STGNET: extending panel coverage in imaging-based spatial transcriptomics using deep generative adversarial networks.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Imaging-based spatial transcriptomics (ST) provides high-resolution gene expression mapping in tissues.
• Current ST technologies are limited by small gene panels, restricting comprehensive analysis.
• This limitation hinders the discovery of crucial biological insights.

Purpose of the Study:

• To introduce STGNET, a deep learning framework to extend gene panel coverage in ST.
• To overcome the limitations of targeted gene panels in spatial transcriptomics.
• To enable deeper and more comprehensive spatial biology studies.

Main Methods:

• Developed STGNET, integrating generative adversarial networks (GANs) and graph neural networks.
• Utilized a multi-stage GAN to learn transcriptomic distribution from single-cell RNA sequencing data.
• Employed a spatially aware graph convolutional network for imputation, considering cell proximity and transcriptional similarity.

Main Results:

• STGNET demonstrated superior performance across nine diverse ST datasets compared to seven state-of-the-art methods.
• Achieved enhanced accuracy in gene imputation and exceptional preservation of cellular topology.
• Successfully reconstructed developmental patterns, identified a novel transitional cell state in breast cancer, and uncovered cell-cell communication networks.

Conclusions:

• STGNET is a powerful solution for expanding gene coverage in targeted ST assays.
• The framework enables deeper biological discovery by overcoming gene panel limitations.
• STGNET facilitates comprehensive spatial biology research and is publicly available.