scRDiT: Generating Single-cell RNA-seq Data by Diffusion Transformers and Accelerating Sampling.

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Single-cell RNA sequencing (scRNA-seq) is vital for understanding cellular heterogeneity.
• Existing analysis tools struggle to capture scRNA-seq data's unique statistical properties and generate realistic virtual datasets.

Purpose of the Study:

• To develop a generative approach for creating high-quality virtual scRNA-seq datasets.
• To address the limitations of current methods in replicating scRNA-seq data characteristics.

Main Methods:

• Introduced scRNA-seq Diffusion Transformer (scRDiT), a neural network based on Denoising Diffusion Probabilistic Models (DDPMs) and Diffusion Transformers (DiTs).
• Employed iterative noise-adding and restoration steps using Gaussian noise on real scRNA-seq data.
• Integrated Denoising Diffusion Implicit Models (DDIMs) to accelerate the sampling process.

Main Results:

• Demonstrated superior performance in generating virtual scRNA-seq data across two distinct datasets.
• Successfully learned essential data features from actual scRNA-seq samples during training.
• Enabled the generation of numerous high-quality synthetic scRNA-seq samples.

Conclusions:

• scRDiT offers a unified methodology for generating high-quality virtual scRNA-seq data.
• The approach empowers researchers to train models on their specific datasets for custom synthetic data generation.
• Facilitates advanced biological research by providing a robust tool for data augmentation and simulation.