CIAlign: A highly customisable command line tool to clean, interpret and visualise multiple sequence alignments

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.

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 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.

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 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.

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.

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, 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 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.

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 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 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 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 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 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 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 (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).

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.

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.

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 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.