Fig. 9
From: Contrastive self-supervised clustering of scRNA-seq data
Model stability across consecutive runs (a1–a4) and to input downsample (b1–b4). The model stability across 3 runs on the real-world datasets has been depicted as the coefficient of variation (for each dataset, the standard deviation across runs divided by the average result). Here, the lower the score, the more stable the model. For this experiment, the input data consisted of all input cells in each dataset. The second analysis assesses how the model performance changes when only a fraction of cells is passed as input. Three different stratified subsets have been generated randomly for each dataset, selecting 25%, 50%, 75% and 100% of all cells in each of the benchmarked real datasets. The annotated values represent the mean score of each experiment. The performance remains relatively stable: providing half of the dataset reduces the ARI score by 5, while only 25% of data attains an ARI score of 0.67