Fig. 2

Performance assessment of feature selection and clustering methods. a Overview of the computational analysis workflow. b Benchmarking of feature selection methods. In each case, the top 10% of features were selected using either a mean-variance trend to find highly variable genes (HVG, left) or a depth-adjusted negative binomial model (DANB) followed by selecting genes with unexpected dropout rates (NBDrop, middle) or dispersions (NBDisp, right). Plots show the percentage of variance explained by each of the four predictors to the total observed variance: cell line, total counts per cell, total detected features per cell, and predicted cell cycle phase. The blue dashed line indicates the average for the predictor cell line. c–e tSNE projections of the full dataset (c) and two sub-sampled datasets with unequal proportions between different cell lines (d, e). f–h Comparison of clustering assignments by different methods on the full dataset (f), subset 1 (g), and subset 2 (h). Stochastic methods (SC3, mclust, pcaReduce) were run 25 times. Bars and indicated values represent mean adjusted rand index (ARI), and dots correspond to results from individual runs. All other methods are deterministic and were run only once