LineageVAE: reconstructing historical cell states and transcriptomes toward unobserved progenitors

Area of Science:

• Computational biology
• Genomics
• Molecular biology

Background:

• Single-cell RNA sequencing (scRNA-seq) offers cell state characterization but is destructive, limiting dynamic process analysis.
• Existing lineage tracing methods provide limited clonal insights due to sparse observations and non-identical progenitors.
• Reconstructing dynamic cell state transitions and historical transcriptomes remains a challenge in single-cell analysis.

Purpose of the Study:

• To develop a novel computational methodology, LineageVAE, to overcome limitations in analyzing dynamic cellular processes using scRNA-seq data.
• To enable the reconstruction of unobservable cell state transitions and infer historical transcriptomes at single-cell resolution.
• To infer progenitor heterogeneity and regulatory dynamics along differentiation trajectories.

Main Methods:

• LineageVAE is a deep generative model utilizing identical lineage barcodes from scRNA-seq data.
• The model transforms observations into sequential trajectories within a latent cell state space.
• Implementation in Python using the PyTorch deep learning library.

Main Results:

• LineageVAE reconstructs unobservable cell state transitions and historical transcriptomes.
• The method infers progenitor heterogeneity and transcription factor activity.
• Successful application to hematopoiesis and reprogrammed fibroblast datasets demonstrated backward cell state transition restoration.

Conclusions:

• LineageVAE provides a powerful computational approach to study dynamic cellular processes previously inaccessible with scRNA-seq.
• The methodology enables single-cell resolution reconstruction of cell state trajectories and regulatory dynamics.
• This work advances the understanding of cell state divergence and differentiation.