Identifying cell states in single-cell RNA-seq data at statistically maximal resolution.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals gene expression variation but faces challenges due to data sparsity and noise.
• Existing clustering methods for scRNA-seq data often lack clear objectives, involve complex preprocessing, and yield results difficult to interpret biophysically.
• Accurate partitioning of cells into distinct gene expression states is crucial for understanding cellular heterogeneity.

Purpose of the Study:

• To develop a mathematically rigorous method for partitioning scRNA-seq data into statistically indistinguishable gene expression states.
• To create a tool that operates directly on raw unique molecule identifier (UMI) counts, automatically determining optimal partitions and cluster numbers.
• To ensure identified cell states reflect biological variation, independent of technical experimental factors.

Main Methods:

• Derived a statistically well-defined solution from first principles, accounting for the known measurement noise structure of scRNA-seq data.
• Implemented the solution in a software tool named Cellstates, which requires zero tunable parameters and processes raw UMI count data.
• Developed supplementary software for hierarchical clustering, differential gene expression analysis, and visualization of identified cell states.

Main Results:

• Cellstates accurately recovers optimal partitions on synthetic datasets.
• On real scRNA-seq data, Cellstates robustly identifies subtle substructure within conventionally annotated cell types.
• The identified diversity of gene expression states is systematically dependent on tissue of origin, not technical experimental features.

Conclusions:

• Cellstates provides a principled and parameter-free approach to deconvolute cellular heterogeneity from scRNA-seq data.
• The method effectively reduces data complexity while preserving meaningful biological structure.
• Cellstates facilitates deeper biological insights by revealing biologically relevant cell states and their hierarchical relationships.