Generative 3D Spatial Reconstruction of Cellular Organization from Single Cell Expression

ORBIS reconstructs whole 3D specimens from single-cell gene expression, generating up to 3.7 million cell positions and achieving lower cell-population shape error than five baselines on eight held-out specimens.

• Why it matters: Single-cell RNA sequencing lacks 3D spatial context, and existing approaches either rely on measured spatial references or treat 2D sections separately, leaving joint 3D reconstruction unresolved.
• What they did: ORBIS is a generative framework trained with flow matching in a shared physical coordinate frame using training views from sparse serial sections, tissue blocks, and whole specimens, then iteratively updates a shared spatial state via dynamically regrouped windows.
• The result: On held-out whole mouse brains, thick brain blocks, and Drosophila melanogaster embryos, ORBIS matches lower shape error across all tests, samples million-cell brains in 0.9–2.8 minutes (~38× faster than LUNA-3D), and improves size/shape error by 3.3-fold/1.8-fold with sampling.

AI summary of the abstract · may contain errors · not medical advice · read the paper.