Abstract
AbstractDuring development and differentiation, transcriptional regulation in the cell often occurs at the level of gene programs (i.e., sets of co-varying genes), rather than isolated genes. It is therefore crucial to identify differential program expression over time, or across case-vs-control samples. However, this has remained difficult: gene programs are inferred by analyzing gene coexpression, and mathematical operations on the latter are nontrivial. Gene coexpression is quantified as a symmetric positive-definite matrix, on which even basic quantities such as arithmetic differences are neither mathematically sound nor biologically interpretable. Here we exploit the structure of the Riemannian manifold of gene coexpression matrices to propose a novel abstraction of gene coexpression that is mathematically well-founded while being computationally tractable and statistically rigorous. Importantly, it also captures biological similarity better than standard coexpression. This conceptual advance enables us to introduce Sceodesic, an algorithm that invokes the log-Euclidean metric from differential geometry to quantify coexpression patterns specific to each cell state, and organizes them into a study-wide panel of interpretable gene programs. Applied to nine single-cell RNA-seq datasets, Sceodesic outperforms existing methods in early detection of cell fate commitment by leveraging differential expression of gene programs, and is also effective in discovering disease-linked programs in multi-sample studies. By respecting the manifold of gene coexpression matrices, Sceodesic resolves a longstanding challenge in relating biological variability to statistical analyses of single-cell RNA-seq data and enables the discovery of gene programs driving differentiation and disease.Software availabilityhttps://singhlab.net/Sceodesic
Publisher
Cold Spring Harbor Laboratory