GRAPHICAL MODELS FOR ZERO-INFLATED SINGLE CELL GENE EXPRESSION.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Traditional bulk gene expression studies aggregate cell data, masking individual cell variations.
• Single-cell sequencing technologies now enable detailed analysis of gene expression at the cellular level.
• Single-cell data often exhibits zero-inflated patterns, where many transcripts are undetectable in individual cells despite population-level abundance.

Purpose of the Study:

• To address the challenge of inferring gene co-regulatory networks from zero-inflated single-cell expression data.
• To propose and validate a novel statistical model for analyzing cell-to-cell variation in gene expression.
• To identify gene interactions not detectable with traditional bulk analysis methods.

Main Methods:

• Development of a multivariate Hurdle model, a mixture of singular Gaussian distributions, to handle zero-inflated expression patterns.
• Application of neighborhood selection with pseudo-likelihood and a group lasso penalty for inferring undirected graphical models.
• Utilizing these methods to capture conditional independencies between genes in single-cell datasets.

Main Results:

• The proposed Hurdle model demonstrates higher sensitivity than existing methods in simulations, even with deviations from model assumptions.
• Application to T follicular helper cells and mouse dendritic cells reveals novel network structures.
• Inferred networks provide insights not obtainable from bulk gene expression data.

Conclusions:

• The multivariate Hurdle model is an effective tool for analyzing zero-inflated single-cell gene expression data.
• This approach enhances the discovery of gene co-regulatory networks, offering a more sensitive alternative to traditional methods.
• The study provides a valuable computational framework for advancing single-cell genomics research.