Polaratio: A magnitude-contingent monotonic correlation metric and its improvements to scRNA-seq clustering

Abstract

Motivation: Single-cell RNA sequencing (scRNA-seq) technologies and analysis tools have allowed researchers to achieve remarkably detailed understandings of the roles and relationships between cells and genes. However, conventional distance metrics, such as Euclidean, Pearson, and Spearman distances, fail to simultaneously take into account the high dimensionality, monotonicity, and magnitude of gene expression data. To address several shortcomings in these commonly used metrics, we present a magnitude-contingent monotonic correlation metric called Polaratio which is designed to enhance the quality of scRNA-seq data analysis. Results: We integrate three interpretable clustering algorithms - Single-Cell Consensus Clustering (SC3), Hierarchical Clustering (HC), and K-Medoids (KM) - through a consensus cell clustering procedure, which we evaluate on various biological datasets to benchmark Polaratio against several well-known metrics. Our results demonstrate Polaratio's ability to improve the accuracy of cell clustering on 5 out of 7 publicly available datasets. Availability: https://github.com/dubai03nsr/Polaratio Contact: pcicalese@uh.edu

Probabilistic clustering of cells using single-cell RNA-seq data

“…In addition, we analyze Worm neuron, Human kidney, 10x pbmc, Mouse retina raw, CITE CBMC and TAM FACS datasets. They are considered silver-standard datasets because in these the cells were labeled according to the authors' computational analysis and biological knowledge [25].…”

“…Six gold-standard datasets and six silver-standard datasets were used for our comparison. Goolam, Kolodziejczyk, Yan, Deng, Pollen, and Biase are considered as gold-standard datasets because in these datasets the cell sub-populations were chosen from distinct biological stages or conditions [25].…”