Resolution Tradeoffs in Modularity Clustering of Single Cell RNA-Seq Datasets

Area of Science:

• Computational Biology
• Network Science
• Genomics

Background:

• Modularity clustering is widely used for community detection in networks and cell type identification in single-cell RNA sequencing (scRNAseq) data.
• The resolution parameter in modularity clustering implicitly determines the number of inferred clusters, but its effect on scRNAseq analysis is not well-described.
• Improper resolution parameter selection can lead to Type I or Type II errors in cell type identification.

Purpose of the Study:

• To explicitly describe clustering as a function of the resolution parameter by introducing the concept of splitting resolution.
• To analyze the behavior of splitting resolutions in k-nearest neighbor graphs derived from cell embeddings.
• To provide insights into the tradeoffs between Type I and Type II errors in cell type inference based on the chosen resolution parameter.

Main Methods:

• Introduced the notion of splitting resolution as the minimum resolution at which a graph or subgraph splits into multiple clusters.
• Derived formulas for splitting resolution in k-nearest neighbor graphs based on normal distributions, considering sample size, embedding dimension, and covariance structure.
• Approximated cell embeddings from seven scRNAseq datasets using Gaussian mixtures to connect theoretical findings to real-world data.

Main Results:

• Demonstrated that the splitting resolution of a disconnected subgraph is inversely proportional to its frequency within the graph.
• Extended the resolution limit theory to general resolution parameter values, highlighting a limitation of modularity clustering for cell type inference.
• Derived explicit formulas for splitting resolutions in graphs formed from normally distributed cell embeddings.

Conclusions:

• The choice of resolution parameter in modularity clustering significantly impacts cell type identification accuracy in scRNAseq data.
• Understanding splitting resolutions provides a framework for analyzing and mitigating errors (Type I and Type II) in cell type inference.
• This work offers a theoretical basis for optimizing the resolution parameter in scRNAseq analysis to improve cell type identification.