scSVA accepts tabular data stored in Comma-Separated Value (CSV) or in Hierarchical Data Format (HDF5) files. The input data consists of (i) X- and Y-coordinates of your favorite embedding (e.g. tSNE or FLE), (ii) gene expression matrix or vectors that contain a gene quantification for each cell, and (iii) metadata information on each cell's group assignment . The file loading depends on a file type and is described separately for each file type. The first thing to do is to open a File tab panel. This panel allows for loading all necessary files to scSVA.
First, select the "Input File Format" and then click on "File Select". This will open a new dialog box with the access to the local file system. Select the file that contains XY coordinates for your plot and click Select button in the dialog box. The output depends on the selected file type:
- HDF5 - A new table opens below the panel. An HDF5 file includes two types of objects: datasets and groups that store datasets and other groups. The table lists both groups and datasets with detailed information like dataset/group name, storage class, dimension and size of datasets. Select the datasets with X and Y coordinates. scSVA supports HDF5-based Loom files. The filename extension must be from the following set [.h5,.hdf5,.h5ad,.loom] or from the same set but with all letters capitalized. H5ad file type can be used if the feature matrix is not in a sparse format.
- CSV - When you load a CSV file you make sure that the file has a header with column names. The column names (feature names) will appear in the new table below the panel. scSVA uses "fread" from data.table package which reads "regular delimited files" i.e. each row in a file has the same number of columns. The field delimiter is recognized automatically and users should use one from the following set [\t ,;:|]. The above field delimiters should not be used in column and row names. scSVA accepts both raw and compressed files (compressed with gzip) and the file type is recognized automatically. The file name extension must be from the following list ['.csv','.tsv','.txt'], or with all letters capitalized or compressed equivalents (.gz).
-
HDF5 - scSVA accepts two dimensional datasets when a "Dataset" is selected or vectors of gene expression values when a "Group" is selected. If the "Dataset" is selected, specify if the rows or columns of the dataset are genes. If the "Group" is selected, scSVA will allow for loading expression vectors (one-dimensional datasets) within a group, with gene symbols as dataset's names. This option is appropriate for very big datasets.
-
CSV - We support only gene expression matrices, where the rows are genes and the columns are cells, and compressed with Gzip. This layout allows for indexing and very fast retrieval of expression vectors. The first column must contain gene names with "Genes" as the column name. The first row contains Cell IDs. Below is an example of an expression matrix:
Genes Cell.1 Cell.2 Cell.3 ... Gene.1 0 1 0 ... Gene.2 2 1 3 ... Gene.3 1 0 0 ... ... ... ... ... ... The file saved as a CSV file with a comma as a delimiter:
Genes,Cell.1,Cell.2,Cell.3,...
Gene.1,0,1,0,...
Gene.2,2,0,3,...
Gene.3,0,1,0,...
...
To compress the expression matrix file use the following command in the terminal:
gzip expression_matrix.csv
The next step is to upload a list with gene names that matches the gene names in the expression matrix. scSVA makes queries on the compressed expression matrix based on that list. The list can be saved at the preprocessing step or extracted from the uncompressed expression matrix file by the command: \cat expression_matrix.csv |cut -d',' -f1 | awk '{if(NR>1) print} >Gene_Names.txt'
or from the compressed file: \gunzip -cd expression_matrix.csv.gz |cut -d',' -f1 | awk '{if(NR>1) print} >Gene_Names.txt'
The index can be created on the compressed file by users using Zindex in the terminal : \zindex expression_matrix.csv.gz --field 1 --tab-delimiter --skip-first 1
for tab-delimited files or \zindex expression_matrix.csv.gz --field 1 --delimiter , --skip-first 1
for e.g. comma-delimited files. scSVA can create the index for you if a tick box "Create Index on Compressed File" is checked. Make sure that Zindex is properly installed on your computer and is visible as an executable file by your operating system. Indexing allows for very efficient retrieval of the rows from compressed files without need for loading of the entire matrix into the memory. Thus, the memory usage is reduced by a factor of 10,000-40,000 for a typical expression matrix which makes possible a fast visualization of gene expression values on a 2D embedding or computing and plotting scores of a custom defined gene signature (set of genes).
Selection of a vector of group IDs is identical to the selection of XY coordinates. Group IDs must be a vector of integers (categorical data).
The tab "Visualize" helps customizing the plot appearance. scSVA support visualization of 2D and 3D cell embeddings.
The default plot is a density plot, where for a given grid resolution,
it shows the number of cells within a square. The grid resolution can be
modified by a scroll bar below the plot. The default resolution is 512
which means that a plot consists of 512x512=262144 square cells. Users
can increase the resolution of the grid to produce higher quality
figures but this results in increasing the plot refreshing time as more
points needs to be plotted. The tradeoff between the figure quality and
the interactive plot responsiveness needs to be exploited by users as it
depends on the number of cells and machine-specific factors. We found
that the default value of 512 is sufficient to produce high quality
figures and the response is fast enough to interactively plot hundreds
of millions of cells.
The two sidebars in the "2D" tab simplify customization of the plot.
Users can modify 1) size of the plotted points, 2) size and type of the
font (x,y-axis labels, title, color bar), 3) transparency of the points
by specifying the transparency value, 4) color palette of the colorbar,
5) colors of the panel and plot background. We included also a
possibility of checking the colorschemes in terms of suitability for
color blind readers (dichromat
package).
Clicking on "Dichromatism" box will open options for various types of
colorblindness (green, red, blue) and users can simulate the effects of
the type of dichromatism which helps selecting a color palette suitable
for people with different types of color perception deficiency. We
provide also color palettes suitable for dichromats from dichromat
package.
These color palette names begin with "dichromat::". The second sidebar
modifies 1) the main title of the plot, 2) x- and y-axis label, 3) color
bar title. Finally, the main plot can be saved in many file formats
(vector or raster) with a custom resolution and size.
Mouse left-click starts a selection of a rectangular region. Zooming in: double-click inside the selected area will zoom in the selected region with the specified resolution ("Grid Resolution" scrollbar). "Reset View" button below the main plot when clicked will restore the original plot xy range with the specified resolution.
To initialize 3D plot, select X, Y, and Z coordinates and load datasets (see File section). Functions for 3D plot customization are similar to those present in 2D plots except the Camera Viewpoint which determines the camera view point about origin (see manual). A panel "Display Feature" changes the feature to be plotted. Users can visualize a density of cells or gene/signature expression values. 2D projection button opens a new window with 3D data projection onto 2D.
Analysis Menu consists of tools for exploratory analysis of uploaded datasets.
Here users can explore number of cells in selected areas, plot gene
expression profiles on 2D embedding, and compute gene expression
distributions and statistics like mean, median, standard deviation in
the selected areas.
Users can use two selection tools: (i) rectangular selection (default),
and (ii) polygonal selection. Mouse left-click on the plot and holding
down the left button lets you create a rectangular selection. To start
polygonal selection check "Polygonal Selection" box. From now, each
mouse left-click on the plot will create a new polygon's vertex and
connect it by edges with the first and the penultimate vertex.
To compute number of cells, select a region of interest and click
"Group A" button. This will mark cells in a selected area and return
number of cells in that area. Users can select two groups of cells.
"Reset Group" button will unselect cells. To plot gene expression values
on 2D or 3D embedding, select a gene in right-hand panel. By default,
the mean expression values are plotted on the grid but users can change
the statistic by selecting "Select Statistic". Three options are
supported: mean, median (the slowest), and maximal expression value. The
mean is suitable for highly expressed genes and the maximum value option
works well for lowly expressed genes (e.g. transcription factors or
surface receptors in scRNA-Seq). For a chosen gene, users can select two
groups of cells as described above and compute the gene expression
statistics in these two groups by clicking on "Get Group Statistic"
button.
Checking box "Compute Gene Expression by Group" starts computation of statistics in groups like number of cells, gene expression mean, median and standard deviation. Once completed, the statistics are listed in the table. The table is fully interactive: 1) groups can be sorted by a mean or number of cells for example, 2) searched by statistic values or group IDs, 3) filtered by statistic values by clicking on boxes below the table. Single-click on a row of the table (group) will show a gene expression distribution for the selected group. Users can mark many groups to compare the expression distributions among them. Each selection opens a new distribution plot on right-hand side. The plots are interactive and colored by the same colors as groups on the embedding. The scroll bar, on the bottom of the plots, changes the y-axis range for all plots. Moving a cursor to the plot bar will show fraction of cells falling in a given expression range. To compare all the distributions on one plot, go to "ECDF" tab to open a plot with empirical cumulative distribution function (users can easily read dropout levels from the ecdf plot or check for bi- or multimodalities that may suggest underclustering) or "Violin" to visualize group expression values by violin plot. "Census" tab contains the fractions of cells in each group.
In order to compute gene signature scores (e.g. from Molecular Signature Database MSigDB) for all cells and visualize them on 2D or 3D embedding, users need to provide a set of gene names to "Enter Gene List" field and specify "Gene Signature Name". The gene names in the list must be in the same format as in gene expression matrix (e.g. gene symbols). The gene expression values across all cells can be retrieved in parallel from an expression matrix by specifying the number of threads on local machine. Click on "Compute Gene Signature Score" box to start retrieving and computing the scores. scSVA computes gene expression z-score across all cells and average over the z-scores across the entire gene set.
scSVA supports interactive analytics and data storage management on the
cloud (Google Cloud Platform). Google Cloud Platform charges you for
storage and running VMs, see
pricing. Stop or delete any GCP
services once you are done with your analysis. Good practice is to
regularly check the billing of your project. scSVA is an R package
distributed as-is, without warranties of any kind. Make sure that you
use the GCP services reasonably and follow best practices. To get
started make sure that your system configuration is ready to use GCP and
in particular to launch virtual machine instances in Google Cloud
Compute Engine (GCE). Follow the instructions below to install and
configure all required tools:
(i) Create and configure a Google Cloud Project, see
instructions.
(ii) Download GCE private key in JSON format, see
instructions.
(iii) Download and install Google Cloud
SDK.
(iv) Initialize and authorize Google Cloud SDK, see
instructions.
See tutorial
for more information. scSVA uses
googleComputeEngineR,
which provides R interface to the GCE API.
To list google buckets from the default project, choose "gs" and click "Explore FS". Select a bucket from the list and click right arrow below the list to show its content. The upper arrow below the list navigates up one directory. To remove the object from the bucket, select it and click "Remove Selected Object". Users can also create new buckets by specifying the name of the bucket and clicking "Create New Bucket" button. Selection of "local" filesystem will list directories and files on your local computer. Users can easily copy objects between the local computer and the bucket by clicking left/right arrow between filesystem panels.
scSVA uses googleComputeEngineR to lunch and operate VMs. In the left corner, there is a link to Google Cloud Console. Clicking it opens the Console in new browser tab. Users can monitor there all services provided by GCP. To get started, go to "GCE Setup" and provide the path to GCE private key in JSON format (see step (ii)), your project name (see step (i)), and the zone you want to launch VM (see instructions). After successful configuration, "GCE List Disks" in "GCE Info" allows to display all disks in the project, and "GCE List Instances" lists all VMs (stopped and running) in the project. Right console displays all the information from GCE. To launch a new VM, or restart a stopped VM, go to "RUN VM" window, and click "GCE List Machine Types". This will list all available machines in your selected zone. Select container with computational tools you want to run on VM, and provide name for your virtual machine. Users need to provide the full name e.g. "mtabaka/scsvatools" to pull scSVAtools container from Docker Hub. Alternatively, pull an image in the VM instance terminal, for example by the command "docker pull mtabaka/scsvatools". To build a new image on Google Cloud, go to terminal, then to directory with a Dockerfile and run the following command: \
gcloud container builds submit --timeout=2h --tag gcr.io/Project_Name/ContainerName
For further information, see instructions to
googleComputeEngineR
package. Usually it takes less than one minute to lunch VM. You can
check the progress in Rstudio console. If the VM instance is
successfully launched you should see its name in the list of VM
instances with status "RUNNING" after clicking the "GCE List Instances"
button. "Get Running VMs" button displays all VM instances with status
"RUNNING". Specify the VM on which you want to run jobs and click
"Assign VM". Notice the change in the name of "Selected VM" above the
right console. From now, you can run commands and containerized tools
and copy files between local machine and VM instance. You can also stop
or delete the VM instance and connect to VM instance through ssh. If you
stop the VM instance, you are charged only for storage, and the
restarted VM instance will have access to all files from previous
session. If you delete VM instance, it will permanently remove a disk
and VM instance (you will not be able to restart it in the future). In
order to list running processes on the VM instance click "Run Top".
Users can easily transfer files from/to the VM instance using the panel
below. Selected files from the list can be modified after clicking
"Edit" button. Users can also create new files on the VM instances. If
your container has tools installed that can run bash scripts selecting a
script on VM and clicking "Run Script" will start running the script on
the VM instance. Make sure that the path to the files used in the script
is correctly specified. The directory mounted by docker is determined
from the script path. For example, if the full path to the script is
/home/user/analysis/script.sh, all files and subdirectories in
/home/user/analysis/ will be mounted in /home/ directory of the running
container. If the input file used in script is in
/home/user/analysis/input/input.csv, it will be available at
/home/input/input.csv in the running container. For that reason, all
input files need to be in the same directory as the script or in its
subdirectories. If multiple containers are available on your VM
instance, select the right one for running the script from the list
"Container to Run". If you have many running VM instances you can switch
between them by selecting and assigning the running instance. We added
also the option of running commands on the VM instance. Write the
command in the test field below the console (as you would do it in the
bash terminal) and click "Run SSH Command". The output of the command is
displayed in the console. To remove all text from the console, click
"Clear" button. As a proof of concept, we implemented a method for the
efficient computation of a 3D single-cell data visualization. The
generation of 3D visualization consists of 3 steps: 1) computation of
diffusion maps; 2) creation of a nearest neighbor graph in diffusion
component space; 3) application of 3D force-directed layout algorithm
(ForceAtlas2) to the nearest neighbor graph. The code is on GitHub in
scSVAtools
repository. scSVA supports generation of 3D visualizations only on GCP
(we plan on extending it to local machines and other cloud platforms).
To get started, transfer a matrix in txt (gzipped) file format where samples are rows and columns are features to VM instance, as described above. Provide the full path to the matrix on VM instance and specify parameters to run diffusion maps:
| Label | Description |
|---|---|
| Path to Matrix | Full path to the input matrix (columns - features, rows - samples) in txt (gzipped) file format |
| Number of NNs | Number of nearest neighbors |
| Number of Diffusion Components | Number of diffusion components to compute |
| Eigendecomposition Package | Package to use for performing eigen decomposition (Irlba, ARPACK) |
| Number of NNs (Local Sigma) | Number of nearest neighbors to use in computing local scaling parameter |
| Number of Threads | Number of threads to use |
| ANN Method | Approximate Nearest Neighbor method to use (Annoy or Nmslib) |
| Number of Trees | Number of trees to generate (if ANNs method == Annoy) |
| M,efC,efS | Parameters to run Nearest Neighbors (if Anns method == Nmslib) |
and click "Run DMaps".
Once the computation of diffusion maps is finished, select parameters for approximate nearest neighbor computation, and click "Run NN Graph". If the path is not specified, diffusion maps generated in the previous step will be used by default.
| Label | Description |
|---|---|
| Path to Matrix | Full path to the input matrix (columns - features, rows - samples) in txt (gzipped) file format |
| Number of NNs | Number of nearest neighbors |
| Number of Threads | Number of threads to use |
| ANN Method | Approximate Nearest Neighbor method to use (Annoy or Nmslib) |
| Number of Trees | Number of trees to generate (if ANNs method == Annoy) |
| M,efC,efS | Parameters to run Nearest Neighbors (if Anns method == Nmslib) |
The last step is to apply 3D force-directed layout (ForceAtlas2) to the nearest neighbor graph. If the path is not specified, the graph from the previous step will be used. Specify parameters:
| Label | Description |
|---|---|
| Path to Graph | Full path to the input graph (as an adjacency list i.e. each ith row should contain Node_i and a list of Nodes connected to Node_i) in txt (gzipped) file format |
| Number of Iterations | Number of iterations |
| Number of Threads | Number of threads |
| Memmory (Gb) | Memory of java virtual machine |
| Scaling Ratio | Scaling parameter, ratio of repulsive to attractive forces. Higher values will result in larger graphs |
| BarnesHut Theta | Theta of the Barnes Hut approximation. The higher theta, the lower accuracy and faster computations |
| BarnesHut Update Every nth Iteration | Update the tree every nth iteration |
| BarnesHut Splits | Split the tree construction at its first level. Number of threads used is 8 to the power barnesHutSplits: 1 - 8 processes, 2 - 64 processes |
| Update Centers | Update Barnes-Hut region centers when not rebuilding Barnes-Hut tree |
| Restart | If TRUE, the simulations will start from the last saved configuration |
and click "RUN FLE" to start the simulation. Once the graph is loaded and processed, users can click "Get Distances" to check the changes in the total distance moved at every iteration. The simulation should be run until a steady state is reached. Progress (percentage of iterations done) can be checked by clicking "Check Progress". Once the simulations are done, click "Load FLE Coordinates" to visualize the results of simulations in the "Visualize->3D" tab.
Here, users can visualize and annotate cell groups on 2D embedding by
selecting color palettes, modify colors of individual groups, show and
modify cluster labels.
The middle panel contains a table with coordinates of group labels and
right-hand panel customizes the plot. First check "Show Groups". This
will compute group positions (cell groups for example) on a grid. "Get
Labels" check box, when selected, computes the position of labels by
taking average x- and y-coordinates for each group. The table in the
middle panel lists all groups (group IDs, x-, and y-coordinates, and
group labels). Users can modify the coordinates of labels plotted on 2D
and group names by double-click on the value or label in the table. The
plot is updated after any modification. "Get Labels" check box opens a
panel in which users can upload group names from a file. The size
(number of rows in case of csv file or number of elements in hdf5
dataset) must be equal to the total number of groups. "Repel Labels"
check box distribute labels in such a way that the labels don't overlap
(ggrepel
package),
which helps distributing the labels in such a way that they don't
overlap. The labels can be further customized by removing the bounding
box or changing font type, size, color, etc. in the "Annotation" tab.
To show a legend with group annotations check "Show Legend". Users can
modify the legend title, position, number of columns, order of groups,
and displayed point size in the panel that opens after checking the box
"Show Legend" Colors of cell groups on 2D embedding can be modified in
"Color Palette" panel. We included support for most of Crayola crayon
colors
which generate distinctive and aesthetically-pleasing color sets to mark
groups of cells. Clicking on Shuffle Colors will permute colors from a
given set. Make sure that none of the groups has a color which is close
to the background color. "Shuffle Cells" will permute the cells on the
grid (important if some groups overlap for example). The individual
group colors can be changed in two ways: (i) checking "Show color List"
will open a table with group colors. Double-click on the column element
will allow for color modification for the selected group. Users can
provide a color in hex format or by giving a color name e.g. "navyblue";
(ii) Specify the "Group ID" whose color you wish to modify (the group
can be highlighted by checking the "Highlight Group" box). Clicking
"Change Group Color" field will open a color picker. Click on a color
and the color of the cell group will change immediately.
Here, users can arrange multiple graphs on a grid. To start click on "Add Plot" button. This will add the main plot to a list of plots. Check the box "Show Multiplots" to make it visible. To add more graphs just click "Add Plot". To change the size of added plot or its resolution go to "Visualize", "2D" tab and modify the plot properties on the right sidebar. scSVA uses R bindings to ImageMagick to arrange graphs on a grid. Multiplots can be zoomed in/out by changing the "Scale" scroll bar, "Number of Columns" specify number of graphs per row.
To start annotating the main plot check the "Annotate" box. From now, any single mouse left-click on the main plot will create a new label. The coordinates and text labels appear in the table in the middle panel. The table is interactive and double click on its element allow for modification of this element. To remove an annotation label, specify the "Annotation ID" you wish to remove and click on "Remove Label". The labels can be further customized by specifying the background color of the label, its transparency, size, and font type of the text. The bounding box can be modified by changing the line width of its border or the radius of corners.
Users can customize their plots by creating their own color palettes. Click on "Select Colors For a New Palette". This will open a color picker widget that can be used to select colors for a custom palette (the colors can be added to a palette also by specifying the hex color code). Specifying the name of the color palette and clicking on "Add New Palette" will make it available in both "Color Palette" selection fields in "Visualize" and "Metadata" -> "Groups" tabs. The new custom color palette will appear at the end of the list of color palettes.
To further customize plots, users can upload their own or open source fonts. scSVA accepts TrueType fonts (filename extension .ttf). This functionality is available on Mac or Linux operating systems. Windows users may consider installing scSVA as a docker container. The default installation path of fonts on linux is /usr/share/fonts, users will be asked to provide a password in the terminal to write the new fonts to this directory. In some cases, to see new fonts on the interactive or saved figure, it may be necessary to restart the current R session. For example, write in RStudio console: .rs.restartR() or use Command/Ctrl + Shift + F10 to restart the R session. We encourage users to import new fonts at the first step and restart R session before loading files with x-, and y- coordinates.
Visualization and exploration of the Immune Cell Atlas dataset of bone marrow immune cells https://preview.data.humancellatlas.org processed using example scripts.
- 2D visualization of 1B cells (upsampled dataset)
- File upload
- 2D plots
- 3D plots, 100M cells
- Basic exploratory analysis: a, b
- Gene Signatures
- Metadata
- Annotation
- Categorical data exploration
- Multiplots
- Cloud
- Creating custom color palletes
- Adding fonts
Created by Marcin Tabaka
Contact me at [email protected]