Fig 1.
Starting from scRNA-seq data (top left), k-NN smoothing is carried out to reduce noise. From the smoothed data, N cells are sampled and GRNs are computed M times. Parallelly, sampling and GRN construction are performed times with shuffling gene names K times, resulting in
shuffled specimens. Using the shuffled data, statistically significant genes and their regulatory relationships are selected. The resulting GRN is subjected to the computation of MDS. At the same time, cells in the original scRNA-seq data are clustered into small clumps by using Louvain clustering with a high resolution. Averaged expression levels of genes in MDS are calculated and used for the clustering of cell clumps. The resulting clusters are regarded as cell types.
Fig 2.
Smoothing and resampling in BootCellNet are important for the robust selection of genes and regulations.
(A) Schematics of the computer experiment. (B) Numbers of genes/nodes in GRNs generated 10 times with or without smoothing and with or without selection. (C) Distribution of number of appearances of edges in 10 specimens, with (red) or without (gray) resampling-based selection.
Fig 3.
BootCellNet robustly identifies MDS.
(A) Distribution of the number of the appearance of genes/nodes in 10 MDS specimens. (B) Numbers of genes in MDS with (red) or without (gray) selection. (C) Relationship between rank of mean degree and number of appearances in 10 MDS specimens of each gene.
Fig 4.
BootCellNet robustly identifies cell types.
(A) Clumps of cells by using Louvain clustering with a resolution of 10. (B) GRN and its MDS (red squares) identified by BootCellNet. (C) An example of the heatmap representation of hierarchical clustering of cell clumps shown in panel A (left) and the resulting clusters (right) based on MDS. (D) An example of the heatmap of hierarchical clustering of cell clumps (left) and the resulting clusters (right) based on 100 genes with variable expression. (E) Adjusted Rand indexes calculated among 10 MDS-based clusterings (left) and 10 variable gene-based ones (right) were shown.
Fig 5.
Cell types in peripheral blood mononuclear cells identified by BootCellNet and its comparison to reference-based labeling and unsupervised clustering.
(A) The resulting clusters obtained by BootCellNet and their annotations described in the main text were shown. (B) Annotation of cells obtained by SingleR with the fine HumanCellAtlas dataset. (C) Clustering result obtained by Louvain clustering implemented in Seurat, with a resolution of 1.0. (D) Heatmap of average expression of MDS genes in each cluster obtained by BootCellNet. Color codes for columns are the same as those in panel A.
Fig 6.
BootCellNet facilitates the inference of the course of hematopoiesis.
(A) GRN and its MDS (red squares) identified by BootCellNet. (B) The resulting cell clusters. (C) Heatmap of average expression of MDS genes in each cluster obtained by BootCellNet. Color codes for columns are the same as those in panel B. (D) Lineage curves were obtained by Slingshot using the clusters by BootCellNet. One lineage curve, which spans from Cluster 1 to Cluster 9 is omitted, since Cluster 9 encompasses lymphoid cells.
Fig 7.
BootCellNet2 and its validation.
(A) Outline of the procedure. Upto the GRN construction, the procedures are the same. From the constructed GRN, the unique set of control nodes consisting of critical nodes and intermittent nodes was computed. Cell clumps are obtained in the same manner as BootCellNet. Averaged expression levels of genes in the unique set are calculated and used for the clustering of cell clumps. Here the clustering was done with multiscale bootstrap to obtain highly confident clusters. The resulting high-confident clustering results are compared to clustering with varying numbers of clusters. The optimal number of clusters is determined as the lowest number of clusters with maximal similarities. (B) Number of appearances of genes in the unique set of control nodes or MDS in 10 specimens. No statistically significant difference was observed by Welch’s two-sided t-test. (C) Adjusted Rand indexes calculated among 10 automated clusterings with (left) or without (right) using the unique set of control nodes are shown. The numbers above the heatmaps represent the number of clusters determined by the BootCellNet2 procedure. (D) Adjusted Rand indexes calculated among 10 automated clusterings with (left) or without (middle) using the unique set of control nodes, along with those without using either automated clustering and the unique set (right) are shown. *p < 5x10^-5 by Welch’s two-sided t-test.
Fig 8.
Identification of cell type clusters by BootCellNet2.
(A) GRN and its unique set of control nodes (critical nodes as red squares, intermittent nodes as red circles) identified by BootCellNet2 are shown. (B) The resulting cell clusters. Annotations for each cluster described in the main text were shown. (C) Heatmap of average expression of genes in the unique set in each cluster obtained by BootCellNet2. Color codes for columns are the same as those in panel B.
Fig 9.
BootCellNet2 gives clustering that comply more with CITE-seq results.
(A) The cell clusters obtained by BootCellNet2. Annotations for each cluster according to GO terms, marker genes, and antibody-derived tags were shown. (B) Heatmap of average expression of antibody-derived tags in each cluster obtained by BootCellNet2. Color codes for columns are the same as those in panel A. (C) Heatmap of average expression of the genes in the unique control set in each cluster obtained by BootCellNet2. Color codes for columns are the same as those in panel A. (D) The cell clusters obtained by the Louvain algorithm with a resolution of 0.7. Annotations for each cluster according to GO terms, marker genes, and antibody-derived tags were shown. (E) Heatmap of average expression of antibody-derived tags in each cluster obtained by the Louvain algorithm. Color codes for columns are the same as those in panel D.
Fig 10.
Identification of cell types in bronchoalveolar lavage fluid scRNA-seq data by BootCellNet2.
(A) GRN and its MDS (red squares) identified by BootCellNet. (B) The resulting cell clusters. (C) Heatmap of average expression of MDS genes in each cluster obtained by BootCellNet. Color codes for columns are the same as those in panel B. (D) Distribution of cells from each class of patients in each cluster. Statistical significances of enrichment of cells from each origin were calculated by Fisher’s exact test with Bonferroni correction. Asterisks were shown when p<5x10^-4.