Exponential-Family Embedding With Application to Cell Developmental Trajectories for Single-Cell RNA-Seq Data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables high-resolution analysis of cellular heterogeneity.
• Nonlinear dimensionality reduction is crucial for visualizing and analyzing scRNA-seq data, particularly for developmental trajectories.
• The statistical properties and performance of existing nonlinear embedding methods are not fully understood.

Purpose of the Study:

• To develop a statistically principled and computationally efficient nonlinear embedding method for scRNA-seq data.
• To jointly embed cells and genes using a random dot product model with exponential-family distributions.
• To investigate and identify distinct developmental trajectories in oligodendrocyte cell populations.

Main Methods:

• Developed the exponential-family SVD (eSVD) method, a nonlinear embedding technique.
• Utilized alternating minimization for computational efficiency and theoretical analysis.
• Proved identifiability conditions and consistency of the eSVD estimator.
• Developed statistically sound procedures for method tuning.
• Applied eSVD with Gaussian distributions to a mouse oligodendrocyte dataset.

Main Results:

• eSVD demonstrates competitive performance against existing embedding methods in simulations.
• The method successfully identified two major, previously indistinguishable, developmental trajectories diverging at mature oligodendrocyte stages.
• Analysis of mouse oligodendrocyte single-cell RNA-seq data revealed complex cell differentiation pathways.

Conclusions:

• eSVD provides a statistically robust and computationally efficient approach for nonlinear embedding of scRNA-seq data.
• The method advances the understanding of single-cell data analysis and trajectory inference.
• eSVD offers novel insights into oligodendrocyte development, highlighting distinct mature cell trajectories.