Next, we isolated each cluster and considered it for further splitting, in the 2D space of the gt-SNE embedding. We centered it using PCA and standardized it by subtracting the mean and dividing by the standard deviation. We marked the cluster for splitting if it now showed three or more outliers based on the median absolute deviation (MAD) with threshold 3.5. We also marked the cluster for splitting if more than 5% of the cells (or 25 cells, whichever is larger) were located at a distance greater than the 70th percentile of the distance to the kth nearest neighbor.
