A Generative Latent Neural Spatial Point Process Model for Host–Microbe Spatial Transcriptomics Data

We study a recently published spatial transcriptomics dataset that jointly profiles host cells and the gut microbiome across ten mouse-gut sections spanning different regions and health states. We propose a marked Gibbs point process whose first-order term, mark head and pairwise potential are parameterized by neural networks conditioned on a latent field that is trained jointly with the point-process pseudo-likelihood. We show that the model captures substantial spatial structure beyond tissue geometry, with strong predictive performance for microbial placement but limited information about taxonomic identity. We further use birth–death sampling to generate new point patterns and demonstrate how the model can reveal differences in host–microbe organization between healthy and tumor-bearing tissue.