UniST: A Unified Computational Framework for 3D Spatial Transcriptomics Reconstruction

Area of Science:

• Computational biology
• Genomics
• Artificial intelligence

Background:

• Spatial transcriptomics (ST) typically yields 2D sections, limiting 3D tissue architecture observation.
• Reconstructing 3D ST from serial sections is challenging due to data sparsity, heterogeneity, and tissue loss.

Purpose of the Study:

• To present UniST, a unified generative AI framework for reconstructing dense, continuous 3D ST landscapes from sparse serial sections.
• To overcome limitations of current 3D ST data acquisition and analysis.

Main Methods:

• UniST integrates kernel point convolution, optical flow-based interpolation, and graph autoencoders with implicit neural representations.
• The framework computationally densifies sparse slices, resolves discontinuities, and imputes gene expression data.

Main Results:

• UniST accurately reconstructed 3D tissue architecture and gene expression patterns across multiple ST platforms and tissue types.
• Demonstrated successful reconstruction of a mouse embryo heart and critical spatial features in human cancer tissues.

Conclusions:

• UniST provides a generalizable computational solution to enhance 3D ST data reconstruction.
• Facilitates cost-efficient and scalable investigation of 3D tissue organization and disease biology.