MetAl is a command-line utilty for calculating metric distances between alternative alignments of the same sequences. The current version is 1.1.
Binaries of MetAl 1.1 are available for Windows, Linux and Mac:
32-bit Linux available on request. If you find any problems with these binaries, please don't hesitate to get in touch. Email me at [email protected].
Run metal -h for usage. There are four metrics, detailed in this paper:
Measuring the distance between multiple sequence alignments Blackburne, B.P. and Whelan, S. Bioinformatics doi:10.1093/bioinformatics/btr701
The default metric, d-pos, should be fine for most uses. Additional information may be generated by using the other metrics too, see the paper for details.
Metal accepts Fasta and "relaxed Phylip" format. In case of doubt, use Fasta.
Some example alignments of the dataset BB12036 from BaliBASE are available in
the example directory.
How different are alignments from ClustalW and Prank, using metric d-pos?
$ metal example/BB12036.prank-fas example/BB12036.clustalw-fas
2636 / 17604 = 0.14973869575096568
So The ClustalW and Prank alignments are ~15% different.
How do different sequences contribute to the distance?
$ metal -a example/BB12036.prank-fas example/BB12036.clustalw-fas
1qf9_A 382 / 2328 = 0.1640893470790378
1uky_ 378 / 2352 = 0.16071428571428573
1zak_A 782 / 2640 = 0.2962121212121212
1zip_ 402 / 2604 = 0.1543778801843318
2ak2_ 472 / 2640 = 0.1787878787878788
2ak3_A 480 / 2712 = 0.17699115044247787
3adk_ 390 / 2328 = 0.16752577319587628
3286 / 17604 = 0.18666212224494433
Clearly, 1zak_A has the greatest discrepency between the two aligners.
What are the differences between the alignments from ProbCons and T-COFFEE?
$ metal example/BB12036.probcons-fas example/BB12036.tcoffee-fas
392 / 17604 = 2.226766643944558e-2
And using the --tree metric with a reference tree (d-evol in the paper):
$ metal -t example/BB12036.tre example/BB12036.probcons-fas example/BB12036.tcoffee-fas
424 / 17604 = 2.4085435128379913e-2
Both ProbCons and T-COFFEE are consistency-based aligners and so we should not be surprised that the alignments they produce are sometimes relatively similar.
The BaliBASE reference alignment is given as example/BB12036.balibase-ref-fas. We can rank the difference alignments in order of distance to the reference like so:
$ for i in example/*fas; do ANS=$(metal $i example/BB12036.balibase-ref-fas); echo $i $ANS ;done | sort -g -k 6
example/BB12036.balibase-ref-fas 0 / 17604 = 0.0
example/BB12036.linsi-fas 1556 / 17604 = 8.838900249943195e-2
example/BB12036.probcons-fas 1602 / 17604 = 9.100204498977506e-2
example/BB12036.muscle-fas 1784 / 17604 = 0.10134060440808908
example/BB12036.tcoffee-fas 1826 / 17604 = 0.10372642581231538
example/BB12036.dialigntx-fas 2044 / 17604 = 0.11610997500568053
example/BB12036.fftnsi-fas 2096 / 17604 = 0.11906384912519882
example/BB12036.clustalw-fas 2636 / 17604 = 0.14973869575096568
example/BB12036.prank-fas 3448 / 17604 = 0.1958645762326744
MAFFT's L-INS-i algorithm is closest, Prank is furthest. Note that although both Prank and ClustalW are 15-20% different to the reference alignment, this is not because they are similar to each other, as we determined they were ~15% different to each other, above. On the other hand, ProbCons and T-COFFEE (9-10% difference from the reference) are similar to each other (~2% difference between them). Being able to compare scores between individual alignments as well as against a reference is one of the advantages of a metric over a non-metric score.
MetAl is written in Haskell.
You need my phytree library first. Neither this nor phyhs are yet on Hackage, so you'll have to use git. You also need Haskell-Platform installed.
git clone git://github.com/benb/phytree.git phytree
cd phytree
cabal update
cabal install
cd ..
git clone git://github.com/benb/MetAl.git metal
cd metal
cabal install
Please contact me at [email protected]
GPL3.
Source code available at http://github.com/benb/MetAl/.