Analysis Algorithms¶
CellEngine provides several machine learning, dimensionality reduction and clustering algorithms for exploratory analysis.
Running Algorithms¶
Howto
 Before starting, check that your scales and compensation are set correctly. These can be viewed and adjusted on the gating page.
 Click algorithms in the left sidebar to open the analysis algorithms page.
 Select common parameters:
 Population selects which population to analyze. Typically this is a singlets or leukocyte gate, excluding debris.
 FCS Files selects which files to analyze.
 If a warning is shown that the number of events exceeds the maximum number of analyzable events, turn on subsample events. You can either select a percentage of events from each file (for proportionate subsampling across files), or an absolute number (for equal subsampling). Specifying a random seed will select the same events every time.
 Analysis Channels selects which channels to use for analysis. Typically these are all of your surface markers, but could include signaling markers and physical parameters (scatter for fluorescence cytometry, DNA and event length for mass cytometry).
 *Rescale channels equally makes each channel have equal weight during analysis. This should usually be left on because it prevents a channel from having too much or too little influence. Rescaling uses your scales' minima and maxima, so again, ensure that your scales are set correctly.
 Compensation selects which compensation matrix to apply.
 Pick which algorithm(s) to run and set their algorithmspecific parameters. See the sections below for information on each. CellEngine lets you select multiple algorithms to run in a pipeline, making it easy, for example, to display clustering results on a UMAP plot.
 Click run. The analysis will run in the background; you can leave the page or close the browser. To monitor the analysis status, click on the activity tracker in the topright corner.
 When the analysis is complete, a new experiment will be created. You can access it from your inbox or the activity tracker.
Which algorithms should I use?
This is largely a matter of personal preference. For dimensionality reduction, while UMAP is generally noted for its speed, the implementation of tSNE in CellEngine is about as fast as UMAP. Through tuning of each algorithm's settings, the results of both can be made to look very similar.
Dimensionality Reduction¶
Principal Component Analysis (PCA)¶
PCA is a linear transformation that preserves as much of the variation in the data as possible in two dimensions. It is deterministic (repeatable) and fast, but because it is a linear transformation, its utility can be limited.
tDistributed Stochastic Neighbor Embedding (tSNE)¶
tSNE is a nonlinear transformation conceived by Hinton and van der Maaten and first applied to cytometry data by Amir et al.
While the original tSNE algorithm is computationally limited to about 10,000 data points, CellEngine uses a fast approximation that can analyze up to 5,000,000 data points in about 12 minutes. If necessary, larger datasets can be subsampled to meet this cap.
To improve reproducibility between runs and enhance the statistical rigor of the algorithm, CellEngine uses the first two principal components instead of random noise to initialize the algorithm.
The following settings are available:
 Perplexity is a smooth measure of the effective number of neighbors, typically between 5 and 50. Lower values emphasize local effects in the data, while higher values emphasize global effects and are necessary for larger datasets.
 Number of nearest neighbors (k) adjusts how many neighboring data points are used to compute the perplexity matrix and has similar effects as perplexity. This should be at least three times the perplexity.
 Number of iterations controls the number of gradient descent (optimization) iterations. In each iteration, the position of each point is adjusted slightly.
 Number of exaggeration iterations controls how many of the gradient descent iterations use the early learning rate, early exaggeration and (internal) early momentum. The goal of this phase is to avoid local minima.
 Early and late learning rate affect the spacing of the points. A rate that is too high can cause the result to “ball up,” while a rate that is too low can cause events to not group together well. The learning rate can be specified separately for the early and late phases, but are usually best set to the same value.
 Early and late exaggeration adjust the attractive forces and thus how tight the groups are. Increase the final exaggeration slightly to make the groups smaller and more spread out, especially for larger datasets.
CellEngine suggests default values for each of these parameters in the algorithm settings page.
It's perfectly valid to run tSNE multiple times with the same or different settings and pick the result that looks the best.
Warning
tSNE results can be misleading and tricky to interpret: using different settings can produce dramatically different results, cluster sizes are meaningless, the distances between clusters might not mean anything, and shapes can appear in both random noise and in real data. See How to Use tSNE Effectively for excellent advice on this topic.
For more information:
 van der Maaten L., Hinton G. Visualizing data using tSNE at https://jmlr.org/papers/volume9/vandermaaten08a/vandermaaten08a.pdf  The “tdistributed” variant of SNE.
 Amir E.D., Davis K.L., Tadmor M.D., Simonds E.F., Levine J.H., Bendall S.C., Shenfeld D.K., Krishnaswamy S., Nolan G.P., Pe'er D. viSNE enables visualization of high dimensional singlecell data and reveals phenotypic heterogeneity of leukemia at https://www.nature.com/articles/nbt.2594  The first application of tSNE to cytometry data.
 Wattenberg M., Viégas F, Johnson I. How to Use tSNE Effectively at https://distill.pub/2016/misreadtsne/  Guidance on using and interpreting tSNE.
 Melville J. smallvis documentation at https://jlmelville.github.io/smallvis/  Thorough investigation of the various tSNE settings.
 Kobak D., Berens P. The art of using tSNE for singlecell transcriptomics at https://www.nature.com/articles/s4146701913056x  Includes guidelines for tuning tSNE settings that apply to both singlecell transcriptomics and cytometry.
Uniform Manifold Approximation and Projection (UMAP)¶
UMAP is a nonlinear transformation developed by McInnes, Jealy and Melville and is similar to tSNE.
CellEngine's implementation of UMAP can analyze up to 10,000,000 data points in about 20 minutes. If necessary, larger datasets can be subsampled to meet this cap.
The following settings are available:
 Number of nearest neighbors (k) adjusts how many neighboring data points are used to compute forces. Lower values emphasize local effects in the data, while higher values emphasize global effects and are necessary for larger datasets.
 Number of iterations controls the number of gradient descent (optimization) iterations. In each iteration, the position of each point is adjusted slightly.
 Minimum distance controls the spread of clusters. Smaller values result in denser clusters, while larger values result in more disperse clusters.
For more information:
 McInnes L., Healy J., Melville J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction at https://arxiv.org/abs/1802.03426
 Becht E., Dutertre C.A., Kwok I.W.H., Ng L.G., Ginhoux F., Newell E.W. Evaluation of UMAP as an alternative to tSNE for singlecell data at https://www.biorxiv.org/content/10.1101/298430v1.
Clustering and Community Detection¶
FlowSOM¶
FlowSOM is a clustering algorithm based on selforganizing maps (SOM) developed by Van Gassen et al. Briefly, the algorithm stages are:
 A SOM is trained. The SOM is a rectangular, nontoroidal grid. In each iteration, FlowSOM refines the map using a randomly selected event from the dataset: First it finds the nearest existing grid point, then it updates that grid point and neighboring grid points' positions in ndimensional space to incorporate the new event.
 The dataset is mapped to the SOM.
 Optionally, consensus clustering is run on the SOM. During this step, a random subset containing 90% of the SOM distance matrix is selected, then hierarchically clustered.
 Either the consensus matrix or the SOM itself is hierarchically clustered to yield "metaclusters."
CellEngine's implementation of FlowSOM can analyze up to 200,000,000 data points in about 1 minute.
The following settings are available:
 X and Y grid size adjust the adjust the number of points in the map. Datasets with more cell populations require larger grid sizes to separate them. x times y must be less than 10,000. Larger grids take more iterations to converge.

Training iterations controls how many iterations are used to train the map.
Note that CellEngine selects one cell per "iteration," whereas Van Gassen's implementation iterates through every cell in the dataset per "iteration." Consequently, the number of iterations in CellEngine is much higher, but also generally does not need to be adjusted according to the size of the dataset.

Start radius affects how many neighboring grid points are updated during the training phase. The end radius is always 0, meaning only the nearest grid point's position is updated without updating neighbors.
 Consensus clustering iterations sets how many times the map is subsetted and clustered. This can be set to 0 to skip consensus clustering and just use standard hierarchical clustering.
 Number of consensus clusters sets how far up the dendrogram is "cut" when observing coclustering. Note that this is not the final number of metaclusters. CellEngine allows you to dynamically select the number of metaclusters after FlowSOM runs, providing more flexibility. However, this number should be close to the number of metaclusters in your dataset.
The output of FlowSOM in CellEngine is a new experiment with FCS files that contain two additional columns: (1) the SOM cluster ID and (2) the Euclidean distance of the cell from its assigned cluster. Additionally, the results of the metaclustering are stored. On the gating page, you can select how many metaclusters to visualize without having to rerun FlowSOM by clicking create cluster gates in the lowerleft.
Key Differences from Van Gassen's Implementation
 CellEngine does not create a graph with a minimumspanning tree (MST) or the "star plots" shown in the original publication. Instead, the results can be viewed on any of CellEngine's other visualization types, including tSNE and UMAP plots, heatmaps and pivot tables.
For more information:
 Kohonen T. SelfOrganized Formation of Topologically Correct Feature Maps at https://doi.org/10.1007%2Fbf00337288
 Van Gassen S., Callebaut B., Van Helden M.J., Lambrecht B.N., Demeester P., Dhaene T., Saeys Y. FlowSOM: Using selforganizing maps for visualization and interpretation of cytometry data at https://onlinelibrary.wiley.com/doi/full/10.1002/cyto.a.22625
Interpreting Algorithm Results¶
There are many approaches to visualizing and interpreting results from dimensionality reduction algorithms, several of which are outlined here. These can be further combined and adjusted according to your experiment design.
Approach 1: Using manual gates and plots colored by markers¶
The first approach uses manual gating to define major cell populations and create a “key” plot to help navigate other plots. The other plots can be colored either by phenotyping markers to further subclassify those populations, or by signaling markers to determine what the populations are doing.
Creating the key plot¶
Howto
 Create a new illustration.
 Insert a pivot table using the button in the toolbar.
 Apply the following settings:
 Dimensions ⮞ Columns Populations: your populations of interest. Check the Overlay checkbox.
 Dimensions ⮞ Rows FCS Files: Generally a single file is sufficient to serve as a key when viewing other plots, but multiple files can be selected if desired (e.g. if some samples are biologically missing a population).
 General ⮞ Y Channel and General ⮞ X Channel UMAP 1 and UMAP 2, tSNE 1 and tSNE 2 or PCA 1 and PCA 2.
 Axes and Legend ⮞ Tick Marks Optionally deselect to tidy up the plot; they are meaningless.
Tip
The order in which the populations are overlaid can be adjusted by following the instructions in annotation, population and channel sorting.
This plot can now serve as a reference to help navigate other plots.
Subclassifying populations using phenotyping markers¶
Here we can see that CD3+ cells are in the topcenter and lowerleft; CD4+ cells are in the topcenter and CD8+ cells are in the lowerleft; and CD45RA+ cells form a continuum within each of those, as well as within the B cell population on the left. This agrees well with the key plot, above.
Howto
 Start from the illustration created earlier.
 Insert another pivot table using the button in the toolbar.
 Apply the following settings:
 Dimensions ⮞ Columns Color Channels: your phenotyping channels of interest. In the example above, CD3, CD4, CD8 and CD45RA are selected.
 Dimensions ⮞ Rows An annotation of your choice, sufficient to select the files you want to view. In the example above, it's our “condition” annotation with “unstim” selected.
 General ⮞ Y Channel and General ⮞ X Channel UMAP 1 and UMAP 2, tSNE 1 and tSNE 2 or PCA 1 and PCA 2.
 Axes and Legend ⮞ Tick Marks Optionally deselect to tidy up the plot; they are meaningless.
 If you have replicate samples, add additional pivot table levels by clicking the add table button, use Filter Annotations to narrow down the displayed files, or enable batched mode to create a multipage illustration.
Tip
The order in which the pivot table rows and columns appear can be adjusted by following the instructions in annotation, population and channel sorting.
This same approach can be used to gate the “blobs” in the gating page:
Howto
 Open the gating page.
 Set the X and Y channels to UMAP 1 and UMAP 2, tSNE 1 and tSNE 2 or PCA 1 and PCA 2.
 Set the plot type in the left panel to “Dot.”
 Select a color channel in the left panel.
 Draw gates as desired, according to the plot colors.
Inspecting cell activity using signaling markers¶
Although subtle, in the above example the CD33+/HLADR+ cells in the lowerright corner of the pSTAT5 plot are a brighter color in the IL2/GMCSF sample, indicating increased STAT5 phosphorylation. (The cells in red are potentially debris.)
Howto
 Start from the illustration created earlier.
 Insert another pivot table using the button in the toolbar.
 Apply the following settings:
 Dimensions ⮞ Columns Color Channels: your signaling channels of interest. In the example above, pSTAT3 and pSTAT5 are selected.
 Dimensions ⮞ Rows An annotation of your choice, across which you want to compare signaling behavior. In the example above, it's our “condition” annotation with “unstim” and “IL2/GMCSF” selected.
 General ⮞ Y Channel and General ⮞ X Channel UMAP 1 and UMAP 2, tSNE 1 and tSNE 2 or PCA 1 and PCA 2.
 Axes and Legend ⮞ Tick Marks Optionally deselect to tidy up the plot; they are meaningless.
 If you have replicate samples, add additional pivot table levels by clicking the add table button, use Filter Annotations to narrow down the displayed files, or enable batched mode to create a multipage illustration.
Tip
The order in which the pivot table rows and columns appear can be adjusted by following the instructions in annotation, population and channel sorting.
Approach 2: Using box plots to compare blobs¶
This approach involves drawing polygon gates around blobs, and then using a box plot to compare the phenotyping or signaling markers between blobs.
In this example, the four plots at the top show the UMAP blobs for each of four replicate samples. We previously drew 10 gates around the blobs, named P1 through P8. Below that is a box plot showing the median of several phenotyping markers for all events (“Ungated”) and for a selected blob (P8). We can use this to see that P8 has relatively high CD3 and CD4 staining, indicating that it represents CD4+ T cells.
Howto
 Draw polygon gates around blobs in the gating page.
 Create a new illustration.
 Insert a pivot table using the button in the toolbar.
 Apply the following settings:
 Dimensions ⮞ Columns An annotation of your choice, sufficient to select the files you want to view. In the example above, it's our “donor” annotation with “LRS046,” “LRS047,” “LRS048” and “LRS049” selected.
 Dimensions ⮞ Rows Populations: Ungated, or whatever population you drew the blob gates on.
 General ⮞ Y Channel and General ⮞ X Channel UMAP 1 and UMAP 2, tSNE 1 and tSNE 2 or PCA 1 and PCA 2.
 Gates ⮞ Gate Labels Set to Name.
 Axes and Legend ⮞ Tick Marks Optionally deselect to tidy up the plot; they are meaningless.
 Insert a box plot using the button in the toolbar.
 Apply the following settings:
 Dimensions ⮞ Axis Labels Channels, and your channels of interest.
 Dimensions ⮞ Legend Entries Populations: Ungated and one or more of your gated blob populations.
 General ⮞ Y Channel and General ⮞ X Channel UMAP 1 and UMAP 2, tSNE 1 and tSNE 2 or PCA 1 and PCA 2.
 Scaling and Range Optionally set Scaling to Log10 or Scaled and/or adjust the Range Min and Range Max to improve visibility of data with a large dynamic range.
Approach 3: Comparing gross differences in plot structure¶
The examples in the approaches above show biological replicates, which are expected to have highly similar embeddings (blob organization or plot structure). However, in some experiments, the embeddings will not be as similar. For example, samples from diseased patients may have more or fewer populations than samples from healthy controls owing to the immune response or presence of cancer cells. In these scenarios, comparing gross differences in the plot structure can be useful.
These comparisons can be performed by inserting pivot tables as described in the first half of Approach 2.
Warning
As noted above, use caution when trying to interpret plot structure: using different algorithm settings can produce dramatically different results, cluster sizes are meaningless, the distances between clusters might not mean anything, and shapes can appear in both random noise and in real data. See How to Use tSNE Effectively for excellent advice on this topic. The same advice generally applies to UMAP. Instead, generally focus on the absence vs. presence of clusters between samples.