Optimization of clustering parameters for single-cell RNA analysis using intrinsic goodness metrics

Area of Science:

• Single-cell RNA sequencing analysis
• Computational biology
• Bioinformatics

Background:

• Accurate clustering of cell subpopulations is essential for single-cell RNA sequencing (scRNA-seq) data analysis.
• Unsupervised clustering performance is highly dependent on chosen algorithms and parameters.
• Predicting clustering accuracy is challenging but crucial for reliable biological insights.

Purpose of the Study:

• To predict the accuracy of unsupervised clustering methods in scRNA-seq data by leveraging intrinsic goodness metrics.
• To evaluate the impact of varying clustering parameters on the accuracy of cell subpopulation identification.
• To identify reliable intrinsic metrics that can serve as proxies for clustering accuracy.

Main Methods:

• Utilized three scRNA-seq datasets with ground truth annotations from distinct anatomical origins.
• Employed the Leiden algorithm and Deep Embedding for Single-cell Clustering (DESC) for unsupervised clustering.
• Implemented linear mixed regression models to analyze parameter impact and ElasticNet regression with 15 intrinsic metrics to predict accuracy.

Main Results:

• The UMAP method for neighborhood graph generation and increased resolution positively impacted clustering accuracy.
• Reduced nearest neighbors amplified the effect of resolution, enhancing preservation of fine-grained cellular relationships.
• Within-cluster dispersion and the Banfield-Raftery index were identified as effective predictors of clustering accuracy.

Conclusions:

• Clustering accuracy in scRNA-seq can be effectively predicted using intrinsic metrics, facilitating parameter optimization.
• Parameter choices, including UMAP, resolution, and number of principal components, significantly influence clustering outcomes.
• Intrinsic metrics like within-cluster dispersion offer a robust method for comparing different clustering configurations and ensuring reliable cell subpopulation identification.