This vignette will demonstrate a full single-cell lineage analysis workflow,
with particular emphasis on the processes of lineage reconstruction and
pseudotime inference. We will make use of the slingshot
package proposed in
(Street et al. 2017) and show how it may be applied in a broad range of settings.
The goal of slingshot
is to use clusters of cells to uncover global structure
and convert this structure into smooth lineages represented by one-dimensional
variables, called “pseudotime.” We provide tools for learning cluster
relationships in an unsupervised or semi-supervised manner and constructing
smooth curves representing each lineage, along with visualization methods for
each step.
The minimal input to slingshot
is a matrix representing the cells in a
reduced-dimensional space and a vector of cluster labels. With these two
inputs, we then:
getLineages
function.getCurves
function.We will work with two simulated datasets in this vignette. The first (referred
to as the “single-trajectory” dataset) is generated below and designed to
represent a single lineage in which one third of the genes are associated with
the transition. This dataset will be contained in a SingleCellExperiment
object (Lun and Risso 2017) and will be used to demonstrate a full
“start-to-finish” workflow.
# generate synthetic count data representing a single lineage
means <- rbind(
# non-DE genes
matrix(rep(rep(c(0.1,0.5,1,2,3), each = 300),100),
ncol = 300, byrow = TRUE),
# early deactivation
matrix(rep(exp(atan( ((300:1)-200)/50 )),50), ncol = 300, byrow = TRUE),
# late deactivation
matrix(rep(exp(atan( ((300:1)-100)/50 )),50), ncol = 300, byrow = TRUE),
# early activation
matrix(rep(exp(atan( ((1:300)-100)/50 )),50), ncol = 300, byrow = TRUE),
# late activation
matrix(rep(exp(atan( ((1:300)-200)/50 )),50), ncol = 300, byrow = TRUE),
# transient
matrix(rep(exp(atan( c((1:100)/33, rep(3,100), (100:1)/33) )),50),
ncol = 300, byrow = TRUE)
)
counts <- apply(means,2,function(cell_means){
total <- rnbinom(1, mu = 7500, size = 4)
rmultinom(1, total, cell_means)
})
rownames(counts) <- paste0('G',1:750)
colnames(counts) <- paste0('c',1:300)
sce <- SingleCellExperiment(assays = List(counts = counts))
The second dataset (the “bifurcating” dataset) consists of a matrix of
coordinates (as if obtained by PCA, ICA, diffusion maps, etc.) along with
cluster labels generated by \(k\)-means clustering. This dataset represents a
bifurcating trajectory and it will allow us to demonstrate some of the
additional functionality offered by slingshot
.
library(slingshot, quietly = FALSE)
data("slingshotExample")
rd <- slingshotExample$rd
cl <- slingshotExample$cl
dim(rd) # data representing cells in a reduced dimensional space
## [1] 140 2
length(cl) # vector of cluster labels
## [1] 140
To begin our analysis of the single lineage dataset, we need to reduce the dimensionality of our data and filtering out uninformative genes is a typical first step. This will greatly improve the speed of downstream analyses, while keeping the loss of information to a minimum.
For the gene filtering step, we retained any genes robustly expressed in at least enough cells to constitute a cluster, making them potentially interesting cell-type marker genes. We set this minimum cluster size to 10 cells and define a gene as being “robustly expressed” if it has a simulated count of at least 3 reads.
# filter genes down to potential cell-type markers
# at least M (15) reads in at least N (15) cells
geneFilter <- apply(assays(sce)$counts,1,function(x){
sum(x >= 3) >= 10
})
sce <- sce[geneFilter, ]
Another important early step in most RNA-Seq analysis pipelines is the choice of
normalization method. This allows us to remove unwanted technical or biological
artifacts from the data, such as batch, sequencing depth, cell cycle effects,
etc. In practice, it is valuable to compare a variety of normalization
techniques and compare them along different evaluation criteria, for which we
recommend the scone
package (Cole and Risso 2018). We also note that the order of these
steps may change depending on the choice of method. ZINB-WaVE (Risso et al. 2018)
performs dimensionality reduction while accounting for technical variables and
MNN (Haghverdi et al. 2018) corrects for batch effects after dimensionality reduction.
Since we are working with simulated data, we do not need to worry about batch effects or other potential confounders. Hence, we will proceed with full quantile normalization, a well-established method which forces each cell to have the same distribution of expression values.
FQnorm <- function(counts){
rk <- apply(counts,2,rank,ties.method='min')
counts.sort <- apply(counts,2,sort)
refdist <- apply(counts.sort,1,median)
norm <- apply(rk,2,function(r){ refdist[r] })
rownames(norm) <- rownames(counts)
return(norm)
}
assays(sce)$norm <- FQnorm(assays(sce)$counts)
The fundamental assumption of slingshot
is that cells which are
transcriptionally similar will be close to each other in some
reduced-dimensional space. Since we use Euclidean distances in constructing
lineages and measuring pseudotime, it is important to have a low-dimensional
representation of the data.
There are many methods available for this task and we will intentionally avoid
the issue of determining which is the “best” method, as this likely depends on
the type of data, method of collection, upstream computational choices, and many
other factors. We will demonstrate two methods of dimensionality reduction:
principal components analysis (PCA) and uniform manifold approximation and
projection (UMAP, via the uwot
package).
When performing PCA, we do not scale the genes by their variance because we do not believe that all genes are equally informative. We want to find signal in the robustly expressed, highly variable genes, not dampen this signal by forcing equal variance across genes. When plotting, we make sure to set the aspect ratio, so as not to distort the perceived distances.
pca <- prcomp(t(log1p(assays(sce)$norm)), scale. = FALSE)
rd1 <- pca$x[,1:2]
plot(rd1, col = rgb(0,0,0,.5), pch=16, asp = 1)
library(uwot)
## Loading required package: Matrix
##
## Attaching package: 'Matrix'
## The following object is masked from 'package:S4Vectors':
##
## expand
rd2 <- uwot::umap(t(log1p(assays(sce)$norm)))
colnames(rd2) <- c('UMAP1', 'UMAP2')
plot(rd2, col = rgb(0,0,0,.5), pch=16, asp = 1)
We will add both dimensionality reductions to the SingleCellExperiment
object,
but continue our analysis focusing on the PCA results.
reducedDims(sce) <- SimpleList(PCA = rd1, UMAP = rd2)
The final input to slingshot
is a vector of cluster labels for the cells. If
this is not provided, slingshot
will treat the data as a single cluster and
fit a standard principal curve. However, we recommend clustering the cells even
in datasets where only a single lineage is expected, as it allows for the
potential discovery of novel branching events.
The clusters identified in this step will be used to determine the global structure of the underlying lineages (that is, their number, when they branch off from one another, and the approximate locations of those branching events). This is different than the typical goal of clustering single-cell data, which is to identify all biologically relevant cell types present in the dataset. For example, when determining global lineage structure, there is no need to distinguish between immature and mature neurons since both cell types will, presumably, fall along the same segment of a lineage.
For our analysis, we implement two clustering methods which similarly assume
that Euclidean distance in a low-dimensional space reflect biological
differences between cells: Gaussian mixture modeling and \(k\)-means. The former
is implemented in the mclust
package (Scrucca et al. 2016) and features an automated
method for determining the number of clusters based on the Bayesian information
criterion (BIC).
library(mclust, quietly = TRUE)
## Package 'mclust' version 5.4.9
## Type 'citation("mclust")' for citing this R package in publications.
##
## Attaching package: 'mclust'
## The following object is masked from 'package:mgcv':
##
## mvn
cl1 <- Mclust(rd1)$classification
colData(sce)$GMM <- cl1
library(RColorBrewer)
plot(rd1, col = brewer.pal(9,"Set1")[cl1], pch=16, asp = 1)
While \(k\)-means does not have a similar functionality, we have shown in (Street et al. 2017) that simultaneous principal curves are quite robust to the choice of \(k\), so we select a \(k\) of 4 somewhat arbitrarily. If this is too low, we may miss a true branching event and if it is too high or there is an abundance of small clusters, we may begin to see spurious branching events.
cl2 <- kmeans(rd1, centers = 4)$cluster
colData(sce)$kmeans <- cl2
plot(rd1, col = brewer.pal(9,"Set1")[cl2], pch=16, asp = 1)
At this point, we have everything we need to run slingshot
on our simulated
dataset. This is a two-step process composed of identifying the
global lineage structure with a cluster-based minimum spanning tree (MST) and
fitting simultaneous principal curves to describe each lineage.
These two steps can be run separately with the getLineages
and getCurves
functions, or together with the wrapper function, slingshot
(recommended). We
will use the wrapper function for the analysis of the single-trajectory dataset,
but demonstrate the usage of the individual functions later, on the bifurcating
dataset.
The slingshot
wrapper function performs both steps of trajectory inference in
a single call. The necessary inputs are a reduced dimensional matrix of
coordinates and a set of cluster labels. These can be separate objects or, in
the case of the single-trajectory data, elements contained in a
SingleCellExperiment
object.
To run slingshot
with the dimensionality reduction produced by PCA and cluster
labels identified by Gaussian mixutre modeling, we would do the following:
sce <- slingshot(sce, clusterLabels = 'GMM', reducedDim = 'PCA')
As noted above, if no clustering results are provided, it is assumed that all
cells are part of the same cluster and a single curve will be constructed. If no
dimensionality reduction is provided, slingshot
will use the first element of
the list returned by reducedDims
.
The output is a SingleCellExperiment
object with slingshot
results
incorporated. All of the results are stored in a PseudotimeOrdering
object,
which is added to the colData
of the original object and can be accessed via
colData(sce)$slingshot
. Additionally, all inferred pseudotime variables (one
per lineage) are added to the colData
, individually. To extract all
slingshot
results in a single object, we can use either the
as.PseudotimeOrdering
or as.SlingshotDataSet
functions, depending on the
form in which we want it. PseudotimeOrdering
objects are an extension of
SummarizedExperiment
objects, which are flexible containers that will be
useful for most purposes. SlingshotDataSet
objects are primarily used for
visualization, as several plotting methods are included with the package. Below,
we visuzalize the inferred lineage for the single-trajectory data with points
colored by pseudotime.
summary(sce$slingPseudotime_1)
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.000 8.631 21.121 21.414 34.363 43.185
library(grDevices)
colors <- colorRampPalette(brewer.pal(11,'Spectral')[-6])(100)
plotcol <- colors[cut(sce$slingPseudotime_1, breaks=100)]
plot(reducedDims(sce)$PCA, col = plotcol, pch=16, asp = 1)
lines(SlingshotDataSet(sce), lwd=2, col='black')
We can also see how the lineage structure was intially estimated by the
cluster-based minimum spanning tree by using the type
argument.
plot(reducedDims(sce)$PCA, col = brewer.pal(9,'Set1')[sce$GMM], pch=16, asp = 1)
lines(SlingshotDataSet(sce), lwd=2, type = 'lineages', col = 'black')
After running slingshot
, we are often interested in finding genes that
change their expression over the course of development. We will demonstrate this
type of analysis using the tradeSeq
package (Van den Berge et al. 2020).
For each gene, we will fit a general additive model (GAM) using a negative
binomial noise distribution to model the (potentially nonlinear)
relationshipships between gene expression and pseudotime. We will then test for
significant associations between expression and pseudotime using the
associationTest
.
library(tradeSeq)
# fit negative binomial GAM
sce <- fitGAM(sce)
# test for dynamic expression
ATres <- associationTest(sce)
We can then pick out the top genes based on p-values and visualize their expression over developmental time with a heatmap. Here we use the top 250 most dynamically expressed genes.
topgenes <- rownames(ATres[order(ATres$pvalue), ])[1:250]
pst.ord <- order(sce$slingPseudotime_1, na.last = NA)
heatdata <- assays(sce)$counts[topgenes, pst.ord]
heatclus <- sce$GMM[pst.ord]
heatmap(log1p(heatdata), Colv = NA,
ColSideColors = brewer.pal(9,"Set1")[heatclus])
Here, we provide further details and highlight some additional functionality of
the slingshot
package. We will use the included slingshotExample
dataset for
illustrative purposes. This dataset was designed to represent cells in a low
dimensional space and comes with a set of cluster labels generated by \(k\)-means
clustering. Rather than constructing a full SingleCellExperiment
object, which
requires gene-level data, we will use the low-dimensional matrix of coordinates
directly and provide the cluster labels as an additional argument.
The getLineages
function takes as input an n
\(\times\) p
matrix and a vector of
clustering results of length n
. It maps connections between adjacent clusters
using a minimum spanning tree (MST) and identifies paths through these
connections that represent lineages. The output of this function is a
PseudotimeOrdering
containing the inputs as well as the inferred MST
(represented by an igraph
object) and lineages (ordered vectors of cluster
names).
This analysis can be performed in an entirely unsupervised manner or in a
semi-supervised manner by specifying known initial and terminal point clusters.
If we do not specify a starting point, slingshot
selects one based on
parsimony, maximizing the number of clusters shared between lineages before a
split. If there are no splits or multiple clusters produce the same parsimony
score, the starting cluster is chosen arbitrarily.
In the case of our simulated data, slingshot
selects Cluster 1 as the starting
cluster. However, we generally recommend the specification of an initial cluster
based on prior knowledge (either time of sample collection or established gene
markers). This specification will have no effect on how the MST is constructed,
but it will impact how branching curves are constructed.
lin1 <- getLineages(rd, cl, start.clus = '1')
lin1
## class: PseudotimeOrdering
## dim: 140 2
## metadata(3): lineages mst slingParams
## pathStats(2): pseudotime weights
## cellnames(140): cell-1 cell-2 ... cell-139 cell-140
## cellData names(2): reducedDim clusterLabels
## pathnames(2): Lineage1 Lineage2
## pathData names(0):
plot(rd, col = brewer.pal(9,"Set1")[cl], asp = 1, pch = 16)
lines(SlingshotDataSet(lin1), lwd = 3, col = 'black')
At this step, slingshot
also allows for the specification of known endpoints.
Clusters which are specified as terminal cell states will be constrained to have
only one connection when the MST is constructed (ie., they must be leaf nodes).
This constraint could potentially impact how other parts of the tree are drawn,
as shown in the next example where we specify Cluster 3 as an endpoint.
lin2 <- getLineages(rd, cl, start.clus= '1', end.clus = '3')
plot(rd, col = brewer.pal(9,"Set1")[cl], asp = 1, pch = 16)
lines(SlingshotDataSet(lin2), lwd = 3, col = 'black', show.constraints = TRUE)
This type of supervision can be useful for ensuring that results are consistent with prior biological knowledge. Specifically, it can prevent known terminal cell fates from being categorized as transitory states.
There are a few additional arguments we could have passed to getLineages
for
greater control:
dist.method
is a character specifying how the distances between clusters
should be computed. The default value is "slingshot"
, which uses the distance
between cluster centers, normalized by their full, joint covariance matrix. In
the presence of small clusters (fewer cells than the dimensionality of the
data), this will automatically switch to using the diagonal joint covariance
matrix. Other options include simple
(Euclidean), scaled.full
,
scaled.diag
, and mnn
(mutual nearest neighbor-based distance).omega
is a granularity parameter, allowing the user to set an upper limit on
connection distances. It represents the distance between every real cluster and
an artificial .OMEGA
cluster, which is removed after fitting the MST. This can
be useful for identifying outlier clusters which are not a part of any lineage
or for separating distinct trajectories. This may be a numeric argument or a
logical value. A value of TRUE
indicates that a heuristic of \(1.5\)
(median edge length of unsupervised MST)
should be used, implying that the
maximum allowable distance will be \(3\) times that distance.After constructing the MST, getLineages
identifies paths through the tree to
designate as lineages. At this stage, a lineage will consist of an ordered set
of cluster names, starting with the root cluster and ending with a leaf. The
output of getLineages
is a PseudotimeOrdering
which holds all of the inputs
as well as a list of lineages and some additional information on how they were
constructed. This object is then used as the input to the getCurves
function.
In order to model development along these various lineages, we will construct
smooth curves with the function getCurves
. Using smooth curves based on all
the cells eliminates the problem of cells projecting onto vertices of piece-wise
linear trajectories and makes slingshot
more robust to noise in the clustering
results.
In order to construct smooth lineages, getCurves
follows an iterative process
similar to that of principal curves presented in (Hastie and Stuetzle 1989). When there is
only a single lineage, the resulting curve is simply the principal curve through
the center of the data, with one adjustment: the initial curve is constructed
with the linear connections between cluster centers rather than the first
prinicpal component of the data. This adjustment adds stability and typically
hastens the algorithm’s convergence.
When there are two or more lineages, we add an additional step to the algorithm: averaging curves near shared cells. Both lineages should agree fairly well on cells that have yet to differentiate, so at each iteration we average the curves in the neighborhood of these cells. This increases the stability of the algorithm and produces smooth branching lineages.
crv1 <- getCurves(lin1)
crv1
## class: PseudotimeOrdering
## dim: 140 2
## metadata(4): lineages mst slingParams curves
## pathStats(2): pseudotime weights
## cellnames(140): cell-1 cell-2 ... cell-139 cell-140
## cellData names(2): reducedDim clusterLabels
## pathnames(2): Lineage1 Lineage2
## pathData names(0):
plot(rd, col = brewer.pal(9,"Set1")[cl], asp = 1, pch = 16)
lines(SlingshotDataSet(crv1), lwd = 3, col = 'black')
The output of getCurves
is an updated PseudotimeOrdering
which now contains
the simultaneous principal curves and additional information on how they were
fit. The slingPseudotime
function extracts a cells-by-lineages matrix of
pseudotime values for each cell along each lineage, with NA
values for cells
that were not assigned to a particular lineage. The slingCurveWeights
function
extracts a similar matrix of weights assigning each cell to one or more
lineages.
The curves objects can be accessed using the slingCurves
function, which will
return a list of principal_curve
objects. These objects consist of the
following slots:
s
: the matrix of points that make up the curve. These correspond to the
orthogonal projections of the data points.ord
: indices which can be used to
put the cells along a curve in order based their projections.lambda
: arclengths along the curve from the beginning to each cell’s
projection. A matrix of these values is returned by the slingPseudotime
function.dist_ind
: the squared distances between data points and their projections
onto the curve.dist
: the sum of the squared projection distances.w
: the vector of weights along this lineage. Cells that were unambiguously
assigned to this lineage will have a weight of 1
, while cells assigned to
other lineages will have a weight of 0
. It is possible for cells to have
weights of 1
(or very close to 1) for multiple lineages, if they are
positioned before a branching event. A matrix of these values is returned by the
slingCurveWeights
function.For large datasets, we highgly recommend using the approx_points
argument with
slingshot
(or getCurves
). This allows the user to specify the resolution of the curves (ie. the number of unique points). While the MST construction operates
on clusters, the process of iteratively projecting all points onto one or more
curves may become computationally burdensome as the size of the dataset grows. For this reason, we set the default value for approx_points
to either \(150\) or the number of cells in the dataset, whichever is smaller. This should greatly ease the computational cost of exploratory analysis while having minimal impact on the resulting trajectories.
For maximally “dense” curves, set approx_points = FALSE
, and the curves will have as many points as there are cells in the dataset. However, note that each projection step in the iterative curve-fitting process will now have a computational complexity proportional to \(n^2\) (where \(n\) is the number of cells). In the presence of branching lineages, these dense curves will also affect the complexity of curve averaging and shrinkage.
We recommend a value of \(100\)-\(200\), hence the default value of \(150\). Note that
restricting the number of unique points along the curve does not impose a
similar limit on the number of unique pseudotime values, as demonstrated below. Even with an unrealistically low value of \(5\) for approx_points
, we still see a
full color gradient from the pseudotime values:
sce5 <- slingshot(sce, clusterLabels = 'GMM', reducedDim = 'PCA',
approx_points = 5)
colors <- colorRampPalette(brewer.pal(11,'Spectral')[-6])(100)
plotcol <- colors[cut(sce5$slingPseudotime_1, breaks=100)]
plot(reducedDims(sce5)$PCA, col = plotcol, pch=16, asp = 1)
lines(SlingshotDataSet(sce5), lwd=2, col='black')
In some cases, we are interested in identifying multiple, disjoint trajectories. Slingshot handles this problem in the initial MST construction by introducing an artificial cluster, called omega
. This artificial cluster is separated from every real cluster by a fixed length, meaning that the maximum distance between any two real clusters is twice this length. Practically, this sets a limit on the maximum edge length allowable in the MST. Setting omega = TRUE
will implement a rule of thumb whereby the maximum allowable edge length is equal to \(3\) times the median edge length of the MST constructed without the artificial cluster (note: this is equivalent to saying that the default value for omega_scale
is \(1.5\)).
rd2 <- rbind(rd, cbind(rd[,2]-12, rd[,1]-6))
cl2 <- c(cl, cl + 10)
pto2 <- slingshot(rd2, cl2, omega = TRUE, start.clus = c(1,11))
plot(rd2, pch=16, asp = 1,
col = c(brewer.pal(9,"Set1"), brewer.pal(8,"Set2"))[cl2])
lines(SlingshotDataSet(pto2), type = 'l', lwd=2, col='black')
After fitting the MST, slingshot
proceeds to fit simultaneous principal curves as usual, with each trajectory being handled separately.
plot(rd2, pch=16, asp = 1,
col = c(brewer.pal(9,"Set1"), brewer.pal(8,"Set2"))[cl2])
lines(SlingshotDataSet(pto2), lwd=2, col='black')
sessionInfo()
## R version 4.1.3 (2022-03-10)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 20.04.4 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.14-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.14-bioc/R/lib/libRlapack.so
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
## [3] LC_TIME=en_GB LC_COLLATE=C
## [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
## [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
## [9] LC_ADDRESS=C LC_TELEPHONE=C
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
##
## attached base packages:
## [1] stats4 stats graphics grDevices utils datasets methods
## [8] base
##
## other attached packages:
## [1] mclust_5.4.9 uwot_0.1.11
## [3] Matrix_1.4-1 mgcv_1.8-40
## [5] nlme_3.1-157 RColorBrewer_1.1-3
## [7] slingshot_2.2.1 TrajectoryUtils_1.2.0
## [9] princurve_2.1.6 SingleCellExperiment_1.16.0
## [11] SummarizedExperiment_1.24.0 Biobase_2.54.0
## [13] GenomicRanges_1.46.1 GenomeInfoDb_1.30.1
## [15] IRanges_2.28.0 S4Vectors_0.32.4
## [17] BiocGenerics_0.40.0 MatrixGenerics_1.6.0
## [19] matrixStats_0.61.0 BiocStyle_2.22.0
##
## loaded via a namespace (and not attached):
## [1] xfun_0.30 bslib_0.3.1
## [3] splines_4.1.3 lattice_0.20-45
## [5] htmltools_0.5.2 yaml_2.3.5
## [7] rlang_1.0.2 jquerylib_0.1.4
## [9] GenomeInfoDbData_1.2.7 stringr_1.4.0
## [11] zlibbioc_1.40.0 codetools_0.2-18
## [13] evaluate_0.15 knitr_1.38
## [15] fastmap_1.1.0 highr_0.9
## [17] Rcpp_1.0.8.3 BiocManager_1.30.16
## [19] DelayedArray_0.20.0 magick_2.7.3
## [21] jsonlite_1.8.0 XVector_0.34.0
## [23] RSpectra_0.16-0 FNN_1.1.3
## [25] digest_0.6.29 stringi_1.7.6
## [27] bookdown_0.25 grid_4.1.3
## [29] cli_3.2.0 tools_4.1.3
## [31] bitops_1.0-7 magrittr_2.0.3
## [33] sass_0.4.1 RCurl_1.98-1.6
## [35] pkgconfig_2.0.3 DelayedMatrixStats_1.16.0
## [37] sparseMatrixStats_1.6.0 rmarkdown_2.13
## [39] R6_2.5.1 igraph_1.3.0
## [41] compiler_4.1.3
Cole, Michael, and Davide Risso. 2018. Scone: Single Cell Overview of Normalized Expression Data.
Haghverdi, Laleh, Aaron T. L. Lun, Michael D. Morgan, and John C. Marioni. 2018. “Batch effects in single-cell RNA-sequencing data are corrected by matching mutual nearest neighbors.” Nat. Biotechnol. 36 (5): 421–27. https://doi.org/10.1038/nbt.4091.
Hastie, Trevor, and Werner Stuetzle. 1989. “Principal Curves.” Journal of the American Statistical Association 84 (406): 502–16.
Lun, Aaron, and Davide Risso. 2017. SingleCellExperiment: S4 Classes for Single Cell Data.
Risso, Davide, Fanny Perraudeau, Svetlana Gribkova, Sandrine Dudoit, and Jean-Philippe Vert. 2018. “A General and Flexible Method for Signal Extraction from Single-Cell RNA-Seq Data.” Nature Communications 9: 284. https://doi.org/10.1038/s41467-017-02554-5.
Scrucca, Luca, Michael Fop, Thomas Brendan Murphy, and Adrian E. Raftery. 2016. “mclust 5: Clustering, Classification and Density Estimation Using Gaussian Finite Mixture Models.” The R Journal 8 (1): 205–33.
Street, Kelly, Davide Risso, Russell B Fletcher, Diya Das, John Ngai, Nir Yosef, Elizabeth Purdom, and Sandrine Dudoit. 2017. “Slingshot: Cell Lineage and Pseudotime Inference for Single-Cell Transcriptomics.” bioRxiv. https://doi.org/10.1101/128843.
Van den Berge, Koen, Hector Roux de Bézieux, Kelly Street, Wouter Saelens, Robrecht Cannoodt, Yvan Saeys, Sandrine Dudoit, and Lieven Clement. 2020. “Trajectory-Based Differential Expression Analysis for Single-Cell Sequencing Data.” Nature Communications 11 (1): 1201. https://doi.org/10.1038/s41467-020-14766-3.