Learning spatial cellular state through 3D reconstruction of the cell
En palabras de los autores
Subcellular-resolution spatial transcriptomics localizes individual RNA transcripts within cells with high-precision. Existing methods for identifying spatially organized cell populations rely on cellular niches defined by neighboring cells, leaving open whether such populations can be identified from individual-cell representations alone. Moreover, representations that jointly capture transcript identity and three-dimensional (3D) spatial organization have not been systematically learned or biologically evaluated. Here we present a representation learning framework that models each cell as a gene-annotated point cloud of transcripts and trains a rotation-equivariant neural network to reconstruct 3D cell shape, yielding spatially informed transcript embeddings. We further aggregate these embeddings into a cell-state representation that summarizes transcript-level spatial patterns. Our framework outperforms existing point-cloud-based 3D reconstruction methods on cell-shape reconstruction. In MERFISH human uterine cancer, combining the learned embeddings with gene expression improved prediction of CD45RO protein intensity over both gene expression alone and niche-based embeddings. Applied to MERFISH human liver cancer and Xenium human lung datasets, our framework identified distinct spatial patterns between cell populations, including spatially organized subpopulations marked by disease-associated gene expression and patterns associated with pathological tissue structures. Together, these results show that learning transcript representations through 3D cell reconstruction provides a novel approach to capture cell-state heterogeneity and biologically relevant spatial organization in subcellular spatial transcriptomics.
Apareció: jueves, 19 de noviembre. bioRxiv. Preprint, todavía sin revisión por pares.