This tutorial presents an outline of the basic series of steps for analyzing a single cell RNA-Seq experiment in Partek Flow starting with the count matrix file.

If you are starting with the raw data (FASTQ files), please begin with our Processing Single Cell RNA-seq FASTQ Files tutorial, which will take you from raw data to a count matrix file. 

This tutorial includes only one sample, but the same steps will be followed when analyzing multiple samples. For notes on a few aspects specific to a multi-sample analysis, please see our Single Cell RNA-Seq Analysis (Multiple Samples) tutorial. 

If you are new to Partek Flow, please see Getting Started with Your Partek Flow Hosted Trial for information about data transfer and import and Creating and Analyzing a Project for information about the Partek Flow user interface.  

Filtering cells

An important step in analyzing single cell RNA-Seq data is to filter out low quality cells. A few examples of low-quality cells are doublets, cells damaged during cell isolation, or cells with too few reads to be analyzed. You can do this in Partek Flow using the Single cell QA/QC task. 

A task node, Single cell QA/QC, is produced. Initially, the node will be semi-transparent to indicate that it has been queued, but not completed. A progress bar will appear on the Single cell QA/QC task node to indicate that the task is running (Figure 1).

 

The Single cell QA/QC report includes interactive violin plots showing the value of every cell in the project on several quality measures (Figure 2).

 

There are three plots: number of read counts per cell, number of detected genes per cell, and the percentage of mitochondrial reads per cell.

Each point on the plots is a cell and the violins illustrate the distribution of values for the y-axis metric. Cells can be filtered either by clicking and dragging to select a region on one of the plots or by setting thresholds using the filters below the plots. Here, we will apply a filter for the number of read counts.

The plot will be shaded to reflect the filter. Cells that are excluded will be shown as black dots on both plots. 

The read counts per cell and number of detected genes per cell are typically used to filter out potential doublets - if a cell as an unusually high number of total counts or detected genes, it may be a doublet. The mitochondrial reads percentage can be used to identify cells damaged during cell isolation - if a cell has a high percentage of mitochondrial counts, it is likely damaged or dying and may need to be excluded. 

Normalization

Because different cells will have a different number of total counts, it is important to normalize the data prior to downstream analysis. For droplet-based single cell isolation and library preparation methods that use a 3' counting strategy, where only the 3' end of each transcript is captured and sequenced, we recommend the following normalization -   1. CPM (counts per million), 2. Add 1, 3. Log2. This accounts for differences in total UMI counts per cell and log transforms the data, which makes the data easier to visualize. 

This adds CPM (counts per million), Add 1, and Log2 to the Normalization order panel. Normalization steps are performed in descending order. 

A new Normalized counts data node will be produced.

For more information on normalizing data in Partek Flow, please see the Normalize Counts section of the user manual.  

Scaling 

For some data sets, it may be necessary to remove technical artifacts or batch effects. To do this, you can use the Scaling task in the Normalization and Scaling section. To configure the scaling task, select the cell or sample attribute effects you would like to regress out of the data set. The scaling task is detailed in our Single Cell Scaling white paper. We will not perform scaling for this data set. 

Filtering genes 

A common task in bulk and single-cell RNA-Seq analysis is to filter the data to include only informative genes. Because there is no gold standard for what makes a gene informative or not and ideal gene filtering criteria depend on your experimental design and research question, Partek Flow has a wide variety of flexible filtering options. 

There are three categories of filter available - noise reduction, statistics based, and feature list. 

The noise reduction filter allows you to exclude genes considered background noise based on a variety of criteria. The statistics based filter is useful for focusing on a certain number or percentile of genes based on a variety of metrics, such as variance. The feature list filter allows you to filter your data set to include or exclude particular genes.

 For example, you can use a noise reduction filter to exclude genes that are not expressed by any cell in the data set, but were included in the matrix file.

This produces a Filtered counts data node. This will be the starting point for the next stage of analysis.

PCA

Principal components (PC) analysis (PCA) is an exploratory technique that is used to describe the structure of high dimensional data by reducing its dimensionality. Because PCA is used to reduce the dimensionality of the data prior to clustering as part of a standard single cell analysis workflow, it is useful to examine the results of PCA for your data set prior to clustering. 

You can choose Features contribute equally to standardize the genes prior to PCA or allow more variable genes to have a larger effect on the PCA by choosing by variance. By default, we take variance into account and focus on the most variable genes. 

If you have multiple samples, you can choose to run PCA for each sample individually or for all samples together using the Split cells by sample option. 

This will generate a Scree plot and PC component loadings table, which are useful for determining how many PCs to use in downstream analysis tasks. 

A new PCA task node will be produced.

The PCA task report includes the PCA plot, the Scree plot, the component loadings table, and the PC projections table. To switch between these elements, use the buttons in the upper right-hand corner of the task report . Each cell is shown as a dot on the PCA scatter plot. 

The Scree plot lists PCs on the x-axis and the amount of variance explained by each PC on the y-axis, measured in Eigenvalue. The higher the Eigenvalue, the more variance is explained by the PC. Typically, after an initial set of highly informative PCs,  the amount of variance explained by analyzing additional PCs is minimal. By identifying the point where the Scree plot levels off, you can choose an optimal number of PCs to use in downstream analysis steps like graph-based clustering and t-SNE. 

In this data set, a reasonable cut-off could be set anywhere between 8 and 25 PCs. We will use 10 in downstream steps. 

Viewing the genes correlated with each PC can be useful when choosing how many PCs to include. 

This table lists genes on rows and PCs on columns. 

In this case, we see a few known marker genes are correlated with PC1 (Figure 7). 

 

Graph-based clustering

Graph-based clustering identifies groups of similar cells using PC values as the input. By including only the most informative PCs, noise in the data set is excluded, improving the results of clustering. 

Clustering can be performed on each sample individually or on all samples together. Here, we are working with a single sample. 

The Number of principal components should be set based on the your examination of the Scree plot and component loadings table. The default value of 100 is likely exhaustive for most data sets, but may introduce noise that reduces the number of clusters that can be distinguished. 

A new Graph-based clustering task node and a Clustering result data node will be generated. 

The Graph-based clustering task report lists the number of clusters and what proportion of cells fall into each cluster. It also includes a cluster biomarkers table. This lists the top-10 genes that distinguish each cluster from the others (Figure 9). These are calculated using an ANOVA test comparing the cells in each group to all the other cells, filtering to genes that are 1.5 fold upregulated, and sorting by ascending p-value. This ensures that the top-10 genes of each cluster are highly and disproportionately expressed in that cluster. 

 

We will use t-SNE to visualize the results of Graph-based clustering.

t-SNE

t-Distributed Stochastic Neighbor Embedding (t-SNE) is a dimensional reduction technique that prioritizes local relationships to build a low-dimensional representation of the high-dimensional data that places objects that are similar in high-dimensional space close together in the low-dimensional representation. This makes t-SNE well suited for analyzing high-dimensional data when the goal is to identify groups of similar objects, such as cell types in single cell RNA-Seq data. 

If you have multiple samples, you can choose to run t-SNE for each sample individually or for all samples together using the Split cells by sample option. Please note that this option will not be present if you are running t-SNE on a clustering result. For clarity, clustering results run with all samples together must be viewed together and clustering results run by sample must be viewed by sample. 

Like Graph-based clustering, t-SNE takes PC values as its input and further reduces the data down to two or three dimensions. For consistency, you should use the same number of PCs as the input for t-SNE that you used for Graph-based clustering. 

A new t-SNE task node will be produced.

The t-SNE scatter plot is interactive and can be viewed for 2D or 3D (calculated separately). Using the t-SNE plot, cells can be classified based on clustering results or differences in gene and pathway expression. By default, the t-SNE plot will be colored by the clusters identified by Graph-based clustering if available. 

Plot controls are located in the control panel on the left. You can adjust plot options, perform selection and filtering, and manage cell type classifications using the control panel. Below the scatter plot is the biomarker table that we saw in the the Graph-based clustering results. The table is interactive and clicking a gene in the table will color the plot by that gene (hold Ctrl on your keyboard and click to color by multiple genes in the table). 

The t-SNE plot is 3D by default. You can rotate the 3D plot by left-clicking and dragging your mouse. You can zoom in and out using your mouse wheel. You can pan by right-clicking and dragging your mouse. The 2D t-SNE is also calculated and you can switch between the 2D and 3D plots using the Plot style radio buttons in the control panel. In 3D, you can switch from points to 3D spheres and also add a fog effect to improve depth perception on the plot. To produce an optimal plot, you can also adjust size of the points using the Point size slider. 

Coloring the t-SNE scatter plot

You can use the Color by options to explore the data. 

The cells on the plot will be colored based on their expression level of CD79A (Figure 13).

 

Coloring by one gene uses the two-color numeric palette, which can be customized by clicking . Clicking the  icon lets you color by an additional gene (up to three genes at a time). If you color by more than one gene, the color palette switches to a Green-Red-Blue color scheme with the balance between the three color channels determined by the values of the three genes. For example, a cell that expresses all three genes would be white, a cell that expresses the first two genes would be yellow, and a cell that expresses none of the genes would be black (Figure 14). 

 

Clicking a cell on the plot shows the expression values of the cell in the legend (Figure 15). 

 

If you want to color by more than three genes at time, for example, by a list of genes that distinguish a particular cell type, you can use the color by list option.

Coloring by a list calculates the first three principal components for the gene list and color the cells on the plot by their values along those three PCs with green for PC1, red for PC2, and blue for PC3 (Figure 16).

 

Typically, the expression of a set of marker genes will be highly correlated, allowing the first PC to account for a large percentage of the variance between cells for that gene list. As a result, the group of cells characterized by their expression of the genes on the list will separate from the rest of the cells along PC1 and will be colored green (Figure 16). If the gene list is more complex, for example, including marker genes for multiple cell types, there may be several sets of correlated genes accounting for significant amounts of variance, leading to groups of cells being distinguishable along PC2 and PC3 as well. In that case, there may be green, blue, and red groups of cells on the plot. If the gene list does not distinguish any group of cells, all cells will have similar PC values, leading to similarly colored cells on the plot. 

In addition to coloring by gene expression and by gene lists, the points can be colored by any cell or sample attribute. Available attributes are listed as options in the Color by drop-down menu. 

Selecting cells on the t-SNE scatter plot

The most basic way to select a point on the scatter plot is to click it with the mouse while in pointer mode. To select multiple cells, you can hold Ctrl on your keyboard and click the cells. To select larger groups of cells, you can switch to Lasso mode by clicking  at the top of the control panel. The lasso lets you freely draw a shape to select a cluster of cells. 

You can also create a lasso with straight lines using Lasso mode by clicking, releasing, and clicking again to draw a shape. 

 

 

By default, selected cells are shown in bold while unselected cells are dimmed (Figure 18). This can be changed to gray selected cells using the Selection/Filtering section of the control panel.

 

Alternatively, you can select cells using any categorical cell or sample attribute. In the Selection/Filtering panel, the Attributes tab drop-down menu lists all categorical attributes. 

This adds check boxes for each of the level of the attribute (i.e., clusters). Click a check box to select the cells with that attribute level. 

This selects cells from Graph-based clusters 3, 7, and 9 (Figure 20). The number of selected cells is listed in the Selection/Filtering section of the control panel.

 

Cells can also be selected based on their gene expression values using the Features tab in the Selection/Filtering section. 

This gives a text-field where you can type any gene to use to filter the cells. 

By default, cells that expression >= 1 of the gene will be selected (Figure 21). The sign and value of the cutoff can be changed for each gene filter. You can add up to 5 gene filters by typing an additional gene names and clicking . Because these filters can be greater than or less than filters, it is possible to select a very specific population of cells using the feature filter. It may be useful to color by your genes of interest to help set the cutoff values. 

 

Filtering cells on the t-SNE scatter plot

Once a cell has been selected on the plot, it can be filtered. The filter controls can exclude or include (only) any selected cell. 

Filtering can be particularly useful when you want to use a gene expression threshold to classify a group of cells, but the gene in question is not exclusively expressed by your cell type of interest. 

Here, for example we can filter to include just cells from a few of the clusters. 

This selects clusters 1, 4, 5, 6, and 8. 

To exclude selected cells, click  (filter exclude). 

The plot will update to show only the included cells (Figure 23). 

 

Cells that are not shown on the plot cannot be selected, allowing you to focus on the visible cells. The number of cells shown on the plot out of the total number of original cells is listed at the top of the control panel. You can adjust the view to focus on only the included cells.

To revert to the original scaling, click the  button again. 

Additional inclusion or exclusion filters can be added to focus on a smaller subset of cells. To remove applied filters, click Clear filters. 

The plot will update to show all cells and return to the original scaling. 

Classifying cells

Classifying cells allows to you assign cells to groups that can be used in downstream analysis and visualizations. Commonly, this is used to describe cell types, such as B cells and T cells, but can be used to describe any group of cells that you want to consider together in your analysis, such as cycling cells or CD14 high expressing cells. Each cell can only belong to one class at a time so you cannot create overlapping classes. 

To classify a cell, just select it then click Classify selection. 

For example, we can classify a cluster of cells expressing high levels of CD79A as B cells. 

 

Because most of these cells express CD79A, a B cell marker, and because they cluster together on the t-SNE, suggesting they have similar overall gene expression, we believe that all these cells are B cells.

The classification, B cells, is added to the Classifications section of the control panel and the number of cells in that classification is listed next to the name (Figure 26).

 

You can edit the name of a classification by selecting  or delete it by selecting . The classifications you have made on the scatter plot are saved as a working draft so if you close the plot and return to it, the classifications will still be there. However, classifications are not available for downstream tasks until you apply them.

Once you have added saved a classification, you can color the t-SNE plot by the new attribute Classification. Here, I classified a few additional cell types using a combination of known marker genes and the clustering results (Figure 27). 

 

The Summary button in the Classifications section links to a summary of the classifications with the number and percentage of cells from each sample that belong to each classification. This is particularly useful when you are classifying cells from multiple samples,.

To use the classifications in downstream tasks and visualizations, you must first apply them.

 

Applying classifications closes the t-SNE scatter plot and runs the Classify cells task. This saves the new cell attributes you created, runs an ANOVA test to identify biomarkers for each classification, and generates a matrix file with the number of cells in each classification for each sample. 

The Classified groups task report opens a t-SNE like the one used to classify the cells, but the newly calculated classifications biomarker table is available below the plot (Figure 30). 

 

To view the cell counts table, open the task report for the Group cell counts data node.

The number of cells for each sample is listed (Figure 31). This data node can be used as the starting point for differential cell count analysis. This is particularly useful in multi-sample experiments, such as immunotherapy studies, where the number of cells of a particular cell type in two or more conditions is of interest. 

 

Comparing gene expression between cell types

A common goal in single cell analysis is to identify genes that distinguish a cell type. To do this, you can use the differential analysis tools in Partek Flow. I will show how to use the Gene Specific Analysis (GSA) test in Partek Flow, which on its default settings is equivalent to limma-trend, a statistical test has been shown to be highly effective for differential analysis of single cell RNA-Seq data (Soneson and Robinson 2018). 

The first page of the configuration dialog asks what attributes you want to include in the statistical test. Here, we only want to consider the Classifications, but in a more complex experiment, you could also include experimental conditions or other sample attributes. 

We will make a comparison between NK cells and all the other cell types to identify genes that distinguish NK cells. You can also use this tool to identify genes that differ between two cell types or genes that differ in the same cell type between experimental conditions. 

The top panel is the numerator for fold-change calculations so the experimental or test groups should be selected in the top panel.

The bottom panel is the denominator for fold-change calculations so the control group should be selected in the bottom panel.

This adds the comparison to the statistical test. 

The GSA task report lists genes on rows and the results of the statistical test (p-value, fold change, etc.) on columns (Figure 34). For more information, please see our documentation page on the GSA task report. 

 

Genes are listed in ascending order by the p-value of the first comparison so the most significant gene is listed first. To view a volcano plot for any comparison, click .  To view a violin plot for a gene, click  next to the Gene ID. 

 The Feature plot viewer will open showing a violin plot for KLRD1 (Figure 35). The violins are density plots with the width corresponding to frequency. 

 

You can switch the grouping of cells using the Group by drop-down menu. The order of groups can be adjusted by dragging groups up and down in the Group order panel. To navigate between genes in the table, click the Next > and Previous > buttons. 

The table lists all of genes in the data set; using the filter control panel on the left, we can filter to just the genes that are significantly different for the comparison.

Here, we are using a very stringent cutoff to focus only on genes that are specific to NK cells, but other applications may require a less stringent cutoff. 

The number of genes at the top of the filter control panel updates to indicate how many genes are left after the filters are applied. 

 

The GSA report will close and a new task, the Differential analysis filter, will run and generate a filtered Feature list data node. 

For more information about the GSA task, please see the Differential Gene Expression - GSA section of our user manual. 

Generating a heat map

Once we have filtered to a list of significantly different genes, we can visualize these genes by generating a heat map. 

The hierarchical clustering task will generate the heat map. You can choose to cluster samples (cells) and features (genes). You will almost always want to cluster features as this generates the clear blocks of color that make heat maps comprehensible. For single cell data sets, you may choose to forgo clustering the cells in favor of ordering them by the attribute of interest. Here, we will not filter the cells, but instead order them by their classification. 

You can filter samples using the Filtering section of the configuration dialog. Here, we will not filter out any samples or cells. 

The heat map will initially appear to be all black. This is common in single cell RNA-Seq data because outlier cells will skew the high and low ends. We can adjust the minimum and maximum of the color scheme to improve the appearance of the heat map.

Distinct blocks of red and green now appear on the plot. Cells are on rows and genes are on columns. Because of the limited number of pixels on the screen, genes are grouped. You can zoom in using the zoom controls or your mouse wheel if you want to view individual gene rows. We can annotate the plot with cell attributes. 

The plot now includes blocks of color along the left edge indicating the classification of the cells. We can transpose the plot to give the cell labels a bit more space.

We can also customize the colors of the plot.

The heat map now shows a teal to yellow gradient with a black midpoint (Figure 38). 

 

As with any visualization in Partek Flow, the image can be saved as a publication-quality image to your local machine by clicking  or sent to a page in the project notebook by clicking . For more information about Hierarchical clustering, please see the Hierarchical Clustering section of the user manual. 

Performing enrichment analysis

While a long list of significantly different genes is important information about a cell type, it can be difficult to identify what the biological consequences of these changes might be just by looking at the genes one at a time. Using enrichment analysis, you can identify gene sets and pathways that are over-represented in a list of significant genes, providing clues to the biological meaning of your results.

We distribute the gene sets from the Gene Ontology Consortium, but Gene set enrichment can work with any custom or public gene set database. 

The Gene set enrichment task report lists gene sets on rows with an enrichment score and p-value for each. It also lists how many genes in the gene set were in the input gene list and how many were not (Figure 40). Clicking the Gene set ID links to the geneontology.org page for the gene set. 

 

In Partek Flow, you can also check for enrichment of KEGG pathways using the Pathway enrichment task. The task is quite similar to the Gene set enrichment task, but uses KEGG pathways as the gene sets. 

The task report is similar to the Gene set enrichment task report with enrichment scores, p-values, and the number of genes in and not in the list (Figure 41). 

 

Clicking the KEGG pathway ID in the Pathway enrichment task report opens a KEGG pathway map (Figure 42). The KEGG pathway maps have fold-change and p-value information from the input gene list overlaid on the map, adding a layer of additional information about whether the pathway was upregulated or downregulated in the comparison.

 

Color are customizable using the control panel on the left and the plot is interactive. Mousing over gene boxes gives the genes accounted for by the box, with genes present in the input list shown in bold, and the coloring gene shown in red (Figure 43).

 

Clicking a pathway box opens the map of that pathway, providing an easy way to explore related gene networks. 

Pipeline

 

For information about automating steps in this analysis workflow, please see our documentation page on Making a Pipeline. 

References

Soneson C and Robinson MD. Bias, robustness and scalability in single-cell differential expression analysis. Nature Methods 2018 Apr;15(4):255-261.