Skip to main content
Back to timeline
bioRxivSource publication:

Cells no longer defined by neighbors: SpaCState learns single-cell spatial state from 3D transcript reconstruction

Synopsis

Identifying spatially organized cell populations has relied mainly on niches built from neighboring cells, leaving open whether a single cell's own molecular representation carries that information. SpaCState models each cell as a gene-labeled 3D point cloud of transcripts, trains a rotation-equivariant network to reconstruct 3D cell shape, and turns the resulting transcript embeddings into a cell-level representation called Spatialstate. In MERFISH human uterine cancer, combining these embeddings with gene expression predicted CD45RO protein intensity better than gene expression alone or niche-based embeddings, and in liver cancer and lung fibrosis data the representation marked cell populations tied to necrotic regions and fibrotic foci.

Source-provided article image: Learning spatial cellular state through 3D reconstruction of the cell
Fig. 1 · bioRxiv · Page 3

Interpretation

Spatial transcriptomics gains a route to comparable cell states that does not pass through the composition of neighboring cells: the arrangement of transcripts inside one cell is enough. Earlier identification of spatially organized populations was built on cellular niches, characterizing each cell by the cell types around it and then clustering; whether a single cell's molecular representation could carry spatial organization on its own was unresolved. On lymphocytes in MERFISH human uterine cancer, combining the embeddings with gene expression predicted CD45RO protein intensity better than gene expression alone and better than niche-based embeddings from NicheScope, CellCharter and BANKSY.

The method treats a cell as a cloud of gene-labeled 3D points and makes the network rebuild the volume the cell occupies, so the reconstruction task itself forces out a useful transcript representation. Transcript embeddings are trained by a geometric objective, reconstructing 3D cell shape, rather than by directly predicting gene identity, which would be nearly trivial because gene identity is already an input feature. The transcript encoder is an e3nn-based rotation-equivariant network that aggregates relative displacements of nearest-neighbor transcripts within a cell; the decoder spreads embeddings onto a regular voxel grid and uses steerable 3D convolutions to predict an occupancy field, trained with binary cross-entropy and Dice losses and scored by voxel IoU.

Perspective

The demonstrated scope is imaging-based spatial transcriptomics: reconstruction was evaluated on MERFISH human liver cancer and Xenium human breast cancer, embeddings on MERFISH human uterine cancer with CD45RO co-staining, and downstream analyses on MERFISH human liver cancer and Xenium human lung fibrosis samples. The intended reader works with subcellular-resolution spatial transcriptomics.

How to choose an appropriate Spatialstate threshold remains unclear; SPP1+ and FABP4+ macrophages occupy different Spatialstate ranges, suggesting a single threshold may not suffice to define biologically meaningful populations. Patterns in the 2D histograms are currently identified by visual inspection, and the individual contribution of higher-order equivariant features to cell state is still hard to interpret.