Citation
The workbench for single cell RNAseq (scRNAseq) is designed to allow biologists meaningful access to single cell data, even with limited informatics training. The workbench begins by selecting a dataset for analysis, and then offers analysis tools following several standard pre-processing steps.
Start by choosing a dataset in the left dropdown.
Learn how to use the single-cell workbench
Initial shape:
Loading initial gene/cell count plots
1 - Dataset selection
2 - Analysis selection
3 - Primary filter
4 - QC: filter out mitochondrial content
5 - Identify highly-variable genes
6 - Perform Principal Component Analysis (PCA)
7 - Perform t-SNE or UMAP analysis
8 - Perform clustering
9 - Find marker genes
10 - Rename, merge, or delete clusters
11 - Compare genes
Primary analysis selected
Enter a gene of interest to see its t-SNE colored both by expression and cluster / cell type.
Filtered shape:
Apply filters as desired.
Press on the "Compute and plot" button to see plots of (a) the number of genes in each cell; (b) number of read counts per cell; and (c) percent mitochondrial content. The general recommendation is to maintain percent mitochondrial content below 0.05% to focus on living cells. Based on the data in these plots you may wish to change some of the criteria in your previous step. Press ‘Save these genes’ before moving to the next step.
Saving genes will filter the dataset in an irreversible way.
Post-QC shape:
No mitochondrial genes with this prefix found in the dataset.
This step will filter the large gene list down to those genes that are more likely to represent the biologically important variability between each of the cells. There are several parameters that can be adjusted here to set the sensitivity versus stringency of what genes are included and you may find that trying different parameters help you identify new features of the dataset. As a brief description, the x-axis represents the average expression of genes across the dataset. The y-axis is a measure called dispersion, which indicates the variance of that gene across the dataset. The workbench will limit your maximum number of highly variable genes to 2,000. By increasing the Min mean you increase the minimal expression value of genes that may be considered as highly variable. We suggest that you change the parameters and observe the plot to further guide your selection parameters. At the completion of this step press on ‘Save these genes’.
Fields below are ignored if this is set
Saving genes will filter the dataset in an irreversible way.
Shape after filtering for highly variable genes:
Suggested highly-variable genes:
The principal component analysis indicates the groups of genes that have the largest contribution to the variability in gene expression in the dataset. For example, PC1 with have the largest effect in dividing the dataset into subtypes of cells or samples. Each principal component is composed of groups of “related” genes. Visualizing your principal component graph and knowing which genes contribute to it (listed in the table) is important for the next step. The number of principal components included will affect your t-SNE (t-distributed stochastic neighbor embedding) plot.
Comma-separated list of gene symbols
Comma-separated list of principal components, ideally between 2 and 5 PCs, just the number only (omit the "PC").
This non-linear dimensional reduction visualization tool is used to visually cells that are similar to each other. The number of principal components to include depends on the result in the PCA step, as listed above. You can vary the number of principal components used and view how it changes the data display. The default recommendation is to look for the point in the PCA curve in which additional components result in minimal added variation.
Comma-separated list of gene symbols
Number of nearest neighbors to consider for each cell
Number of principal components to be embedded into the plot
The Leiden clustering step is used to find the most likely groups of associated cells within a network. Here they are color-coded. The number of neighbors from the previous step will have an effect on the smallest possible size for a cluster. If you are interested in groups of cells that are all larger than 20 cells, for example (based on the gene coloring in the initial PCA) – then you can try 6, 10 or 15 neighbors, for example. However, if this is a smaller dataset of regular RNA-seq, for example, with only biological triplicates, starting with two neighbors makes more sense – because the smallest ‘natural’ group should be 3 replicates. Alternatively, if some of the populations in a single cell dataset are very small, again, 3 neighbors could be a useful approach. However, the smaller the number of neighbors, the larger the number of clusters. The resolution determines how granular the clustering will be. It is set to 0.5 by default. To decrease resolution you can drop it to a lower value, for example. Or increase the number for higher resolution.
This analysis step will show you the marker genes within each group of cells as defined by the clustering method used. You can adjust the number of genes you would like to see for each cluster by adjusting the 'N genes' value.
Click to select genes of interest. You can select individual genes as well as entire rows and columns.
Select desired marker genes in the table above and/or type gene symbols in the field below to visualize
This step allows you to rename, merge, or delete clusters. You can also assign a new label to a group of clusters. To merge clusters, check the box and select the clusters you want to merge. To delete a cluster, select the cluster and press the delete button. To rename a cluster, select the cluster and type the new name in the text box. To assign a new label to a group of clusters, check the box and type the new label in the text box. Press the 'Rerun with labels' button to apply the changes and view the revised plots.
| Group | Num Cells | Markers | New label | Keep |
|---|
Merging and/or deleting clusters is irreversible.
The method dropdown menu lists the available statistical tests for your comparison. The “t-test overestimated variance” option may help in situations where clusters contain few cells and variance is difficult to estimate directly, otherwise for robust clusters choose “t-test” (Assumes normally distributed data). The “Wilcoxon-Rank-Sum” option is a non-parametric test and may be helpful when a dataset includes a few genes with very high expression (outliers) or data distribution is not known. Multiple testing correction can be performed either by Benjamini-Hochberg or Bonferroni methods, where Bonferroni is more conservative.
| Name | log2FC | P-value | FDR | No data to display |
|---|
| Name | log2FC | P-value | FDR | No data to display |
|---|