Over the years, I've found useful packages for analyzing single-cell data. As a result, there are many different approaches to handle problems one might have when analyzing these data. The Seurat package has become a popular choice for researchers to perform end to end analysis. However, many tools may not be ready to accept Seurat objects out of the box. I designed this package to keep track of the different tools I have used and found useful, and to allow users to easily apply these tools to data stored in Seurat objects.
R packages from CRAN that are required to run scutils
install.packages(c('dplyr', 'Seurat', 'data.table', 'reticulate', 'SoupX', 'DUBStepR', 'openxlsx', 'HGNChelper'))
The sceasy package is required to run scVI
devtools::install_github('cellgeni/sceasy')
This package can be installed using the following code in R. Please note that some functions will require external dependencies to be downloaded and installed.
devtools::install_github('shahrozeabbas/scutils')
Many machine learning algorithms rely on a set of variable or informative features to create models. This becomes especially necessary for single-cell data which are sparse. Many genes do not provide enough information to be used for downstream clustering. Seurat calculates a set a variable features for this purpose, which can sometimes retain mitochondrial or ribosomal genes. Depending on what biological question is, this may inform downstream clustering in a negative way. The CleanVarGenes function can be used to remove these genes from the variable feature list.
Ranjan et al. created a novel way to calculate variable features for single-cell data. The authors use pairwise gene-gene correlations to produce a list of genes that are robust across different conditions for the same cell types. It can be a useful tool in your arsenal and can be run using the dupstepR function. For more information about the method please refer to their paper. To install DUBstepR, please use CRAN.
K-means clustering is a classic machine learning approach to group data into similar categories or clusters. The FindClustersKmeans function in this package is applied typically to PCA dimensions. There are many resources available online to learn more about k-means, but you can start here.
Droplet-based Single-cell experiments often capture mRNA that is present in the assay solution and encapsulated along with a given cell. Before sequencing, these cells are lysed and 'cell-free' RNA is added to the mixture, contaminating the downstream expression counts. The SoupX method attempts to correct this problem and the SoupCorrect wrapper function included in this package helps to facilitate this process. Please read more about this awesome package here.
Batch effect is a major issue when combining multiple datasets in single-cell analyses. There are currently many packages for addressing this problem along with ways to assess how well the integration performed. Lakkis et al. created the CarDEC method for this purpose. The authors also define a metric known as the CV Score to assess how well their method and others are able to correct for batch effect. This package implements the metric in an easy-to-use CVScore function that can be included in any single-cell analysis.
This is another method that is part of a toolkit and stands for Single-Cell Variational Inference and is a personal favorite. scVI uses probabilistic models to integrate and remove batch effect. Similar to Seurat, the toolkit provides methods for many different types of data such as scATAC-seq and spatial sequencing data. The function in this package is written to accept a Seurat object and allows the user to supply batch labels present in the dataset. It's extremely fast and can be used to build atlases with reasonable runtimes. The website linked above contains tutorials for all their methods. The scVI package must be installed seperately and can be run using the RunSCVI function.
One of the many great challenge of scRNA-seq is clustering then annotating for cell types. Most of the time, it requires meticulous curation of markers from literature that are compared to calculated markers from your pipeline before a final annotation is decided upon. This process can be confusing to say the least. The ScType algorithm is fast, flexible, and lightweight and has been used widely for it's annotating capabilities. Chances are that you will still have to manually go through your data to confirm annotations, but the AnnotateCellTypes wrapper makes the job a bit easier as it requires only a Seurat object as input. You can read more about the algorithm here.