Supplementary Figure 4: Scalability, accuracy and rare cell-type detection rate of SC3 and benchmarking of the hybrid SC3
(a) Run times for different clustering methods as a function of the number of cells (N). All methods were run on a MacBook Pro (Mid 2014), OS X Yosemite 10.10.5 with 2.8 GHz Intel Core i7 processor, 16 GB 1600 MHz DDR3 of RAM. Two results shown for SC3 correspond to nstart=1000 and nstart=50, where nstart is the number of starting points for k-means clustering; (b) Reducing the number of k-means runs (nstart) from 1,000 to 50 results only in a slightly worse performance for SC3, yet with significant computational savings, as shown in (a). The black line indicates ARI = 0.8; (c) Using the hybrid SC3 based on reference labels provided by the authors. Same as Fig. 2c in the main text, but using the reference labels provided by the authors as inputs to the SVM. Dots represent outliers higher (lower) than the highest (lowest) value within 1.5 x IQR, where IQR is the interquartile range. The black line indicates ARI = 0.8; (d) Robustness of SC3 for the detection of rare cell-types. For two of the datasets, we remove different percentages of the cells in the rare cell-types. The figure shows the mean fraction of SC3 runs in which all the rare cells were clustered together as a function of the total number of cells in the rare cell-type; (e) Sensitivity of SC3 for identifying rare cell-types when the hybrid SC3 approach is used with 30% of cells to train the SVM. This figure was derived from (d) by correcting the mean fraction of times that the rare cells were located in the same cluster using the probability of drawing rare cells within the 30% of all cells (Methods).
