Automatic detection of anchor points for multiple sequence alignment
 Florian Pitschi^{1},
 Claudine Devauchelle^{2} and
 Eduardo Corel^{2, 3}Email author
https://doi.org/10.1186/1471210511445
© Pitschi et al; licensee BioMed Central Ltd. 2010
Received: 10 December 2009
Accepted: 2 September 2010
Published: 2 September 2010
Abstract
Background
Determining beforehand specific positions to align (anchor points) has proved valuable for the accuracy of automated multiple sequence alignment (MSA) software. This feature can be used manually to include biological expertise, or automatically, usually by pairwise similarity searches. Multiple local similarities are be expected to be more adequate, as more biologically relevant. However, even good multiple local similarities can prove incompatible with the ordering of an alignment.
Results
We use a recently developed algorithm to detect multiple local similarities, which returns subsets of positions in the sequences sharing similar contexts of appearence. In this paper, we describe first how to get, with the help of this method, subsets of positions that could form partial columns in an alignment. We introduce next a graphtheoretic algorithm to detect (and remove) positions in the partial columns that are inconsistent with a multiple alignment. Partial columns can be used, for the time being, as guide only by a few MSA programs: ClustalW 2.0, DIALIGN 2 and TCoffee. We perform tests on the effect of introducing these columns on the popular benchmark BAliBASE 3.
Conclusions
We show that the inclusion of our partial alignment columns, as anchor points, improve on the whole the accuracy of the aligner ClustalW on the benchmark BAliBASE 3.
Background
Multiple sequence alignment (MSA) appears as the initial step of almost all biological sequence analyses. However, MSA is well known to be a difficult problem, both from the algorithmic point of view and with respect to the biological relevance of the output. The local alignment is a classical paradigm in sequence analysis [1, 2]. The idea of including local alignment information into global alignment tools, like in DIALIGN [3], represented an important step in alignment accuracy, and is also at work in more recent tools like TCoffee [4], MUSCLE [5] or MAFFT [6]. A latest trend is to include homology information retrieved from existing databases, such as, e.g., in DbClustal [7]. For recent reviews on MSA programs see [8–10]. Another way to improve the accuracy of existing MSA is to include userspecified anchor points, which are specific positions that should turn out to be aligned in the output [11]. This information can be composed of a small number of expertbased constraints, or can be used to include additional information, such as secondary structure predictions (like in [12]) or other information derived from external resources [7].
Multiple sequence alignment, being NPhard under any reasonable optimisation scheme, must consistently rely on heuristics. The inclusion of anchor points can result in a dramatic improvement on the relevance of the alignment, if it constrains the search of the local optimum to a region that contains the "true" alignment. The number of MSA programs that currently accept the inclusion of userspecified anchor points is unfortunately limited. To our knowledge, only DIALIGN 2 has such an explicit option [13], while it is also possible to include anchor points in ClustalW 2.0 [14], by using the format developed for the BLASTbased BALLAST [7] tool, and in TCoffee, by including the anchor points as library files.
Anchored alignment is also widely used for whole genome alignment strategies, for which it is almost required, by the sheer size of the input, to start by detecting strong pairwise local similarities for the efficiency of the subsequent algorithm (see e.g., [15]). For instance, exact maximal repeated substrings (like multiMUMs or MEMs [16]) can prove to be sufficiently informative, although more recent methods use spaced seeds (see [17]). In this paper, we follow likewise a combinatorial approach, but our focus concerns however not whole genomes, but sequences that are amenable to traditional multiple alignment methods, such as protein or genesized nucleic sequences.
We introduce a method to determine automatically a set of such "anchor points" for multiple alignment software. We base ourselves on a previously introduced algorithm, the Nlocal decoding, introduced by Didier [18] that clusters together positions in the sequences whose contexts of appearance of a given length N are similar but exhibit an a priori unspecified number of mismatches. More precisely, we use a method called MS4 [19], which selects multiple local similarities resulting from the Nlocal decoding, but for an adaptive value of N.
However, specifying contradicting anchor points can prove deleterious. Indeed, suggesting or imposing that some positions be aligned while these positions are incompatible with the ordering induced by the sequence can altogether destroy the relevance of the alignment. The simplest kind of incompatibility arises from internal repeats inside sequences. The MS4 method is here tuned to accept only similarities occurring at most once in any sequence. Usually, anchor points are specified as pairs of aligned residues, or possibly, of aligned segments. In order not to confuse the reader, we will call the output of this procedure partial columns, because they would look as such in a multiple alignment display.
The core of the present paper consists in a graphtheoretic algorithm to tackle the global issues of consistency with a multiple alignment. To do this, we consider the ordertheoretic definition of consistency (as used, e.g., in DIALIGN [20]). Each sequence is seen as an abstract ordered sequence of positions (from left to right). A collection of subsets of positions can be added to a given multiple alignment under a technical condition which ensures that the elements of different subsets never appear in contradicting orders. This condition is readily encoded in a directed graph, and the consistency problem amounts to getting a directed acyclic graph (DAG) from it. Our algorithm starts by implementing a heuristic solution to the NPhard problem known as the minimum feedback arcset problem. Once a DAG has been identified, positions that contradict the induced partial order are removed from the corresponding partial column. We call the output of this procedure consistent partial columns.
As a validation of the method, we introduced the partial columns, and the consistent partial columns, into the programs accepting anchoring options. We tested the effect of introducing these two types of anchor points on the performance of these MSA tools on the global benchmark BAliBASE 3 [21]. The results show that the use of either type of partial columns induce improvements of performance for ClustalW 2.0 on BAliBASE, which are better and stabler with the consistent ones. By contrast, we get a consistent degradation of performance for DIALIGN, and almost no variation for TCoffee.
Although we used our method only with the MS4based partial columns, this algorithm can be applied to any other set of partial columns. The MS4 approach has the advantage to detect directly multiple similarities with only linear complexity. Virtually any scheme for detecting local similarities could produce an input for our method, provided that all internal repetitions be removed. It is for example possible to use pairwise similarities, such as used by most MSA programs, and select among them those that involve more than two sequences to construct the partial columns, albeit at some computational cost. In [22], we used the pairwise optimal fragments for DIALIGN, and the consistency algorithm described in the present paper, this time with satisfactory results.
Methods
where ℓ(i) is the length of the ith sequence, is the abstract set of positions in the sequences, and is endowed with a natural partial ordering "≼" such that (i, p) ≼ (i', p') holds if and only if i = i' and p ≤ p'. Let S_{ i } be the set of sites of the ith sequence, i.e. the set {(i, p)1≤ p ≤ ℓ(i)}. In the following, we identify S_{ i } with the ith sequence.
We call a subset $C\subset \mathcal{S}$ambiguous if it contains a repetition, that is, there is a sequence ${\mathcal{S}}_{i}$ such that the intersection $C\cap {\mathcal{S}}_{i}$contains at least two distinct elements (i, p) and (i, p'), which are then also called ambiguous with respect to C. This definition is extended to an equivalence relation E on $\mathcal{S}$ by calling E itself ambiguous, if it contains an equivalence class which is an ambiguous subset.
A nonambiguous subset $C\subset \mathcal{S}$ will be called a partial alignment column. A nonambiguous equivalence relation consists therefore only of partial alignment columns. If an equivalence relation is consistent, it is obviously nonambiguous. The converse is however in general not true.
The MS4 method
Our partial column detection scheme is called MS4, and is described in [19]. It relies on a fast algorithm for producing partitions of sites, the Nlocal decoding, that we briefly recall.
The MS4 method combines the different equivalence classes from various values of N by introducing a new construction, the partition tree, which encodes how the equivalence classes for successive values of N are related.
The leaves of the partition tree are the sites in the sequences. Let us say that a node is ambiguous if the leaves of the partition tree that are children of this node form an ambiguous subset of sites.
Consistent Partial Columns
We present now the algorithm that resolves the inconsistencies among a set of partial columns.
The succession graph of a set $\mathcal{C}$ of partial columns is the edgeweighted directed graph $SG(\mathcal{C})=(\mathcal{C},E,w)$ where we have an edge e = (C, C') if and only if there exists a sequence i and sites (i, p) ∈ C and (i, p') ∈ C' that satisfy p <p'. An edge from C to C' means that there exists at least one sequence where C occurs to the left of C'. The weight ( C,C' ) of the edge (C, C') is then defined as the number of sequences i with this property. For convenience purposes, we also add an initial vertex v_{start} and a terminal one v_{end}. The following result is quite easy to establish.
Lemma 1. The set $\mathcal{C}$ is consistent if and only if SG($\mathcal{C}$) is a directed acyclic graph (DAG).
 1.
delete some edges of the succession graph G = SG($\mathcal{C}$) to turn it into a DAG,
 2.
transform the subsets C_{ i } themselves so that the succession graph of this new set of partial columns is itself a DAG.
For our applications, we will take $\mathcal{C}={\mathcal{C}}_{k}$ described in the previous section, but the procedure we introduce here would work starting with any set of disjoint nonambiguous subsets of $\mathcal{S}$.
Getting a Directed Acyclic Graph
and k* = min{k > 0(V, E_{ k } ) is a DAG}.
Removing Inconsistencies
The acyclic graph $(V,{E}_{{k}^{*}})$ can turn out to be disconnected, so we reconnect it by adding all the necessary edges of the form (v_{start}, u) or (u, v_{end}), and denote with G* the corresponding graph. Let ≤* be the partial order defined on C by the DAG G*. For each sequence s, let ${\mathcal{C}}_{s}$ be the set of partial columns C of $\mathcal{C}$having a (necessarily unique by definition) site (s, j_{ C } ) in s. There are two order relations on Vs = ${\mathcal{C}}_{s}$ U { v_{start},v_{end}}, namely

the total order ≼_{ s }induced by the natural order ≼ of $\mathcal{S}$ defined in section Methods,

the partial order ${\le}_{s}^{*}$ induced by the order ≤*defined by G*
The relation $R={\preccurlyeq}_{s}\cap {\le}_{s}^{*}$ is the largest order which is a subrelation of both ≼ _{ s } and ${\le}_{s}^{*}$ The total suborders, or chains, of the relation R are those subsets of occurrences of partial columns that are consistent. To minimise the number of lost sites, we choose a maximal chain.
Chains of R correspond to paths in G_{ s } . Let g _{ s } = (v_{start},u_{1},...,u_{ n } , v_{end}) be a path from v_{start} to v_{end} in G_{ s }of maximal length. For all partial columns $C\in {\mathcal{C}}_{s}$ such that C ∉ g_{ s }, remove the site (s, j_{ C } ) from C. Let ${\mathcal{C}}^{\circ}$ be the set of partial columns obtained after applying this procedure for all sequences s ∈ S. The order in which they have been selected does not matter. If we wish to stress the difference between consistent and nonconsistent partial columns, we will sometimes refer to the latter as raw partial columns.
Lemma 2. The succession graph SG(${\mathcal{C}}^{\circ}$) of the resulting partial column set is a DAG.
Proof. Every direct transition between occurrences of partial columns in ${\mathcal{C}}^{\circ}$ is encoded as an edge appearing in some longest path g in some graph G_{ s } . Therefore, every edge of the succession graph ${G}^{\circ}=SG({\mathcal{C}}^{\circ})$ corresponds to a path in the graph G^{+}. Since G^{+} is a DAG, the graph G°cannot have any cycle.
All current implementations of anchoring options take as input a list of pairs of matching positions. To obtain a set of anchor points from a set C of partial alignment columns, we consider all maximal segments of consecutive pairs of sites (i, p),...,(i, p+k) and (i', p'),..., (i', p' +k) such that every pair of sites (i, p+l) and (i', p' +l),1 ≤ l ≤ k, belongs to some partial alignment column ${C}_{j}\in \mathcal{C}$.
Results and Discussion
In order to evaluate the effect of introducing the MS4 partial columns in multiple alignments, we have used the reference protein multiple alignment benchmark database BAliBASE (release 3) [21]. As is usually done, we have only considered the core regions to assess the effect of the introduction of the partial columns in the MSA software. In order to do this, we have slightly waylaid DbClustal from its usual function, by including our MS4based partial columns as anchors points encoded in BALLAST files, as explained in [7]. We have also used the anchoring option of DIALIGN 2 and included the partial columns as library les in TCoffee.
For each of the reference sets in BAliBASE 3, we have examined and analyzed the performances of the aligners that accept anchors before and after the inclusion of two types of position subsets: (1) raw MS4 partial columns, computed according to section MS4 method (2) consistent MS4 partial columns, as obtained after applying the algorithm described in section Consistent Partial Columns.
The partial columns must be split into segments of pairwise matching positions, and attributed a weight. For a pair of segments of length l we set the weight to 10l for ClustalW and Dialign, and a uniform value of 100M for TCoffee, where M is the number of sequences in the dataset. For each of the obtained alignments, we have computed the sumofpairs (SP) and totalcolumn (TC) scores, and compared it to the scores obtained by the aligner alone. On DIALIGN, the results proved disappointing. With TCoffee, no improvement nor degradation whatsoever was observed in the overwhelming majority of cases: there is a variation on less than 25 datasets over the whole BAliBASE3 (which consisting of 218), and a substantial one on about 5 only. These results are after all not so surprising, since both DIALIGN and TCoffee already rely on local strategies. We will henceforth focus our discussion on the results obtained with ClustalW 2.0 alone. We omit "MS4" in what follows.
Sumofpairs score of ClustalW with anchor points on Balibase 3, compared to other usual aligners
Alignment  RV11  RV12  RV20  RV30  RV40  RV50 

CLUSTALW 2.0  49.27  86.89  86.23  70.71  79.65  70.56 
CW+pc. (2)  34.78  81.92  83.02  64.52  71.45  67.78 
CW+pc. (6)  52.62  89.84  85.90  70.66  80.70  77.28 
CW+pc. (12)  53.13  87.63  85.71  73.14  80.31  74.35 
CW+c. pc. (2)  50.66  86.35  86.21  74.53  80.37  77.51 
CW+c. pc. (6)  52.62  90.11  86.32  74.21  80.55  77.46 
CW+c. pc. (12)  53.13  87.63  86.21  74.36  80.42  74.75 
DIALIGNTX  51.52  89.18  87.88  76.18  83.64  82.28 
MAFFT 6.717b  66.19  93.36  92.72  87.08  92.19  90.25 
MUSCLE 3.7  57.16  91.54  88.91  81.45  86.49  83.52 
PROBCONS 1.12  66.97  94.12  91.68  84.53  90.34  89.41 
TCOFFEE 7.81  66.77  94.08  91.62  83.81  89.96  89.43 
TotalColumn score of ClustalW with anchor points on Balibase 3, compared to other usual aligners
Alignment  RV11  RV12  RV20  RV30  RV40  RV50 

CLUSTALW 2.0  24.00  72.32  20.44  26.87  40.04  34.21 
CW+pc. (2)  13.48  62.48  16.15  19.41  29.21  27.14 
CW+pc. (6)  21.34  74.95  24.18  27.01  39.07  41.67 
CW+pc. (12)  17.33  66.36  24.96  28.69  33.01  31.71 
CW+c. pc. (2)  25.89  70.42  26.18  30.24  37.28  40.06 
CW+c. pc. (6)  21.34  75.59  26.18  29.65  37.34  38.56 
CW+c. pc. (12)  17.33  66.36  25.25  27.27  32.92  31.71 
DIALIGNTX  26.81  75.69  30.78  38.90  45.17  47.05 
MAFFT 6.717b  44.13  83.83  45.46  58.90  60.56  59.52 
MUSCLE 3.7  32.06  80.90  35.30  41.19  45.32  46.39 
PROBCONS 1.12  41.96  86.05  41.15  54.73  53.61  57.89 
TCOFFEE 7.81  42.65  85.71  39.21  49.99  56.30  59.11 
Specificity scores for ClustalW with anchor points on Balibase 3
Alignment  RV11  RV12  RV20  RV30  RV40  RV50 

CLUSTALW 2.0  71.50/28.97  93.73/76.67  94.87/21.26  87.39/29.84  93.18/41.99  87.83/37.59 
CW+c. pc. (6)  75.50/27.18  95.60/79.56  95.73/26.97  89.23/33.12  92.76/39.18  91.96/41.17 
SP scores
We can observe from Figure 7 that, as a general trend, the inclusion of raw partial columns induces a general degradation of performance for s_{min} < 5, and a global improvement above this threshold (except for RV30). The score degradation for s_{min} < 5 shows that the raw partial columns include inconsistent similarities for these values, which are eliminated by requiring that a column span a minimum number of sequences in order to be considered.
Figure 8 illustrates the effect on the SP scores of introducing consistent partial columns. A perceptible improvement (from 0.5 to 2 points) is then observed, and no degradation of the SP scores for weak values of s_{min} (except for RV12) is to be seen. This indicates that the consistency algorithm manages to suppress inconsistent similarities even when they only concern as few as 2 sequences. More generally, the improvement due to the inclusion of the consistent partial columns is clearer for RV40 and RV50.
This result fits with the expectations, since these two datasets contain respectively large (C or N) extensions and large internal deletions. It is well known that supplying local information help global aligners to deal with large indels.
TC scores
Figures 9 and 10 show the variation of TC score with respectively raw and consistent columns. The TC score is much more stringent, since a single mistake in a column as compared to the reference alignment results in a score of 0 for the considered column. As for the SP score, the degradation that can be observed for s_{min} < 5 with raw partial columns disappears as soon as the columns have been filtered by the consistency algorithm, and gives on the contrary a perceptible improvement (with the notable exception of RV40). This means that, although the actual number of correctly aligned pairs does not greatly increase (see Figure 8), the improvement concerns essential columns of the core reference alignment. If the consistent partial columns are able to improve the TC scores, it shows that they can find previously undetected local similarities for a subset of sequences where the similarity was missed and now can be included for all sequences, because the TC score will only raise if a column is aligned correctly in all sequences. The improvement is more perceptible for RV30 and RV50. The dataset RV30 contains highly divergent sequences and RV50 large indels, as we recalled before.
Discussion
Time performances for the steps of the consistency algorithm
Algorithm step  RV11  RV12  RV20  RV30  RV40  RV50 

MS4  0.83  1.61  18.83  34.08  10.91  13.25 
Consistency  0.73  5.22  134.10  160.07  51.76  65.14 
We have also used both types of partial columns with DIALIGN. However, as mentioned, probably since this aligner is already based on local similarities, we didn't observe any improvements on BAliBASE 3. Further investigations seem to show that MS4 is here to blame. When partial alignment columns are constructed from the pairwise similarities computed by DIALIGN, we have shown in [22] that the consistency algorithm successfully removes inconsistencies, resulting this time in an improvement of performance with DIALIGN 2. This supports the idea that a more refined criterion for selecting the nodes in the partition tree than the one currently implemented in the MS4 method is required to be successfully applied as a local similarity detector that performs well on more modern aligners. At any rate, MS4 seems more adapted to alignmentfree classification. According to our experience, the partial columns obtained by MS4 are nevertheless useful for the visual expertise of alignments, for they highlight local homologies (for instance when used with a multiple alignment editor like Jalview), which are easier to visualise than the usual simple substitutions schemes used by these editors.
Conclusions
The introduction of our MS4based partial columns give therefore encouraging results. The overall influence of their inclusion can be summed up in two principal observations. The introduction of local information results in an improvement of the correctness of ClustalW, as already observed by the authors themselves, who developed DbClustal for this goal. Initially, DbClustal uses local fragments based on BLAST searches (local similarities with sequences stored in generalist protein databanks). The inclusion of userdefined anchor points being also possible, we have in this way been able to assess the improvement of performance that results from the inclusion of these local primary sequencebased similarities, constructed without score matrices or sliding window of predefined length. With the local aligners for which the inclusion of anchor points is possible, the results are not conclusive, especially with DIALIGN, although they happen to have quite a neutral effect on TCoffee. It is unfortunate that the anchoring option is not featured in any other aligner, especially any other global aligner, to be able to give more insight on the usefulness of the construction presented here.
The improvement obtained for ClustalW is most perceptible for datasets containing sequences of unequal lengths, and the computation of MS4 partial columns seems then justified in view of the gain in accuracy they provide. In other respects, the computation of consistent partial columns can help the eyeexpertise of multiple alignments, for the number of obtained position subsets is quite reasonable, and, as the TC score performance seems to indicate, their visualization allows to correct whole columns in the alignment, since they appear to correspond to conserved zones in the considered sequences (like in Figure 12 for instance). We have moreover introduced an algorithmic approach that can be further explored. The consistency algorithm can be used with other local similarities as input, as already tested with success on DIALIGN [22]. These results encourage us to improve our approach on several points. In particular, the mere filtering of edges of the succession graph by their weight to get a DAG in section Getting a Directed Acyclic Graph is overly simplistic (although effective). We are currently exploring more refined ways of getting a DAG, in order to reduce the number of erased edges. Another interesting feature would consist in splitting the contradicting partial columns into subsets of similarly behaved sites. These algorithmic improvements could then fit in a general tool for making local similarities consistent.
Declarations
Acknowledgements
We wish to thank M. Hoebeke, M. Baudry and G. Grasseau for assistance with the code, B. Morgenstern for guidance on the assessment of the anchor performance, R. Steinkamp for assistance in producing the results, and K. Hoff for some valuable help with the R package. EC acknowledges partial financial support from C.N.R.S., University of EvryVald'Essonne (France), and DFG Project MO 1048/61. FP acknowledges support from PICB (MaxPlanckGesellschaft and the Chinese Academy of Sciences).
Authors’ Affiliations
References
 Bailey TL, Elkan C: Fitting a mixture model by expectation maximization to discover motifs in biopolymers. Proceedings of the Second International Conference on Intelligent Systems for Molecular Biology 1994, 28–36.Google Scholar
 Smith RF, Smith TF: PatternInduced Multisequence Alignment (PIMA) algorithm employing secondary structuredependent gap penalties for comparitive protein modelling. Protein Engineering 1992, 5: 35–41. 10.1093/protein/5.1.35View ArticlePubMedGoogle Scholar
 Subramanian AR, Kaufmann M, Morgenstern B: DIALIGNTX: greedy and progressive approaches for the segmentbased multiple sequence alignment. Algorithms for Molecular Biology 2008, 3: 6. 10.1186/1748718836View ArticlePubMedPubMed CentralGoogle Scholar
 Notredame C, Higgins D, Heringa J: TCoffee: a novel algorithm for multiple sequence alignment. J Mol Biol 2000, 302: 205–217. 10.1006/jmbi.2000.4042View ArticlePubMedGoogle Scholar
 Edgar R: MUSCLE: Multiple sequence alignment with high score accuracy and high throughput. Nuc Acids Res 2004, 32: 1792–1797. 10.1093/nar/gkh340View ArticleGoogle Scholar
 Katoh K, Misawa K, Kuma K, Miyata T: MAFFT: a novel method for rapid multiple sequence alignment based on fast Fourier transform. Nuc Acids Research 2002, 30: 3059–3066. 10.1093/nar/gkf436View ArticleGoogle Scholar
 Thompson JD, Plewniak F, Thierry JC, Poch O: DbClustal: rapid and reliable global multiple alignments of protein sequences detected by database searches. Nucleic Acids Research 2000, 28: 2919–2926. 10.1093/nar/28.15.2919View ArticlePubMedPubMed CentralGoogle Scholar
 Kemena K, Notredame C: Upcoming challenges for multiple sequence alignment methods in the highthroughput era. Bioinformatics 2009, 25(19):2455–2465. 10.1093/bioinformatics/btp452View ArticlePubMedPubMed CentralGoogle Scholar
 Notredame C: Recent evolutions of multiple sequence alignment algorithms. PLoS Comput Biol 2007., 3(8): 10.1371/journal.pcbi.0030123Google Scholar
 Edgar RC, Batzoglou S: Multiple Sequence Alignment. Current Opinion in Structural Biology 2006, 16: 368–373. 10.1016/j.sbi.2006.04.004View ArticlePubMedGoogle Scholar
 Morgenstern B, Prohaska SJ, Pöhler D, Stadler PF: Multiple sequence alignment with userdefined anchor points. Algorithms for Molecular Biology 2006, 1: 6. 10.1186/1748718816View ArticlePubMedPubMed CentralGoogle Scholar
 Morgenstern B, Subramanian A, Hiran S, Steinkamp R, Meinicke P, Corel E: DIALIGNTX and multiple protein alignment using secondary structure information at GOBICS. Nucl Acids Res 2010, 38: W19W22. 10.1093/nar/gkq442View ArticlePubMedPubMed CentralGoogle Scholar
 Morgenstern B, Werner N, Prohaska SJ, Schneider RSI, Subramanian AR, Stadler PF, WeyerMenkhoff J: Multiple sequence alignment with userdefined constraints at GOBICS. Bioinformatics 2005, 21: 1271–1273. 10.1093/bioinformatics/bti142View ArticlePubMedGoogle Scholar
 Larkin MA, Blackshields G, P BN, Chenna R, McGettigan PA, McWilliam H, Valentin F, Wallace IM, Wilm A, Lopez R, Thompson JD, Gibson TJ, Higgins DG: ClustalW and ClustalX version 2.0. Bioinformatics 2007, 23(21):2947–48. 10.1093/bioinformatics/btm404View ArticlePubMedGoogle Scholar
 Aaron CE Darling FRB Bob Mau, Perna NT: Mauve: multiple alignment of conserved genomic sequence with rearrangements. Genome Research 2004, 14(7):1394–1403. 10.1101/gr.2289704View ArticleGoogle Scholar
 Höhl M, Kurtz S, Ohlebusch E: Efficient multiple genome alignment. Bioinformatics 2002, 18: 312S320S.View ArticleGoogle Scholar
 Kucherov G, Noé L, Roytberg M: A unifying framework for seed sensitivity and its application to subset seeds. J Bioinform Comput Biol 2006, 4(2):553–569. 10.1142/S0219720006001977View ArticlePubMedPubMed CentralGoogle Scholar
 Didier G, Laprevotte I, Pupin M, Hénaut A: Local Decoding of sequences and alignmentfree comparison. J Computational Biology 2006, 13: 1465–1476. 10.1089/cmb.2006.13.1465View ArticleGoogle Scholar
 Corel E, Pitschi F, Laprevotte I, Grasseau G, Didier G, Devauchelle C: MS4  MultiScale Selector of Sequence Signatures: An alignmentfree method for the classification of biological sequences. BMC Bioinformatics 2010, 11: 406. 10.1186/1471210511406View ArticlePubMedPubMed CentralGoogle Scholar
 Morgenstern B, Dress A, Werner T: Multiple DNA and protein sequence alignment based on segmenttosegment comparison. Proc Natl Acad Sci USA 1996, 93: 12098–12103. 10.1073/pnas.93.22.12098View ArticlePubMedPubMed CentralGoogle Scholar
 Thompson JD, Koehl P, Ripp R, Poch O: BAliBASE 3.0: latest developments of the multiple sequence alignment benchmark. Proteins: Structure, Function, and Bioinformatics 2005, 61: 127–136. 10.1002/prot.20527View ArticleGoogle Scholar
 Corel E, Pitschi F, Morgenstern B: A mincut algorithm for the consistency problem in multiple sequence alignment. Bioinformatics 2010, 26(8):1015–1021. 10.1093/bioinformatics/btq082View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.