Quickly Creating Pseudobulks
Quickly Creating Pseudobulks from Single Cell Gene Expression Data with Linear Algebra
Introduction
Single cell gene expression analysis often involves identifying transcriptionally distinct clusters of cells that may represent different cell-types or cell-states. After identifying these clusters of cells, we may want to summarize the gene expression profiles for each cluster by aggregating them into pseudobulks ie. the summed up gene expression profile for all cells in each cluster. I find pseudobulks really helpful in heatmap visualizations to look at top marker or differentially expressed genes, particularly if you have cell-types or cell-states across many patients or conditions. Here, I will compare the runtimes of two different ways to create these pseudobulks that I recently learned from Kamil Slowikowski using simulated data.
The comparison
To begin, let’s simulate a gene expression matrix with 5,000 genes and 10,000 cells.
library(Matrix)
set.seed(0) ## enable reproducibility
## 5000 genes
nrow <- 5000
## 10000 cells
ncol <- 10000
## simulate sparse matrix with ~0.01% of values as 1s and rest as 0s
mat <- matrix(round(rbinom(nrow*ncol, 1, 0.01/100)), nrow, ncol)
dim(mat)
## [1] 5000 10000
## how many 1s (rest are 0s)
sum(mat)
## [1] 4933
rownames(mat) <- paste0('gene', 1:nrow)
colnames(mat) <- paste0('cell', 1:ncol)
## take a look
mat[1:5, 1:5]
## cell1 cell2 cell3 cell4 cell5
## gene1 0 0 0 0 0
## gene2 0 0 0 0 0
## gene3 0 0 0 0 0
## gene4 0 0 0 0 0
## gene5 0 0 0 0 0
## convert to sparse matrix for faster computing
mat.sparse <- Matrix(mat, sparse = TRUE)
class(mat.sparse)
## [1] "dgCMatrix"
## attr(,"package")
## [1] "Matrix"
## take a look
mat.sparse[1:5, 1:5]
## 5 x 5 sparse Matrix of class "dgCMatrix"
## cell1 cell2 cell3 cell4 cell5
## gene1 . . . . .
## gene2 . . . . .
## gene3 . . . . .
## gene4 . . . . .
## gene5 . . . . .
Let’s say these cells fall into 10 different arbitrary cell-types. Let’s simulate a cell-type annotation vector.
## make cell-type annotation information
celltype <- as.factor(sample(1:10, ncol, replace=TRUE))
levels(celltype) <- paste0('celltype', 1:10)
names(celltype) <- colnames(mat)
## take a look
head(celltype)
## cell1 cell2 cell3 cell4 cell5 cell6
## celltype8 celltype5 celltype4 celltype10 celltype2 celltype3
## 10 Levels: celltype1 celltype2 celltype3 celltype4 celltype5 ... celltype10
## distribution of cells per cell-types
table(celltype)
## celltype
## celltype1 celltype2 celltype3 celltype4 celltype5 celltype6
## 1046 994 1007 998 968 1007
## celltype7 celltype8 celltype9 celltype10
## 1012 970 988 1010
To create a pseudobulk gene expression profile for each cell-type, we
may loop through the cell-types using the lapply function and sum up
the gene expression values for all cells belonging to each cell-type. We
can then column bind these summed up gene expression values using
cbind and end up with a matrix that has the total expression for our
5,000 gene for each of our 10 cell-types. Note, we will also use
Sys.time() to keep track of how it takes to run this code for
comparative purposes later.
## pseudobulk gene expression per cell-type
getPseudobulk <- function(mat, celltype) {
mat.summary <- do.call(cbind, lapply(levels(celltype), function(ct) {
cells <- names(celltype)[celltype==ct]
pseudobulk <- rowSums(mat[, cells])
return(pseudobulk)
}))
colnames(mat.summary) <- levels(celltype)
return(mat.summary)
}
## test runtime
start_time1 <- Sys.time()
## call function
mat.summary <- getPseudobulk(mat.sparse, celltype)
end_time1 <- Sys.time()
## take a look
dim(mat.summary)
## [1] 5000 10
head(mat.summary)
## celltype1 celltype2 celltype3 celltype4 celltype5 celltype6
## gene1 0 0 0 1 0 0
## gene2 0 0 0 0 0 0
## gene3 0 0 0 0 0 0
## gene4 0 0 1 0 0 0
## gene5 0 1 0 0 0 0
## gene6 0 0 0 0 0 0
## celltype7 celltype8 celltype9 celltype10
## gene1 0 0 0 0
## gene2 0 1 0 0
## gene3 0 0 0 0
## gene4 0 0 0 0
## gene5 0 0 0 0
## gene6 0 0 0 0
Alternatively, we can use linear algebra to come up with the same pseudobulk gene expression matrix. We can generate a model matrix where each row is a cell and each column is a cell-type. The model matrix entry is 1 if the cell belong to the cell-type, and 0 otherwise.
## make model matrix
mm <- model.matrix(~ 0 + celltype)
colnames(mm) <- levels(celltype)
## take alook
dim(mm)
## [1] 10000 10
head(mm)
## celltype1 celltype2 celltype3 celltype4 celltype5 celltype6 celltype7
## 1 0 0 0 0 0 0 0
## 2 0 0 0 0 1 0 0
## 3 0 0 0 1 0 0 0
## 4 0 0 0 0 0 0 0
## 5 0 1 0 0 0 0 0
## 6 0 0 1 0 0 0 0
## celltype8 celltype9 celltype10
## 1 1 0 0
## 2 0 0 0
## 3 0 0 0
## 4 0 0 1
## 5 0 0 0
## 6 0 0 0
Now, to come up with the same pseudobulk gene expression matrix, all we need to do is perform a little matrix multiplication! Recall: “row onto column” for multiplying matrices.
## test runtime
start_time2 <- Sys.time()
## multiple row of mat.sparse (gene) onto column of the model matrix (cell-type annotation)
mat.summary.mm <- mat.sparse %*% mm
end_time2 <- Sys.time()
## take a look
dim(mat.summary.mm)
## [1] 5000 10
head(mat.summary.mm)
## 6 x 10 Matrix of class "dgeMatrix"
## celltype1 celltype2 celltype3 celltype4 celltype5 celltype6
## gene1 0 0 0 1 0 0
## gene2 0 0 0 0 0 0
## gene3 0 0 0 0 0 0
## gene4 0 0 1 0 0 0
## gene5 0 1 0 0 0 0
## gene6 0 0 0 0 0 0
## celltype7 celltype8 celltype9 celltype10
## gene1 0 0 0 0
## gene2 0 1 0 0
## gene3 0 0 0 0
## gene4 0 0 0 0
## gene5 0 0 0 0
## gene6 0 0 0 0
We can check that the results using these two approaches are indeed the same.
all.equal(as.matrix(mat.summary), as.matrix(mat.summary.mm))
## [1] TRUE
But the matrix multiplication is faster!
## cbind, lapply runtime
end_time1-start_time1
## Time difference of 0.01023602 secs
## matrix multiplication runtime
end_time2-start_time2
## Time difference of 0.006520987 secs
Additional Tips
Such a model matrix approach can also more easily accomodate additional ways of splitting the cells. For example, consider if our data came from 3 different patients.
## more complex model matrices
patient <- as.factor(sample(1:3, ncol, replace=TRUE))
levels(patient) <- paste0('patient', 1:3)
names(patient) <- colnames(mat)
## take a look
head(patient)
## cell1 cell2 cell3 cell4 cell5 cell6
## patient2 patient1 patient2 patient1 patient2 patient1
## Levels: patient1 patient2 patient3
## distribution of cells per patient
table(patient)
## patient
## patient1 patient2 patient3
## 3270 3324 3406
Perhaps we want to look at the pseudobulk gene expression profile for each cell-type within each patient. To do this, we can just add to our model matrix. Again each row is a cell but now each column is a combination of patient and cell-type. The model matrix entry is 1 if the cell came from the patient and is of the cell-type, and 0 otherwise.
mm2 <- model.matrix(~ 0 + patient:celltype)
dim(mm2)
## [1] 10000 30
mm2[1:10,1:4]
## patientpatient1:celltypecelltype1 patientpatient2:celltypecelltype1
## 1 0 0
## 2 0 0
## 3 0 0
## 4 0 0
## 5 0 0
## 6 0 0
## 7 0 0
## 8 0 1
## 9 0 0
## 10 0 0
## patientpatient3:celltypecelltype1 patientpatient1:celltypecelltype2
## 1 0 0
## 2 0 0
## 3 0 0
## 4 0 0
## 5 0 0
## 6 0 0
## 7 0 0
## 8 0 0
## 9 0 0
## 10 0 0
Now we end up with a pseudobulk gene expression matrix that has the total expression for our 5,000 gene for each of our 10 cell-types within each of our 3 patients.
mat.summary.mm2 <- mat.sparse %*% mm2
dim(mat.summary.mm2)
## [1] 5000 30
mat.summary.mm2[1:10,1:4]
## 10 x 4 Matrix of class "dgeMatrix"
## patientpatient1:celltypecelltype1 patientpatient2:celltypecelltype1
## gene1 0 0
## gene2 0 0
## gene3 0 0
## gene4 0 0
## gene5 0 0
## gene6 0 0
## gene7 0 0
## gene8 0 0
## gene9 0 1
## gene10 0 0
## patientpatient3:celltypecelltype1 patientpatient1:celltypecelltype2
## gene1 0 0
## gene2 0 0
## gene3 0 0
## gene4 0 0
## gene5 0 0
## gene6 0 0
## gene7 0 0
## gene8 0 0
## gene9 0 0
## gene10 0 0
Caveats
However, note that matrix multiplication is only faster in this setting since our gene expression matrix is sparse.
Try it out for yourself!
- What happens if we use a regular dense expression matrix instead of a sparse matrix?
- Does the improvement in speed depend on the number of genes or number of cells?
Additional reading
Recent Posts
- Using AI to find heterogeneous scientific speakers on 04 November 2024
- The many ways to calculate Moran's I for identifying spatially variable genes in spatial transcriptomics data on 29 August 2024
- Characterizing spatial heterogeneity using spatial bootstrapping with SEraster on 23 July 2024
- I use R to (try to) figure out which hospital I should go to for shoppable medical services by comparing costs through analyzing Hospital Price Transparency data on 22 April 2024
- Cross modality image alignment at single cell resolution with STalign on 11 April 2024
Related Posts
- Using AI to find heterogeneous scientific speakers
- The many ways to calculate Moran's I for identifying spatially variable genes in spatial transcriptomics data
- Characterizing spatial heterogeneity using spatial bootstrapping with SEraster
- I use R to (try to) figure out which hospital I should go to for shoppable medical services by comparing costs through analyzing Hospital Price Transparency data