CIAlign: A highly customisable command

line tool to clean, interpret and visualise

multiple sequence
alignments
Charlotte Tumescheit, Andrew E. Firth and Katherine Brown
Department of Pathology, University of Cambridge, Cambridge, United Kingdom

ABSTRACT
Background. Throughout biology, multiple sequence alignments (MSAs) form the
basis of much investigation into biological features and relationships. These alignments
are at the heart of many bioinformatics analyses. However, sequences in MSAs are
often incomplete or very divergent, which can lead to poor alignment and large gaps.
This slows down computation and can impact conclusions without being biologically
relevant. Cleaning the alignment by removing common issues such as gaps, divergent
sequences, large insertions and deletions and poorly aligned sequence ends can
substantially improve analyses. Manual editing of MSAs is very widespread but is time-
consuming and difficult to reproduce.
Results. We present a comprehensive, user-friendly MSA trimming tool with multiple
visualisation options. Our highly customisable command line tool aims to give
intervention power to the user by offering various options, and outputs graphical
representations of the alignment before and after processing to give the user a clear
overview of what has been removed. The main functionalities of the tool include
removing regions of low coverage due to insertions, removing gaps, cropping poorly
aligned sequence ends and removing sequences that are too divergent or too short. The
thresholds for each function can be specified by the user and parameters can be adjusted
to each individual MSA. CIAlign is designed with an emphasis on solving specific and
common alignment problems and on providing transparency to the user.
Submitted 6 April 2021
Accepted 1 February 2022
Conclusion. CIAlign effectively removes problematic regions and sequences from
Published 15 March 2022 MSAs and provides novel visualisation options. This tool can be used to fine-tune
Corresponding author
alignments for further analysis and processing. The tool is aimed at anyone who wishes
Katherine Brown, [email protected] to automatically clean up parts of an MSA and those requiring a new, accessible way of
Academic editor
visualising large MSAs.
Burkhard Morgenstern
Additional Information and
Subjects Bioinformatics, Computational Biology, Evolutionary Studies, Genomics
Declarations can be found on
page 25 Keywords Multiple sequence alignment, Alignment quality, Python tool, Comparative genomics,
Transcriptomics, Phylogenetics
DOI 10.7717/peerj.12983

Copyright
2022 Tumescheit et al. INTRODUCTION
Distributed under Throughout biology, multiple sequence alignments (MSAs) of DNA, RNA or amino acid
Creative Commons CC-BY 4.0 sequences are often the basis of investigation into biological features and relationships.
OPEN ACCESS Applications of MSAs include, but are not limited to, transcriptome analysis, in which

How to cite this article Tumescheit C, Firth AE, Brown K. 2022. CIAlign: A highly customisable command line tool to clean, interpret
and visualise multiple sequence alignments. PeerJ 10:e12983 http://doi.org/10.7717/peerj.12983
 transcripts may need to be aligned to genes; RNA structure prediction, in which an MSA
improves results significantly compared to predictions based on single sequences; and
phylogenetics, where trees are usually created based on MSAs. There are many more
applications of MSA at a gene, transcript and genome level, involved in a huge variety of
traditional and new approaches to genetics and genomics, many of which could benefit
from the tool presented here.
An MSA typically represents three or more DNA, RNA or amino acid sequences,
which represent partial or complete gene, transcript, protein or genome sequences. These
sequences are aligned by inserting gaps between residues to bring more similar residues
(either based on simple sequence similarity or an evolutionary model) into the same
column, allowing insertions, deletions and differences in sequence length to be taken
into account (Boswell, 1987; Higgins & Sharp, 1988). The first widely used automated
method for generating MSAs was Clustal (Higgins & Sharp, 1988) and more recent
versions of this tool are still in use today, along with tools such as MUSCLE (Edgar,
2004), MAFFT (Katoh et al., 2002), T-coffee (Notredame, Higgins & Heringa, 2000) and
many more. The majority of tools are based upon various heuristics used to optimise
progressive sequence alignment using a dynamic programming based algorithm such as
the Needleman-Wunsch algorithm (Needleman & Wunsch, 1970).
It has been shown previously that removing divergent regions from an MSA can improve
the resulting phylogenetic tree (Talavera & Castresana, 2007). Various tools are available
to identify or remove poorly aligned columns, including trimAl (Capella-Gutiérrez, Silla-
Martínez & Gabaldón, 2009), Gblocks (Talavera & Castresana, 2007) and ZORRO (Wu,
Chatterji & Eisen, 2012). These four tools use various algorithms to assign confidence scores
for each column in an MSA. Gblocks (Talavera & Castresana, 2007) identifies and removes
stretches of contiguous columns with low conservation. All positions with gaps, or adjacent
to gaps, are also removed (Talavera & Castresana, 2007). With trimAl (Capella-Gutiérrez,
Silla-Martínez & Gabaldón, 2009), poorly aligned columns are identified using proportion
of gaps, residue similarity and consistency across multiple alignments, either column-by-
column or based on a sliding window across the alignment. ZORRO uses hidden Markov
models to model sequence evolution and calculates posterior probabilities that columns
are correctly aligned (Wu, Chatterji & Eisen, 2012). All of these tools have been shown
to improve the accuracy of phylogenetic analysis under some circumstances and all can
be valuable (Talavera & Castresana, 2007; Capella-Gutiérrez, Silla-Martínez & Gabaldón,
2009; Wu, Chatterji & Eisen, 2012). However, poorly aligned columns are not the only issue
found in MSAs. All of these tools are designed to identify problematic columns, but none
are able to identify problematic rows which are disrupting an alignment. They also cannot
distinguish which gaps are the result of insertions within sequences and which are the
result of partial sequences. Column-wise tools can also be too stringent when working with
highly divergent alignments. Gblocks, trimAl and ZORRO are specifically tailored towards
phylogenetic analysis rather than other applications such as building consensus sequences,
scaffolding of contigs or secondary structure analysis.
Various refinement methods incorporated into alignment software can also improve
MSAs (Edgar, 2004; Katoh et al., 2002). Some tree building software can also take into

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 2/29

 account certain discrepancies in the alignment, for example RaXML (Stamatakis, 2014) can
account for missing data in some columns and check for duplicate sequence names and gap-
only columns; similarly GUI based toolkits for molecular biology such as MEGA (Kumar
et al., 2018) sometimes have options to delete or ignore columns containing gaps.
Several common issues affect the speed, complexity and reliability of specific downstream
analyses but are not addressed by these existing tools. Clean and Interpret Alignments
(CIAlign) is primarily intended to address four such issues and to be used (where
appropriate) in combination with existing tools which remove unreliable alignment
columns. Researchers in many fields regularly edit MSAs by hand to address these issues,
however, as well as being extremely time consuming, ensuring reproducibility with this
approach is almost impossible and it cannot be incorporated into an automated analysis
pipeline. CIAlign automatically removes full columns and full or partial rows from user
generated MSAs to address these issues in a fast, reproducible manner and can be easily
added to an automated pipeline. The downstream applications of alignments cleaned with
CIAlign are not limited to phylogenetic analysis and are too numerous to list, but CIAlign
as an alignment cleaning tool is particularly targetted towards users working with complex
or highly divergent alignments, partial sequences and problematic assemblies and towards
those developing complex pipelines requiring fine-tuning of parameters to meet specific
criteria.
The first issue we intend to address is that it is common for an MSA to contain more
gaps towards either end than in the body of the alignment. This problem occurs at
both the sequencing and alignment stage. For example, the ends of de novo assembled
transcripts tend to have lower read coverage (Bushmanova et al., 2019) and so have a
higher probability of mis-assembly and therefore mis-alignment. MSAs created using these
sequences therefore also have regions of lower reliability towards either end. Similarly, both
Sanger sequences and sequences generated with Oxford Nanopore’s long read sequencing
technology, which are often used directly in MSAs, tend to have lower quality scores at
either the beginning or the end (Richterich, 1998; Tyler et al., 2018; Magi, Giusti & Tattini,
2017). Automated removal of these regions from MSAs would therefore increase the
reliability of downstream analyses. As sequences are often partial, poor quality sequence
ends can be scattered throughout the alignment, and so do not necessarily result in whole
columns which are unreliable. A tool such as CIAlign, which identifies gaps at the ends
of sequences on a row-by-row basis, is therefore needed in these cases, rather than a
tool which works on whole columns only. Also, while generating an MSA, terminal gaps
complicate analysis, and the weighting of terminal gaps relative to internal gap opening
and gap extension penalties can make a large difference to the resulting alignment (Fitch &
Smith, 1983). This again leads to regions of ambiguity and therefore gaps towards the ends
of sequences within the alignment, which can be rectified with CIAlign.
Secondly, insertions or other stretches of sequence can be present in a minority of
sequences in an MSA, leading to large gaps in the remaining sequences. For example,
alignments of sections of bacterial genomes often result in long gaps representing genes
which are absent in the majority of species. These gaps can be observed, for example, in
multiple genome alignments shown in Tettelin et al. (2005) for Streptococcus agalactiae

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 3/29

 and Hu et al. (2011) for Burkholderia, amongst others, which show many genes which
are present in only a few genomes. While these regions are of interest in themselves and
certainly should not be excluded from all further analysis, they are not relevant for every
downstream analysis. For example, a consensus sequence for these bacteria would exclude
these regions and their presence would increase the time required for phylogenetic analysis
without necessarily adding any additional information. Large gaps in some sequences may
also result from missing data, rather than true biological differences and, if this is known to
be the case, it is often appropriate to remove these regions before performing phylogenetic
analysis (Sayyari, Whitfield & Mirarab, 2017). Unlike other available tools, CIAlign can
distinguish between gaps within the body of a sequence, which users may wish to remove,
and gaps padding the ends of sequences of different lengths, which occur, for example,
when aligning overlapping partial sequences, and remove the internal insertions only.
Thirdly, one or a few highly divergent sequences can heavily disrupt the alignment and
therefore complicate downstream analysis. It is very common for an MSA to include one
or a few outlier sequences which do not align well with the majority of the alignment. One
example of this is in analyses identifying novel sequences in large numbers of datasets. It is
common to manually remove phylogenetic outliers which are unlikely to truly represent
members of a group of interest (see for example Schulz et al., 2020; Käfer et al., 2019;
Bäckström et al., 2019) but this is not feasible when processing large numbers of alignments.
Alignment masking tools such as trimAl and Gblocks work column-by-column, and so,
unlike CIAlign, are not able to remove divergent rows.
Finally, very short partially overlapping sequences cannot always be reliably aligned
using standard global alignment algorithms. It is very common to remove these sequences,
manually or otherwise, prior to further analysis.
There are also several common issues in alignment visualisation. Large alignments can be
difficult to visualise and a small and concise but accurate visualisation can be useful when
presenting results, so this has been incorporated into the software. With many alignment
trimming tools it can be difficult to track exactly which changes the software has made, so
a visual output showing these changes could be helpful.
Transparency is often an issue with bioinformatics software, with poor reporting of
exactly how a file has been processed (Petyuk, Gatto & Payne, 2019; Brito et al., 2020;
Langille, Ravel & Fricke, 2018). CIAlign has been developed to process alignments in a
transparent manner, to allow the user to clearly and reproducibly report their methodology.
CIAlign is freely available at http://github.com/KatyBrown/CIAlign.

MATERIALS & METHODS

CIAlign is a command line tool implemented in Python 3. It can be installed either via
pip3 or from GitHub and is independent of the operating system. It has been designed to
enable the user to remove specific issues from an MSA, to visualise the MSA (including a
markup file showing which regions and sequences have been removed) and to interpret
the MSA in several ways. CIAlign works on nucleotide or amino acids alignments and will
detect which of these is provided. A log file is generated to show exactly which sequences

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 4/29

 and positions have been removed from the alignment and why they were removed. Users
can then adjust the software parameters according to their needs.
CIAlign takes as its input any pre-computed alignment in FASTA format containing
at least two sequences (for some cleaning functions three sequences are required). Most
MSAs created with standard alignment software will be of an appropriate scale, for example
single or multi-gene alignments and whole genome alignments for many microbial species.
The path to the alignment file is the only mandatory parameter. Every function is run
only if specified in the parameters and many function-specific parameters allow options
to be fine-tuned. Using the parameter option all will turn on all the available functions
and run them with the default parameters, unless otherwise specified. The clean option
will run all cleaning functions, visualise all the visualisation functions and interpret
all the interpretation functions, again with the default parameters. Additionally, the user
can provide parameters via a configuration file instead of via the command line.
CIAlign has been designed to maximise usability, reproducibility and reliability. The
code is written to be as readable as possible and all functions are fully documented. All
functions are covered by unit tests. CIAlign is freely available, open source and fully version
controlled.

Cleaning alignments
CIAlign consists of several functions to clean an MSA by removing commonly encountered
alignment issues. All of these functions are optional and can be fine-tuned using user
parameters. All parameters have default values. The available functions are presented here
in the order they are executed by the program. The order can have a direct impact on the
results, the functions removing positions that lead to the greatest disruptions in the MSA
should be run first, as they potentially make removing more positions unnecessary and
therefore keep processing to a minimum. For example, divergent sequences often contain
many insertions compared to the consensus, so removing these sequences first reduces the
number of insertions which need to be removed. Sequences can be made shorter during
processing with CIAlign and therefore too short sequences are removed last.
Figure 1 shows a graphical representation of an example toy alignment before (Fig. 1A)
and after (Figs. 1B–1F) using each function individually. The remove gap only function is
run by default after every cleaning step, unless otherwise specified by the user.

Remove divergent
For each column in the alignment, this function finds the most common nucleotide
or amino acid and generates a temporary consensus sequence. Each sequence is then
compared individually to this consensus sequence. Sequences which match the consensus
at a proportion of positions less than a user-defined threshold (default: 0.65) are excluded
from the alignment (Fig. 1B). It is recommended to run the make similarity matrix function
to calculate pairwise similarity before removing divergent sequences, in order to adjust the
parameter value for more or less divergent alignments. This function requires an alignment
of three or more sequences.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 5/29

 Figure 1 Mini alignments showing the main functionalities of CIAlign based on Example 1. (A)
Input alignment before application of CIAlign, generated using the command ‘‘CIAlign infile
example1.fasta plot_input’’. (B) Output alignment showing the functionality of the remove divergent
function, generated using the command ‘‘CIAlign infile example1.fasta remove_divergent
plot_output’’. (C) Output alignment showing the functionality of the remove insertions function,
generated using the command ‘‘CIAlign infile example1.fasta remove_insertions plot_output’’.
(D) Output alignment showing the functionality of the crop ends function, generated using the command
‘‘CIAlign infile example1.fasta crop_ends plot_output’’. (E) Output alignment showing the
functionality of the remove short sequences function, generated using the command ’’CIAlign infile
example1.fasta remove_short plot_output’’. (F) Output alignment showing the functionality of
the remove gap only function, generated using the command ‘‘CIAlign infile example1.fasta
plot_output’’. Subplots were generated using the draw mini alignment function of CIAlign. In all
subplots sequences are labelled according to their position in the input alignment.
Full-size DOI: 10.7717/peerj.12983/fig-1

Remove insertions
In order for CIAlign to define a region as an insertion, an alignment gap must be present
in the majority of sequences and flanked by a minimum number of non-gap positions on
either side, which can be defined by the user (default: 5). This pattern can be the result
of an insertion in a minority of sequences or a deletion in a majority of sequences. The
minimum and maximum size of insertion to be removed can also be defined by the user
(default: 3 and 200, respectively) (Fig. 1C). This function requires an alignment of three or
more sequences.

Crop ends
The crop ends function redefines where each sequence starts and ends, based on the
ratio of the numbers of gap and non-gap positions observed up to a given position in
the sequence. It then replaces all non-gap positions before and after the redefined start
and end, respectively, with gaps. This will be described for redefining the sequence start,
however crop ends is also applied to the reverse of the sequence to redefine the sequence

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 6/29

 Figure 2 Crop ends diagram. This manually created example illustrates how the crop ends function
works internally. The length of the sequence shown is 111 including gaps and 80 excluding gaps (1). With
a threshold of 10% for the proportion of non-gap positions to consider for change in end positions, 8 po-
sitions at the start and at the end, respectively, are being considered (illustrated by red crossbars). For each
of these, the number of preceding gaps is calculated (2). Then the change in gap numbers (3) for every
two consecutive non-gap positions is compared to the gap number change threshold, which is 5%, i.e. 4
gaps, as a default value. Looking at the change in gap numbers, the last change at each end equal to or big-
ger than the threshold is coloured in red. This leads to redefining the start and the end of this example se-
quence to be where the nucleotides are coloured in green.
Full-size DOI: 10.7717/peerj.12983/fig-2

end. The number of gap positions separating every two consecutive non-gap positions is
compared to a threshold and if that difference is higher than the threshold, the start of the
sequence will be reset to that position. This threshold is defined as a proportion of the total
sequence length, excluding gaps, and can be defined by the user (default: 0.05) (Figs. 1D,
2). The user can set a parameter that defines the maximum proportion of the sequence for
which to consider the change in gap positions (default: 0.1) and therefore the innermost
position at which the start or end of the sequence may be redefined. It is recommended to
set this parameter no higher than 0.1, since even if there are a large number of gap positions
beyond this point, this is unlikely to be the result of incomplete sequences (Fig. 2). This
function requires an alignment of three or more sequences.

Remove short sequences

The remove short function removes sequences which have less than a specified number of
non-gap positions,which can be set by the user (default: 50) (Fig. 1E).

Remove gap only columns

The remove gap only function only removes columns that contain only gaps. These could
be introduced by manual editing of the MSA before using CIAlign or by running the
functions above (Fig. 1F). The main purpose of the function is to clean the gap only
columns that are likely to be introduced after running any of the cleaning functions.

Visualisation
There are several ways of visualising the alignment, which both allow the user to interpret
the alignment and clearly show which positions and sequences CIAlign has removed.
CIAlign can also be used simply to visualise an alignment, without running any of the
cleaning functions. All visualisations can be output as publication ready image files.

Mini alignments
CIAlign provides functionality to generate mini alignments, in which an MSA is visualised
using coloured rectangles on a single x and y axis, with each rectangle representing a single
nucleotide or amino acid (e.g., Fig. 1, 3–5). Even for large alignments, this function provides

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 7/29

 Figure 3 Mini alignments and legends showing further functionalities of CIAlign based on Example
1. (A) Alignment showing the functionality of the plot markup function, generated using the command
‘‘CIAlign infile example1.fasta all’’. The areas that have been removed are marked up in differ-
ent colours, each corresponding to a certain function of CIAlign. (B) Output alignment after application
of all functions of CIAlign combined, generated using the command ‘‘CIAlign infile example1.fasta
all’’. Subplots were generated using the draw mini alignment function.
Full-size DOI: 10.7717/peerj.12983/fig-3

a visualisation that can be easily viewed and interpreted. Many properties of the resulting
file (dimensions, DPI, file type) are parameterised. In order to minimise the memory and
time required to generate the mini alignments, the matplotlib imshow function (Hunter,
2007) for displaying images is used. Briefly, each position in each sequence in the alignment
forms a single pixel in an image object and a custom dictionary is used to assign colours.
The image object is then stretched to fit the axes.

Sequence logos
CIAlign can generate traditional sequence logos (Schneider & Stephens, 1990) or sequence
logos using rectangles instead of letters to show the information and base/amino acid
content at each position, which can increase readability in less conserved regions. Sequence
logos can also be generated for sections of the alignment if a set of boundary coordinates
is provided.

Interpretation
Some additional functions are provided to further interpret the alignment, for example
plotting the number of sequences with non-gap residues at each position (the coverage),
calculating a pairwise similarity matrix and generating a consensus sequence with various
options.
Given the toy example shown in Fig. 1A, running all possible cleaning functions will
lead to the markup plot shown in Fig. 3A and the result shown in Fig. 3B. In the markup
plot each removed part is highlighted in a different colour corresponding to the function
with which it was removed.

Example alignments
Four example alignments are provided within the software directory to demonstrate the
functionality of CIAlign. Examples 1 and 2 use simulated sequences, Examples 3 and 4
use real biological sequences and are designed to resemble the type of complex alignment
many researchers encounter.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 8/29

 Figure 4 Mini alignments showing the main functionalities of CIAlign based on Example 2. (A)
Input alignment before application of CIAlign, generated using the command ‘‘CIAlign infile
example2.fasta plot_input’’. (B) Alignment markup showing areas that were removed by CIAlign,
generated using the command ‘‘CIAlign infile example2.fasta all’’. (C) Output alignment after
application of CIAlign, generated using the command ‘‘CIAlign infile example2.fasta all’’.
Subplots were generated using the draw mini alignment function.
Full-size DOI: 10.7717/peerj.12983/fig-4

Example 1 is a very short alignment of six sequences which was generated manually
by creating arbitrary sequences of nucleotides that would show every cleaning function
while being as short as possible. This alignment contains an insertion, gaps at the ends of
sequences, a very short sequence and some highly divergent sequences.
Example 2 is a larger alignment based on randomly generated amino acid sequences
using RandSeq (a tool from ExPASy (Gasteiger et al., 2003)) with an average amino acid
composition, which were aligned with MAFFT v7.407, under the default settings (Katoh et
al., 2002). The sequences were adjusted manually to reflect an alignment that would fully
demonstrate the functionalities of CIAlign. It consists of many sequences that align well,

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 9/29

 Figure 5 Mini alignments showing the main functionalities of CIAlign based on Example 3. (A)
Input alignment before application of CIAlign, generated using the command ‘‘CIAlign infile
example3.fasta plot_input’’. (B) Output alignment after application of CIAlign, generated using the
command ‘‘CIAlign infile example3.fasta all remove_divergent_minperc 0.5’’. Subplots were
generated using the draw mini alignment function.
Full-size DOI: 10.7717/peerj.12983/fig-5

however there are again a few problems: one sequence has a large insertion, one is very
short, one is extremely divergent and some have multiple gaps at the start and at the end.
For Example 3, putative mitochondrial gene cytochrome C oxidase I (COI) sequences
were identified by applying TBLASTN v2.9.0 (Camacho et al., 2009) to the human COI
sequence (GenBank accession NC_012920.1, positions 5,904–7,445, translated to amino
acids), querying against 1,565 transcriptomic datasets from the NCBI transcriptome
shotgun assembly (TSA) database (Transcriptome Shotgun Assembly Sequence Database,
https://www.ncbi.nlm.nih.gov/genbank/tsa/) under the default settings. 2,855 putative COI
transcripts were reverse complemented where required, and those corresponding to the
COI gene of the primary host of the TSA dataset were identified using the BOLD online
specimen identification engine (Ratnasingham & Herbert, 2007) (accessed 07/10/2019)
querying against the species level barcode records. The resulting 232 sequences were then
aligned with MAFFT, under the default settings.
For Example 4, 91 sequences were selected from Example 3 to be representative of as
many taxonomic families as possible and to exclude families with unclear phylogeny in
the literature. These sequences were aligned with MAFFT under the default settings and

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 10/29

 the alignment was refined with 1000 iterations. Robinson-Foulds (RF) distances (Robinson
& Foulds, 1981) of the resulting trees were calculated using ete3 compare (Huerta-Cepas,
Serra & Bork, 2016).
Materials and methods for benchmarking and for larger scale examples with biological
data are provided as Materials and Methods S1.

RESULTS
Here, an example is presented and the visualisation functions are used to illustrate the
functionality of CIAlign. Results will differ when using different parameters and thresholds.
CIAlign was applied to the Example 2 alignment with the following options:
python3 CIAlign.py infile INFILE outfile_stem OUTFILE_STEM all
Using these settings on the alignment in Fig. 4A results in the markup shown in Fig. 4B
and the output shown in Fig. 4C. The markup shows which function has removed each
sequence or position. The benefits of CIAlign are clear in this simulation—the single poorly
aligned sequence, the large insertion, very short sequences and gap-only columns have been
removed and the unreliably aligned end segments of the sequences have been cropped. The
resulting alignment is significantly shorter, which will speed up and simplify any further
analysis. The clear graphical representation makes it easy to see what has been removed, so
in the case of over-trimming the user can intervene and adjust functions and parameters.
In order to demonstrate the use of CIAlign on real biological sequences, an alignment
was generated based on the COI gene commonly used in phylogenetic analysis and DNA
barcoding (Ratnasingham & Herbert, 2007). As CIAlign addresses some common problems
encountered when generating an MSA based on de novo assembled transcripts, which tend
to have a higher error rate at transcript ends, gaps due to difficult to assemble regions and
divergent sequences due to chimeric connections between unrelated regions (Bushmanova
et al., 2019; Liao et al., 2019), COI-like transcripts were identified by searching the NCBI
transcriptome shotgun assembly database. Aligning these transcripts demonstrated several
common problems—multiple insertions, poor alignment at the starts and ends of sequences
and a few divergent sequences resulting in excessive gaps (Fig. 5A). This alignment was
cleaned using the default CIAlign settings except the threshold for removing divergent
sequences was reset to 50%, as some of the sequences are from evolutionarily distant
species. Cleaning this alignment with CIAlign took an average of 68.1 s and used on
average a maximum of 1.13GB of RAM (mean across 10 runs, on one Intel Core i7-7560U
core with 4 GB of RAM, running at 2.40 GHz, RAM measured as maximum resident set
size, this machine and 10 replicates were also used for all subsequent measurements of
CIAlign resource requirements in this section). Under these settings, CIAlign resolved
several of the problems with the alignment: the insertions and highly divergent sequences
were removed and the poorly aligned regions at the starts and ends of sequences were
cropped (Fig. 5B). One sequence and 6,029 positions were removed from the alignment
and a total of 2,446 positions were cropped from the ends of 112 sequences. The processed
alignment is 26.6% of the size of the input alignment. However, a minimal amount of
actual sequence data (as opposed to gaps) was removed, with 85.7% of bases remaining.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 11/29

 A subset of this sequence set was selected to demonstrate the functionality of CIAlign
in streamlining phylogenetic analysis. 91 COI-like transcripts from different taxonomic
families of metazoa were selected from Example 3, incorporated into an MSA and cleaned
using CIAlign with the same settings as above (Fig. S1). CIAlign took an average of 20.8 s to
clean this alignment and used, on average, a maximum of 486 MB of RAM. 1,437 positions
were removed from the alignment and a total of 289 positions were cropped from the ends
of 17 sequences. The processed alignment is 70.7% of the size of the input alignment and
96.5% of bases remain. Phylogenetic trees were generated for the input alignment and for
the alignment processed with CIAlign, using PhyML (Guindon & Gascuel, 2003) under the
GTR model plus the default settings. For the input alignment, PhyML used 138 MB of
memory and took 532 s . For the cleaned alignment PhyML used 109 MB of memory and
took 243 s. The tree generated with the input alignment (Fig. S1D) had an RF distance
from a ‘‘correct’’ tree (generated manually based on the literature, Fig. S1D, literature
listed in Materials and Methods S1) of 100 (normalised Robinson-Foulds (n-RF) 0.570,
(QD) (Smith, 2019) 0.159). The tree generated with the cleaned alignment (Fig. S1E) had
an RF distance from the correct tree of 90 (n-RF 0.520, QD 0.073) Therefore, the tree based
on the CIAlign cleaned alignment was generated more quickly and was more similar to the
expected tree.

Testing with simulated and benchmark data

EvolvAGene, INDELible and BAliBASE—alignment and phylogeny
We performed a series of benchmarking analyses on simulated and benchmark data, in
order to test and demonstrate the utility of the CIAlign cleaning functions, confirm the
validity of our default parameter settings and ensure that running these functions does
not have unexpected negative effects on downstream analyses. Running any tool which
removes residues from an alignment has a potential cost, so these tests are intended to
allow users to weigh this against the benefit of running CIAlign for their intended use case.
First, CIAlign was tested using three tools: EvolvAGene (Hall, 2008), INDELible (Fletcher
& Yang, 2009) and BAliBASE (Bahr et al., 2001). EvolvAGene and INDELible generate
sets of unaligned sequences alongside ‘‘true’’ alignments and phylogenies expected to
accurately represent the relationship between the sequences (Hall, 2008; Fletcher & Yang,
2009). BAliBASE is a set of alignments designed for benchmarking sequence alignment
tools (Bahr et al., 2001). We used these tools to determine if cleaning a user generated
alignment with CIAlign affects its distance from the true alignment.
Test alignments were created using four common alignment algorithms—Clustal
Omega (Sievers & Higgins, 2018), MUSCLE (Edgar, 2004), MAFFT global (FFT-NS-i)
(Katoh et al., 2002) and MAFFT local (L-NS-i) (Katoh et al., 2002). These alignments
were then cleaned with CIAlign with relaxed, moderate or stringent parameter settings
(Table S1). With relaxed CIAlign settings, a median of 0.400% of correct pairs of aligned
residues (POARs) (Thompson, Plewniak & Poch, 1999) were removed, for moderate settings
2.31% were removed and for stringent settings 6.06% were removed (Fig. 6A, Table 1).
For comparison, the median total proportion of residues removed was 2.38% for relaxed,
3.24% for moderate and 5.36% for stringent (Fig. 6A, Table 1). The median proportions of

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 12/29

 Figure 6 Metrics from benchmarking CIAlign with simulated data. (A) Box plots showing the impact
of running CIAlign cleaning functions with relaxed (green, R, left box), moderate (blue, M, middle box)
and stringent (pink, S, right box) parameter values on alignments of sequences simulated using either
EvolvAGene (Bahr et al., 2001) or INDELible (Sievers & Higgins, 2018) and on the BAliBASE (Thompson,
Plewniak & Poch, 1999) benchmark alignments (plots are combined for the three tools, for separated plots
see Fig. S3). From left to right, the y-axis represents proportion of correctly aligned pairs of residues (Siev-
ers et al., 2013) removed (identified by comparison with a benchmark alignment), proportion of total nu-
cleotides (i.e. non-gap positions) removed, proportion of gaps removed, proportion of positions (gap or
non-gap) removed. (continued on next page. . . )
Full-size DOI: 10.7717/peerj.12983/fig-6

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 13/29

 Figure 6 (...continued)
(B) Scatter plot showing a linear regression analysis of the impact of the total proportion of positions re-
moved on the proportion of correctly aligned pairs of residues removed by CIAlign for relaxed, mod-
erate and stringent parameter values. The statistic m is the slope of the regression line. (C) Violin plots
showing the distribution of normalised Robinson-Foulds distances (Hall, 2008) (left column) and Quar-
tet divergence (right column) (Fletcher & Yang, 2009) between benchmark trees and test trees without
running CIAlign cleaning functions (yellow) and after running CIAlign with the three sets of parameter
values, for trees based on simulated sequences generated with EvolvAGene (Bahr et al., 2001) (top row)
and INDELible (Sievers & Higgins, 2018) (bottom row). Red and black lines show the median and mean,
respectively. (D) Density plot showing the distribution of the percentage identity between the input se-
quence to EvolvAGene (Bahr et al., 2001) and a consensus sequence based on an alignment of the sim-
ulated sequences generated by this tool, without running CIAlign (yellow) and after running CIAlign
cleaning functions with the three sets of parameter values. (E) Density plots showing the distribution of
the percentage identity between the input sequence to BadRead (Sievers & Higgins, 2020) and a consen-
sus sequences generated with (blue) and without (yellow) running CIAlign cleaning functions for align-
ments of good (top), medium (middle) and poor (bottom) quality simulated reads. (F) Box plot show-
ing the proportion of correct positions removed by the CIAlign cleaning functions for alignments of good,
medium and bad quality simulated reads (left) and scatter plot showing a linear regression analysis of
the impact of the total proportion of positions removed on the proportion of correct residues removed
by CIAlign for each read quality level (right). The statistic m is the slope of the regression line. (G) Box
plots showing the impact of running CIAlign on the mean ZORRO (Wu, Chatterji & Eisen, 2012) col-
umn confidence score (top) and the proportion of columns with high ZORRO column confidence scores
(>0.4) for EvolvAGene (Bahr et al., 2001) (left), INDELible (Sievers & Higgins, 2018) (centre) and BAl-
iBASE (Thompson, Plewniak & Poch, 1999) (right) alignments.

gap positions removed were much higher: 51–56% for all sets of parameters (Fig. 6A, Fig.
S2, Table 1). This shows that with relaxed and moderate settings, running CIAlign has a very
minimal impact on correctly aligned residues in the alignment, while a considerable amount
of gaps and noise are removed. The more stringent settings should be used cautiously,
however, even with high stringency, a large majority of correctly aligned residues remain
and the majority of gaps are removed. These results are separated by simulation tool
(EvolvAGene, INDELible or BAliBASE) and alignment tool (MUSCLE, MAFFT global,
MAFFT local and Clustal Omega) in Fig. S2.
To directly compare the impact of CIAlign on correctly aligned pairs of residues to
its overall impact, we fitted a linear regression line to show how, on average, the overall
proportion of positions removed from the alignment impacts the proportion of correctly
aligned residues reoved (Fig. 6B). The resulting line had a gradient of 0.281 for relaxed
parameters, 0.361 for moderate parameters and 0.554 for stringent parameters. In other
words, for every 1% of material removed from the alignment by CIAlign with relaxed
settings, an average of only 0.281% of correctly aligned residue pairs will be removed,
with moderate settings 0.361% and with stringent settings 0.554% (Fig. 6B). This will vary
depending on the input alignment and the use case. These results are shown separately
for MUSCLE, MAFFT and Clustal Omega in Fig. S2E. The impact of CIAlign on correctly
aligned pairs is most severe on the Clustal Omega EvolvAGene alignments, which have
lower pairwise identity than the alignments generated with the other tools and so have
more sequences removed entirely by the remove divergent function (discussed below).
In most cases, CIAlign is not intended or expected to change the phylogenetic tree
resulting from an alignment, although in many cases it will make building phylogenetic

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 14/29

 Table 1 Table showing the impact of running CIAlign cleaning functions with relaxed, moderate and stringent parameter values on alignments
of sequences simulated using either EvolvAGene (Bahr et al., 2001) or INDELible (Sievers & Higgins, 2018) and on the BAliBASE (Thompson,
Plewniak & Poch, 1999) benchmark alignments (results are combined for the three tools). For each stringency level, the median percentage
of correctly aligned pairs of residues (Sievers et al., 2013) removed (identified by comparison with a benchmark alignment), proportion of
total nucleotides (i.e., non-gap positions) removed, proportion of gaps removed and proportion of positions (gap or non-gap) removed have
been calculated for EvolvAGene, INDELible and BAliBASE. The mean normalised Robinson-Foulds (RF) distance (Hall, 2008) and Quartet
divergence (Fletcher & Yang, 2009) are based on comparison with benchmark trees for EvolvAGene and INDELible. Consensus percentage identity
is between the input sequence to EvolvAGene and a consensus sequence based on an alignment of the simulated sequences generated by this
tool. Confidence scores are the mean ZORRO (Wu, Chatterji & Eisen, 2012) column confidence scores and the proportion of columns with high
ZORRO column confidence scores (>0.4) for EvolvAGene, INDELible (Sievers & Higgins, 2018) and BAliBASE (Thompson, Plewniak & Poch, 1999)
alignments. All statistics are two-sided Mann Whitney U tests comparing the alignment without running CIAlign to the alignment after running
CIAlign with the specified parameters.

Metric Statistic CIAlign stringency

None Relaxed Moderate Stringent
Correct Pairs Removed Median % – 0.400 2.31 6.06
Nucleotides Removed Median % – 2.38 3.24 5.36
Gaps Removed Median % – 51.7 52.0 55.9
Positions Removed Median % – 9.62 10.6 13.3
Mean 0.241 0.240 0.246 0.250
MWU Test Statistic – 320490 316553 312115
Normalised RF Distance
MWU P-value – 0.955 0.695 0.394
Significance – – – –
Mean 0.162 0.163 0.167 0.171
MWU Test Statistic – 320125 316179 311455
Quartet Divergence
MWU P-value – 0.989 0.665 0.356
Significance – – – –
Mean 67.2 71.5 71.5 71.5
Consensus Percentage MWU Test Statistic – 23294 22924 23258
Identity MWU P-value – 1.89E -67 2.61E -68 1.56E -67
*** *** ***
Significance –
Mean 3.66 4.68 4.63 4.72
MWU Test Statistic – 688583 700927 688059
Confidence Score
MWU P-value – 8.65E -31 7.84E -28 3.61E -33
*** *** ***
Significance –
Mean 69.1 84.3 84.0 85.6
Percentage High MWU Test Statistic – 465471 477660 462908
Confidence Columns MWU P-value – 2.44E -111 1.31E -105 6.89E -116
*** *** ***
Significance –
Notes.
Significance is shown as *** if the p-value is less than 0.001, ** if the p-value is less than 0.01, * if the p-value is less than 0.05 and–if the p-value is greater than 0.05.

trees faster. To test this, phylogenetic trees were generated for each of the EvolvAGene
and INDELible alignments (BAliBASE does not provide reference trees) to determine
if cleaning with CIAlign impacts the distance between the true phylogenetic tree and a
phylogenetic tree based on a test alignment (Fig. 6C, Table 1). For the EvolvAGene and
INDELible alignments, the mean n-RF distance and QD between the test trees and true trees
were virtually unchanged by running CIAlign and none of the changes were statistically

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 15/29

 significant (n-RF p = 0.955, 0.695, 0.394, QD p = 0.989, 0.665, 0.356 for relaxed, moderate,
stringent, respectively, Mann Whitney U test) (Fig. 6C, Table 1).
We also compared the input sequence for our EvolvAGene simulations to consensus
sequences based on alignments with and without CIAlign cleaning. For all three stringency
levels, CIAlign increased the percentage nucleotide identity between the consensus sequence
and the input sequence by between 4% and 5% (Fig. 6D, Table 1). All of these changes
are statistically significant (relaxed: p = 1.89E−67, moderate: p = 2.61E−68, stringent,
p = 1.56E−67, Mann–Whitney U test).
The long-read sequencing simulation tool BadRead (Wick, 2019) was used to
demonstrate the use of CIAlign to remove common sources of error in long read sequencing
data. Sequences were generated to represent low, moderate and high quality Oxford
Nanopore reads based on an input genome, then aligned and cleaned with CIAlign with
moderate settings (Table S1). Using CIAlign increased the identity between the alignment
consensus and the input sequence significantly for all read quality levels—by 6.57% for high
quality reads, 9.51% for moderate quality reads and 12.3% for poor quality reads (Fig. 6E,
Table S2) (p = 2.22E−35, 1.37E−13, 1.55E−9, respectively, Mann–Whitney U test). For
the high quality reads, the reads cleaned with CIAlign generated consensus sequences
almost identical to the input sequence, with a mean of 99.2% identity (Fig. 6E, Table S2).
The proportion of the positions removed from the alignment which were correct (in this
case positions in the alignment which match the input sequence used to generate the reads)
was calculated in order to demonstrate the potential cost of running CIAlign. For the good
quality simulated reads, a median of 3.99% of the positions which were removed match the
input sequence, for medium quality 5.03% and for low quality 7.31% (Fig. 6F, Table S2).
A linear regression analysis showed that, on average, removing 1% of total positions with
CIAlign removes 0.0740% of correct positions for good quality simulated reads, 0.504%
for medium quality reads and 0.491% for bad quality reads (Fig. 6F).
The alignment masking tool ZORRO (Wu, Chatterji & Eisen, 2012) provides a
confidence score (maximum 10) for each column in the MSA, representing a measure
of uncertainty in that column. This confidence score was measured for each column of
each of the EvolvAGene, INDELible and BAliBASE alignments. The mean confidence score
increased by 1.02 for relaxed, 0.970 for moderate and 1.06 for stringent CIAlign settings, all
of which are significant improvements (p = 8.65E−31, 7.84e−28, 3.61E−33, respectively,
Mann–Whitney U test) (Fig. 6G). The proportion of columns with a confidence score
greater than 0.4 (the minimum suggested in the ZORRO documention (Wu, Chatterji
& Eisen, 2012)) was also measured and increased by 15.2%, 14.9% and 16.5% for
relaxed, moderate and stringent CIAlign settings (p = 2.44E−111, 1.31E−105, 6.88E−116,
respectively, Mann–Whitney U test) (Fig. 6G, Table 1).

HomFam—alignment and phylogeny

CIAlign was also benchmarked using the HomFam (Sievers et al., 2013) set of benchmark
alignments, for which a small set of sequences which can be reliably aligned (referred to
henceforth as the seed sequences) are provided alongside a much larger set of sequences
which are variably distant from the seed (the test sequences). The seed sequences were

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 16/29

 aligned with (‘‘seed + test alignment’’) and without (‘‘seed-only alignment’’) the test
sequences. We used these benchmark datasets to determine if running the CIAlign cleaning
functions can bring the alignment of the seed sequences in the seed + test alignment closer
to that of the seed sequences in the seed-only alignment.
A median of 2.10% of correctly aligned residue pairs and 8.22% of residues were removed
from the seed sequences in the seed + test alignments, while 92.1% of gaps introduced into
the seed sequences were removed (Fig. 7A, Table 2). Regression analysis showed an average
loss of 0.130% of correctly aligned residue pairs for every 1% of the alignment removed
with CIAlign (Fig. 7B). There was no significant change in seed sequence phylogeny
from the seed + test alignment before and after running CIAlign (nRF, p = 0.928, QD,
p = 0.672, Mann–Whitney U test) (Fig. 7C, Table 2). Comparing the consensus for the
seed sequences in the seed-only alignment with the consensus for the same sequences in
the seed + reference alignment, the mean percentage identity increased dramatically by
28.8% after running CIAlign (p = 2.35E−17, Mann–Whitney U test) (Fig. 7D, Table 2).

QuanTest2—protein structure prediction

The tool Quantest2 (Sievers & Higgins, 2020) allows benchmarking of alignment quality
in terms of its impact on protein secondary structure prediction. We therefore tested the
impact of CIAlign on the percentage similarity between reference secondary structures
and those predicted based on an alignment with multiple other sequences. We aligned
the sequence sets provided in this benchmark and cleaned the alignments with CIAlign
(Table S1). A mean of 76.0% of positions in the secondary structure of the reference
sequences in the CIAlign cleaned alignment were consistent with the reference structure,
compared to 67.9% of positions in the original alignments, a significant improvement
of 8.13% (Fig. 7E, Table 2) (p = 9.35E−20, Mann–Whitney U test). A linear regression
demonstrated that any cleaning with CIAlign increases, on average, the percentage of
correct positions in the resulting structure but that the benefit decreases linearly with the
amount of material removed by CIAlign (Fig. 7F).

Benchmarking data availability

Full output tables for the simulations with EvolvAGene, INDELible, BAliBASE, BadRead,
HomFam and QuanTest2 are available in Online Tables 1–4 at http://github.com/
KatyBrown/CIAlign/benchmarking/tables and the simulated data and alignments at
https://github.com/KatyBrown/benchmarking_data_CIAlign.

Comparing alignment tools

In addition to our primary analyses using MAFFT (Edgar, 2004), MUSCLE (Katoh et
al., 2002) and Clustal Omega (Sievers & Higgins, 2018), we measured the performance of
CIAlign with a number of other alignment tools, including progressive, iterative, non-
heuristic, consistency based, HMM-based, context based and phylogeny aware methods
(Supplemental Information 1, Table S3).
CIAlign performed similarly with most alignment tools in terms of not excessively
removing correctly aligned residues. The mean proportion of correctly aligned pairs
removed was 2.80% across all simulations, tools and stringency levels, with a standard

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 17/29

 Figure 7 Metrics from benchmarking CIAlign using HomFam and QuanTest2. (A) Box plot show-
ing the impact of running CIAlign with moderate settings (Table S2) on the seed sequences in combined
alignments of seed and test sequences from the HomFam benchmark set (Wright, 2015), from left to right,
the y-axis represents proportion of correctly aligned pairs of residues (Sievers et al., 2013) removed (iden-
tified by comparison with alignments of the seed sequences only), proportion of total nucleotides (i.e.,
non-gap positions) removed, proportion of gaps removed, proportion of positions (gap or non-gap) re-
moved. (B) Scatter plot showing a linear regression analysis of the impact of the total proportion of posi-
tions removed on the proportion of correctly aligned pairs of residues removed by CIAlign (identified by
comparison with alignments of the seed sequences only) for the HomFam benchmark set. The statistic m
is the slope of the regression line (C) Violin plot showing the distribution of normalised Robinson-Foulds
distances (Hall, 2008) (nRF) and Quartet divergence (qD) (Fletcher & Yang, 2009) between maximum
likelihood trees generated based on seed sequences in alignments of seed sequences only and alignments
of seed sequences plus test sequences from the HomFam benchmark set (Wright, 2015), with (blue) and
without (orange) cleaning with CIAlign. (D) Density plot showing the distribution of the percentage iden-
tity (top), Needleman-Wunsch score (middle) (Needleman & Wunsch, 1970) and alignment width be-
tween consensus sequences generated from seed sequence only alignments and consensus sequences gen-
erated from combined seed and test sequences in the HomFam benchmark set (Wright, 2015). (E) Den-
sity plot showing the distribution of the percentage similarity between reference secondary structures and
secondary structures based on alignments before (orange) and after (blue) running CIAlign with moder-
ate stringency settings (Table S2), calculated using QuanTest2 (Finn et al., 2014) and using the QuanTest2
reference structures and test alignments. (F) Scatter plot showing a linear regression analysis of the impact
of the percentage of the original sequence length remaining after running CIAlign, with moderate parame-
ter values (Table S2), on the change in the percentage of correct positions in the structure prediction after
running CIAlign. The statistic m is the slope of the regression line.
Full-size DOI: 10.7717/peerj.12983/fig-7

deviation of 5.36% (Fig. S3A, Table S3). There was one particular outlier for this metric,
with Clustal Omega (Sievers & Higgins, 2018), a HMM-based method, using stringent
settings removes a higher proportion of correctly aligned residues for the EvolvAGene
nucleotide simulations (median 24.5%). This is the result of a higher proportion of
sequences being removed by the remove divergent function, as the mean percentage
identity between pairs of sequences in the Clustal Omega alignments is lower (with a mean

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 18/29

 Table 2 Table showing the impact of running CIAlign with moderate settings (Table S2) on the seed sequences in combined alignments of
seed and test sequences from the HomFam benchmark set (Wright, 2015). The median proportion of correctly aligned pairs of residues (Sievers et
al., 2013) removed (identified by comparison with alignments of the seed sequences only), proportion of total nucleotides (i.e., non-gap positions)
removed, proportion of gaps removed, proportion of positions (gap or non-gap) removed were calculated for all HomFam datasets. Normalised
Robinson-Foulds distances and Quartet divergences are between maximum likelihood trees generated based on seed sequences in alignments of seed
sequences only and alignments of seed sequences plus test sequences from the HomFam benchmark set (Wright, 2015), before and after running
CIAlign. Consensus percentage identity is between consensus sequences generated from seed sequence only alignments and consensus sequences
generated from combined seed and test sequences in the HomFam benchmark set (Wright, 2015). QuanTest2 percentage similarity is the percent-
age similarity between reference secondary structures and secondary structures based on alignments before and after running CIAlign with moder-
ate stringency settings (Table S2), calculated using QuanTest2 (Finn et al., 2014) and using the QuanTest2 reference structures and test alignments.
All statistics are two-sided Mann Whitney U tests comparing alignments before and after running CIAlign.

Metric Statistic Before/After CIAlign

cleaning
Before After
Correct Pairs Removed Median % – 2.1
Nucleotides Removed Median % – 8.22
Gaps Removed Median % – 92.13
Positions Removed Median % – 70.38
Mean 0.19 0.19
MWU Test Statistic – 3542
Normalised RF Distance
MWU P-value – 0.93
MWU Significance – –
Mean 0.11 0.11
MWU Test Statistic – 3693
Quartet Divergence
MWU P-value – 0.67
MWU Significance – –
Mean 19.77 48.58
MWU Test Statistic – 6264
Consensus Percentage Identity
MWU P-value – 2.35E -17
***
MWU Significance –
Mean 67.86 75.99
MWU Test Statistic – 17650
QuanTest2 Percentage Similarity
MWU P-value – 9.35E -20
***
MWU Significance –
Notes.
Significance is shown as *** if the p-value is less than 0.001, ** if the p-value is less than 0.01, * if the p-value is less than 0.05 and—if the p-value is greater than 0.05.

of 57.9% identity) than the threshold of 65% identity used to remove divergent sequences
under the stringent CIAlign settings (Table S1, Fig. S3B).
Otherwise, the extent to which CIAlign will remove positions from an alignment is
primarily related to the number of gaps introduced by the alignment software. Amino acid
alignments generated with the tool DECIPHER (Wright, 2015) are outliers because this
tool introduces fewer and shorter internal gaps (as opposed to terminal gaps) into these
alignments than any other tool (under the default settings), which reduces the number
of positions meeting the criteria to be removed with either the crop ends or the remove
insertions functions (Fig. S3C, Table S3). Across all tools, there is a positive correlation
between the proportion of gaps in the input alignment and the proportion of residues

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 19/29

 (r = 0.793, p = 1.01E−33, Spearman’s ρ), gaps (r = 0.480, p = 4.99E−10), positions
(r = 0.890, p = 1.84E−52) and correctly aligned pairs (r = 0.461, p = 3.00E−9) removed
(Fig. S3D).
CIAlign does not significantly change the distance between the true phylogenetic tree
and the alignment phylogenetic tree for any of the alignment tools (Table S3). It does
however improve the consensus sequence significantly (mean 4.68% improvement) in
every case except for the DECIPHER amino acid alignments (Fig. S3E) (Mann–Whitney
U test, p < 0.05, exact p-values are available in Table S3).
Additional figures showing a full breakdown of the comparisons between alignment
tools are available on the CIAlign GitHub page in the benchmarking/Online_Figures
directory. These results are summarised in Fig. S3 and Table S3.
Full results for all alignment tools are available in Online Table 5 at http://github.
com/KatyBrown/CIAlign/benchmarking/tables and the simulated data and alignments at
https://github.com/KatyBrown/benchmarking_data_CIAlign.

Comparison with Gblocks, trimAl and ZORRO

It is not appropriate to compare CIAlign directly with tools intended specifically to
identify and remove poorly aligned columns, as it is intended to be complementary to
(and, where appropriate, used alongside) such tools. However, we have calculated the
proportion of correctly aligned pairs, gaps and residues removed using the default settings
for Gblocks (Talavera & Castresana, 2007), trimAl (Capella-Gutiérrez, Silla-Martínez &
Gabaldón, 2009) and ZORRO (Wu, Chatterji & Eisen, 2012) as it may be informative for
users familiar with another tool to visualise the relative impact of CIAlign on an alignment.
All p-values for this section are available in Table S4.
Across the EvolvAGene and INDELible alignments, CIAlign removed a median of
0.188% of correctly aligned pairs with the most relaxed settings, 0.749% with moderate
settings and 3.76% with stringent settings (Fig. S4A, Table S4). To compare, Gblocks
removed 22.4%, trimAl 1.42% and ZORRO 0.148% (Fig. S4A, Table S4). CIAlign is
therefore significantly less deleterious of correctly aligned material than Gblocks at all three
stringency levels ,while trimAl falls between the moderate and stringent CIAlign settings
for this measure. ZORRO removes slightly less correctly aligned pairs than CIAlign with
relaxed settings (Fig. S4A, Table S4). CIAlign removes significantly less positions (7.41%,
8.10% and 9.96% for relaxed, moderate and stringent settings) overall than Gblocks
(38.2%) and trimAl (12.8%) at all stringency settings and a similar proportion to ZORRO
(7.64%) when run with moderate settings (Fig. S4A, Table S4). A linear regression, showing
the relationship between the total proportion of positions removed with each tool and
the proportion of correctly aligned residue pairs removed, shows CIAlign with relaxed
settings has a similar trade-off between gain and loss of signal to ZORRO (Fig. S4B). For
moderate CIAlign settings trimAl and CIAlign are comparable, except with Clustal Omega
alignments, where, as discussed above, CIAlign removes a large proportion of divergent
sequences and therefore a greater proportion of correct positions. Highly stringent CIAlign
settings are between trimAl and Gblocks for this metric, again with the exception of Clustal
Omega alignments (Fig. S4B).

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 20/29

 None of these tools significantly increased or decreased the distance between trees
generated with the test alignments and the true trees except Gblocks, which significantly
increased the distance from the true tree with both divergence measures (Fig. S4C,
Table S4). Cleaning with CIAlign generates a consensus sequence with 71.5% identity to
the true consensus with all three sets of CIAlign parameters, this is significantly higher than
any of the other tools (Table S4).
The exact aligned residue pairs removed by CIAlign and the other tools were also
compared, to demonstrate the extent to which CIAlign overlaps with and differs from
the other tools (Fig. S4D). As Gblocks removes a very large proportion of the alignment,
including all gaps, inevitably a large majority of the positions removed by CIAlign are also
removed by Gblocks (Fig. S4D). However, CIAlign precisely targets only positions meeting
its criteria, removing much less material than Gblocks overall. Compared with trimAl,
the most stringent CIAlign settings remove 30.4% unique material (Fig. S4D). At lower
stringency settings the majority of pairs removed by CIAlign are also removed by trimAl,
but trimAl again has a much more severe impact on the alignment. With ZORRO, while
there is a moderate overlap with CIAlign (33.5%, 48.7% and 58.5% for relaxed, moderate
and stringent settings, respectively), there is also a large proportion of material (49.5%,
30.7% and 18.0%) which is uniquely removed by CIAlign (Fig. S4D). When comparing
ZORRO, Gblocks and trimAl directly with each other, the overlap is much greater, with
ZORRO, the most precise of the three tools, removing primarily a subset of the positions
removed by trimAl, which are a subset of those removed by Gblocks (Fig. S4D).
These results demonstrate that CIAlign is performing a different role to these three
tools, as the locations targetted by CIAlign are only removed by other tools at the expense
of large sections of the alignment which CIAlign would leave intact.
Full results for Gblocks, trimAl and ZORRO compared to CIAlign are are available in
Online Table 6 at http://github.com/KatyBrown/CIAlign/benchmarking/tables and the data
at https://github.com/KatyBrown/benchmarking_data_CIAlign.

Realignment
As alignment tools take into account all the sequences and columns in the input file, the
most scrupulous option will always be to unalign and then realign sequences after running
a tool such as CIAlign, rather than using the CIAlign output directly in downstream
analysis. To test the extent to which using CIAlign outputs directly without realignment
could impact results, we removed gaps from the EvolvAGene alignments cleaned with
CIAlign with relaxed, moderate and stringent parameter settings and then reran the
original alignment tool on the result. We then calculated the sum-of-pairs score (Bahr et
al., 2001) treating the realigned file as the true alignment and the CIAlign output as the test
alignment. The mean sum-of-pairs score was 0.984, meaning 98.4% of pairs of nucleotides
aligned realigned MSA were also aligned in the CIAlign output (Fig. S5). This suggests that
while realigning the MSA cleaned with CIAlign is diligent, the effect is likely to be minimal.
The full results of this analysis are available in Online Table 7.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 21/29

 Resource and time requirements
Memory and runtime measurements were conducted by randomly drawing alignments
from the HomFam benchmark set (Sievers et al., 2013) and measuring the time and memory
used for each of the core CIAlign functions. Further measurements were taken by running
the CIAlign core functions on an MSA of constant size with different numbers of gaps. The
runtime decreases linearly with an increasing proportion of gaps. The results are shown in
Fig. S6.
It should be noted that, besides the size of the MSA and its gap content, the runtime is
impacted by which combination of functions is applied. For very long MSAs the size of
the final image becomes a limiting factor when creating a sequence logo, as the matplotlib
library (Hunter, 2007) has restrictions on the number of pixels in one object. We have
provided detailed instructions about this limit in the ‘‘Guidelines for using CIAlign’’ on
the CIAlign GitHub.

Examples of using CIAlign with biological data

We also used CIAlign to clean real biological data from several online databases, in order to
test and demonstrate its usefulness in automated processing of different types of sequencing
data.

Cleaning Pfam alignments

The Pfam database provides manually curated seed alignments for over 17,000 protein
families, plus much larger automatically generated full alignments containing sequences
identified by database searching (Finn et al., 2014). CIAlign cleaning functions were applied
to seed and full alignments for 500 Pfam domains and consensus sequences were generated
for both alignments, before and after cleaning. Randomly selected sequences from the
full alignment were then compared to each consensus. For the full alignments, the mean
identity between the consensus sequence and the alignment sequences increased by 10.7%
(p = 0.00, Mann–Whitney U test) after cleaning with CIAlign (Fig. 8A). For the seed
alignments identity also increased significantly, by 4.89% (p = 0.00, Mann–Whitney U
test) (Fig. 8A). After running CIAlign, the full alignment consensus approaches the level
of similarity to the alignment sequences which is seen for seed alignment consensus,
despite the full alignment having undergone no manual curation (Fig. 8A). Even for the
curated seed alignments, cleaning with CIAlign further increases the similarity between the
consensus and the aligned sequences. Full results are listed in Online Table 8.

Removing insertions and deletions from human genes

To demonstrate the ability of CIAlign to remove non-majority indels, we used data for
50 indels across over 150 individuals from the 1000 genomes project (Auton et al., 2015),
which has annotated insertions and deletions for individual human genomes. In all cases,
CIAlign removed all insertions present in a majority of samples and ignored all insertions
present in a minority of samples (Fig. 8B). Full results are listed in Online Table 9.

Removing outliers
CIAlign can also be used to remove clear outliers from an alignment, for example prior to
phylogenetic analysis. To illustrate this, we ran the CIAlign cleaning functions on data from

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 22/29

 the mammalian 10K trees project (Arnold, Matthews & Nunn, 2010). Three single-gene
trees were identified with clear outliers, the 12S ribosomal gene from primates and the
APOB and RAG1 genes from Carnivora. The issues with these trees are shown in Fig. 8C
and Fig. S7. CIAlign successfully removed the outlying group, without removing any other
sequences, in all three of these cases.

DISCUSSION
We have demonstrated that CIAlign can successfully mitigate the alignment issues caused
by non-majority insertions, poorly aligned sequence ends, highly divergent sequences and
short sequences and demonstrated this capability on specific examples, simulated and
benchmark datasets and large biological datasets. CIAlign has been shown to significantly
improve the accuracy of consensus sequences and secondary structure predictions generated
from MSAs (Figs. 6C and 7D) It also minimises the detrimental effect of adding additional
poorer quality sequences to both benchmark and real alignments (Fig. 7C and 8A). In most
cases, the proportion of correctly aligned material removed by CIAlign is minimal.
It is important to note that while CIAlign is helpful in mitigating alignment issues, using
an appropriate alignment tool and parameters to generate the original alignment is still
essential.

Comparison with other software

While the functionality of CIAlign has some overlaps with other software, for example
Gblocks (Talavera & Castresana, 2007), ZORRO (Wu, Chatterji & Eisen, 2012) and
trimAl (Capella-Gutiérrez, Silla-Martínez & Gabaldón, 2009), the presented software can
be seen as complementary to these, with some different features and applications. Our
analyses have shown that CIAlign can precisely remove insertions, divergent sequences
and poor quality sequence ends without an excessive impact on the rest of the alignment.
CIAlign is much more precise than Gblocks and, except under the most stringent settings,
also removes substantially less positions than trimAl. Therefore, although a side effect of
using these tools may be to remove the specific features targetted by CIAlign, it would be
unnecessarily deleterious for users only wanting to target these features to choose Gblocks
or trimAl. CIAlign removes slightly more material than ZORRO, but much of the material
removed by both tools is unique, indicating that these tools, while similarly precise, are
performing different roles. The impact of CIAlign on the structure of trees generated
from the cleaned alignments was shown to insignificant. ZORRO and trimAl also had an
insignificant impact, while Gblocks had a significant negative impact on tree accuracy.
Compared to non-automated tools, for example Jalview (Waterhouse et al., 2009), CIAlign
both saves time and increases reproducibility. The visualisation options provided by
CIAlign are not, to our knowledge, available in other tools.

Parameters
Having as many parameters as possible to allow as much user control as possible gives
greater flexibility. However, this also means that these parameters should be adjusted, which
requires a good understanding of the cleaning functions and the MSA in question. CIAlign

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 23/29

 Figure 8 Metrics from using CIAlign with biological data. (A) Left, density plots showing the distri-
bution of percentage identity (top) and normalised Needleman-Wunsch score (Needleman & Wunsch,
1970) (bottom) between samples of sequences from the Pfam (Finn et al., 2014) full alignments and con-
sensus sequences generated based on Pfam seed alignments without (dark blue) and with (light blue)
CIAlign cleaning and Pfam full alignments without (pink) and with (orange) CIAlign cleaning. Right, box
plots showing the alignment total size (top) and number of gaps (bottom) for these four alignments. (B)
Left, bar chart showing the size of insertions from the 1000 Genomes data (Auton et al., 2015) used to test
the ability of CIAlign to remove insertions and deletions. Right, bar chart showing the proportion of se-
quences in which these insertions were present in data from 162 individuals and whether they were (pink)
or were not (blue) removed by the CIAlign remove insertions function. (continued on next page. . . )
Full-size DOI: 10.7717/peerj.12983/fig-8

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 24/29

 Figure 8 (...continued)
(C) Left, phylogenetic tree based on an alignment of sequences from the 10k trees project (Arnold,
Matthews & Nunn, 2010) for the 12s ribosomal gene in primates. Colours represent known monophyletic
groups of primates. Nodes have been collapsed where multiple sequences from the same group formed a
monophyletic clade. Sequences annotated with circles were removed by CIAlign. Top-right, tree based on
the same alignment after cleaning with CIAlign, which removed the outlying group. Bottom-right, mini
alignments showing the effect of running CIAlign on this alignment.

offers default parameters selected to be often applicable based on our benchmarking

simulations and testing with different types of data. However, parameter choice highly
depends on MSA divergence and the downstream application. To choose appropriate
values it is recommended to first run CIAlign with all default parameters and then adjust
these parameters based on the results. Since the mini alignments show what has been
removed by which functions it is straightforward to identify the effect of each function and
any changes to the parameters which may be required.

Future work
New features are in progress to be added in the future, such as collapsing very similar
sequences, removing divergent columns, and making the colour scheme for the bases or
amino acids customisable. CIAlign is currently not parallelised, as the most time limiting
function, remove insertions, requires information from the entire alignment. However, a
future release will incorporate the ability to process more than one alignment in parallel.

CONCLUSIONS
CIAlign is a highly customisable tool which can be used to clean multiple sequence
alignments and address several common alignment problems. Due to its multiple user
options it can be used for many applications. CIAlign provides clear visual output showing
which positions have been removed and for what reason, allowing the user to adjust the
parameters accordingly. A number of additional visualisation and interpretation options
are provided.

ADDITIONAL INFORMATION AND DECLARATIONS

Funding
This work was supported by the Wellcome Trust (106207) and the European Research
Council (646891). The funders had no role in study design, data collection and analysis,
decision to publish, or preparation of the manuscript.

Grant Disclosures
The following grant information was disclosed by the authors:
Wellcome Trust: 106207.
European Research Council: 646891.

Competing Interests
The authors declare there are no competing interests.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 25/29

 Author Contributions
• Charlotte Tumescheit and Katherine Brown conceived and designed the experiments,
performed the experiments, analyzed the data, prepared figures and/or tables, authored
or reviewed drafts of the paper, and approved the final draft.
• Andrew E. Firth conceived and designed the experiments, authored or reviewed drafts
of the paper, and approved the final draft.

Data Availability
The following information was supplied regarding data availability:
The code is available at Zenodo: Katy Brown & Charlotte Tumescheit. (2022).
KatyBrown/CIAlign: v1.0.15 (v1.0.15). Zenodo. DOI:10.5281/zenodo.6330781.
The benchmarking data is available at GitHub: https://github.com/KatyBrown/
benchmarking_data_CIAlign.

Supplemental Information
Supplemental information for this article can be found online at http://dx.doi.org/10.7717/
peerj.12983#supplemental-information.

REFERENCES
Arnold C, Matthews LJ, Nunn CL. 2010. The 10kTrees website: a new online resource for
primate phylogeny. Evolutionary Anthropology: Issues, News, and Reviews 19:114–118
DOI 10.1002/evan.20251.
Auton A, Abecasis GR, Altshuler DM, Durbin RM, Abecasis GR, Bentley DR.
2015. A global reference for human genetic variation. Nature 526:68–74
DOI 10.1038/nature15393.
Bäckström D, Yutin N, Jørgensen SL, Dharamshi J, Homa F, Zaremba-Niedwiedzka
K, Spang A, Wolf YI, Koonin EV, Ettema TJG. 2019. Virus genomes from deep sea
sediments expand the ocean megavirome and support independent origins of viral
gigantism. mBio 10(2):e02497-18 DOI 10.1128/mBio.02497-18.
Bahr A, Thompson JD, Thierry J-C, Poch O. 2001. BAliBASE (Benchmark Alignment
dataBASE): enhancements for repeats, transmembrane sequences and circular
permutations. Nucleic Acids Research 29:323–326 DOI 10.1093/nar/29.1.323.
Boswell RD. 1987. Sequence alignment by word processor. Trends in Biochemical Sciences
12:279–280 DOI 10.1016/0968-0004(87)90135-6.
Brito JJ, Li J, Moore JH, Greene CS, Nogoy NA, Garmire LX, Mangul S. 2020. Recom-
mendations to enhance rigor and reproducibility in biomedical research. GigaScience
9(6):giaa056 DOI 10.1093/gigascience/giaa056.
Bushmanova E, Antipov D, Lapidus A, Prjibelski AD. 2019. rnaSPAdes: a de novo tran-
scriptome assembler and its application to RNA-Seq data. GigaScience 8(9):giz100
DOI 10.1093/gigascience/giz100.
Camacho C, Coulouris G, Avagyan V, Ma N, Papadopoulos J, Bealer K, Madden
TL. 2009. BLAST+: architecture and applications. BMC Bioinformatics 10:421
DOI 10.1186/1471-2105-10-421.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 26/29

 Capella-Gutiérrez S, Silla-Martínez JM, Gabaldón T. 2009. trimAl: a tool for auto-
mated alignment trimming in large-scale phylogenetic analyses. Bioinformatics
25:1972–1973 DOI 10.1093/bioinformatics/btp348.
Edgar RC. 2004. MUSCLE: a multiple sequence alignment method with reduced time
and space complexity. BMC Bioinformatics 5:113 DOI 10.1186/1471-2105-5-113.
Finn RD, Bateman A, Clements J, Coggill P, Eberhardt RY, Eddy SR, Heger A, Hether-
ington K, Holm L, Mistry J, Sonnhammer ELL, Tate J, Punta M. 2014. Pfam: the
protein families database. Nucleic Acids Research 42(Database issue):D222–D230
DOI 10.1093/nar/gkt1223.
Fitch WM, Smith TF. 1983. Optimal sequence alignments. Proceedings of the
National Academy of Sciences of the United States of America 80:1382–1386
DOI 10.1073/pnas.80.5.1382.
Fletcher W, Yang Z. 2009. INDELible: a flexible simulator of biological sequence evolu-
tion. Molecular Biology and Evolution 26:1879–1888 DOI 10.1093/molbev/msp098.
Gasteiger E, Gattiker A, Hoogland C, Ivanyi I, Appel RD, Bairoch A. 2003. ExPASy:
the proteomics server for in-depth protein knowledge and analysis. Nucleic Acids
Research 31:3784–3788 DOI 10.1093/nar/gkg563.
Guindon S, Gascuel O. 2003. A simple, fast, and accurate algorithm to estimate
large phylogenies by maximum likelihood. Systematic Biology 52:696–704
DOI 10.1080/10635150390235520.
Hall BG. 2008. Simulating DNA coding sequence evolution with EvolveAGene 3.
Molecular Biology and Evolution 25:688–695 DOI 10.1093/molbev/msn008.
Higgins DG, Sharp PM. 1988. CLUSTAL: a package for performing multiple sequence
alignment on a microcomputer. Gene 73:237–244
DOI 10.1016/0378-1119(88)90330-7.
Hu B, Xie G, Lo C-C, Starkenburg SR, Chain PSG. 2011. Pathogen compara-
tive genomics in the next-generation sequencing era: genome alignments,
pangenomics and metagenomics. Briefings in Functional Genomics 10:322–333
DOI 10.1093/bfgp/elr042.
Huerta-Cepas J, Serra F, Bork P. 2016. ETE 3: reconstruction, analysis, and visual-
ization of phylogenomic data. Molecular Biology and Evolution 33:1635–1638
DOI 10.1093/molbev/msw046.
Hunter JD. 2007. Matplotlib: a 2D graphics environment. Computing in Science &
Engineering 9:90–95.
Käfer S, Paraskevopoulou S, Zirkel F, Wieseke N, Donath A, Petersen M, Jones TC, Liu
S, Zhou X, Middendorf M, Junglen S, Misof B, Drosten C. 2019. Re-assessing the
diversity of negative strand RNA viruses in insects. PLOS Pathogens 15(12):e1008224
DOI 10.1371/journal.ppat.1008224.
Katoh K, Misawa K, Kuma K, Miyata T. 2002. MAFFT: a novel method for rapid
multiple sequence alignment based on fast Fourier transform. Nucleic Acids Research
30:3059–3066 DOI 10.1093/nar/gkf436.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 27/29

 Kumar S, Stecher G, Li M, Knyaz C, Tamura K. 2018. MEGA X: molecular evolutionary
genetics analysis across computing platforms. Molecular Biology and Evolution
35:1547–1549 DOI 10.1093/molbev/msy096.
Langille MGI, Ravel J, Fricke WF. 2018. Available upon request: not good enough for
microbiome data!. Microbiome 6:8 DOI 10.1186/s40168-017-0394-z.
Liao X, Li M, Zou Y, Wu F-X, Pan Y, Wang J. 2019. Current challenges and solutions of
de novo assembly. Quantitative Biology 7:90–109 DOI 10.1007/s40484-019-0166-9.
Magi A, Giusti B, Tattini L. 2017. Characterization of MinION nanopore data for
resequencing analyses. Briefings in Bioinformatics 18:940–953.
NCBI. 2012. Transcriptome Shotgun Assembly Sequence Database. Bethesda: National
Center for Biotechnology Information. Available at https://www.ncbi.nlm.nih.gov/
genbank/tsa/ (accessed on 08 October 2019).
Needleman SB, Wunsch CD. 1970. A general method applicable to the search for
similarities in the amino acid sequence of two proteins. Journal of Molecular Biology
48:443–453 DOI 10.1016/0022-2836(70)90057-4.
Notredame C, Higgins DG, Heringa J. 2000. T-coffee: a novel method for fast and
accurate multiple sequence alignment11Edited by J Thornton. Journal of Molecular
Biology 302:205–217 DOI 10.1006/jmbi.2000.4042.
Petyuk VA, Gatto L, Payne SH. 2019. Reproducibility and transparency by design.
Molecular & Cellular Proteomics 18(8 suppl 1):S202–S204
DOI 10.1074/mcp.IP119.001567.
Ratnasingham S, Herbert PD. bold: The Barcode of Life Data System (http://www.barcodinglife.o
Molecular Ecology Notes 7:335–364.
Richterich P. 1998. Estimation of errors in raw DNA sequences: a validation study.
Genome Research 8:251–259 DOI 10.1101/gr.8.3.251.
Robinson DF, Foulds LR. 1981. Comparison of phylogenetic trees. Mathematical
Biosciences 53:131–147 DOI 10.1016/0025-5564(81)90043-2.
Sayyari E, Whitfield JB, Mirarab S. 2017. Fragmentary gene sequences negatively
impact gene tree and species tree reconstruction. Molecular Biology and Evolution
34:3279–3291 DOI 10.1093/molbev/msx261.
Schneider TD, Stephens RM. 1990. Sequence logos: a new way to display consensus
sequences. Nucleic Acids Research 18:6097–6100 DOI 10.1093/nar/18.20.6097.
Schulz F, Roux S, Paez-Espino D, Jungbluth S, Walsh DA, Denef VJ, McMahon KD,
Konstantinidis KT, Eloe-Fadrosh EA, Kyrpides NC, Woyke T. 2020. Giant virus
diversity and host interactions through global metagenomics. Nature 578:432–436
DOI 10.1038/s41586-020-1957-x.
Sievers F, Dineen D, Wilm A, Higgins DG. 2013. Making automated multiple align-
ments of very large numbers of protein sequences. Bioinformatics 29:989–995
DOI 10.1093/bioinformatics/btt093.
Sievers F, Higgins DG. 2018. Clustal Omega for making accurate alignments of many
protein sequences. Protein Science 27:135–145 DOI 10.1002/pro.3290.
Sievers F, Higgins DG. 2020. QuanTest2: benchmarking multiple sequence alignments
using secondary structure prediction. Bioinformatics 36:90–95.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 28/29

 Smith MR. 2019. Bayesian and parsimony approaches reconstruct informative
trees from simulated morphological datasets. Biology Letters 15:20180632
DOI 10.1098/rsbl.2018.0632.
Stamatakis A. 2014. RAxML version 8: a tool for phylogenetic analysis and post-analysis
of large phylogenies. Bioinformatics 30:1312–1313
DOI 10.1093/bioinformatics/btu033.
Talavera G, Castresana J. 2007. Improvement of phylogenies after removing divergent
and ambiguously aligned blocks from protein sequence alignments. Systematic
Biology 56:564–577 DOI 10.1080/10635150701472164.
Tettelin H, Masignani V, Cieslewicz MJ, Donati C, Medini D, Ward NL, Angiuoli SV,
Crabtree J, Jones AL, Durkin AS, DeBoy RT, Davidsen TM, Mora M, Scarselli M,
Ros I, Peterson JD, Hauser CR, Sundaram JP, Nelson WC, Madupu R, Brinkac
LM, Dodson RJ, Rosovitz MJ, Sullivan SA, Daugherty SC, Haft DH, Selengut
J, Gwinn ML, Zhou L, Zafar N, Khouri H, Radune D, Dimitrov G, Watkins K,
O’Connor KJB, Smith S, Utterback TR, White O, Rubens CE, Grandi G, Madoff
LC, Kasper DL, Telford JL, Wessels MR, Rappuoli R, Fraser CM. 2005. Genome
analysis of multiple pathogenic isolates of Streptococcus agalactiae: implications for
the microbial pan-genome. Proceedings of the National Academy of Sciences of the
United States of America 102:13950–13955 DOI 10.1073/pnas.0506758102.
Thompson JD, Plewniak F, Poch O. 1999. A comprehensive comparison of mul-
tiple sequence alignment programs. Nucleic Acids Research 27:2682–2690
DOI 10.1093/nar/27.13.2682.
Tyler AD, Mataseje L, Urfano CJ, Schmidt L, Antonation KS, Mulvey MR, Corbett CR.
2018. Evaluation of Oxford nanopore’s MinION sequencing device for microbial
whole genome sequencing applications. Scientific Reports 8:1–12.
Waterhouse AM, Procter JB, Martin DMA, Clamp M, Barton GJ. 2009. Jalview Version
2—a multiple sequence alignment editor and analysis workbench. Bioinformatics
25:1189–1191 DOI 10.1093/bioinformatics/btp033.
Wick RR. 2019. Badread: simulation of error-prone long reads. Journal of Open Source
Software 4:1316 DOI 10.21105/joss.01316.
Wright ES. 2015. DECIPHER: harnessing local sequence context to improve protein
multiple sequence alignment. BMC Bioinformatics 16:322
DOI 10.1186/s12859-015-0749-z.
Wu M, Chatterji S, Eisen JA. 2012. Accounting for alignment uncertainty in phyloge-
nomics. PLOS ONE 7:e30288 DOI 10.1371/journal.pone.0030288.

Tumescheit et al. (2022), PeerJ, DOI 10.7717/peerj.12983 29/29