Cell Bender Functionality & Plotting

Compiled: February 28, 2024

Source: vignettes/articles/Cell_Bender_Functions.Rmd

CellBender Functionality

CellBender is software for elimination of technical artifacts in scRNA-seq/snRNA-seq that uses deep generative model for unsupervised removal of ambient RNA and chimeric PCR artifacts. You can find out more info about CellBender from the bioRxiv Preprint, GitHub Repo, and Documentation.

Following completion of CellBender scCustomize contains a couple of functions that may be helpful when creating and visualizing data in Seurat.

Importing Cell Bender H5 Outputs

The output from CellBender is an H5 file that is styled and can be read like 10X Genomics H5 file. In CellBender pre-v3 this file could be read by Seurat::Read10X_h5() (although that function assumes that the file contains no name prefix). However, v3+ contains extra information that causes Read10X_h5 to fail.

scCustomize contains function Read_CellBender_h5_Mat() can be used when reading in CellBender h5 files regardless of the version of CellBender used.

cell_bender_mat <- Read_CellBender_h5_Mat(file_name = "PATH/SampleA_out_filtered.h5")

Importing CellBender data based on STARsolo or pre-V3 Cell Ranger inputs

If the input that CellBender uses is based on STARsolo or pre-V3 Cell Ranger data then some of the slot name which stores feature/gene ids in the H5 file is different. These files can be read by specifying the optional feature_slot_name parameter.

cell_bender_starsolo_mat <- Read_CellBender_h5_Mat(file_name = "PATH/SampleA_out_filtered.h5", feature_slot_name = "genes")

Importing CellBender data with non-standard group names

If CellBender H5 file contains non-standard H5 group names then Read_CellBender_h5_Mat will error. To circumvent this simply supply the name of the H5 group that contains the count data.

cell_bender_name_mat <- Read_CellBender_h5_Mat(file_name = "PATH/SampleA_out_filtered.h5", h5_group_name = "background_removed")

Reading multiple CellBender files with single function

scCustomize also contains two wrapper functions to easily read multiple CellBender files stored either in single directory or in multiple sub-directories Read_CellBender_h5_Multi_File and Read_CellBender_h5_Multi_Directory.

Read in H5 outputs from multiple sub-directories with Read_CellBender_h5_Multi_Directory().

Cell Bender Output Structure

The following is typical output directory structure for Cell Bender files with sub-directory labeled with sample name and each file also prefixed with the sample name.

Parent_Directory
├── sample_01
│   └── sample_01_out_cell_barcodes.csv
│   └── sample_01_out_filtered.h5
│   └── sample_01_out.h5
│   └── sample_01_out.log
│   └── sample_01_out.pdf
└── sample_02
│   └── sample_02_out_cell_barcodes.csv
│   └── sample_02_out_filtered.h5
│   └── sample_02_out.h5
│   └── sample_02_out.log
│   └── sample_02_out.pdf
Read Files

All we have to do is adjust the parameters to account for cell bender file names and directory structure.

  • secondary_path In this case we can leave as NULL because samples are located in immediate subdirectory (secondary path must be the same for all samples).
  • custom_name = "_out.h5" This specifies what the common file suffix of all files are.

Optional Parameters
* parallel and num_cores to use multiple core processing. * sample_list By default Read_CellBender_h5_Multi_Directory will read in all sub-directories present in parent directory. However a subset can be specified by passing a vector of sample directory names. * sample_names As with other functions by default Read_CellBender_h5_Multi_Directory will use the sub-directory names within parent directory to name the output list entries. Alternate names for the list entries can be provided here if desired. These names will also be used to add cell prefixes if merge = TRUE (see below).
* merge logical (default FALSE). Whether to combine all samples into single sparse matrix and using sample_names to provide sample prefixes.

cell_bender_merged <- Read_CellBender_h5_Multi_Directory(base_path = "assets/Cell_Bender_Example/",
    custom_name = "_out.h5", sample_names = c("WT1", "WT2"), merge = TRUE)

Read in H5 outputs from single directory with Read_CellBender_h5_Multi_File().

If all output files are in single directory you can use Read_CellBender_h5_Multi_File to read in all of the files with single function.

Parent_Directory
├── CellBender_Outputs
│   └── sample_01_out.h5
│   └── sample_02_out.h5
cell_bender_merged <- Read_CellBender_h5_Multi_File(data_dir = "assets/Cell_Bender_Example/", custom_name = "_out.h5",
    sample_names = c("WT1", "WT2"))

Create Seurat Object with Corrected Counts

Creating a Seurat object from merged CellBender matrices then is identical to creating any other Seurat object.

cell_bender_seurat <- CreateSeuratObject(counts = cell_bender_merged, names.field = 1, names.delim = "_")

Creating Dual Assay Objects

Sometimes it can be helpful to create object that contains both the cell ranger values and cell bender values (we’ll come to why below). scCustomize contains a helper function Create_CellBender_Merged_Seurat() to handle object creation in one quick step.

For this function we assume that we will use the cell calling algorithm of Cell Ranger with the modified counts for Cell Bender.

Read in both sets of data 

cell_bender_merged <- Read_CellBender_h5_Multi_Directory(base_path = "assets/Cell_Bender_Cell_Ranger_Data/Cell_Bender_Example/",
    custom_name = "_out.h5", sample_names = c("WT1", "WT2"), merge = TRUE)

cell_ranger_merged <- Read10X_h5_Multi_Directory(base_path = "assets/Cell_Bender_Cell_Ranger_Data/Cell_Ranger_Example/",
    default_10X_path = FALSE, h5_filename = "filtered_feature_bc_matrix.h5", merge = TRUE, sample_names = c("WT1",
        "WT2"), parallel = TRUE, num_cores = 2)

Create Dual Assay Seurat Object

To run the function the user simply needs to provide the names of the two matrices and a name for assay containing the Cell Ranger counts (by default this is named “RAW”).

dual_seurat <- Create_CellBender_Merged_Seurat(raw_cell_bender_matrix = cell_bender_merged, raw_counts_matrix = cell_ranger_merged,
    raw_assay_name = "RAW")

Optional Parameters

Users can specify any additional parameters normally passed to Seurat::CreateSeuratObject() when using this function.

dual_seurat <- Create_CellBender_Merged_Seurat(raw_cell_bender_matrix = cell_bender_merged, raw_counts_matrix = cell_ranger_merged,
    raw_assay_name = "RAW", min_cells = 5, min_features = 200)

Pre/Post Cell Bender Analysis

It can be very important with tools like Cell Bender to analyze how much the process has effected data on a per cell basis.

Add Pre/Post to Meta Data

scCustomize includes function Add_CellBender_Diff() to help with this process. This function will take the nCount and nFeature statistics from both assays in the object and calculate the difference and return 2 new columns (“nCount_Diff” and “nFeature_Diff”) to the object meta.data.

dual_seurat <- Add_CellBender_Diff(seurat_object = dual_seurat, raw_assay_name = "RAW", cell_bender_assay_name = "RNA")

head(dual_seurat@meta.data, 5)
orig.ident nCount_RNA nFeature_RNA nCount_RAW nFeature_RAW nFeature_Diff nCount_Diff
WT1_AAACCCAAGCGACCCT-1 WT1 9844 3423 10032 3481 58 188
WT1_AAACCCAAGGAGGCAG-1 WT1 9248 3477 9431 3531 54 183
WT1_AAACCCACACACCAGC-1 WT1 3207 1699 3472 1852 153 265
WT1_AAACCCACATCTGTTT-1 WT1 14328 3930 14572 3993 63 244
WT1_AAACCCAGTATTCTCT-1 WT1 12034 3911 12223 3958 47 189

Calculate per sample averages

We can then use Median_Stats() to calculate per sample averages across all cells by supplying the new variables to the median_var parameter.

median_stats <- Median_Stats(seurat_object = dual_seurat, group_by_var = "orig.ident", median_var = c("nCount_Diff",
    "nFeature_Diff"))
orig.ident Median_nCount_RNA Median_nFeature_RNA Median_nCount_Diff Median_nFeature_Diff
WT1 7724.5 2936 216 73
WT2 7404.5 2803 212 61
Totals (All Cells) 7565.0 2869 214 67

Examine most changed features

It can also be helpful to understand what features may have changed the most. scCustomize provides the function CellBender_Feature_Diff to determine changes in features. This will return a data.frame with rowSums for each feature, the difference in each feature, and the percent change of each feature.

feature_diff <- CellBender_Feature_Diff(seurat_object = dual_seurat, raw_assay = "RAW", cell_bender_assay = "RNA")
Raw_Counts CellBender_Counts Count_Diff Pct_Diff
Gm42418 256834 3198 253636 98.75484
Uqcr11 1315 20 1295 98.47909
Gm48099 277 5 272 98.19495
Tmsb4x 6010 116 5894 98.06988
Rps21 2597 61 2536 97.65114

Plot feature differences

In addition to returning the data.frame it can be useful to visually examine the changes/trends after running CellBender. The function CellBender_Diff_Plot takes the data.frame from CellBender_Feature_Diff as input and plots the results.

Optional Parameters

  • pct_diff_threshold plot genes that exhibit a change equal to or greater than this threshold.
  • num_features instead of plotting genes above a threshold simply plot the top X changed genes.
  • num_labels change how many genes are labeled.
  • label logical, whether or not to label features.
  • custom_labels specify vector of specific features to label.
p1 <- CellBender_Diff_Plot(feature_diff_df = feature_diff)
p2 <- CellBender_Diff_Plot(feature_diff_df = feature_diff, pct_diff_threshold = 50)
p3 <- CellBender_Diff_Plot(feature_diff_df = feature_diff, num_features = 500, pct_diff_threshold = NULL)
p4 <- CellBender_Diff_Plot(feature_diff_df = feature_diff, num_labels = 10)
p5 <- CellBender_Diff_Plot(feature_diff_df = feature_diff, label = F)
p6 <- CellBender_Diff_Plot(feature_diff_df = feature_diff, custom_labels = "Gm48099")

wrap_plots(p1, p2, p3, p4, p5, p6, ncol = 2)

Dual Assay Plotting

For Cell Bender especially, but also potentially for other assays as well, it can be helpful during analysis to plot the corrected and uncorrected counts for given feature. scCustomize contains function FeaturePlot_DualAssay() to make easy.

Users just need to supply the names of the two assays to plot and the features. NOTE: Make sure both assays have been normalized before plotting. The function will attempt to check and make sure both assays have been normalized but has not been tested in all scenarios.

Example Plotting

For this example I’m using unpublished single nucleus RNA dataset from mouse cortex and have subsetted the astrocytes.

Ambient RNA gene

First let’s plot gene that represents ambient RNA as it’s restricted in expression to neurons (synaptic gene). If Cell Bender has worked well we expect that expression of this gene will be very different between the two assays.

Non-Ambient RNA gene

Now let’s plot normally astrocyte restricted gene. If Cell Bender has worked well we expect that expression of this gene shouldn’t be very different between the two assays.