bioRxivJClub
Learning spatial cellular state through 3D reconstruction of the cell
bioRxiv · · Preprint · Open access
Lee, Anad + 2 more
Abstract ↗AI summary
The abstract is read at the publisher; the summary is JClub's.
A rotation-equivariant model that reconstructs 3D cell shape from gene-annotated transcript point clouds learns spatially informed transcript embeddings that improve CD45RO protein intensity prediction in MERFISH uterine cancer.
- Why it matters: Spatially organized cell populations are often inferred from neighboring-cell niches, leaving uncertainty about whether single-cell 3D transcript representations alone can capture cell-state heterogeneity and biologically meaningful spatial organization.
- What they did: The framework treats each cell as a gene-annotated point cloud of transcripts and trains a rotation-equivariant neural network to reconstruct 3D cell shape, then aggregates transcript embeddings into cell-state representations.
- The result: It outperforms existing point-cloud 3D reconstruction methods for cell-shape reconstruction and, in MERFISH human uterine cancer, embedding-based features improve CD45RO protein intensity prediction beyond gene expression alone and niche-based embeddings; in MERFISH human liver cancer and Xenium human lung, it identifies distinct spatial patterns between cell populations.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.
- Open access
Free. We’ll email you when a new paper cites this one or any other paper you follow on JClub — at most one email a day. Every email has a one-tap unsubscribe.