I would ask your question again in the bioinformatics group. Good luck.

Alignment using multalin would be useful. Try it out

On nucleotides, all these software are equally limited and if your sequences are coding, align proteins and thread DNA back onto your MSA. I would recommend using the M-Coffee mode of T-Coffee that makes it possible to combine the output of several aligners. In theory, the resulting alignments should be a bit more accurate, but the real point is that this approach will show you where the aligners agree on the final MSA, thus making it possible to cut out columns.

On the command line, use: t_coffee -seq -method clustalw_msa mafft_msa muscle_msa -output aln,score_html, score_ascii

The score_ascii output can the be used to programmatically extract the columns having the highest level of agreement among the methods (see documentation).

Overall, in terms of accuracy, mafft, probcons, and T-Coffee are among the most accurate. ClustalW has been shown to have a lower accuracy. It is still disputed wether prank or Sate and other so-called 'phylogenetic aligners' produce improved sequence alignments. On protein sequences they do not, as judged from structure based references.

If you look up any recent MSA paper (type multiple sequence alignment in medline), you will find lots of recent benchmarks. I tend to only focus on the 'difficult ones' that is the low similarity in prefab or the BB11 in Balibase. On these sets, methods like T-Coffee (consistency based) have an edge (they include T-Coffee, Mummals, probcons, MSAProb), on simpler datasets everything is more or less equal.

If the sequences are coding sequences then the protein sequences can be aligned using PRALINE. This software takes in to account the consensus of accessible surface area and secondary structure conservation and hence most of time i have seen the alignment would be accurate.