Package 'scBubbletree' reference manual

Title:	Quantitative visual exploration of scRNA-seq data
Description:	scBubbletree is a quantitative method for visual exploration of scRNA-seq data. It preserves biologically meaningful properties of scRNA-seq data, such as local and global cell distances, as well as the density distribution of cells across the sample. scBubbletree is scalable and avoids the overplotting problem, and is able to visualize diverse cell attributes derived from multiomic single-cell experiments. Importantly, Importantly, scBubbletree is easy to use and to integrate with popular approaches for scRNA-seq data analysis.
Authors:	Simo Kitanovski [aut, cre]
Maintainer:	Simo Kitanovski <[email protected]>
License:	GPL-3 + file LICENSE
Version:	1.7.17
Built:	2024-07-19 02:50:05 UTC
Source:	https://github.com/bioc/scBubbletree

The R package scBubbletree

Description

Method for quantitative visualization of single cell RNA-seq data

Details

This package contains functions for clustering, hierarchical grouping of clusters and visualization of scRNA-seq data.

Author(s)

Authors and maintainers:

Simo Kitanovski [email protected] (ORCID)

Comparison of two bubbletrees generated from the same scRNA-seq data

Description

compare_bubbletrees takes as its main input two bubbletrees generated from the same input data but potentially with different input parameters (e.g. clustering method or resolutions).

It then does the following two operations:

1. computes the Jaccard distance (JD) and the intersection between paris of clusters from the two bubbletrees. This is visualized as a heatmap.

2. it visualizes the two bubbletrees together with the heatmap.

Usage

compare_bubbletrees(btd_1,
                    btd_2,
                    tile_bw = FALSE,
                    tile_text_size = 3,
                    ratio_heatmap = 0.5)
compare_bubbletrees(btd_1,
                    btd_2,
                    tile_bw = FALSE,
                    tile_text_size = 3,
                    ratio_heatmap = 0.5)

Arguments

`btd_1`	bubbletree object
`btd_2`	bubbletree object
`tile_text_size`	integer, size of tile labels (default = 3)
`tile_bw`	logical, tile grayscale (tile_bw = TRUE) vs. color (tile_bw = FALSE, default)
`ratio_heatmap`	nummeric, probability (default = 0.5) that dictates the relative width and height of the heatmap and the bubbletrees

Details

compare_bubbletrees takes as its main input two bubbletrees generated from the same input data but potentially with different input parameters (e.g. clustering method or resolutions).

It then does the following two operations:

1. computes the Jaccard distance and the intersection between paris of clusters from the two bubbletrees. This is visualized as a heatmap.

2. it visualizes the two bubbletrees together with the heatmap.

Value

`comparison`	ggplot2 objects assembled by R-package `patchwork`
`m`	data.frame object with JD and intersection for each pair of clusters from the two bubbletrees

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

btd_1 <- get_bubbletree_graph(x = A,
                              r = 1,
                              n_start = 20,
                              iter_max = 100,
                              algorithm = "original",
                              knn_k = 50,
                              hclust_method = "average",
                              hclust_distance = "euclidean",
                              cores = 1,
                              round_digits = 2,
                              show_simple_count = FALSE)


btd_2 <- get_bubbletree_kmeans(x = A,
                               k = 8,
                               cores = 1,
                               round_digits = 1,
                               show_simple_count = FALSE,
                               kmeans_algorithm = "MacQueen",
                               hclust_distance = "euclidean",
                               hclust_method = "average")

btd_comparison <- compare_bubbletrees(btd_1 = btd_1,
                                      btd_2 = btd_2,
                                      tile_bw = FALSE,
                                      tile_text_size = 3,
                                      ratio_heatmap = 0.5)


# plot
btd_comparison$tree_comparison

# data.frame of heatmap data
btd_comparison$m
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

btd_1 <- get_bubbletree_graph(x = A,
                              r = 1,
                              n_start = 20,
                              iter_max = 100,
                              algorithm = "original",
                              knn_k = 50,
                              hclust_method = "average",
                              hclust_distance = "euclidean",
                              cores = 1,
                              round_digits = 2,
                              show_simple_count = FALSE)


btd_2 <- get_bubbletree_kmeans(x = A,
                               k = 8,
                               cores = 1,
                               round_digits = 1,
                               show_simple_count = FALSE,
                               kmeans_algorithm = "MacQueen",
                               hclust_distance = "euclidean",
                               hclust_method = "average")

btd_comparison <- compare_bubbletrees(btd_1 = btd_1,
                                      btd_2 = btd_2,
                                      tile_bw = FALSE,
                                      tile_text_size = 3,
                                      ratio_heatmap = 0.5)


# plot
btd_comparison$tree_comparison

# data.frame of heatmap data
btd_comparison$m

Dataset: 500 PBMCs

Description

d_500 is a list with 3 elements:

1. A = numeric matrix $A^500x15$ with n=500 rows for PBMCs and f=15 principal components.

2. f = character vector $f$ of length 500. Each element in $f$ represents the predicted cell type of a specific cell.

3. fs = numeric matrix containing normalized gene expressions of 12 marker genes in 500 cells.

Usage

data("d_500", package = "scBubbletree")data("d_500", package = "scBubbletree")

Format

Format of d_500: list

Details

This data is a sample drawn from a larger dataset of 2,700 PBMCs. The original dataset was processed as described in vignette (accessed 23, Sep, 2022):

https://satijalab.org/seurat/articles/multimodal_reference_mapping.html

See R script inst/script/get_d_500.R to see how this dataset was created.

Source

https://satijalab.org/seurat/articles/multimodal_reference_mapping.html

Examples

data("d_500", package = "scBubbletree")

A <- d_500$A
base::dim(A)

f <- d_500$f
base::table(f)

fs <- d_500$fs
base::dim(fs)

data("d_500", package = "scBubbletree")

A <- d_500$A
base::dim(A)

f <- d_500$f
base::table(f)

fs <- d_500$fs
base::dim(fs)

Dataset: scRNA-seq data of 3,918 cells from 5 adenocarcinoma cell lines

Description

d_ccl is a list with 3 elements:

1. A = numeric matrix with n=3,918 rows for cells and f=15 principal components

2. m = data.frame meta data

3. e = numeric matrix containing normalized gene expressions of 5 marker genes

Usage

data("d_ccl", package = "scBubbletree")data("d_ccl", package = "scBubbletree")

Format

Format of d_ccl: list

Details

d_ccl is a scRNA-seq dataset containing a mixture of 3,918 cells from five human lung adenocarcinoma cell lines (HCC827, H1975, A549, H838 and H2228). The dataset is available here:

https://github.com/LuyiTian/sc_mixology/blob/master/data/ sincell_with_class_5cl.RData

The library has been prepared with 10x Chromium platform and sequenced with Illumina NextSeq 500 platform. Raw data has been processed with Cellranger. The tool demuxlet has been used to predict the identity of each cell based on known genetic differences between the different cell lines.

See R script inst/script/get_d_ccl.R to see how this dataset was created.

Source

https://github.com/LuyiTian/sc_mixology/blob/master/data/ sincell_with_class_5cl.RData

References

Tian, Luyi, et al. "Benchmarking single cell RNA-sequencing analysis pipelines using mixture control experiments." Nature methods 16.6 (2019): 479-487

Examples

data("d_ccl", package = "scBubbletree")

A <- d_ccl$A
base::dim(A)

m <- d_ccl$m
utils::head(m)

e <- d_ccl$e
base::dim(e)

data("d_ccl", package = "scBubbletree")

A <- d_ccl$A
base::dim(A)

m <- d_ccl$m
utils::head(m)

e <- d_ccl$e
base::dim(e)

Build bubbletree from matrix $A$ of low-dimensional projections and vector $cs$ of externally generated cluster IDs

Description

get_bubbletree_dummy takes two main inputs:

1. numeric matrix $A^{n\times f}$ , which represents a low-dimensional projection (obtained e.g. by PCA) of the original high-dimensional scRNA-seq data, with $n$ rows as cells and $f$ columns as low-dimension features.

2. vector $cs$ of cluster IDs of each cell

The function get_bubbletree_dummy performs one main operation. It organizes the bubbles (defined by $cs$ ) in a hierarchical dendrogram (bubbletree) which represents the hierarchical relationships between the clusters (bubbles).

Usage

get_bubbletree_dummy(x,
                     cs,
                     B = 200,
                     N_eff = 100,
                     hclust_distance = "euclidean",
                     hclust_method = "average",
                     cores = 1,
                     round_digits = 2,
                     show_simple_count = FALSE,
                     verbose = TRUE)
get_bubbletree_dummy(x,
                     cs,
                     B = 200,
                     N_eff = 100,
                     hclust_distance = "euclidean",
                     hclust_method = "average",
                     cores = 1,
                     round_digits = 2,
                     show_simple_count = FALSE,
                     verbose = TRUE)

Arguments

`x`	numeric matrix ( $A^{n\times f}$ with $n$ cells, and $f$ low-dimensional projections of the original single cell RNA-seq dataset)
`cs`	vector, cluster IDs
`B`	integer, number of bootstrap iterations to perform in order to generate bubbletree
`N_eff`	integer, number of cells to draw randomly from each cluster when computing inter-cluster distances
`hclust_distance`	distance measure to be used: euclidean (default) or manhattan, see documentation of `stats::dist`
`hclust_method`	agglomeration method to be used, default = average. See documentation of `stats::hclust`
`cores`	integer, number of PC cores for parallel execution
`round_digits`	integer, number of decimal places to keep when showing the relative frequency of cells in each bubble
`show_simple_count`	logical, if `show_simple_count`=T, cell counts in each bubble will be divided by 1,000 to improve readability. This is only useful for samples that are composed of millions of cells.
`verbose`	logical, progress messages

Details

This function is similar to get_bubbletree_kmeans and get_bubbletree_graph but skips the clustering step. See the documentation of the respective functions.

Value

`A`	input x matrix
`k`	number of clusters
`km`	NULL
`ph`	boot_ph: bootstrap dendrograms $H_b$ ; main_ph: bubbletree $H$
`pair_dist`	inter-cluster distances used to generate the dendrograms
`cluster`	cluster assignments of each cell
`input_par`	list of all input parameters
`tree`	ggtree bubbletree object
`tree_meta`	meta-data associated with the bubbletree

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

cs <- base::sample(x = LETTERS[1:5], size = nrow(A), replace = TRUE)

db <- get_bubbletree_dummy(x = A,
                           cs = cs,
                           B = 200,
                           N_eff = 100,
                           hclust_distance = "euclidean",
                           hclust_method = "average",
                           cores = 1)
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

cs <- base::sample(x = LETTERS[1:5], size = nrow(A), replace = TRUE)

db <- get_bubbletree_dummy(x = A,
                           cs = cs,
                           B = 200,
                           N_eff = 100,
                           hclust_distance = "euclidean",
                           hclust_method = "average",
                           cores = 1)

Louvain clustering and hierarchical grouping of $k'$ clusters (bubbles)

Description

get_bubbletree_graph has two main inputs:

2. clustering resolution $r$

The function get_bubbletree_graph performs two main operations. First, it performs Louvain clustering to identify groups (bubbles) of transcriptionally similar cells; second, it organizes the bubbles in a hierarchical dendrogram (bubbletree) which adequatly represents inter-cluster relationships.

Usage


get_bubbletree_graph(x,
                     r,
                     B = 200,
                     N_eff = 100,
                     n_start = 20,
                     iter_max = 100,
                     algorithm = "original",
                     knn_k = 20,
                     hclust_method = "average",
                     hclust_distance = "euclidean",
                     cores = 1,
                     round_digits = 2,
                     show_simple_count = FALSE,
                     verbose = TRUE)
get_bubbletree_graph(x,
                     r,
                     B = 200,
                     N_eff = 100,
                     n_start = 20,
                     iter_max = 100,
                     algorithm = "original",
                     knn_k = 20,
                     hclust_method = "average",
                     hclust_distance = "euclidean",
                     cores = 1,
                     round_digits = 2,
                     show_simple_count = FALSE,
                     verbose = TRUE)

Arguments

`x`	numeric matrix ( $A^{n\times f}$ with $n$ cells, and $f$ low-dimensional projections of the original single cell RNA-seq dataset)
`r`	number, clustering resolution
`B`	integer, number of bootstrap iterations to perform in order to generate bubbletree. If $B=200$ (default), cluster centroids are used to compute inter-cluster distances and $N_{eff}$ is ignored, i.e. all cells are used to compute centroids.
`N_eff`	integer, number of cells to draw randomly from each cluster when computing inter-cluster distances.
`n_start`, `iter_max`	parameters for Louvain clustering, see documentation of function `FindClusters`, R-package `Seurat`
`algorithm`	character, four clustering algorithms: 'original', 'LMR', 'SLM' and 'Leiden', see documentation of function `FindClusters`, R-package `Seurat`
`knn_k`	integer, defines $k$ for the k-nearest neighbor algorithm, see documentation of function `FindClusters`, R-package `Seurat`
`hclust_method`	the agglomeration method to be used (default = average). See documentation of `stats::hclust`
`hclust_distance`	distance measure to be used: euclidean (default) or manhattan, see documentation of `stats::dist`
`cores`	integer, number of PC cores for parallel execution
`round_digits`	integer, number of decimal places to keep when showing the relative frequency of cells in each bubble
`show_simple_count`	logical, if `show_simple_count`=T, cell counts in each bubble will be divided by 1,000 to improve readability. This is only useful for samples that are composed of millions of cells.
`verbose`	logical, progress messages

Details

For Louvain clustering get_bubbletree_graph uses the function FindClusters implemented in R-package Seurat. For additional information on the clustering procedure see the documentation of FindClusters. To organize the resulting clusters in a hierarchical dendrogram, then the following steps are performed:

1. In bootrap iteration $b$ from $1:B$

2. draw up to $N_{eff}$ number of cells at random from each cluster without replacement

3. compute distances (in space $A^{n\times f}$ ) between all pairs of cells in cluster $i$ and cluster $j$

4. compute mean distance between cluster $i$ and $j$ and populate inter-cluster distance matrix $D_{b}^{k \times k}$

5. perform hierarchical clustering with user-specified agglomeration method based on $D_{b}^{k \times k}$ to generate dendrogram $H_b$

6. quantify branch robustness in $H$ by counting how many times each branch is found among bootrap dendrograms $H_b$

Value

`A`	input x matrix
`k`	number of clusters
`r`	clustering resolution
`ph`	boot_ph: bootstrap dendrograms $H_b$ ; main_ph: bubbletree $H$
`pair_dist`	inter-cluster distances used to generate the dendrograms
`cluster`	cluster assignments of each cell
`input_par`	list of all input parameters
`tree`	ggtree bubbletree object
`tree_meta`	meta-data associated with the bubbletree

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

b <- get_bubbletree_graph(x = A,
                          r = 1,
                          B = 200,
                          N_eff = 100,
                          n_start = 20,
                          iter_max = 100,
                          algorithm = "original",
                          knn_k = 20,
                          hclust_method = "average",
                          hclust_distance = "euclidean",
                          cores = 1,
                          round_digits = 2,
                          show_simple_count = FALSE)

b$tree
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

b <- get_bubbletree_graph(x = A,
                          r = 1,
                          B = 200,
                          N_eff = 100,
                          n_start = 20,
                          iter_max = 100,
                          algorithm = "original",
                          knn_k = 20,
                          hclust_method = "average",
                          hclust_distance = "euclidean",
                          cores = 1,
                          round_digits = 2,
                          show_simple_count = FALSE)

b$tree

k-means clustering and hierarchical grouping of $k$ clusters (bubbles)

Description

get_bubble_kmeans takes two main inputs:

1. numeric matrix $A^{n \times f}$ , which represents a low-dimensional projection (obtained e.g. by PCA) of the original high-dimensional scRNA-seq data, with $n$ rows as cells and $f$ columns as low-dimension features.

2. number $k$ of clusters

The function get_bubble_kmeans performs two main operations. First, it performs k-means clustering to identify groups (bubbles) of transcriptionally similar cells. Second, it organizes the bubbles in a hierarchical dendrogram (bubbletree) which adequatly represents inter-cluster relationships.

Usage

get_bubbletree_kmeans(x,
                      k,
                      B = 200,
                      N_eff = 100,
                      n_start = 1000,
                      iter_max = 300,
                      kmeans_algorithm = "MacQueen",
                      hclust_distance = "euclidean",
                      hclust_method = "average",
                      cores = 1,
                      round_digits = 2,
                      show_simple_count = FALSE,
                      verbose = TRUE)
get_bubbletree_kmeans(x,
                      k,
                      B = 200,
                      N_eff = 100,
                      n_start = 1000,
                      iter_max = 300,
                      kmeans_algorithm = "MacQueen",
                      hclust_distance = "euclidean",
                      hclust_method = "average",
                      cores = 1,
                      round_digits = 2,
                      show_simple_count = FALSE,
                      verbose = TRUE)

Arguments

`x`	numeric matrix ( $A^{n\times f}$ with $n$ cells, and $f$ low-dimensional projections of the original single cell RNA-seq dataset)
`k`	integer, number of clusters
`B`	integer, number of bootstrap iterations to perform in order to generate bubbletree
`N_eff`	integer, number of cells to draw randomly from each cluster when computing inter-cluster distances
`n_start`, `iter_max`, `kmeans_algorithm`	parameters for k-means clustering, see documentation of function `k-means`, R-package `stats`
`hclust_distance`	distance measure to be used: euclidean (default) or manhattan, see documentation of `stats::dist`
`hclust_method`	the agglomeration method to be used, default = average. See documentation of `stats::hclust`
`cores`	integer, number of PC cores for parallel execution
`round_digits`	integer, number of decimal places to keep when showing the relative frequency of cells in each bubble
`show_simple_count`	logical, if `show_simple_count`=T, cell counts in each bubble will be divided by 1,000 to improve readability. This is only useful for samples that are composed of millions of cells.
`verbose`	logical, progress messages

Details

For k-means clustering get_bubble_kmeans uses the function kmeans implemented in R-package stats (version 4.2.0). For additional information on the clustering procedure see the documentation of kmeans. To organize the resulting clusters in a hierarchical dendrogram these steps are performed:

1. In bootrap iteration $b$ from $1:B$

2. draw up to $N_{eff}$ number of cells at random from each cluster without replacement

3. compute distances (in space $A^{n\times f}$ ) between pairs of cells in cluster $i$ and cluster $j$

4. compute mean distance between cluster $i$ and $j$ and populate inter-cluster distance matrix $D_{b}^{k \times k}$

5. perform hierarchical clustering with user-specified agglomeration method based on $D_{b}^{k \times k}$ to generate dendrogram $H_b$

6. quantify branch robustness in $H$ by counting how many times each branch is found among the bootrap dendrograms $H_b$

Value

`A`	input matrix x
`k`	number of clusters
`km`	k-means clustering results identical to those generated by function k-means from R-package stats
`ph`	boot_ph: bootstrap dendrograms $H_b$ ; main_ph: bubbletree $H$
`pair_dist`	inter-cluster distances used to generate the dendrograms
`cluster`	cluster assignments of each cell
`input_par`	list of all input parameters
`tree`	ggtree bubbletree object
`tree_meta`	meta-data associated with the bubbletree

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

b <- get_bubbletree_kmeans(x = A,
                           k = 8,
                           B = 200,
                           N_eff = 100,
                           cores = 1,
                           round_digits = 1,
                           show_simple_count = FALSE,
                           kmeans_algorithm = "MacQueen",
                           hclust_distance = "euclidean",
                           hclust_method = "average")

b$tree
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

b <- get_bubbletree_kmeans(x = A,
                           k = 8,
                           B = 200,
                           N_eff = 100,
                           cores = 1,
                           round_digits = 1,
                           show_simple_count = FALSE,
                           kmeans_algorithm = "MacQueen",
                           hclust_distance = "euclidean",
                           hclust_method = "average")

b$tree

Visualization of categorical cell features using tile plots

Description

get_cat_tiles creates tile plot to visualize the relative frequency of categorical cell features between and within the bubbles of a bubbletree

Usage

get_cat_tiles(btd,
              f,
              integrate_vertical,
              round_digits = 2,
              tile_text_size = 3,
              tile_bw = FALSE,
              x_axis_name = "Feature",
              rotate_x_axis_labels = TRUE)
get_cat_tiles(btd,
              f,
              integrate_vertical,
              round_digits = 2,
              tile_text_size = 3,
              tile_bw = FALSE,
              x_axis_name = "Feature",
              rotate_x_axis_labels = TRUE)

Arguments

`btd`	bubbletree object
`f`	character vector, categorical cell features
`integrate_vertical`	logical, if integrate_vertical=TRUE: relative frequency of the features is shown in each bubble, if integrate_vertical=FALSE: relative frequencies of the features is shown within each bubble
`round_digits`	integer, number of decimal places to keep when showing the relative frequency of cells in each bubble
`tile_text_size`	integer, size of tile labels
`x_axis_name`	character, x-axis title
`rotate_x_axis_labels`	logical, should the x-axis labels be shown horizontally (rotate_x_axis_labels = FALSE) or vertically (rotate_x_axis_labels = TRUE)
`tile_bw`	logical, tile grayscale (tile_bw = TRUE) vs. color (tile_bw = FALSE, default)

Details

get_cat_tiles uses two main inputs:

1. bubbletree object

2. character vector of categorical cell features.

The order of the cells used to generat the bubbletree (input 1.) should correspond to the order of cells in the vector of categorical cell features (input 2.)

This function computes:

1. with integrate_vertical=T: relative frequencies of each feature across the different bubbles

2. with integrate_vertical\=F: within-bubble relative frequencies (composition) of different features

Value

`plot`	ggplot2, tile plot
`table`	data.frame, raw data used to generate the plot

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
f <- d_500$f


b <- get_bubbletree_graph(x = A,
                          r = 0.8,
                          N_eff = 100)


g_v <- get_cat_tiles(btd = b,
                     f = f,
                     integrate_vertical = TRUE,
                     round_digits = 2,
                     tile_text_size = 3,
                     x_axis_name = "Feature",
                     rotate_x_axis_labels = TRUE)


g_h <- get_cat_tiles(btd = b,
                    f = f,
                    integrate_vertical = FALSE,
                    round_digits = 2,
                    tile_text_size = 3,
                    x_axis_name = "Feature",
                    rotate_x_axis_labels = TRUE)

b$tree|g_v$plot|g_h$plot
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
f <- d_500$f


b <- get_bubbletree_graph(x = A,
                          r = 0.8,
                          N_eff = 100)


g_v <- get_cat_tiles(btd = b,
                     f = f,
                     integrate_vertical = TRUE,
                     round_digits = 2,
                     tile_text_size = 3,
                     x_axis_name = "Feature",
                     rotate_x_axis_labels = TRUE)


g_h <- get_cat_tiles(btd = b,
                    f = f,
                    integrate_vertical = FALSE,
                    round_digits = 2,
                    tile_text_size = 3,
                    x_axis_name = "Feature",
                    rotate_x_axis_labels = TRUE)

b$tree|g_v$plot|g_h$plot

Gini impurity index computed for a clustering solution and a vector of categorical cell feature labels

Description

How well is a set of categorical feature labels (e.g. cell type predictions) partitioned accross the different clusters of a clustering solution? We can assess this using the Gini impurity index (see details below).

Inputs are two equal-sized vectors:

1) clusters IDs

2) labels

Output:

1) cluster-specific purity -> Gini impurity (GI) index

2) clustering solution impurity -> Weighted Gini impurity (WGI) index

Usage

get_gini(labels, clusters)
get_gini(labels, clusters)

Arguments

`labels`	character or numeric vector of labels
`clusters`	character or numeric vector of cluster IDs

Details

To quantify the purity of a cluster (or bubble) $i$ with $n_i$ number of cells, each of which carries one of $L$ possible labels (e.g. cell type), we computed the Gini impurity index:

$GI_i=\sum_{j=1}^{L} \pi_{ij}(1-\pi_{ij}),$

with $\pi_{ij}$ as the relative frequency of label $j$ in cluster $i$ . In homogeneous ('pure') clusters most cells carry a distinct label. Hence, the $\pi$ 's are close to either 1 or 0, and GI takes on a small value close to zero. In ‘impure’ clusters cells carry a mixture of different labels. In this case most $\pi$ are far from either 1 or 0, and GI diverges from 0 and approaches 1. If the relative frequencies of the different labels in cluster $i$ are equal to the (background) relative frequencies of the labels in the sample, then cluster $i$ is completely ‘impure’.

To compute the overall Gini impurity of a bubbletree, which represents a clustering solution with $k$ bubbles, we estimated the weighted Gini impurity (WGI) by computing the weighted (by the cluster size) average of the GIs:

$WGI=\sum_{i=1}^{k} GI_i {n_i}/{n},$

with $n_i$ as the number of cells in cluster $i$ and $n=\sum_i n_i$ .

Value

`gi`	Gini impurity of each bubble
`wgi`	Weighted Gini impurity index of the bubbletree

Author(s)

Simo Kitanovski [email protected]

Examples

get_gini(labels = base::sample(x = LETTERS[1:4], size = 100, replace = TRUE),
         clusters = base::sample(x = letters[1:4], size = 100, replace = TRUE))
get_gini(labels = base::sample(x = LETTERS[1:4], size = 100, replace = TRUE),
         clusters = base::sample(x = letters[1:4], size = 100, replace = TRUE))

Gini impurity index computed for a list of clustering solutions obtained by functions `get_k` or `get_r` and a vector of categorical cell feature labels

Description

Given The Gini impurity (GI) index allows us to quantitatively evaluate how well a set of labels (categorical features) are split across a set of bubbles. We have a completely perfect split (GI = 0) when each bubble is 'pure', i.e. each bubble contains labels coming from distinct a class. In contrast to this, we have completely imperfect split (GI = 1) when the relative frequency distribution of the labels in each bubble is identical to the background relative frequency distribution of the labels.

Cell type predictions are a type of categorical features that are often used to evaluate the goodness of the clustering. get_gini_k takes as input: 1) a vector of labels for each cell (e.g. cell types) and 2) object returned by function get_k or get_r. Then it computes for each $k$ or $r$ the cluster purity and weightred gini impurity of each clustering solution mean GI, which is another way of finding an optimal clustering resolution.

Usage

get_gini_k(labels, obj)
get_gini_k(labels, obj)

Arguments

`labels`	character/factor vector of labels
`obj`	object returned by functions `get_k` or `get_r`

Details

To quantify the purity of a cluster (or bubble) $i$ with $n_i$ number of cells, each of which carries one of $L$ possible labels (e.g. cell type), we computed the Gini impurity index:

$GI_i=\sum_{j=1}^{L} \pi_{ij}(1-\pi_{ij}),$

$WGI=\sum_{i=1}^{k} GI_i {n_i}/{n},$

with $n_i$ as the number of cells in cluster $i$ and $n=\sum_i n_i$ .

Value

`gi_summary`	GI for each bubble of a clustering solution with clustering resolution $k$ or $r$
`wgi_summary`	WGI for each clustering solution with clustering resolution $k$ or $r$

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
f <- d_500$f


b_k <- get_k(x = A,
           ks = 1:5,
           B_gap = 5,
           n_start = 100,
           iter_max = 200,
           kmeans_algorithm = "MacQueen",
           cores = 1)


b_r <- get_r(x = A,
            rs = c(0.1, 0.5, 1),
            B_gap = 5,
            n_start = 20,
            iter_max = 100,
            algorithm = "original",
            cores = 1)

get_gini_k(labels = f, obj = b_k)
get_gini_k(labels = f, obj = b_r)
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
f <- d_500$f


b_k <- get_k(x = A,
           ks = 1:5,
           B_gap = 5,
           n_start = 100,
           iter_max = 200,
           kmeans_algorithm = "MacQueen",
           cores = 1)


b_r <- get_r(x = A,
            rs = c(0.1, 0.5, 1),
            B_gap = 5,
            n_start = 20,
            iter_max = 100,
            algorithm = "original",
            cores = 1)

get_gini_k(labels = f, obj = b_k)
get_gini_k(labels = f, obj = b_r)

Finding optimal number $k$ of clusters

Description

To perform k-means clustering we must specify a number $k$ of clusters. Data-driven metrics, such as the Gap statistic or the within-cluster sum of squares (WCSS), can be used to infer appropriate $k$ from the data. get_k computes the Gap statistic and WCSS for a number of clusters $ks$ .

Usage


get_k(x,
     ks,
     B_gap = 20,
     n_start = 1000,
     iter_max = 300,
     kmeans_algorithm = "MacQueen",
     cores = 1,
     verbose = TRUE)
get_k(x,
     ks,
     B_gap = 20,
     n_start = 1000,
     iter_max = 300,
     kmeans_algorithm = "MacQueen",
     cores = 1,
     verbose = TRUE)

Arguments

`x`	numeric matrix $A^{n x f}$ with $n$ cells, and $f$ low-dimensional projections
`ks`	integer vector, $k$ values to consider
`B_gap`	integer, number of Monte Carlo ("bootstrap") samples taken when computing the Gap statistic (see documentation of function `clusGap`, R-package `cluster`)
`n_start`, `iter_max`, `kmeans_algorithm`	parameters for k-means clustering, see documentation of function `kmeans`, R-package `stats`
`cores`	integer, number of PC cores for parallel execution
`verbose`	logical, progress messages

Details

To compute the Gap statistic get_k adapts the algorithm of function clustGap from R-package cluster (version 2.1.3). For k-means clustering get_k uses the function kmeans implemented in R-package stats (version 4.2.0). For additional information see the respective documentations.

Value

`boot_obj`	The results: k-means clustering solutions, the Gap statistic and WCSS
`gap_stats_summary`, `wcss_stats_summary`	main results; Gap statistic and WCSS estimates. Means, standard errors and 95% confidence intervals are provided for each $k$
`gap_stats`, `wcss_stats`	intermediate results; Gap statistic and WCSS estimates for each $k$ and bootstrap iteration $b$

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

b <- get_k(x = A,
           ks = 1:5,
           B_gap = 10,
           n_start = 100,
           iter_max = 200,
           kmeans_algorithm = "MacQueen",
           cores = 1,
           verbose = TRUE)

b$gap_stats_summary
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A

b <- get_k(x = A,
           ks = 1:5,
           B_gap = 10,
           n_start = 100,
           iter_max = 200,
           kmeans_algorithm = "MacQueen",
           cores = 1,
           verbose = TRUE)

b$gap_stats_summary

Visualization of numeric cell features using tile plots

Description

get_num_tiles creates tile plot to visualize a summary (e.g. mean, median or sum) of a numeric cell feature (e.g. gene expression of a specific gene) in each bubble of a bubbletree

Usage

get_num_tiles(btd,
              fs,
              summary_function,
              round_digits = 2,
              tile_text_size = 3,
              tile_bw = FALSE,
              x_axis_name = "Feature",
              rotate_x_axis_labels = TRUE)
get_num_tiles(btd,
              fs,
              summary_function,
              round_digits = 2,
              tile_text_size = 3,
              tile_bw = FALSE,
              x_axis_name = "Feature",
              rotate_x_axis_labels = TRUE)

Arguments

`btd`	bubbletree object
`fs`	numeric vector or matrix, numeric cell features
`summary_function`	character, "mean", "median" or "sum", "pct nonzero", "pct zero", summaries are allowed
`round_digits`	integer, number of decimal places to keep when showing the relative frequency of cells in each bubble
`tile_text_size`	integer, size of tile labels
`x_axis_name`	character, x-axis title
`rotate_x_axis_labels`	logical, should the x-axis labels be shown horizontally (`rotate_x_axis_labels` = FALSE) or vertically (`rotate_x_axis_labels` = TRUE)
`tile_bw`	logical, tile grayscale (tile_bw = TRUE) vs. color (`tile_bw` = FALSE, default)

Details

get_num_tiles uses two main inputs:

1. bubbletree object

2. numeric vector or matrix of numeric cell features.

The order of the cells used to generat the bubbletree (input 1.) should correspond to the order of cells in the vector/matrix of numeric cell features (input 2.)

This function computes summaries of numeric cell feature in each bubble: 1. mean = mean of feature 2. median = median of feature 3. sum = sum of feature 4. pct nonzero = sum of cells with feature > 0 5. pct zero = sum of cells with feature = 0

Important note: NA and NULL values are omitted.

Value

`plot`	ggplot2, tile plot
`table`	data.frame, raw data used to generate the plot

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
fs <- d_500$fs

b <- get_bubbletree_kmeans(x = A, k = 8)

g <- get_num_tiles(btd = b,
                   fs = fs,
                   summary_function = "mean",
                   round_digits = 2,
                   tile_text_size = 3,
                   tile_bw = TRUE,
                   x_axis_name = "Gene expression",
                   rotate_x_axis_labels = TRUE)


b$tree|g$plot
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
fs <- d_500$fs

b <- get_bubbletree_kmeans(x = A, k = 8)

g <- get_num_tiles(btd = b,
                   fs = fs,
                   summary_function = "mean",
                   round_digits = 2,
                   tile_text_size = 3,
                   tile_bw = TRUE,
                   x_axis_name = "Gene expression",
                   rotate_x_axis_labels = TRUE)


b$tree|g$plot

Visualization of numeric cell features using violin plots

Description

get_num_violins creates violin plot to visualize the distribution of of numeric cell features (e.g. gene expressions) in each bubble of a bubbletree

Usage


get_num_violins(btd,
                fs,
                x_axis_name = "Feature distribution",
                rotate_x_axis_labels = TRUE)
get_num_violins(btd,
                fs,
                x_axis_name = "Feature distribution",
                rotate_x_axis_labels = TRUE)

Arguments

`btd`	bubbletree object
`fs`	numeric vector or matrix, numeric cell features
`x_axis_name`	character, x-axis title
`rotate_x_axis_labels`	logical, should the x-axis labels be shown horizontally (`rotate_x_axis_labels` = FALSE) or vertically (`rotate_x_axis_labels` = TRUE)

Details

get_num_violins uses two main inputs:

1. bubbletree object

2. numeric vector or matrix of numeric cell features.

The order of the cells used to generat the bubbletree (input 1.) should correspond to the order of cells in the vector/matrix of numeric cell features (input 2.)

This function visualizes densities of numeric cell feature in the different bubble.

Value

`plot`	ggplot2, violin plot
`table`	data.frame, raw data used to generate the plot

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
fs <- d_500$fs

b <- get_bubbletree_graph(x = A, r = 0.8)

g <- get_num_violins(btd = b,
                     fs = fs,
                     x_axis_name = "Feature distribution",
                     rotate_x_axis_labels = TRUE)


b$tree|g$plot
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A
fs <- d_500$fs

b <- get_bubbletree_graph(x = A, r = 0.8)

g <- get_num_violins(btd = b,
                     fs = fs,
                     x_axis_name = "Feature distribution",
                     rotate_x_axis_labels = TRUE)


b$tree|g$plot

Finding optimal clustering resulution $r$ and number of communities $k'$

Description

To perform Louvain clustering we must specify a clustering resulution $r$ . Data-driven metrics, such as the Gap statistic or the within-cluster sum of squares (WCSS) can be used to infer appropriate $r$ from the data. get_r computes the Gap statistic and WCSS for a vector of clustering resolutions $rs$ .

Usage


get_r(x,
     rs,
     B_gap = 20,
     n_start = 20,
     iter_max = 100,
     algorithm = "original",
     knn_k = 20,
     cores = 1,
     verbose = TRUE)
get_r(x,
     rs,
     B_gap = 20,
     n_start = 20,
     iter_max = 100,
     algorithm = "original",
     knn_k = 20,
     cores = 1,
     verbose = TRUE)

Arguments

`x`	numeric matrix $A^{n x f}$ with $n$ cells, and $f$ low-dimensional projections
`rs`	number vector, $r$ values to consider
`B_gap`	integer, number of Monte Carlo ("bootstrap") samples taken when computing the Gap statistic (see documentation of function `clusGap`, R-package `cluster`)
`n_start`, `iter_max`	parameters for Louvain clustering, see documentation of function `FindClusters`, R-package `Seurat`
`algorithm`	character, four clustering algorithms: 'original', 'LMR', 'SLM' and 'Leiden', see documentation of function `FindClusters`, R-package `Seurat`
`knn_k`	integer, defines k for the k-nearest neighbor algorithm, see documentation of function `FindClusters`, R-package `Seurat`
`cores`	integer, number of PC cores for parallel execution
`verbose`	logical, progress messages

Details

To compute the Gap statistic get_r adapts the algorithm of function clustGap from R-package cluster (version 2.1.3). For Louvain clustering get_r uses the function FindClusters implemented in the R-package Seurat. For additional information see the respective documentations.

Value

`boot_obj`	The results: k-means clustering solutions, the Gap statistic and WCSS
`gap_stats_summary`, `wcss_stats_summary`	main results; Gap statistic and WCSS estimates. Means, standard errors and 95% confidence intervals are provided for each $r$ and $k'$
`gap_stats`, `wcss_stats`	intermediate results; Gap statistic and WCSS estimates for each $r$ and $k'$ and bootstrap iteration $b$

Author(s)

Simo Kitanovski [email protected]

Examples

# input data
data("d_500", package = "scBubbletree")
A <- d_500$A


b <- get_r(x = A,
           rs = c(0.1, 0.5, 1),
           B_gap = 10,
           n_start = 20,
           iter_max = 100,
           algorithm = "original",
           cores = 1,
           verbose = TRUE)

b$gap_stats_summary
# input data
data("d_500", package = "scBubbletree")
A <- d_500$A


b <- get_r(x = A,
           rs = c(0.1, 0.5, 1),
           B_gap = 10,
           n_start = 20,
           iter_max = 100,
           algorithm = "original",
           cores = 1,
           verbose = TRUE)

b$gap_stats_summary

Package 'scBubbletree'

Help Index

The R package scBubbletree

Description

Details

Author(s)

See Also

Comparison of two bubbletrees generated from the same scRNA-seq data

Description

Usage

Arguments

Details

Value

Author(s)

See Also

Examples

Dataset: 500 PBMCs

Description

Usage

Format

Details

Source

Examples

Dataset: scRNA-seq data of 3,918 cells from 5 adenocarcinoma cell lines

Description

Usage

Format

Details

Source

References

Examples

Build bubbletree from matrix AAA of low-dimensional projections and vector cscscs of externally generated cluster IDs

Description

Usage

Arguments

Details

Value

Author(s)

See Also

Examples

Louvain clustering and hierarchical grouping of k′k'k′ clusters (bubbles)

Description

Usage

Arguments

Details

Value

Author(s)

See Also

Examples

k-means clustering and hierarchical grouping of kkk clusters (bubbles)

Description

Usage

Arguments

Details

Value

Author(s)

See Also

Examples

Visualization of categorical cell features using tile plots

Description

Usage

Arguments

Details

Value

Author(s)

See Also

Examples

Gini impurity index computed for a clustering solution and a vector of categorical cell feature labels

Description

Usage

Arguments

Details

Value

Author(s)

See Also

Examples

Gini impurity index computed for a list of clustering solutions obtained by functions get_k or get_r and a vector of categorical cell feature labels

Description

Usage

Arguments

Build bubbletree from matrix $A$ of low-dimensional projections and vector $cs$ of externally generated cluster IDs

Louvain clustering and hierarchical grouping of $k'$ clusters (bubbles)

k-means clustering and hierarchical grouping of $k$ clusters (bubbles)

Gini impurity index computed for a list of clustering solutions obtained by functions `get_k` or `get_r` and a vector of categorical cell feature labels

Finding optimal number $k$ of clusters

Finding optimal clustering resulution $r$ and number of communities $k'$