Integrating Multiple Clustering Techniques and Performance Measures via Ranking for scRNA-Seq Data

As single-cell gene expression data analysis continues to grow, the need for reliable clustering techniques has become increasingly important. This growth has also led to a rise in heuristic means for technique choice, which could lead to inaccurate reports if a comprehensive evaluation of the resulting clusters is omitted. Previous work in the field of microarray data provided measures of stability to evaluate clustering techniques for highly correlated data. Additional work on aggregation in the same era presented a way to combine ranked lists using several performance measures. In this paper, we adapt the stability measures from the microarray era to function on single-cell data and employ them alongside several existing measures to evaluate characteristics of clustering techniques at varying parameter choices. We then compare individual measures alongside an aggregation method (AM) on all measures to rank these techniques for five datasets with known groups and one without known groups. Our findings show that techniques ranked highest by a single validation measure tend to have mediocre performance for other measures for multiple datasets. Alternatively, our method of aggregating techniques based on their ranked performance on all measures ensures that techniques with a strong overall performance are ranked highest. The techniques selected in this fashion had above average performance in most of the measured clustering characteristics and perform well in other ways, such as consistently estimating cluster counts that closely match the true number of biological groups.