From 2.5D Stacking to Continuous 3D

Reconstructing True 3D
Spatial Omics

Capturing the 3D organization of cells is essential for deciphering complex biological processes, yet severely hindered by the destructive nature of physical sectioning. DeepSpatial is a generative framework driven by Optimal Transport Flow Matching that learns a continuous dynamic vector fieldenabling the direct extraction of uninterrupted, infinitely resolvable tissue states at any arbitrary spatial depth.

Paper Docs GitHub
import scanpy as sc
import deepspatial as ds

adatas = [sc.read_h5ad(f"slice_{i}.h5ad")
          for i in range(5)]

model = ds.DeepSpatial()
model.setup_data(adatas)
model.build_model()
model.fit()

adata_3d = model.reconstruct_full_volume(
    adatas, thickness=10)
Continuous Z-Reconstruction
AnnData Compatible
Unbalanced OT Cell-to-cell coupling
•
Flow Matching Probability flow ODE
•
GiT Gene Diffusion Transformer
•
39,168,380 Cells in whole-brain atlas
•
STARmap / RIBOmap 3D ground-truth validated
•
openST / IMC Cross-modality generalization
•
129 Slices Allen Brain CCFv3 atlas
•
Unbalanced OT Cell-to-cell coupling
•
Flow Matching Probability flow ODE
•
GiT Gene Diffusion Transformer
•
39,168,380 Cells in whole-brain atlas
•
STARmap / RIBOmap 3D ground-truth validated
•
openST / IMC Cross-modality generalization
•
129 Slices Allen Brain CCFv3 atlas
The Pipeline

From Discrete Slices to Continuous Volume

DeepSpatial reframes 3D reconstruction as a continuous probability density evolution. Rather than merely inserting 2.5D planes, it models the multi-modal transitions between physical slices as mixed continuous-discrete dynamics via a scalable Transformer architecture.

Input Slides

Input Slides

Multi-layer 2D spatial omics sections with spatial coordinates, omics expressions (genes & proteins), optional cell-type labels, and physical slice metadata including thickness and interval distances.

Output Virtual Tissue

Output Virtual Tissue

A fully continuous 3D spatial atlas with precise 3D coordinates, cell-type identities, and high-dimensional omics expressions assigned to each generated virtual cell.

UOT Coupling

i. UOT Coupling

Unbalanced Optimal Transport constructs a cost matrix integrating spatial proximity, transcriptomic similarity, and cell-type penalty to establish biologically rigorous cell-to-cell correspondences across adjacent slices.

Flow Matching

ii. Flow Matching (GiT)

A scalable multi-modal Transformer trained via flow matching learns a shared continuous vector field over spatial coordinates, transcriptomic expressions, and categorical cell-type identities.

3D Volume Generation

iii. 3D Volume Generation

Solving Probability Flow ODEs with density-preserving bidirectional sampling to synthesize tissue at any arbitrary depth, guided by interpolated macroscopic cell density fields for biologically faithful volume reconstruction.

Interactive Explorer

3D Cell Atlas

Explore the reconstructed spatial omics volumes. Rotate, zoom and filter across annotated cell types. Each point represents a single cell in its true 3D spatial context.

merfish_3d_mesh.html
WebGL Accelerated
merfish_3d_pointcloud.html
WebGL
her2_breast_3d_pointcloud.html
WebGL
Applications

Downstream Analysis

By providing an infinitely resolvable 3D virtual tissue, DeepSpatial catalyzes the shift from traditional planar analyses to true volumetric quantificationenabling 3D cell-cell communication mapping, spatial domain identification, and whole-organ atlas construction.

Virtual Slide

i. Virtual Slide

Generate virtual tissue slides at arbitrary Z-depths, revealing continuous 3D microenvironments destroyed by physical sectioning.

3D Spatial Domain

ii. 3D Spatial Domain

Apply CellCharter or similar methods to delineate continuous 3D spatial domains and architectures across the reconstructed volume.

Whole-Brain 3D Atlas

iii. Whole-Brain 3D Atlas

Scalable reconstruction of a 39.2M-cell mouse whole-brain spatial atlas from 129 CCF-aligned sections.

In Silico Sectioning

iv. In Silico Sectioning

Computationally slice the volume along coronal or sagittal planes for cross-platform comparison and topological analysis.

Open Source

Built for Researchers

DeepSpatial is built on AnnData and fully compatible with the Scanpy ecosystem. GPU-accelerated PyTorch implementation enables efficient 3D manifold recovery of large-scale spatial datasets.

Tutorials & API Reference

Step-by-step guides for MERFISH Mouse Hypothalamus, Human Breast Cancer IMC, and Deep-STARmap validationplus full API documentation.

Read the Docs

Citation (BibTeX)

If DeepSpatial is useful in your research, please cite our paper.

@article{yang2026deepspatial, author = {Yang, Yuhang and Luo, Yiming and Zhang, Kai and Bu, Yonggan and Xia, Zheng and Peng, Haoxin and Yan, Rui and Liu, Qi and Chen, Yang and Shen, Lin and Chen, Enhong}, title = {Reconstructing True 3D Spatial Omics at Single-Cell Resolution}, year = {2026}, doi = {10.64898/2026.04.28.721395}, journal = {bioRxiv} }