Data preprocessing and the first-time unsupervised clustering

After QC, we used Seurat package (Version 3.1.5) in R (Version 3.6.3) to perform the same analysis pipeline for all scRNA-seq data sets. By default, we employed a global-scaling normalization method “LogNormalize” that normalized the feature expression measurements for each cell by the total expression, multiplied this by a scale factor (10,000), and log-transformed the result. Second, to avoid the interference from doublet cells, we identified and removed these doublet cells by using DoubletFinder [25] package (Version 2.0.3) in R. Third, we calculated CV-rank for each gene in all cells and used the top 2000 genes with the highest CV-rank for the downstream analyses, including principal component analysis (PCA) and unsupervised clustering (the Louvain algorithm) [26]. Then, we performed PCA to identify the true dimension of data sets, and we chose as many principal components as possible for the downstream analyses. As for the unsupervised clustering, we chose 0.6 as the default resolution parameter, and this clustering result was defined as the first-time unsupervised clustering result (Fig. 1a–c).

Workflow for sensitive gene identification. a After the single-cell sequencing, we obtained expression profiles of various cell types, with different colors representing different cell types. We used Seurat to calculate the CV-rank for all genes in all cells, and the top 2000 genes were defined as HVGs (red); b Based on the results of the first-time unsupervised clustering, we detected high CV-rank genes in each cluster; c Shannon entropy based on the average expressions of these genes (with high CV-rank in more than half of clusters) among cells in each cluster. The genes with high entropy (higher than the median entropy) were regarded as the sensitive genes; d We re-selected the top 2000 HVGs with sensitive genes removed from the expression matrix