CausalGenDiff: Generative causal diffusion bridges scRNA-seq and spatial transcriptomics.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate gene expression analysis in spatial context is crucial.
• Current single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) integration methods show suboptimal performance (structural similarity <60%).
• Existing methods often overlook critical causal gene relationships.

Purpose of the Study:

• To develop a novel computational model for improved integration of scRNA-seq and ST data.
• To address the limitations of existing methods by incorporating causal gene dependencies.
• To enhance the understanding of gene regulatory mechanisms within spatial contexts.

Main Methods:

• Proposed CausalGenDiff, a model integrating diffusion and autoregressive processes.
• Extended the Causal Attention Transformer for high-dimensional gene expression data.
• Incorporated Variational Autoencoder (VAE)-based pretraining and fine-tuning strategies.

Main Results:

• CausalGenDiff consistently outperformed state-of-the-art methods on 10 tissue datasets.
• Achieved 5%-32% improvement in Pearson correlation and structural similarity metrics.
• Demonstrated effective capture of gene regulatory mechanisms without predefined relationships.

Conclusions:

• CausalGenDiff significantly advances scRNA-seq and ST data integration.
• The model's ability to leverage causal gene relationships offers deeper biological insights.
• This approach provides a robust framework for spatial transcriptomics analysis.