Reproduce “SpaDiT: Diffusion Transformer for Spatial Gene Expression Prediction using scRNA-seq” on OpenResearch

About this paper

Xiaoyu Li, Fangfang Zhu, Wenwen Min

SpaDiT is a diffusion transformer model developed by researchers at Yunnan University that integrates scRNA-seq and spatial transcriptomics data to predict unmeasured spatial gene expression. This deep learning method achieved state-of-the-art performance across ten diverse datasets, demonstrating superior accuracy in gene expression prediction and spatial pattern restoration, while also exhibiting robustness to data sparsity.

Key insights

• A conditional diffusion generative model with a Transformer architecture can effectively learn complex relationships between scRNA-seq and spatial transcriptomics data, enabling precise inference of missing spatial gene expression.
• Leveraging comprehensive scRNA-seq data as a dynamic conditioning factor, particularly through a Flash-attention mechanism, is crucial for mitigating data sparsity and potential batch effects in ST predictions.
• The strategic design of latent and condition embedding modules, including the integration of common genes, is essential for preserving high-dimensional similarity, gene-gene relationships, and accurately restoring spatial patterns.

Summary by alphaXiv.