scVAEDer: integrating deep diffusion models and variational autoencoders for single-cell transcriptomics analysis.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Single-cell data analysis requires effective dimensionality reduction for downstream tasks.
• Existing generative models struggle to capture both global and local variations in embeddings.
• Meaningful embeddings are crucial for understanding complex biological systems.

Purpose of the Study:

• To introduce scVAEDer, a scalable deep learning model for learning comprehensive low-dimensional representations of single-cell data.
• To develop a model that effectively integrates global structure and local variations.
• To demonstrate the utility of learned embeddings in various biological applications.

Main Methods:

• Developed scVAEDer, a hybrid model combining variational autoencoders and deep diffusion models.
• Trained the model on single-cell RNA sequencing (scRNA-seq) data.
• Utilized learned embeddings for generative tasks, perturbation response prediction, and gene expression analysis.

Main Results:

• scVAEDer successfully learns embeddings that retain both global and local data structures.
• The model generates novel, realistic scRNA-seq data.
• scVAEDer accurately predicts perturbation responses and identifies gene expression changes during dedifferentiation.
• Master regulators in biological processes were effectively detected using the learned embeddings.

Conclusions:

• scVAEDer offers a scalable and powerful approach for single-cell data embedding.
• The learned representations improve downstream analyses, including data generation and biological interpretation.
• This model advances the application of deep learning in single-cell genomics.