SPADE: Inferring Transcriptional Dynamics from Spatial Transcriptomics with Physics-Informed Deep Learning

In situ sequencing–based spatial transcriptomics technologies, such as 10x Genomics Xenium and Vizgen MERSCOPE, have recently emerged as powerful platforms that enable subcellular-resolution mapping of RNA transcripts within intact tissues. While existing computational models developed for pixel-based spatial transcriptomics can be applied to in situ sequencing data, these approaches overlook molecule-level information and thus underutilize the full potential of the high-resolution measurements. Recognizing that post-transcriptional mRNA localization arises from a hybrid process of active transport and diffusion, we hypothesized that the spatial distribution of transcripts relative to the transcription start site encodes information about transcriptional activity within short time windows, offering a new paradigm for inferring transcriptional dynamics.To realize this capability, we present SPADE, a physics- and systems biology–informed deep learning framework that leverages the spatial organization of RNA molecules to infer transcriptional dynamics. SPADE first constructs a trajectory for each cell, ordered along a pseudo-time axis defined by local shifts in molecule distributions, and then employs a recurrent neural network to disentangle RNA synthesis from drift–diffusion processes under a bistate transcriptional regulation model. Extensive evaluations on both simulated and in-house spatial transcriptomics datasets demonstrate that SPADE accurately reveals gene-specific bursting patterns, recovers dynamic transcription rates, and uncovers regulatory delays between genes. As the first framework to estimate temporal variations in transcription rates from static spatial transcript distributions, SPADE establishes a novel paradigm for studying transcriptional dynamics and their underlying biological mechanisms.