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

Single-cell RNA sequencing offers a comprehensive view of cellular states, but it discards crucial three-dimensional (3D) spatial organization. To reconstruct this organization, a method must capture individual cell positions, physical size, geometry, and the adjacent cells within tissues. Existing techniques either align cells onto measured spatial references or rebuild 2D sections independently, leaving the reconstruction of a full 3D specimen unaddressed.

Additionally, estimating the probable position of each cell can minimize per-cell error while simplifying tissue geometry. ORBIS, a generative framework, addresses this challenge by jointly reconstructing whole 3D specimens from single-cell gene expression data without the need for spatial references during inference.

ORBIS achieves spatial organization through a process called flow matching in a shared physical coordinate frame, leveraging training views sampled from sparse serial sections, tissue blocks, and whole specimens. In the inference phase, a shared spatial state is iteratively refined through dynamically grouped windows, allowing the generation of millions of cell positions without individually reconstructing and piecing together tissue regions.

To evaluate the effectiveness of ORBIS, researchers measured physical size, cell-population shape, per-cell placement, and neighborhood organization separately. This distinction helps differentiate accurate tissue geometry from low per-cell error.

In tests involving eight previously unseen specimens, such as whole mouse brains, substantial brain blocks, and Drosophila embryos, ORBIS demonstrated superior performance compared to five existing baselines across all metrics. It achieved the lowest error in cell-population shape, surpassing all competitors. Furthermore, ORBIS could reconstruct specimens containing up to 3.7 million cells.

Notably, on a single GPU, ORBIS processes million-cell brains in just 0.9-2.8 minutes, which is about 38 times faster than LUNA-3D, a 3D extension of the state-of-the-art generative spatial reconstruction method LUNA.

On two additional brain tests using held-out data, ORBIS reduced size error by 3.3-fold and improved shape error by 1.8-fold compared to a single-call variant of LUNA-3D. Importantly, when applied to an independent dissociated single-cell brain atlas, ORBIS successfully returned on average 37.9% of each region's cells to their original dissection region across 23 distinct regions, a significant improvement over LUNA-3D's 28.5%.