GigaScience, 2020, 1–9
doi: xx.xxxx/xxxx
Manuscript in Preparation
Paper
P A P E R
A single-cell RNA-seq Training and Analysis Suite
using the Galaxy Framework
Mehmet Tekman
1
*
†
, Bérénice Batut
1
*
†
, Alexander Ostrovsky
2
, Christophe
Antoniewski
3
, Dave Clements
2
, Fidel Ramirez
4
, Graham J Etherington
5
,
Hans-Rudolf Hotz
6
, Jelle Scholtalbers
7
, Jonathan R Manning
8
, Lea
Bellenger
3
, Maria A Doyle
9
, Mohammad Heydarian
2
, Ni Huang
8,10
, Nicola
Soranzo
5
, Pablo Moreno
8
, Stefan Mautner
1
, Irene Papatheodorou
8
, Anton
Nekrutenko
11
, James Taylor
2
, Daniel Blankenberg
12
, Rolf Backofen
1
and
Björn Grüning
1
*
1
Chair of Bioinformatics, University of Freiburg, Freiburg, Germany, and
2
Department of Biology, Johns
Hopkins University, Baltimore, Maryland, USA and
3
ARTbio, Sorbonne Université, CNRS FR 3631, Inserm US
037, Paris, France and Institut de Biologie Paris Seine, Paris, France and
4
Boehringer Ingelheim, Biberach,
Germany and
5
Earlham Institute, Norwich Research Park, Norwich, NR4 7UZ, United Kingdom and
6
Friedrich
Miescher Institute for Biomedical Research, Basel, Switzerland and Swiss Institute of Bioinformatics, Basel,
Switzerland and
7
European Molecular Biology Laboratory, Genome Biology Unit, Heidelberg, Germany and
8
European Molecular Biology Laboratory, European Bioinformatics Institute, EMBL-EBI, Hinxton, United
Kingdom and
9
Research Computing Facility, Peter MacCallum Cancer Centre, Melbourne, Victoria 3000,
Australia and Sir Peter MacCallum Department of Oncology, The University of Melbourne, Victoria 3010,
Australia and
10
Wellcome Sanger Institute, Cambridge, United Kingdom and
11
Department of Biochemistry
and Molecular Biology, The Pennsylvania State University, University Park, Pennsylvania, USA and
12
Genomic
Medicine Institute, Lerner Research Institute, Cleveland Clinic, Cleveland, Ohio, USA
*
tekman@informatik.uni-freiburg.de; berenice.batut@gmail.com; gruening@informatik.uni-freiburg.de
†
These authors contributed equally to this work.
Abstract
Background The vast ecosystem of single-cell RNA-seq tools has until recently been plagued by an excess of diverging
analysis strategies, inconsistent le formats, and compatibility issues between dierent software suites. The uptake of 10x
Genomics datasets has begun to calm this diversity, and the bioinformatics community leans once more towards the large
computing requirements and the statistically-driven methods needed to process and understand these ever-growing
datasets.
Results Here we outline several Galaxy workows and learning resources for scRNA-seq, with the aim of providing a
comprehensive analysis environment paired with a thorough user learning experience that bridges the knowledge gap
between the computational methods and the underlying cell biology. The Galaxy reproducible bioinformatics framework
provides tools, workows and trainings that not only enable users to perform one-click 10x preprocessing, but also
empowers them to demultiplex raw sequencing from custom tagged and full-length sequencing protocols. The downstream
analysis supports a wide range of high-quality interoperable suites separated into common stages of analysis: inspection,
ltering, normalization, confounder removal and clustering. The teaching resources cover an assortment of dierent
concepts from computer science to cell biology. Access to all resources is provided at the singlecell.usegalaxy.eu portal.
Conclusions
The reproducible and training-oriented Galaxy framework provides a sustainable HPC environment for users
to run exible analyses on both 10x and alternative platforms. The tutorials from the Galaxy Training Network along with
the frequent training workshops hosted by the Galaxy Community provide a means for users to learn, publish and teach
scRNA-seq analysis.
Key words: scRNA; Galaxy; resources; HPC; single-cell; 10x; Training; Web;
1
.CC-BY 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 28, 2020. ; https://doi.org/10.1101/2020.06.06.137570doi: bioRxiv preprint
2
|
GigaScience, 2020, Vol. 00, No. 0
Key Points
• Single-cell RNA-seq has stabilised towards 10x Genomics datasets.
• Galaxy provides rich and reproducible scRNA-seq workows with a wide range of robust tools.
• The Galaxy Training Network provides tutorials for the processing of both 10x and non-10x datasets.
Background
Single-Cell RNA-seq and cellular heterogeneity. The continuing
rise in single cell technologies has led to previously unprece-
dented levels of analysis into cell heterogeneity within tissue
samples, providing new insights into developmental and dif-
ferentiation pathways for a wide range of disciplines. Gene
expression studies are now performed at a cellular level of res-
olution, which compared to bulk RNA-seq methods, allows re-
searchers to model their tissue samples as distributions of dif-
ferent expressions instead of as an average.
Pathways from Single-cell data. The various expression proles
uncovered within tissue samples infer discrete cell types which
are related to one another across an “expression landscape”.
The relationships between the more distinct proles are in-
ferred via distance-metrics or manifold learning techniques.
Ultimately, the aim is to model the continuous biological pro-
cess of cell dierentiation from multipotent stem cells to dis-
tinct mature cell types, and infer lineage and dierentiation
pathways between transient cell types [1].
Elucidating Cell Identity. Trajectory analysis which integrates
the up or down regulation of signicant genes along lineage
branches can then be performed in order to uncover the factors
and extracellular triggers that can coerce a pluripotent cell to
become biased towards one cell fate outcome compared to an-
other. This undertaking has created a new frontier of explo-
ration in cell biology, where researchers have assembled refer-
ence maps for dierent cell lines for the purpose of fully record-
ing these cell dynamics and their characteristics in which to
create a global “atlas” of cells [2, 3].
Pitfalls and Technical Challenges
Sequencing sensitivity and Normalization. With each new protocol
comes a host of new technical problems to overcome. The rst
wave of software utilities to deal with the analysis of single cell
datasets were statistical packages, aimed at tackling the issue
of “dropout events” during sequencing, which would manifest
as a high prevalence of zero-entries in over 80% of the feature-
count matrix. These zeroes were problematic, since they could
not be trivially ignored as their presence stated that either the
cell did not produce any molecules for that transcript, or that
the sequencer simply did not detect them. Normalisation tech-
niques originally developed for bulk RNA-seq had to be adapted
to accommodate for this uncertainty, and new ones were cre-
ated that harnessed hurdle models, data imputation via mani-
fold learning techniques, or by pooling subsets of cells together
and building general linear models [4].
Improvements in sequencing. With the downstream analysis
packages attempting to solve the dropouts via stochastic meth-
ods, the upstream sequencing technologies also aspired to solve
the capture eciency via new well, droplet, and ow cytome-
try based protocols, all of which lend importance to the process
of barcoding sequencing reads.
In each protocol, cells are tagged with cell barcodes such
that any reads derived from them can be unambiguously as-
signed to the cell of origin. The inclusion of unique molecular
identiers (UMIs) are also employed to mitigate the eects of
amplication bias of transcripts within the same cell. The de-
tection, extraction, and (de-)multiplexing of cell barcodes and
UMIs is therefore one of the rst hurdles researchers encounter
when receiving raw FASTQ data from a sequencing facility.
The Burgeoning Software Ecosystem
Since its conception, several dierent packages and many
pipelines have been developed to assist researchers in the anal-
ysis of scRNA-seq [5, 6]. The vast majority of these pack-
ages were written for the R programming language since many
of the novel normalisation methods developed to handle the
dropout events depended on statistical packages that were
primarily R-based [7]. Standalone analysis suites emerged
as the dierent authors of these packages rapidly expanded
their methods to encapsulate all facets of the single-cell anal-
ysis, often creating compatibility issues with previous package
versions. The Bioconductor repository provided some much-
needed stability in this regard by hosting stable releases, but
researchers were still prone to building directly from reposi-
tory sources in order to reap the benets of new features in the
upstream versions [8, 9].
Nonexchangeable Data Formats. Another issue was the prolif-
eration of the many dierent and quickly evolving R-based
le formats for processing and storing the data, such as
SingleCellExperiment as used by the Scater suite, SCSeq from
RaceID, and SeuratObject from Seurat [10, 11]. Many new pack-
ages would cater only towards one format or suite, leading to
the common problem that data processed in one suite could not
be reliably processed by methods in another. This incompati-
bility between packages fuelled a choice of one analysis suite
over another, or conversely required researchers to dig deeper
into the internal semantics of R S4 objects in order to manually
slot data components together [12]. These problems only accel-
erated the rapid development of these suites, leading to further
version instability. As a result of this analysis diversity, there
are many tutorials on how to perform scRNA-seq analysis each
oriented around one of these pipelines [13].
Error propagation and Analysis Uncertainty. Dierent pipelines
produce dierent results, where the stochastic nature of the
analyses means that any uncertainty in a crucial quality control
step upstream, such as ltering or the removal of unwanted
variability, can propagate forward into the downstream sec-
tions to yield wildly dierent results on the same data. This
uncertainty, and the statistically-driven methods to overcome
Compiled on: August 28, 2020.
Draft manuscript prepared by the author.
.CC-BY 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 28, 2020. ; https://doi.org/10.1101/2020.06.06.137570doi: bioRxiv preprint
Tekman et al.
|
3
them, leaves a wide knowledge gap for researchers simply try-
ing to understand the underlying dynamics of cell identity.
Rise of 10x Genomics
10x Launch. In 2015, 10x Genomics released their GemCode prod-
uct, which was a droplet-seq based protocol capable of se-
quencing tens of thousands of cells with an average cell quality
higher than other facilities [14]. This unprecedented level of
throughput steadily gained traction amongst researchers and
startups seeking to perform single-cell analysis, and thus 10x
datasets began to prevail in the eld.
10x Analysis Software. 10x Genomics provided software that was
able to perform much of the pre-processing, and provided
feature-count matrices in a transparent HDF5-based format
which provided a means of ecient matrix storage and ex-
change, and conclusively removed the restriction for down-
stream analysis modules to be written in R.
ScanPy, a popular alternative. The ScanPy suite [15], written in
Python, using its own HDF5-based AnnData format became a
valid alternative for analysing 10x datasets. The Seurat devel-
opers had similar aspirations and soon adopted the LOOM format,
another HDF5 variant. However, the popularity of ScanPy rose
as it began to integrate the methods of other standalone pack-
ages into its codebase, making it the natural choice for users
who wanted to achieve more without compromising on com-
patibility between dierent suites [9].
Solutions in the Cloud
Accessible Science. As the size of datasets scaled, so did the
computing resources required to analyse them, both in terms
of the processing power and in storage. Galaxy is an open-
source biocomputing infrastructure that exemplies the three
main tenets of science: reproducibility, peer review, and open-
access - all freely accessible within the web-browser [16]. It
hosts a wide range of highly-cited bioinformatics tools with
many dierent versions, and enables users to freely create their
own workows via a seamless drag-and-drop interface.
Reproducible Workows. Galaxy can make use of Conda or Con-
tainers to setup tool environments in order to ensure that the
bioinformatics tools will always be able to run, even when the
library dependencies for a tool have changed, by building tools
under locked version dependencies and bundling them together
in a self-contained environment [17]. These environments pro-
vide a concise solution for the package version instability that
plagues scRNA-seq analysis notebooks, both in terms of repro-
ducibility and analysis exibility. A user could keep the quality
control results obtained from an older version of ScanPy, whilst
running a newer ScanPy version at the clustering stage to reap
the benets of the later improvements in that algorithm. By
allowing the user to select from multiple versions of the same
tool, and by further permitting dierent versions of the tools
within a workow, Galaxy enables an unprecedented level of
free-ow analysis by letting researchers pick and choose the
best aspects of a tool without worrying about the underlying
software libraries [18]. The burden of software incompatibility
and choice of programming language that plagued the scRNA-
seq analysis ecosystem before, are now completely alleviated
from the user.
User-driven Custom Workows. Analyses are not limited to one
pipeline either, as the datasets which are passed between tools
can easily be interpreted by a dierent tool that is capable of
reading that dataset. In the case of scRNA-seq, Galaxy can
convert between CSV, MTX, LOOM and AnnData formats. This inter-
exchange of modules from dierent tools further extends the
exibility of the analysis by again letting the user decide which
component of a tool would be best suited for a specic part of
an analysis.
Training Resources. Galaxy also provides a wide range of learn-
ing resources, with the aim of guiding users step-by-step
through an analysis, often reproducing the results of published
works. The teaching and training materials are part of the
Galaxy Training Network (GTN), which is a worldwide collabo-
rative eort to produce high-quality teaching material in order
to educate users in how to analyse their data, and in turn to
train others of the same materials via easily deployable work-
shops backed by monthly stable releases of the GTN materials
[19]. Training materials are provided on a wide variety of dier-
ent topics, and workshops are hosted regularly, as advertised
on the Galaxy Events web portal. The GTN has grown rapidly
since its conception and gains new volunteers every year who
each contribute and coordinate training and teaching events,
maintain topic and subtopics, translate tutorials into multiple
languages, and provide peer review on new material [20].
Methods
Stable Workows in Galaxy. The analysis of scRNA-seq within
Galaxy was a two-pronged eort concentrated on bringing
high quality single-cell tools into Galaxy, and providing the
necessary workows and training to accompany them. As men-
tioned in the previous section, this eort needed to overcome
incompatible le format issues, unstable packages due to rapid
development, and needed to establish a standardised basis for
the analysis.
Tutorials. The tutorials are split into two main parts as out-
lined in Figure 1: rst, the pre-processing stage which con-
structs a count matrix from the initial sequencing data; sec-
ond, the cluster-based downstream analysis on the count ma-
trix. These stages are very dierent from one another in terms
of their information content, since the pre-processing stage re-
quires the researcher to be more familiar with wetlab sequenc-
ing protocols than your average bioinformatician would nor-
mally know, and the downstream analysis stage requires the
researcher to be familiar with statistics concepts that a wet-
lab scientist might not be too familiar with. The tutorials are
designed to broadly appeal to both the biologist and the statis-
tician, as well as complete beginners to the entire topic.
Pre-processing Workows
The pre-processing scRNA-seq materials tackle the two most
common use-cases that researchers will encounter when they
rst begin the eld: processing scRNA-seq data from 10x Ge-
nomics, and processing data generated from alternative pro-
tocols. For instance, microwell-based protocols have been
known to yield more features and display lower levels of
dropouts compared to 10x, and so we accommodate for them
by providing a more customizable path through the pre-
processing stage [21].
Barcode Extraction. Before the era of 10x Genomics, scRNA-seq
data had to be demultiplexed, mapped, and quantied. The de-
multiplexing stage entails an intimate knowledge of cell bar-
codes and Unique Molecular Identiers (UMIs) which are pro-
tocol dependent, and expects that the bioinformatician knows
.CC-BY 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 28, 2020. ; https://doi.org/10.1101/2020.06.06.137570doi: bioRxiv preprint
4
|
GigaScience, 2020, Vol. 00, No. 0
Figure 1. The main stages of single-cell analysis, separated broadly into the up-
per and lower stages of pre-processing and downstream analysis, respectively.
The upper part illustrates the two main routes to generating a count-matrix
from sequencing data; via one-click quantication solutions, or through man-
ual demultiplexing. The lower part describes the four main stages required
to perform cluster-based analysis from the count-matrix, through ltering,
normalisation, confounder removal, and embedding.
exactly where and how the data was generated. One common
pitfall at this very rst stage is estimating how many cells to
expect from the FASTQ input data, and this requires three cru-
cial pieces of information: which reads contain the barcodes (or
precisely, which subset of both the forward and reverse reads
contains the barcodes); of these barcodes, which specic ones
were actually used for the analysis; and how to resolve barcode
mismatches/errors.
Barcode Estimation. Naive strategies involve using a known bar-
code template and querying against the FASTQ data to prole the
number of reads that align to a specic barcode, often employ-
ing ’knee’ methods to estimate this amount [22]. However,
this approach is not robust, since certain cells are more likely
to be over-represented compared to others, and some cell bar-
codes may contain more unmappable reads compared to others,
meaning that the metric of higher library read depth is not nec-
essarily correlated with a better-dened cell. Ultimately, the
bioinformatician must inquire directly with the sequencing lab
as to which cell barcodes were used, as these are often not spe-
cic to the protocol but to the technician who designed them,
with the idea that they should not align to a specic reference
genome or transcriptome.
One-click Pre-processing
Quantication with Cell Ranger. 10x Genomics simplied the
scRNA package ecosystem by using a language independent le
format, and streamlining much of the barcode particularities
with their Cell Ranger pipeline, allowing researchers to focus
more on the internal biological variability of their datasets [23].
Quantication with STARsolo. The pre-processing workow (ti-
tled "10X StarSolo Workow") in Galaxy uses RNA STARsolo util-
ity as a drop-in replacement for Cell Ranger, because not only
is it a feature update of the already existing RNA STAR tool in
Galaxy, but because it boasts a ten-fold speedup in comparison
to Cell Ranger and does not require Illumina lane-read informa-
tion to perform the processing [24, 25].
Other Approaches. The pre-processing workows for these
“one-click” solutions consume the same datasets and yield
approximately the same count matrices by following simi-
lar modes of barcode discovery and quantication. Within
Galaxy, there is also Alevin (paired with Salmon) and scPipe
which can both also perform the necessary demultiplexing,
(alignment-free) mapping, and quantication stages in a sin-
gle step [26, 27, 28].
Flexible Pre-processing
CELSeq2 Barcoding. The custom pre-processing workow (ti-
tled "CELSeq2: Single Batch mm10") is modelled after the CEL-
seq2 protocol using the barcoding strategies of the Freiburg
Max-Planck Institute laboratory as its main template, but the
workow is actually exible to accommodate any droplet or
well-based protocol such as SMART-seq2, and Drop-seq [29].
Manual Demultiplexing and Quantication. The training picto-
graphically guides users through the concepts of extracting cell
barcodes from the protocol, explains the signicance of UMIs
in the process of read deduplication with illustrative examples,
and instructs the user in the process of performing further
quality controls on their data during the post-mapping process
via RNA STAR and other tools that are native to Galaxy.
Training the User. At each stage, the user’s knowledge is queried
via question prompts and expandable answer box dialogs, as
well as helpful hints for future processing in comment boxes,
all written in the transparent Markdown specication devel-
oped for contributing to the GTN.
Downstream Workows
Common Stages of Analysis. The downstream modules are de-
ned by the ve main stages of downstream scRNA-seq anal-
ysis: ltering, normalisation, confounder removal, clustering,
and trajectory inference. There are three workows to aid in
this process (two of which are shown in Figure 2), each sport-
ing a dierent well-established scRNA-seq pipeline tool.
Quality Control with Scater. The Scater pipeline follows a
visualise-lter-visualise paradigm which provides an intuitive
means to perform quality control on a count matrix by use of
repeated incremental changes on a dataset through the use of
PCA and library size based metrics [30]. Once this pre-analysis
stage is complete, the full downstream analysis (comprising
the ve stages mentioned above) can be performed by work-
ows based on the following suites: RaceID and ScanPy.
Downstream Analysis with the RaceID Suite. RaceID was developed
initially to analyse rare cell transcriptomes whilst being ro-
bust against noise, and thus is ideal for working with smaller
datasets in the range of 300 to 1000 cells. Due to its complex
cell lineage and fate predictions models, it can also be used on
larger datasets with some scaling costs.
.CC-BY 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 28, 2020. ; https://doi.org/10.1101/2020.06.06.137570doi: bioRxiv preprint
Tekman et al.
|
5
Downstream Analysis with the ScanPy Suite. ScanPy was devel-
oped as the Python alternative to the innumerable R-based
packages for scRNA-seq which was the dominant language for
such analyses, and it was one of the rst packages with na-
tive 10x genomics support. Since then it has grown substan-
tially, and has been re-implementing much of the newer R-
based methods released in BioConductor as “recipe” modules,
thereby providing a single source to perform many dierent
types of the same analysis.
The workows derived from both these suites emulate the
ve main stages of analysis mentioned previously, where lter-
ing, normalisation, and confounder removal are typically sep-
arated into distinct stages, though sometimes merged into one
step depending on the tool.
Filtering
Cell and Gene Removal. During the ltering stage, the initial
count matrix removes low-quality or unwanted cells using
commonly used parameters such as minimum gene detection
sensitivity and minimum library size, and low-quality genes
are also removed under similar metrics, where the minimum
number of cells for a gene to be included is decided. The Scater
pre-analysis workow can also be used here to provide a PCA-
based method of feature selection so that only the highly vari-
able genes are left in the analysis.
Disadvantages of Filtering. There is always the danger of over-
ltering a dataset, whereby setting overzealous lower-bound
thresholds on gene variability, can have the undesired eect of
removing essential housekeeping genes. These relatively uni-
formly expressed genes are often required for setting a baseline
to which the more desired dierentially expressed genes can
be selected from. It is therefore important that the user rst
performs a naive analysis and only later rene their ltering
thresholds to boost the biological signal.
Normalisation
Library Size Normalisation. The normalisation step aims to re-
move any technical factors that are not relevant to the analysis,
such as the library size, where cells sharing the same identity
are likely to dier from one another more by the number of
transcripts they exhibit, than due to more relevant biological
factors.
Intrinsic Cell Factors. The rst and foremost is cell capture e-
ciency, where dierent cells produce more or less transcripts
based on the amplication and coverage conditions they are
sequenced in. The second is the presence of dropout events
which manifest as a prevalence of “zeroes” in the nal count
matrix. Whether a “zero” is imputable to the lack of detection
of an existing molecule or to the absence of the molecule in the
cell is uncertain. This uncertainty alone has led to a wide se-
lection of dierent normalisation techniques that try to model
this expression either via hurdle models, or imputing the data
via manifold learning techniques, or working around the issue
by pooling subsets of cells together [31].
In this regard, both the RaceID and ScanPy workows of-
fer many dierent normalisation techniques, and users are en-
couraged to take advantage of the branching workow model
of Galaxy to explore all possible options.
Confounder Removal
Regression of Cell Cycle Eects. Other sources of variability stem
from unwanted biological contributions known as confounder
eects, such as cell cycle eects and transcription. Depending
on what stage of the cell cycle a cell was sequenced at, two cells
of the same type might cluster dierently because one might
have more transcripts due to it being in the M-phase of the cell
Figure 2. Downstream analysis workows as shown in the Galaxy Workow
Editor for (top) RaceID and (bottom) ScanPy, each displaying modules symbol-
izing the ve main stages of analysis.
cycle. Library sizes notwithstanding, it is the variability in spe-
cic cell cycle genes that can be the main driving factor in the
overall variability. Thankfully, these eects are easy to regress
out, and we replicate an entire standalone ScanPy workow
dedicated to detecting and visualising the eects based on the
original notebook [32].
Transcriptional Bursting. The transcription eects are harder to
model, as these are semi-stochastic and are as of yet still not
well understood. In bulk RNA-seq the expression of genes un-
dergoing transcription are averaged to give “high” or “low”
signals producing a global eect that gives the false impres-
sion that transcription is a continuous process. The reality is
more complex, where cells undergo transcription in “bursts” of
activity followed by periods of no activity, at irregular intervals
[33]. At the bulk level these discrete processes are smoothed to
give a continuous eect, but at the cell level it could mean that
even two directly adjacent cells of the same type normalised to
the same number of transcripts can still have dierent levels
of expression for a gene due to this process. This is not some-
thing that can be countered for, but it does educate the users
in which factors they can or cannot control in an analysis, and
how much variability they can expect to see.
Clustering and Projection
Dimension Reduction and Clustering. Once a user has obtained a
count matrix they are condent with, they are then guided
through the process of dimension reduction (with choice of dif-
ferent distance metrics), choosing a suitable low-dimensional
embedding, and performing clustering through commonly-
used techniques such as k-means, hierarchical, and neighbour-
hood community detection. These complex techniques are il-
lustrated in layman’s terms through the use of helpful images
and community examples. For example, the GTN ScanPy tu-
torial explains the Louvain clustering approach[34] via a stan-
dalone slide deck to assist in the workow [35].
Commonly-used Embeddings. The clustering and the cluster in-
spection stages are notably separated into distinct utilities here,
with the understanding that the same initial clustering can ap-
pear dissimilar under dierent projections, e.g. t-distributed
Stochastic Network Embedding (tSNE) against Uniform Mani-
fold Approximation and Projection (UMAP) [36, 37]. Ultimately
the user is encouraged to play with the plotting parameters to
yield the best looking clusters.
Static Plots or Interactive Environments. Cluster inspection tools
are available that allow users to easily generate static plots
tailored to pipeline-specic information as originally dened
.CC-BY 4.0 International licensemade available under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is
The copyright holder for this preprintthis version posted August 28, 2020. ; https://doi.org/10.1101/2020.06.06.137570doi: bioRxiv preprint
