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

Abstract

Single-cell RNA sequencing profiles cellular states at scale but removes their three-dimensional (3D) spatial organization. Reconstructing this organization requires recovering not only individual cell positions but also the physical size, geometry, and neighborhoods of an entire tissue. Existing methods either map cells onto measured spatial references or reconstruct two-dimensional sections independently, leaving the joint reconstruction of a new 3D specimen unresolved. Moreover, predicting the average plausible position of each cell can reduce per-cell error while compressing tissue geometry. We introduce ORBIS, a generative framework that jointly reconstructs whole 3D specimens from single-cell gene expression without requiring a spatial reference at inference. ORBIS learns spatial organization through flow matching in a shared physical coordinate frame, using training views sampled from sparse serial sections, tissue blocks, and whole specimens. At inference, a shared spatial state is iteratively updated through dynamically regrouped windows, enabling the generation of millions of cell positions without independently reconstructing and stitching tissue regions. We evaluate physical size, cell-population shape, per-cell placement, and neighborhood organization separately to distinguish accurate tissue geometry from low per-cell error. Across eight held-out specimens spanning whole mouse brains, thick brain blocks, and Drosophila embryos, ORBIS achieves lower cell-population shape error than all five evaluated baselines on every test and reconstructs specimens containing up to 3.7 million cells. On a single GPU, ORBIS samples million-cell brains in 0.9-2.8 minutes, approximately 38 times faster than the sampler of LUNA-3D, our 3D extension of the state-of-the-art generative spatial reconstruction method LUNA. On two held-out brain tests, sampling reduces size error by 3.3-fold and shape error by 1.8-fold compared with a single-call variant. On an independent dissociated single-cell brain atlas, ORBIS returns on average 37.9% of each region's cells to their dissection region across 23 regions, compared with 28.5% for LUNA-3D.