Introduction to single-cell RNA-seq analysis

Clustering

library(scater) # scRnaSeq QC

Loading required package: SingleCellExperiment
Loading required package: SummarizedExperiment
Loading required package: MatrixGenerics
Loading required package: matrixStats

Attaching package: ‘MatrixGenerics’

The following objects are masked from ‘package:matrixStats’:

    colAlls, colAnyNAs, colAnys, colAvgsPerRowSet, colCollapse, colCounts, colCummaxs, colCummins, colCumprods,
    colCumsums, colDiffs, colIQRDiffs, colIQRs, colLogSumExps, colMadDiffs, colMads, colMaxs, colMeans2, colMedians,
    colMins, colOrderStats, colProds, colQuantiles, colRanges, colRanks, colSdDiffs, colSds, colSums2, colTabulates,
    colVarDiffs, colVars, colWeightedMads, colWeightedMeans, colWeightedMedians, colWeightedSds, colWeightedVars,
    rowAlls, rowAnyNAs, rowAnys, rowAvgsPerColSet, rowCollapse, rowCounts, rowCummaxs, rowCummins, rowCumprods,
    rowCumsums, rowDiffs, rowIQRDiffs, rowIQRs, rowLogSumExps, rowMadDiffs, rowMads, rowMaxs, rowMeans2, rowMedians,
    rowMins, rowOrderStats, rowProds, rowQuantiles, rowRanges, rowRanks, rowSdDiffs, rowSds, rowSums2, rowTabulates,
    rowVarDiffs, rowVars, rowWeightedMads, rowWeightedMeans, rowWeightedMedians, rowWeightedSds, rowWeightedVars

Loading required package: GenomicRanges
Loading required package: stats4
Loading required package: BiocGenerics
Loading required package: parallel

Attaching package: ‘BiocGenerics’

The following objects are masked from ‘package:parallel’:

    clusterApply, clusterApplyLB, clusterCall, clusterEvalQ, clusterExport, clusterMap, parApply, parCapply,
    parLapply, parLapplyLB, parRapply, parSapply, parSapplyLB

The following objects are masked from ‘package:stats’:

    IQR, mad, sd, var, xtabs

The following objects are masked from ‘package:base’:

    anyDuplicated, append, as.data.frame, basename, cbind, colnames, dirname, do.call, duplicated, eval, evalq,
    Filter, Find, get, grep, grepl, intersect, is.unsorted, lapply, Map, mapply, match, mget, order, paste, pmax,
    pmax.int, pmin, pmin.int, Position, rank, rbind, Reduce, rownames, sapply, setdiff, sort, table, tapply, union,
    unique, unsplit, which.max, which.min

Loading required package: S4Vectors

Attaching package: ‘S4Vectors’

The following objects are masked from ‘package:base’:

    expand.grid, I, unname

Loading required package: IRanges
Loading required package: GenomeInfoDb
Loading required package: Biobase
Welcome to Bioconductor

    Vignettes contain introductory material; view with 'browseVignettes()'. To cite Bioconductor, see
    'citation("Biobase")', and for packages 'citation("pkgname")'.


Attaching package: ‘Biobase’

The following object is masked from ‘package:MatrixGenerics’:

    rowMedians

The following objects are masked from ‘package:matrixStats’:

    anyMissing, rowMedians

Loading required package: scuttle
Loading required package: ggplot2

library(scran) # scRnaSeq normalisation
library(bluster) # scRnaSeq clustering
library(cluster) # for silhouette
library(igraph) # for graph-based clustering and plotting networks

Warning: package ‘igraph’ was built under R version 4.1.2

Attaching package: ‘igraph’

The following object is masked from ‘package:scater’:

    normalize

The following object is masked from ‘package:GenomicRanges’:

    union

The following object is masked from ‘package:IRanges’:

    union

The following object is masked from ‘package:S4Vectors’:

    union

The following objects are masked from ‘package:BiocGenerics’:

    normalize, path, union

The following objects are masked from ‘package:stats’:

    decompose, spectrum

The following object is masked from ‘package:base’:

    union

library(pheatmap) # for heatmap plotting
library(patchwork) # to combine plots
library(tidyverse) # data wrangling and plotting (ggplot2)

Registered S3 methods overwritten by 'dbplyr':
  method         from
  print.tbl_lazy     
  print.tbl_sql      
── Attaching packages ─────────────────────────────────────────────────────────────────────────────────────────── tidyverse 1.3.1 ──
✓ tibble  3.1.6     ✓ dplyr   1.0.8
✓ tidyr   1.2.0     ✓ stringr 1.4.0
✓ readr   2.1.2     ✓ forcats 0.5.1
✓ purrr   0.3.4     
Warning: package ‘tidyr’ was built under R version 4.1.2
Warning: package ‘readr’ was built under R version 4.1.2
Warning: package ‘dplyr’ was built under R version 4.1.2
── Conflicts ────────────────────────────────────────────────────────────────────────────────────────────── tidyverse_conflicts() ──
x dplyr::as_data_frame() masks tibble::as_data_frame(), igraph::as_data_frame()
x dplyr::collapse()      masks IRanges::collapse()
x dplyr::combine()       masks Biobase::combine(), BiocGenerics::combine()
x purrr::compose()       masks igraph::compose()
x dplyr::count()         masks matrixStats::count()
x tidyr::crossing()      masks igraph::crossing()
x dplyr::desc()          masks IRanges::desc()
x tidyr::expand()        masks S4Vectors::expand()
x dplyr::filter()        masks stats::filter()
x dplyr::first()         masks S4Vectors::first()
x dplyr::groups()        masks igraph::groups()
x dplyr::lag()           masks stats::lag()
x ggplot2::Position()    masks BiocGenerics::Position(), base::Position()
x purrr::reduce()        masks GenomicRanges::reduce(), IRanges::reduce()
x dplyr::rename()        masks S4Vectors::rename()
x purrr::simplify()      masks igraph::simplify()
x dplyr::slice()         masks IRanges::slice()

library(DT) # For nice printing of tables in the html

Warning: package ‘DT’ was built under R version 4.1.2
Registered S3 method overwritten by 'htmlwidgets':
  method           from         
  print.htmlwidget tools:rstudio

Overview

Some of the materials originate in the Hemberg group course material with some of the text copied with a few edits. Also see the OSCA book’s “Basic” and “Advanced” chapters on clustering. In particular, please read the overview with regard to the comments on the “correctness” of any given clustering result.

Once we have normalized the data and removed confounders we can carry out analyses that are relevant to the biological questions at hand. The exact nature of the analysis depends on the data set. One of the most promising applications of scRNA-seq is de novo discovery and annotation of cell-types based on transcription profiles. This requires the identification of groups of cells based on the similarities of the transcriptomes without any prior knowledge of the labels, or unsupervised clustering. To avoid the challenges caused by the noise and high dimensionality of the scRNA-seq data, clustering is performed after feature selection and dimensionality reduction, for data that has not required batch correction this would usually on the PCA output. As our data has required batch correction we will use the “corrected” reducedDims data.

We will focus graph-based clustering, however, it is also possible to apply hierarchical clustering and k-means clustering on smaller data sets - see the OSCA book for details. Graph-base clustering is a more recent development and better suited for scRNA-seq, especially large data sets.

Load data

We will use the data set generated in the previous session. This contains 7 samples from the Caron data set. For the purposes of these materials, in the interests of time, each sample has been downsampled to only contain 500 cells.

sce <- readRDS("../Robjects/DataIntegration_mnn.out.Rds")

Graph-based clustering overview

Graph-based clustering entails building a nearest-neighbour (NN) graph using cells as nodes and their similarity as edges, then identifying ‘communities’ of cells within the network. A graph-based clustering method has three key parameters:

• How many neighbors are considered when constructing the graph
  
• What scheme is used to weight the edges
  
• Which community detection algorithm is used to define the clusters

In the NN graph two nodes (cells), say A and B, are connected by an edge if the distance between them is amongst the k smallest distances from A to other cells (‘KNN’), and from B to other cells (shared-NN, ‘SNN’). The value of k can be roughly interpreted as the anticipated size of the smallest subpopulation” (see scran’s buildSNNGraph() manual).

The edges between nodes (cells) can be weighted based on the similarity of the cells; edges connecting cells that are more closely related will have a higher weight. The three common methods for this weighting are (see the bluster package documentation for the makeSNNGraph function):

• rank - the weight is based on the highest rank of the shared nearest neighbours
  
• number - the weight is based the number of nearest neighbours in common between the two cells
  
• jaccard - the Jaccard index of the two cells sets of nearest neighbours.

Clusters are then identified using an algorithms that interprets the connections of the graph to find groups of highly interconnected cells. A variety of different algorithms are available to do this, in these materials we will focus on three methods: walktrap, louvain and leiden. See the OSCA book for details of other available in scran.

• Pros:
  • fast and memory efficient (avoids the need to construct a distance matrix for all pairs of cells)
  • no assumptions on the shape of the clusters or the distribution of cells within each cluster
  • no need to specify a number of clusters to identify (but the size of the neighbourhood used affects the size of clusters)
• Cons:
  • loss of information beyond neighboring cells, which can affect community detection in regions with many cells.

The plot below shows the same data set as a network built using three different numbers of neighbours: 5, 15 and 25 (from here).

Implementation

The implementation of clustering in R is carried out using functions from a number of different packages, in particular the bluster and igraph packages. scran provides a handy “wrapper” function clusterCells that allows us use a variety of different algorithms with one simple command.

By default clusterCells just returns a vector containing the cluster number for each cell. We can also retrieve the intermediate statistics (varying according to the algorithm used) and the SNN graph by specifying the bluster argument full = TRUE. If you are only interested in retrieving the clusters, this isn’t necessary but in this first instance we will retrieve the graph and visualise it.

clustering1 <- clusterCells(sce, use.dimred="corrected", full=TRUE)

This has defined 14 clusters with varying numbers of cells:

table(clustering1$clusters)

   1    2    3    4    5    6    7    8    9   10   11   12   13   14 
  92  151  496  253   47  117 1230   31   51   16  629  201   74  112 

The number of cells in the data set is large and plotting all the cells would take too long, so we randomly choose 1000 nodes (cells) in the network before plotting the resulting smaller network. Adding sample data to the graph and plotting the results are done using the igraph package. Cells can be color-coded by sample type:

# extract the graph
snn.gr <- clustering1$objects$graph

# Add Sample group to vertices (nodes, ie cells)
V(snn.gr)$SampleGroup <- as.character(colData(sce)$SampleGroup)

# pick 1000 nodes randomly
set.seed(1423)
selectedNodes <- sample(3500, 1000)

# subset graph for these 1000 randomly chosen nodes
snn.gr.subset <- subgraph(snn.gr, selectedNodes)

# set colors for clusters
grps <-  V(snn.gr.subset)$SampleGroup
cols <- c("dodgerblue", "lightyellow")[as.numeric(factor(grps))]
names(cols) <- grps

# plot graph
plot.igraph(snn.gr.subset,
  layout = layout_with_fr(snn.gr.subset),
  vertex.size = 3, 
  vertex.label = NA,
  vertex.color = cols,
  frame.color = cols,
  main = "default parameters"
)

# add legend
legend('bottomright',
       legend=unique(names(cols)),
       pch=21,
       pt.bg=unique(cols),
       pt.cex=1, cex=.6, bty="n", ncol=1)

More commonly we will visualise the clusters by superimposing them on a tSNE or UMAP plot. We can store the clusters in the sce object colData.

sce$Clusters1 <- clustering1$clusters
plotTSNE(sce, colour_by="Clusters1", text_by = "Clusters1")

Modularity

Several methods to detect clusters (‘communities’) in networks rely on the ‘modularity’ metric. For a given partition of cells into clusters, modularity measures how separated clusters are from each other, based on the difference between the observed and expected weight of edges between nodes. For the whole graph, the closer to 1 the better.

Walktrap method

The walktrap method relies on short random walks (a few steps) through the network. These walks tend to be ‘trapped’ in highly-connected regions of the network. Node similarity is measured based on these walks. Nodes are first each assigned their own community. Pairwise distances are computed and the two closest communities are grouped. These steps are repeated a given number of times to produce a dendrogram. Hierarchical clustering is then applied to the distance matrix. The best partition is that with the highest modularity. The original article describing the algorithm is Pons P, Latapy M (2006) Computing communities in large networks using random walks. J Graph Algorithms Appl 10(2):191–218

Walktrap is the default algorithm for clusterCells, k is set to 10 by default and the default edge weighting is “rank”. To explicitly request a specific algorithm and to set the k to a different number of nearest neighbours, we use a SNNGraphParam object from the bluster package (which is the package clusterCells is using under the hood).

Let’s set the k to 15 but keep the other parameters the same. This time we will just return the clusters:

sce$walktrap15 <- clusterCells(sce, 
                           use.dimred = "corrected", 
                           BLUSPARAM = SNNGraphParam(k = 15, cluster.fun = "walktrap"))

This time we have defined 12 clustering. As a general rule, increasing k will tend to decrease the number of clusters (not always, but generally).

table(sce$walktrap15)

   1    2    3    4    5    6    7    8    9   10   11   12 
 223   72  451  200  151   62   53  266 1304   36   87  595 

We can visualise the assignment of cells from different samples to the clusters using a heatmap.

table(sce$walktrap15, sce$SampleName) %>% 
  pheatmap(cluster_rows = FALSE, cluster_cols = FALSE)

Most clusters comprise cells from several replicates of a same sample type, cluster 9 appears to be predominantly cells from the ETV6-RUNX samples.

We can visualise this on the TSNE:

plotTSNE(sce, colour_by="walktrap15", text_by = "walktrap15")

plotTSNE(sce, colour_by="walktrap15") +
  facet_wrap(sce$SampleGroup)

Additional clustering parameters (Advanced)

The different clustering algorithms may have additional parameters, specific to the algorithm, that can be adjusted. With the walktrap algorithm we could also tweak the number of “steps” in each walk. The default is 4, but we could, for example, change this to 10 by adding the parameter cluster.args = list(steps = 10) to the SNNGraphParam object in the clusterCells command.

The Louvain method

With the Louvain method, nodes are also first assigned their own community. This hierarchical agglomerative method then progresses in two-step iterations:

1. nodes are re-assigned one at a time to the community for which they increase modularity the most, if at all.
2. a new, ‘aggregate’ network is built where nodes are the communities formed in the previous step.

These two steps are repeated until modularity stops increasing. The diagram below is copied from this article.

We now apply the Louvain approach, store its outcome in the SCE object and show cluster sizes.

sce$louvain15 <- clusterCells(sce, 
                           use.dimred = "corrected", 
                           BLUSPARAM = SNNGraphParam(k = 15, cluster.fun = "louvain"))

table(sce$louvain15)

  1   2   3   4   5   6   7 
532 806 646 747 221 384 164 

The t-SNE plot shows cells color-coded by cluster membership:

plotTSNE(sce, colour_by="louvain15", text_by = "louvain15")

If we split by sample type we can see differences in the clusters between the sample groups:

plotTSNE(sce, colour_by="louvain15") + 
  facet_wrap(~ sce$SampleGroup)

Leiden method

The Leiden method improves on the Louvain method by guaranteeing that at each iteration clusters are connected and well-separated. The method includes an extra step in the iterations: after nodes are moved (step 1), the resulting partition is refined (step2) and only then the new aggregate network made, and refined (step 3). The diagram below is copied from this article.

sce$leiden20 <- clusterCells(sce, 
                           use.dimred = "corrected", 
                           BLUSPARAM = SNNGraphParam(k = 20, cluster.fun = "leiden"))

table(sce$leiden20)

   1    2    3    4    5    6    7    8    9   10   11 
 490 1256  198   22  632    1  115  236  270  114  166 

The t-SNE plot shows cells color-coded by cluster membership:

plotTSNE(sce, colour_by="leiden20", text_by = "leiden20")

Assessing cluster behaviour

A variety of metrics are available to aid us in assessing the behaviour of a particular clustering method on our data. These can help us in assessing how well defined different clusters within a single clustering are in terms of the relatedness of cells within the cluster and the how well separated that cluster is from cells in other clusters, and to compare the results of different clustering methods or parameter values (e.g. different values for k).

We will consider “Silhouette width” and “Modularity”. Further details and other metrics are described in the “Advanced” section of the OSCA book.

Silhouette width

The silhouette width (so named after the look of the traditional graph for plotting the results) is a measure of how closely related cells within cluster are to one another versus how closely related cells in the cluster are to cells in other clusters. This allows us to assess cluster separation.

For each cell in the cluster we calculate the the average distance to all other cells in the cluster and the average distance to all cells not in the cluster. The cells silhouette width is the difference between these divided by the maximum of the two values. Cells with a large silhouette are strongly related to cells in the cluster, cells with a negative silhouette width are more closely related to other clusters.

We will use the approxSilhouette function from the bluster package. The resulting table gives us the silhouette width for each cell, the cluster it belongs to, and which other cluster it is most closely related to.

sil.approx <- approxSilhouette(reducedDim(sce, "corrected"), clusters=sce$louvain15)
sil.approx

DataFrame with 3500 rows and 3 columns
                       cluster    other     width
                      <factor> <factor> <numeric>
1_AAACGGGCAGTTCATG-1         1        3 0.0637850
1_AAACGGGGTTCACCTC-1         1        3 0.2470117
1_AAAGATGAGCGATGAC-1         2        3 0.0802831
1_AAAGATGCAGCCAATT-1         3        1 0.2701720
1_AAAGTAGCAATGCCAT-1         3        2 0.2765963
...                        ...      ...       ...
12_TTTCCTCGTCCAAGTT-1        4        5 0.3715681
12_TTTGCGCCAGGCTCAC-1        2        3 0.0525437
12_TTTGGTTTCCAAGCCG-1        5        3 0.1446539
12_TTTGTCACAATGAAAC-1        4        3 0.3645356
12_TTTGTCATCAGTTGAC-1        4        5 0.3713531

We can view the results in as a beeswarm plot. We colour each cell according to either its current cluster or, if the cell has a negative silhouette width, the cluster that it is closest to.

silPlot.dat <- sil.approx %>% 
  as.data.frame() %>% 
  mutate(closestCluster = ifelse(width > 0, cluster, other) %>% factor())

ggplot(silPlot.dat, aes(x=cluster, y=width, colour=closestCluster)) +
  ggbeeswarm::geom_quasirandom(method="smiley")

We could also look at the correspondence between different clusters by plotting these numbers on a grid showing for each cluster number of cells in that cluster that are closer to another cluster, colouring each tile by the proportion of the total cells in the cluster that it contains. Ideally we would like to see a strong diagonal band and only a few off-diagonal tiles containing small number of cells.

plotSilGrid <- function(dat){
  dat %>% 
    count(cluster, closestCluster,  name="olap") %>% 
    group_by(cluster) %>% 
    mutate(total  = sum(olap)) %>% 
    mutate(proportion = olap / total) %>% 
    mutate(proportion = ifelse(cluster == closestCluster, proportion, -proportion)) %>% 
    ggplot(aes(x = cluster, y = closestCluster, fill = proportion)) +
      geom_tile(col = "grey") +
      geom_text(aes(label = olap), size=5) +
      scale_fill_gradientn(colors = c("#fc8d59", "#ffffbf", "#91cf60"),
                            limits = c(-1, 1)) +
      geom_vline(xintercept=seq(0.5, 30.5, by=1)) +
      geom_hline(yintercept=seq(0.5, 30.5, by=1), colour="lightgrey", linetype=2) +
      guides(fill = "none") +
      theme(
          aspect.ratio = 1,
          axis.text.x = element_text(angle = 90, 
                                     hjust = 1,  
                                     vjust = 0.5, 
                                     size = 13),
          axis.text.y = element_text(size = 13),
          plot.title = element_text(hjust = 0.5),
          panel.background = element_blank())
}
plotSilGrid(silPlot.dat)

From these two plots we can see that clusters 3 to 8 appear to have a good degree of separation, however, clusters 1 and 2 have a large number of cells that appear to be closer to cluster 3. It is possible clusters 1-3 contain a degree of heterogeneity that suggests further splitting of these clusters may be required - perhaps there should more than 7 clusters to adequately reflect the biology.

Let’s do the same plots with the walktrap clusters

sil.approx <- approxSilhouette(reducedDim(sce, "corrected"), clusters=sce$walktrap15)
silPlot.dat <- sil.approx %>% 
  as.data.frame() %>% 
  mutate(closestCluster = ifelse(width > 0, cluster, other) %>% factor())

ggplot(silPlot.dat, aes(x=cluster, y=width, colour=closestCluster)) +
  ggbeeswarm::geom_quasirandom(method="smiley")

plotSilGrid(silPlot.dat)

This clustering appears to have generated a set of clusters with better separatedness than the Louvain method with a k of 15.

And again with the leiden clusters

sil.approx <- approxSilhouette(reducedDim(sce, "corrected"), clusters=sce$leiden20)
silPlot.dat <- sil.approx %>% 
  as.data.frame() %>% 
  mutate(closestCluster = ifelse(width > 0, cluster, other) %>% factor())

ggplot(silPlot.dat, aes(x=cluster, y=width, colour=closestCluster)) +
  ggbeeswarm::geom_quasirandom(method="smiley")

plotSilGrid(silPlot.dat)

This seems similar to the walktrap clustering.

Modularity to assess clusters quality

As mentioned early, the modularity metric is used in evaluating the separatedness between clusters. Some of the clustering algorithms, e.g. Louvain, seek to optimise this for the entire NN graph as part of their cluster detection. Modularity is a ratio between the observed weights of the edges within a cluster versus the expected weights if the edges were randomly distributed between all nodes. Rather than calculating a single modularity value for the whole graph, we can instead calculate a pair-wise modularity value between each pair of clusters using the pairwiseModularity function from the bluster package. For this we need to have the graph from the clustering, so we will rerun the walktrap clustering with k=15 to obtain this. We can plot the resulting ratios on a heatmap. We would expect the highest modularity values to be on the diagonal.

walktrap15 <- clusterCells(sce, 
                           use.dimred = "corrected", 
                           BLUSPARAM = SNNGraphParam(k = 15, cluster.fun = "walktrap"),
                           full = TRUE)
g <- walktrap15$objects$graph
ratio <- pairwiseModularity(g, walktrap15$clusters, as.ratio=TRUE)
pheatmap(log2(ratio+1), cluster_rows=FALSE, cluster_cols=FALSE,
    color=colorRampPalette(c("white", "blue"))(100))

Largely, this reflects what we saw from the silhouette widths, but also reveals some additional inter-connectedness between other clusters e.g. cluster 10 and cluster 2. We can also visualise this as network graph where nodes are clusters and the edge weights are the modularity.

cluster.gr <- igraph::graph_from_adjacency_matrix(log2(ratio+1),
                                                  mode="upper", 
                                                  weighted=TRUE, diag=FALSE)

set.seed(11001010)
plot(cluster.gr, 
     edge.width=igraph::E(cluster.gr)$weight*5,
     layout=igraph::layout_with_lgl)

Comparing two sets of clusters

The wide variety of clustering algorithms available and their differing outputs reflect the different ways that it is possible to view out data in high-dimensional space. It may often be useful to assess the concordance between different clustering methods in order to assess how clusters relate to each other - e.g. does one cluster from one method equate to just one cluster in the other or is it a combination of different clusters. This may be revealing about the underlying biology. We will use the Jaccard index as measure of concordance between clusters. A value of 1 represents perfect concordance between clusters (i.e. they contain exactly the same cells).

jacc.mat <- linkClustersMatrix(sce$louvain15, sce$walktrap15)
rownames(jacc.mat) <- paste("Louvain", rownames(jacc.mat))
colnames(jacc.mat) <- paste("Walktrap", colnames(jacc.mat))
pheatmap(jacc.mat, color=viridis::viridis(100), cluster_cols=FALSE, cluster_rows=FALSE)

We can see that Louvain clusters 2, and 5 are equivalent to walktrap clusters 9 and 4. The remaining Louvain clusters are combinations of cells from various walktrap clusters. We may want to look at marker genes for these clusters to assess what these two different views are telling us about the biology.

Cluster sweep

As we have seen, there are a number of different parameters we can change to alter the final clustering result - primarily the edge weighting method, the k and the clustering algorithm. There is no one gold standard that will fit all data, so, in most cases, it is necessary to assess a number of different clusterings to obtain one that provides a view of the data that suits our biological interpretations. The clusterSweep function allows us to apply a range of different parameters in one go and obtain the clustering for each.

For example, suppose we wish to assess the effect of different values of k on the walktrap clustering. We can parallelize this process to make it faster.

out <- clusterSweep(reducedDim(sce, "corrected"),
                    BLUSPARAM = NNGraphParam(),
                    k = as.integer(c(5, 10, 15, 20, 25)),
                    cluster.fun = "walktrap",
                    BPPARAM=BiocParallel::MulticoreParam(5))

The resulting object is list containing a DataFrame with the clusters for each combination of the clustering parameters and a corresponding DataFrame showing the parameters used to generate each of these:

out$clusters[,1:4]

DataFrame with 3500 rows and 4 columns
     k.5_cluster.fun.walktrap k.10_cluster.fun.walktrap k.15_cluster.fun.walktrap k.20_cluster.fun.walktrap
                     <factor>                  <factor>                  <factor>                  <factor>
1                          4                          4                         8                         7
2                          4                          4                         8                         7
3                          6                          7                         9                         6
4                          6                          7                         9                         3
5                          11                         3                         3                         3
...                       ...                       ...                       ...                       ...
3496                       18                        11                        12                        10
3497                       6                         7                         9                         6 
3498                       10                        12                        4                         9 
3499                       18                        11                        12                        10
3500                       8                         11                        5                         10

out$parameters

DataFrame with 5 rows and 2 columns
                                  k cluster.fun
                          <integer> <character>
k.5_cluster.fun.walktrap          5    walktrap
k.10_cluster.fun.walktrap        10    walktrap
k.15_cluster.fun.walktrap        15    walktrap
k.20_cluster.fun.walktrap        20    walktrap
k.25_cluster.fun.walktrap        25    walktrap

We can then combine this cluster sweep with the metrics for assessing cluster behaviour in order to get a overview of the effects of these parameter changes that may enable us to make some decisions as to which clustering or clusterings we may wish to investigate further.

Here we will just look at the mean silhouette width and the number of clusters.

df <- as.data.frame(out$parameters)
df$num.clusters <- apply(out$clusters, 2, max)
df$silhouette <- map_dbl(as.list(out$clusters), function(cluster) {
    sil <- approxSilhouette(reducedDim(sce, "corrected"), cluster)
    mean(sil$width)
})

nclPlot <- ggplot(df, aes(x = k, y = num.clusters)) + 
                  geom_line(lwd=2)
silPlot <- ggplot(df, aes(x = k, y = silhouette)) + 
                  geom_line(lwd=2)
nclPlot + silPlot

Based on our previous analysis and knowledge of the biology we may feel that 12 clusters represents a good number clusters, but we can see here that both k = 15 and k = 20 provide this, but k = 20 gives us a better silhouette score. On the other had, perhaps k = 10 provides greater resolution of cell types, with more clusters with only a slight decrease in the silhouette score.

Earlier we looked at the Jaccard index as a means of comparing two different clusterings. We could apply the same method here:

jacc.mat <- linkClustersMatrix(out$clusters$k.5_cluster.fun.walktrap, 
                               out$clusters$k.15_cluster.fun.walktrap)
rownames(jacc.mat) <- paste("Walktrap_5", rownames(jacc.mat))
colnames(jacc.mat) <- paste("Walktrap_15", colnames(jacc.mat))
pheatmap(jacc.mat, color=viridis::viridis(100), cluster_cols=FALSE, cluster_rows=FALSE)

Finally, we can add all (or a subset) of the clusterings from clusterSweep to our SCE object.

colData(sce) <- cbind(colData(sce), DataFrame(out$clusters))

The OSCA book provides some additional methods for comparing different clusterings that can be combined with the cluster sweep results to assess cluster behaviour under different parameters.

In this section, we have just done a sweep changing the k, but it is also possible to combine this with multiple clustering algorithms and multiple edge weightings.

Note: In practice, on a full data set, with multiple algorithms and values of k, this can take a long time to run. Usually I do not run this interactively in R studio, but instead write an R script that will end by exporting the final output of clusterSweep to an RDS object which I can later load into R. This R script can then be sent as a job to the cluster. This is also means the job can be massively more parallelized by using more cores. We will see this in the exercise for this session.

Finalise clustering selection

When you have come to a decision about which clustering to use it is convenient to add it to colData column called “label” using the colLabels function. This means downstream code does not need to be changed should you later decide to switch to a different clustering (and makes the code easily re-usable for different analyses).

For now we will use the walktrap k=15 clustering.

colLabels(sce) <- sce$walktrap15

Expression of known marker genes

If we expect our clusters to represent known cell types for which there are well established marker genes, we can now start to investigate the clusters by plotting in parallel the expression of these genes. This can also help us in assessing if our clustering has satisfactorily partitioned our cells.

plotTSNE(sce, colour_by = "label", text_by = "label") +
  ggtitle("Walktrap clusters")

Having identified clusters, we now display the level of expression of cell type marker genes to quickly match clusters with cell types. For each marker we will plot its expression on a t-SNE, and show distribution across each cluster on a violin plot.

We will be using gene symbols to identify the marker genes, so we will switch the rownames in the SCE object to be gene symbols. We use the scater function uniquifyFeatureNames to do this as there are a few duplicated gene symbols.

rownames(sce) <- uniquifyFeatureNames(rowData(sce)$ID, rowData(sce)$Symbol)

B-cells markers

p1 <- plotTSNE(sce, by_exprs_values = "reconstructed",
               colour_by = "MS4A1",
               text_by = "label")
p2 <- plotTSNE(sce, by_exprs_values = "reconstructed"
               , colour_by = "CD79A",
               text_by = "label")
p1 + p2

plotExpression(sce, 
               exprs_values = "reconstructed",
               x = "label", 
               colour_by = "label",
               features=c("MS4A1", "CD79A"))

Monocyte markers

p1 <- plotTSNE(sce, by_exprs_values = "reconstructed",
               colour_by = "FCGR3A",
               text_by = "label")
p2 <- plotTSNE(sce, by_exprs_values = "reconstructed", 
               colour_by = "MS4A7",
               text_by = "label")
p1 + p2

plotExpression(sce, 
               exprs_values = "reconstructed",
               x = "label", 
               colour_by = "label",
               features=c("FCGR3A", "MS4A7"))

Dendritic cell markers

p1 <- plotTSNE(sce, by_exprs_values = "reconstructed", 
               colour_by = "FCER1A",
               text_by = "label")
p2 <- plotTSNE(sce, by_exprs_values = "reconstructed", 
               colour_by = "CST3",
               text_by = "label")
p1 + p2

plotExpression(sce, 
               exprs_values = "reconstructed",
               x = "label", 
               colour_by = "label",
               features=c("FCER1A", "CST3"))

Save data

Write SCE object to file.

saveRDS(sce, file="../Robjects/Caron_clustering_material.rds")

Session information

sessionInfo()

R version 4.1.1 (2021-08-10)
Platform: x86_64-apple-darwin17.0 (64-bit)
Running under: macOS Big Sur 11.5.2

Matrix products: default
LAPACK: /Library/Frameworks/R.framework/Versions/4.1/Resources/lib/libRlapack.dylib

locale:
[1] en_GB.UTF-8/en_GB.UTF-8/en_GB.UTF-8/C/en_GB.UTF-8/en_GB.UTF-8

attached base packages:
[1] parallel  stats4    stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
 [1] DT_0.21                     forcats_0.5.1               stringr_1.4.0               dplyr_1.0.8                
 [5] purrr_0.3.4                 readr_2.1.2                 tidyr_1.2.0                 tibble_3.1.6               
 [9] tidyverse_1.3.1             patchwork_1.1.1             pheatmap_1.0.12             igraph_1.2.11              
[13] cluster_2.1.2               bluster_1.2.1               scran_1.20.1                scater_1.20.1              
[17] ggplot2_3.3.5               scuttle_1.2.1               SingleCellExperiment_1.14.1 SummarizedExperiment_1.22.0
[21] Biobase_2.52.0              GenomicRanges_1.44.0        GenomeInfoDb_1.28.4         IRanges_2.26.0             
[25] S4Vectors_0.30.2            BiocGenerics_0.38.0         MatrixGenerics_1.4.3        matrixStats_0.61.0         

loaded via a namespace (and not attached):
 [1] ggbeeswarm_0.6.0          colorspace_2.0-3          ellipsis_0.3.2            XVector_0.32.0           
 [5] BiocNeighbors_1.10.0      fs_1.5.2                  rstudioapi_0.13           farver_2.1.0             
 [9] fansi_1.0.2               lubridate_1.8.0           xml2_1.3.3                sparseMatrixStats_1.4.2  
[13] knitr_1.37                jsonlite_1.8.0            broom_0.7.12              dbplyr_2.1.1             
[17] compiler_4.1.1            httr_1.4.2                dqrng_0.3.0               backports_1.4.1          
[21] assertthat_0.2.1          Matrix_1.4-0              fastmap_1.1.0             limma_3.48.3             
[25] cli_3.2.0                 BiocSingular_1.8.1        htmltools_0.5.2           tools_4.1.1              
[29] rsvd_1.0.5                gtable_0.3.0              glue_1.6.2                GenomeInfoDbData_1.2.6   
[33] Rcpp_1.0.8                jquerylib_0.1.4           cellranger_1.1.0          vctrs_0.3.8              
[37] crosstalk_1.2.0           DelayedMatrixStats_1.14.3 xfun_0.30                 beachmat_2.8.1           
[41] rvest_1.0.2               lifecycle_1.0.1           irlba_2.3.5               statmod_1.4.36           
[45] edgeR_3.34.1              zlibbioc_1.38.0           scales_1.1.1              hms_1.1.1                
[49] RColorBrewer_1.1-2        yaml_2.3.5                gridExtra_2.3             sass_0.4.0               
[53] stringi_1.7.6             ScaledMatrix_1.0.0        BiocParallel_1.26.2       rlang_1.0.2              
[57] pkgconfig_2.0.3           bitops_1.0-7              evaluate_0.15             lattice_0.20-45          
[61] labeling_0.4.2            htmlwidgets_1.5.4         cowplot_1.1.1             tidyselect_1.1.2         
[65] magrittr_2.0.2            R6_2.5.1                  generics_0.1.2            metapod_1.0.0            
[69] DelayedArray_0.18.0       DBI_1.1.2                 pillar_1.7.0              haven_2.4.3              
[73] withr_2.5.0               RCurl_1.98-1.6            modelr_0.1.8              crayon_1.5.0             
[77] utf8_1.2.2                tzdb_0.2.0                rmarkdown_2.12            viridis_0.6.2            
[81] locfit_1.5-9.5            grid_4.1.1                readxl_1.3.1              reprex_2.0.1             
[85] digest_0.6.29             munsell_0.5.0             beeswarm_0.4.0            viridisLite_0.4.0        
[89] bslib_0.3.1               vipor_0.4.5