Synthetic single cell RNA sequencing data from small pilot studies using deep generative models.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Deep generative models (DGMs) like VAEs and DBMs generate synthetic data.
• Their utility for single-cell RNA sequencing (scRNA-seq) data is underexplored.
• DGMs could aid in planning scRNA-seq experiments using pilot data.

Purpose of the Study:

• To evaluate VAEs and DBMs for generating synthetic scRNA-seq data.
• To assess if synthetic data reflects properties relevant for downstream analysis.
• To investigate the impact of pilot dataset size and sequencing technology.

Main Methods:

• Implemented and compared two VAE variants (scVI posterior and prior sampling).
• Developed and applied single-cell deep Boltzmann machines (scDBMs).
• Evaluated clustering accuracy, cell type proportions, and gene expression distributions on synthetic data.

Main Results:

• VAE prior sampling showed high variability, likely amplifying small dataset artifacts.
• All models struggled with cell type abundance, overrepresenting common types.
• Synthetic data cluster proportions improved with larger pilot datasets.
• Univariate gene distributions were learned, but bimodality posed challenges.
• 10x Genomics data presented greater inference challenges than Smart-seq2 due to sparsity.

Conclusions:

• Generative deep learning models show potential for supporting scRNA-seq experimental design.
• Model performance is sensitive to pilot dataset size and data sparsity.
• Further development is needed to address limitations in representing cell type abundance and complex gene expression patterns.