Single-Cell Consensus Clustering (SC3
) is a tool for unsupervised clustering of scRNA-seq data. SC3
achieves high accuracy and robustness by consistently integrating different clustering solutions through a consensus approach. An interactive graphical implementation makes SC3
accessible to a wide audience of users. In addition, SC3
also aids biological interpretation by identifying marker genes, differentially expressed genes and outlier cells. A manuscript describing SC3
in details is published in Nature Methods.
SingleCellExperiment
, QC and scater
SC3
is a purely clustering tool and it does not provide functions for the sequencing quality control (QC) or normalisation. On the contrary it is expected that these preprocessing steps are performed by a user in advance. To encourage the preprocessing, SC3
is built on top of the Bioconductor’s SingleCellExperiment class and uses functionality of scater package for QC.
SC3
InputIf you already have a SingleCellExperiment
object created and QCed using scater
then proceed to the next chapter.
If you have a matrix containing expression data that was QCed and normalised by some other tool, then we first need to form an SingleCellExperiment
object containing the data. For illustrative purposes we will use an example expression matrix provided with SC3
. The dataset (yan
) represents FPKM gene expression of 90 cells derived from human embryo. The authors (Yan et al.) have defined developmental stages of all cells in the original publication (ann
data frame). The rows in the yan
dataset correspond to genes and columns correspond to cells.
library(SingleCellExperiment)
library(SC3)
library(scater)
head(ann)
## cell_type1
## Oocyte..1.RPKM. zygote
## Oocyte..2.RPKM. zygote
## Oocyte..3.RPKM. zygote
## Zygote..1.RPKM. zygote
## Zygote..2.RPKM. zygote
## Zygote..3.RPKM. zygote
yan[1:3, 1:3]
## Oocyte..1.RPKM. Oocyte..2.RPKM. Oocyte..3.RPKM.
## C9orf152 0.0 0.0 0.0
## RPS11 1219.9 1021.1 931.6
## ELMO2 7.0 12.2 9.3
The ann
dataframe contains just cell_type1
column which correspond to the cell labels provided by authors of the original publication. Note that in general it can also contain more information about the cells, such as plate, run, well, date etc.
Now we can create a SingleCellExperiment
object from yan
expression matrix.
Note that SC3
requires both counts
and logcounts
slots to exist in the input SingleCellExperiment
object. The counts
slot is used for gene filtering, which is based on gene dropout rates. logcounts
slot, which is supposed to contain both normalised and log-transformed expression matrix, is used in the main clustering algorithm. In the case of the yan
dataset even though the counts
are not available (we only have FPKM values) we can use the FPKM values for gene dropout rate calculations since FPKM normalisation does not change the dropout rate.
SC3
also requires the feature_symbol
column of the rowData
slot of the input SingleCellExperiment
object to contain preferable feature names (genes/transcript) which will be used in the futher visualisations.
# create a SingleCellExperiment object
sce <- SingleCellExperiment(
assays = list(
counts = as.matrix(yan),
logcounts = log2(as.matrix(yan) + 1)
),
colData = ann
)
# define feature names in feature_symbol column
rowData(sce)$feature_symbol <- rownames(sce)
# remove features with duplicated names
sce <- sce[!duplicated(rowData(sce)$feature_symbol), ]
scater
allows a user to quickly visualize and assess any SingleCellExperiment
object, for example using a PCA plot:
sce <- runPCA(sce)
## Warning in (function (A, nv = 5, nu = nv, maxit = 1000, work = nv + 7, reorth =
## TRUE, : You're computing too large a percentage of total singular values, use a
## standard svd instead.
plotPCA(sce, colour_by = "cell_type1")
If you would like to explore clustering of your data in the range of k
s (the number of clusters) from 2 to 4, you just need to run the main sc3
method and define the range of k
s using the ks
parameter (here we also ask SC3
to calculate biological features based on the identified cell clusters):
sce <- sc3(sce, ks = 2:4, biology = TRUE)
## Setting SC3 parameters...
## Calculating distances between the cells...
## Performing transformations and calculating eigenvectors...
## Performing k-means clustering...
## Calculating consensus matrix...
## Calculating biology...
By default
SC3
will use all but one cores of your machine. You can manually set the number of cores to be used by setting then_cores
parameter in thesc3
call.
To quickly and easily explore the SC3
solutions using an interactive Shiny application use the following method:
sc3_interactive(sce)
Visual exploration can provide a reasonable estimate of the number of clusters k
. Once a preferable k
is chosen it is also possible to export the results into an Excel file:
sc3_export_results_xls(sce)
This will write all results to sc3_results.xls
file. The name of the file can be controlled by the filename
parameter.
SC3
writes all its results obtained for cells to the colData
slot of the sce
object by adding additional columns to it. This slot also contains all other cell features calculated by the scater
package either automatically during the sce
object creation or during the calculateQCMetrics
call. One can identify the SC3
results using the "sc3_"
prefix:
col_data <- colData(sce)
head(col_data[ , grep("sc3_", colnames(col_data))])
## DataFrame with 6 rows and 6 columns
## sc3_2_clusters sc3_3_clusters sc3_4_clusters
## <factor> <factor> <factor>
## Oocyte..1.RPKM. 2 2 2
## Oocyte..2.RPKM. 2 2 2
## Oocyte..3.RPKM. 2 2 2
## Zygote..1.RPKM. 2 2 2
## Zygote..2.RPKM. 2 2 2
## Zygote..3.RPKM. 2 2 2
## sc3_2_log2_outlier_score sc3_3_log2_outlier_score
## <numeric> <numeric>
## Oocyte..1.RPKM. 0 3.34623
## Oocyte..2.RPKM. 0 3.37839
## Oocyte..3.RPKM. 0 3.16723
## Zygote..1.RPKM. 0 0.00000
## Zygote..2.RPKM. 0 0.00000
## Zygote..3.RPKM. 0 0.00000
## sc3_4_log2_outlier_score
## <numeric>
## Oocyte..1.RPKM. 3.34623
## Oocyte..2.RPKM. 3.37839
## Oocyte..3.RPKM. 3.16723
## Zygote..1.RPKM. 0.00000
## Zygote..2.RPKM. 0.00000
## Zygote..3.RPKM. 0.00000
Additionally, having SC3
results stored in the same slot makes it possible to highlight them during any of the scater
’s plotting function call, for example:
sce <- runPCA(sce)
## Warning in (function (A, nv = 5, nu = nv, maxit = 1000, work = nv + 7, reorth =
## TRUE, : You're computing too large a percentage of total singular values, use a
## standard svd instead.
plotPCA(
sce,
colour_by = "sc3_3_clusters",
size_by = "sc3_3_log2_outlier_score"
)
SC3
writes all its results obtained for features (genes/transcripts) to the rowData
slot of the sce
object by adding additional columns to it. This slot also contains all other feature values calculated by the scater
package either automatically during the sce
object creation or during the calculateQCMetrics
call. One can identify the SC3
results using the "sc3_"
prefix:
row_data <- rowData(sce)
head(row_data[ , grep("sc3_", colnames(row_data))])
## DataFrame with 6 rows and 13 columns
## sc3_gene_filter sc3_2_markers_clusts sc3_2_markers_padj
## <logical> <numeric> <numeric>
## C9orf152 FALSE NA NA
## RPS11 FALSE NA NA
## ELMO2 TRUE 2 3.43043e-06
## CREB3L1 TRUE 2 1.00000e+00
## PNMA1 FALSE NA NA
## MMP2 TRUE 1 1.00000e+00
## sc3_2_markers_auroc sc3_3_markers_clusts sc3_3_markers_padj
## <numeric> <numeric> <numeric>
## C9orf152 NA NA NA
## RPS11 NA NA NA
## ELMO2 0.905833 2 8.75271e-08
## CREB3L1 0.635833 2 3.65118e-04
## PNMA1 NA NA NA
## MMP2 0.549722 1 1.00000e+00
## sc3_3_markers_auroc sc3_4_markers_clusts sc3_4_markers_padj
## <numeric> <numeric> <numeric>
## C9orf152 NA NA NA
## RPS11 NA NA NA
## ELMO2 0.969697 2 8.97773e-08
## CREB3L1 0.827020 2 3.85412e-04
## PNMA1 NA NA NA
## MMP2 0.549722 3 1.00000e+00
## sc3_4_markers_auroc sc3_2_de_padj sc3_3_de_padj sc3_4_de_padj
## <numeric> <numeric> <numeric> <numeric>
## C9orf152 NA NA NA NA
## RPS11 NA NA NA NA
## ELMO2 0.969697 3.33801e-06 7.88682e-10 1.86624e-09
## CREB3L1 0.827020 1.00000e+00 2.03905e-03 6.16236e-03
## PNMA1 NA NA NA NA
## MMP2 0.543929 1.00000e+00 1.00000e+00 1.00000e+00
Because the biological features were also calculated for each k
, one can find ajusted p-values for both differential expression and marker genes, as well as the area under the ROC curve values (see ?sc3_calc_biology
for more information).
The default settings of SC3
allow to cluster (using a single k
) a dataset of 2,000 cells in about 20-30 minutes.
For datasets with more than 2,000 cells SC3
automatically adjusts some of its parameters (see below). This allows to cluster a dataset of 5,000 cells in about 20-30 minutes. The parameters can also be manually adjusted for datasets with any number of cells.
For datasets with more than 5,000 cells SC3
utilizes a hybrid approach that combines unsupervised and supervised clusterings (see below). Namely, SC3
selects a subset of cells uniformly at random, and obtains clusters from this subset. Subsequently, the inferred labels are used to train a Support Vector Machine (SVM), which is employed to assign labels to the remaining cells. Training cells can also be manually selected by providing their indeces.
SC3
also provides methods for plotting all figures from the interactive session.
The consensus matrix is a N by N matrix, where N is the number of cells in the input dataset. It represents similarity between the cells based on the averaging of clustering results from all combinations of clustering parameters. Similarity 0 (blue) means that the two cells are always assigned to different clusters. In contrast, similarity 1 (red) means that the two cells are always assigned to the same cluster. The consensus matrix is clustered by hierarchical clustering and has a diagonal-block structure. Intuitively, the perfect clustering is achieved when all diagonal blocks are completely red and all off-diagonal elements are completely blue.
sc3_plot_consensus(sce, k = 3)