TemporalVAE: atlas-assisted temporal mapping of time-series single-cell transcriptomes during embryogenesis

Area of Science:

• Computational Biology
• Developmental Biology
• Genomics

Background:

• Single-cell time-series atlases are crucial for biomedical research but challenging to analyze due to complex cell states and batch effects.
• Existing methods struggle with temporal analysis of individual datasets, limiting their utility.

Purpose of the Study:

• To introduce TemporalVAE, a deep generative model for accurate inference of biological time from single-cell data.
• To demonstrate the model's scalability, accuracy, and interpretability in diverse biomedical applications.

Main Methods:

• Developed TemporalVAE, a dual-objective deep generative model for inferring biological time from a compressed latent space.
• Validated the model on large-scale mouse embryogenesis datasets (millions of cells).
• Applied TemporalVAE to human and non-human primate peri-implantation data.

Main Results:

• TemporalVAE demonstrated scalability with millions of cells and accurate cell staging across platforms.
• Identified temporally sensitive genes through in silico perturbation, showcasing interpretability.
• Successfully staged human peri-implantation cells (in vivo and in vitro) and enabled cross-primate embryo comparisons.

Conclusions:

• TemporalVAE offers a robust solution for temporal analysis of single-cell data, overcoming batch effects and complex cell states.
• The model shows broad applicability in developmental biology, disease modeling, and comparative genomics across species.