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

Understanding gene expression within a spatial context requires the effective integration of single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) data. However, existing approaches often perform suboptimally, with structural similarity typically falling below 60%. We identify the neglect of causal gene relationships as a major limiting factor. To address this, we propose CausalGenDiff, a model that integrates diffusion and autoregressive processes to exploit these underlying causal dependencies. Our approach extends the Causal Attention Transformer originally designed for image generation to handle high-dimensional gene expression data, enabling the capture of gene regulatory mechanisms without relying on predefined relationships. We further incorporate VAE-based pretraining and fine-tuning strategies to enhance performance, supported by thorough ablation studies. Evaluated on 10 tissue datasets, our method consistently outperforms state-of-the-art baselines across four standard metrics, achieving improvements of 5%-32% in Pearson correlation and structural similarity, thereby contributing to both technical advancement and biological insight.