are available at https://github.com/VariaanZhou/SpaTRACE.