Multiple sequence alignments

5. A standard multiple alignment program: ClustalW

From what we have learned in previous sections, doing a simultaneous N-wise alignment is not a realistic option if we have, say, 50 sequences to align. What to do? The obvious alternative is to use a so-called progressive alignment method: The alignment is built up in stages where a new sequence is added to an existing alignment, using some rules to determine in which order the sequences should be added, and how.

ClustalW (Thompson, Higgins & Gibson, 1994) is one of the standard programs implementing one variant of the progressive method in wide use today for multiple sequence alignment. The W denotes a specific version that has been developed from the original Clustal program.

The basic steps of the algorithm implemented in ClustalW are:

Compute the pairwise alignments for all against all sequences. The similarities are stored in a matrix (sequences versus sequences).
Convert the sequence similarity matrix values to distance measures, reflecting evolutionary distance between each pair of sequences.
Construct a tree (the so-called guide tree) for the order in which pairs of sequences are to be aligned and combined with previous alignments. This is done using a neighbour-joining clustering algorithm. In the case of ClustalW, a method by Saitou & Nei is used.
Progressively align the sequences/alignments together into each branch point of the guide tree, starting with the least distant pairs of sequences. At each branch point, one must do either a sequence-sequence, sequence-profile, or profile-profile alignment.

Note that the original idea of using a simultaneous N-wise dynamic programming alignment method had the algorithmic complexity O(c²ⁿ), whereas this method has something like O(n²) (which comes from the all-against-all pairwise comparison step).

A number of rules (tricks, some would say) are used to increase the success rate of the procedure:

Each sequence is weighted according to how different it is from the other sequences. This accounts for the case where one specific subfamily is overrepresented in the data set.
The substitution matrix used for each alignment step depends on the similarity of the sequences (a somewhat circular argument, but what the hell...).
Position-specific gap-open penalties are modified according to residue type using empirical observations in a set of alignments based on 3D structures. In general, hydrophobic residues have higher gap penalties than hydrophilic, since they are more likely to be in the hydrophobic core, where gaps should not occur.
Gap-open penalties are decreased if the position is spanned by a consecutive stretch of five or more hydrophilic residues.
Both gap-open and gap-extend penalties are increased if there are no gaps in a column, but gaps occur nearby in the alignment.
The guide tree can in some circumstances be overriden, for instance by deferring joining two branches if they are too dissimilar, until more information has been added by processing other branches.

There are some specific cases where ClustalW is know to have problems.

If the sequences are similar only in some smaller regions, while the larger parts are not recognisably similar, then ClustalW may have problems aligning all sequences properly. This is because ClustalW tries to find global alignments, not local. In such a case, it may be wise to cut out the similar parts with some other tool (text editor).
If one sequence contains a large insertion compared to the rest, then there may be problems, for much the same reason as the previous point.
If one sequence contains a repetitive element (such as a domain), while another sequence only contains one copy of the element, then ClustalW may split the single domain into two half-domains to try to align the first half with the first the domain in the first sequence, and the other half to the second domain in the first sequence. There are many proteins that contain multiple, very similar copies of a domain, so one should watch out for this.

ClustalW is an example of an algorithm that has given up on trying to be perfect (because it takes too much time), and instead uses an approximation strategy, combined with more-or-less intelligent tricks that guide the computation towards a successful (but not necessarily optimal) result. This is called a heuristic algorithm.

One important point to keep in mind is that since ClustalW is a heuristic algorithm, it cannot produce a solution that is guaranteed to be optimal. But in practice, the results it produces are good enough, and one should perhaps worry more about the quality of the input data. For example, if one has sequences that are just barely significantly similar, one should worry more about if all of them really belong in the alignment at all, rather than if the alignment is perfect or not (which it in such a case almost certainly isn't).

ClustalW has a number of parameters that the user can change. This will affect the exact manner in which the computation proceeds, and it may be useful to compare runs with different parameters; the near-perfect parameter set varies with the specific case.

One shouldn't really be content with just ClustalW. There are other multiple-alignment software out there, and it is by no means clear that ClustalW is the best. It is usually very difficult to make good, unbiased comparisons between multiple-alignment software. And of course, one still has the basic problem of how to choose the objective criterion to determine what the best multiple alignment really is.