scRDiT: Generating single-cell RNA-seq data by diffusion transformers and accelerating sampling

TL;DR

scRDiT introduces a diffusion-transformer–based approach to generate high-fidelity synthetic scRNA-seq data from real datasets, addressing the scarcity of samples. By combining a DDPM framework with Diffusion Transformers and a zero-negation preprocessing step, it trains cell-type–specific models and uses DDIM sampling to achieve 10–20× faster generation. The method demonstrates robust performance across multiple datasets, improving zero-proportion and coefficient-of-variation realism while producing samples that closely match real distributions per cell type, as shown by KL, Wasserstein, and MMD metrics. This approach provides a practical tool for augmenting sparse scRNA-seq datasets, enabling more reliable downstream analyses and method benchmarking.

Abstract

Motivation: Single-cell RNA sequencing (scRNA-seq) is a groundbreaking technology extensively utilized in biological research, facilitating the examination of gene expression at the individual cell level within a given tissue sample. While numerous tools have been developed for scRNA-seq data analysis, the challenge persists in capturing the distinct features of such data and replicating virtual datasets that share analogous statistical properties. Results: Our study introduces a generative approach termed scRNA-seq Diffusion Transformer (scRDiT). This method generates virtual scRNA-seq data by leveraging a real dataset. The method is a neural network constructed based on Denoising Diffusion Probabilistic Models (DDPMs) and Diffusion Transformers (DiTs). This involves subjecting Gaussian noises to the real dataset through iterative noise-adding steps and ultimately restoring the noises to form scRNA-seq samples. This scheme allows us to learn data features from actual scRNA-seq samples during model training. Our experiments, conducted on two distinct scRNA-seq datasets, demonstrate superior performance. Additionally, the model sampling process is expedited by incorporating Denoising Diffusion Implicit Models (DDIM). scRDiT presents a unified methodology empowering users to train neural network models with their unique scRNA-seq datasets, enabling the generation of numerous high-quality scRNA-seq samples. Availability and implementation: https://github.com/DongShengze/scRDiT

Figures (4)

• Figure 1: This figure shows how scRDiT process on scRNA-seq data.After data preprocessing, the real scRNA-seq data were fed into the model for training to fit the real data. After training, the user specifies the batch size of simulated data to be synthesized, and the model randomly generates Gaussian noise and reverts it to simulated scRNA-seq data.The generated simulated data can be used for downstream tasks or to test the reliability of the data.
• Figure 2: Illustration of the DiT architecture.
• Figure 3: scRDiT generating scRNA-seq data. a. The fitting process of the model. t-SNE plots of real data and generated scRNA-seq samples from beginning of training to the final performance. b. Dependence of MMD with the increase of training epochs. c. t-SNE plots of virtual scRNA-seq samples and real data.
• Figure 4: Comparison with scRDiT and UNet. a. Number of epochs to fit real data based on DiT and UNet. b. Violin plots for the coefficient of variation (cv) and zero proportion (zero prop) of scRNA-seq data for the two methods. c. Effect of different acceleration rates on the quality of the synthesized samples.