SpaTRACE: Spatiotemporal recurrent auto-encoder for reconstructing signaling and regulatory networks from spatiotemporal transcriptomics data

Cell-cell communication (CCC) orchestrates coordinated cellular behaviors underlying development, regeneration, and disease. Recent advances in spatiotemporal transcriptomics enable simultaneous measurement of gene expression across spatial contexts and developmental progression. However, most existing CCC inference methods rely heavily on curated ligand-receptor (LR) databases and implicitly assume steady-state gene expression, limiting applicability to understudied species and hindering robust inference of dynamic signaling cascades. Here, we introduce SpaTRACE, a transformer-based temporal-causal framework for pathway-free CCC inference from developmental spatial transcriptomics data. SpaTRACE trains a temporal causal attention-based transformer along pseudotime-sampled trajectories to model time-lagged dependencies across spatially resolved cell populations. Attention weights enable de novo reconstruction of intercellular signaling interactions (ligand-target gene; LR-TG) and intracellular gene regulatory relationships (transcription factor-target gene; TF-TG), which are integrated to infer dynamic CCC networks without reliance on predefined LR databases. Across synthetic datasets, SpaTRACE accurately recovers LR-TG interactions, TF-TG regulation, and correct LR pairings, outperforming existing CCC methods, particularly under noisy pathway settings. Applied to mouse midbrain development and axolotl brain regeneration, SpaTRACE recovers canonical signaling modules, identifies stage-specific transitions, and uncovers previously under-characterized interactions. Together, SpaTRACE provides a general and statistically powerful framework for dissecting dynamic intercellular communication and regulatory programs from spatiotemporal transcriptomics data. Code is available at https://github.com/VariaanZhou/SpaTRACE.