scDiffusion: conditional generation of high-quality single-cell data using diffusion model

• DOI: 10.48550/arxiv.2401.03968
• Publication Date: 2024-01-01
• Authors: Erpai Luo; Minsheng Hao; Lei Wei; Xuegong Zhang

Single-cell RNA sequencing (scRNA-seq) data are important for studying the laws of life at single-cell level. However, it is still challenging to obtain enough high-quality scRNA-seq data. To mitigate limited availability data, generative models have been proposed computationally generate synthetic Nevertheless, generated with current not very realistic yet, especially when we need controlled conditions. In meantime, Diffusion shown their power in generating high fidelity, providing a new opportunity generation. this study, developed scDiffusion, model combining diffusion and foundation We designed multiple classifiers guide process simultaneously, enabling scDiffusion under condition combinations. also control strategy called Gradient Interpolation. This allows continuous trajectories cell development from given state. Experiments showed that can gene expression closely resembling real Also, conditionally produce on specific types including rare types. Furthermore, could use multiple-condition generation type was out training Leveraging Interpolation strategy, developmental trajectory mouse embryonic cells. These experiments demonstrate powerful tool augmenting provide insights into fate research.