Generative Modeling of Spatial Transcriptomics via Gaussian Mixture Flow Matching

Spatial transcriptomics enables the joint analysis of cellular gene expression and spatial organization, offering new insights into tissue development and function. Existing graph- and embedding-based methods capture spatial patterns but remain descriptive and non-generative. Spatial transcriptomics is still an expensive and resource-intensive assay, limiting its widespread application. Consequently, there is a growing need for generative models capable of accurately simulating gene expression within spatial contexts, particularly in settings where experimental data acquisition is impractical or cost-prohibitive. In this work, we revisit the generative modeling of spatially informed cell embeddings derived from gene expression and spatial information and apply a Gaussian Mixture Flow model (GMFlow) to explicitly model the multinomial nature of cell type distributions during generation. Using a mouse embryonic development dataset, we find that both single-Gaussian conditional flow and GMFlow models learn cell type distributions over different developmental timepoints with comparable accuracy, while GMFlow results in generations that are closer to ground truth cell representations, suggesting that GMFlow may enable more realistic simulation of cellular landscape in spatial transcriptomics data.