Aligner Reference¶
The homepage gives a brief summary of supported aligners. This page discusses which to select and why, based on a benchmark of 9 of the 13 aligners against curated HIV-1 env reference alignments, conducted as part of the author's M.Sc. thesis.
Choosing an aligner¶
On the standard alignment-quality metrics (fD, fM, CSS, TC), all 9 benchmarked aligners scored similarly (0.96–0.98 on fD/fM/CSS), indicating that these metrics do not meaningfully separate the tools. What does separate them is how similar their output alignments are to one another, measured pairwise and independent of any reference. Two groups emerge from this comparison:
- MAFFT, MUSCLE, PROBCONS, Clustal Omega, and T-Coffee (in its default and regressive modes) produce alignments that are close to each other and to the curated reference; any of these can reasonably be chosen on the basis of speed or familiarity, since the choice has little effect on the resulting alignment.
- ClustalW, PRANK, and T-Coffee's
fmcoffee/quickalnmodes produce alignments that diverge noticeably from the main group and from each other, and should only be selected where there is a specific reason to do so, such as PRANK's phylogeny-aware indel placement.
MAFFT is the pipeline default and a reasonable choice for most datasets. Where the effect
of aligner choice on downstream results is a concern, alignment with one tool from the main
group alongside one of the outliers, followed by comparison of the two results, is
recommended.
Aligner comparison table¶
| Aligner | Notes |
|---|---|
| MAFFT (default) | Fast, well-tested, recommended starting point. |
| MUSCLE | MUSCLE 5 reimplements the PROBCONS algorithm, so its output tends to closely resemble PROBCONS. |
| PROBCONS | Similar output to MUSCLE. Slower. |
| Clustal Omega | Fast and accurate, uses HMMs. |
| T-Coffee (default, regressive) | Other T-Coffee modes (fmcoffee, a meta-method combining several aligners; quickaln, a faster heuristic mode) can produce noticeably different alignments — see below. |
| ClustalW | Legacy aligner; use only with a clear understanding of the tradeoffs involved. |
| PRANK | Phylogeny-aware; attempts real alignment of variable regions, producing much longer alignments than other tools. |
| VIRULIGN | Modifies and discards sequences during alignment (frameshift handling) — see caveats. |
| MACSE | Also modifies sequences during alignment, though less aggressively than VIRULIGN and without discarding sequences. |
| MAFFT-SEED | Appropriate when seeding alignment with a curated panel_alignment. |
| MUSCLE-FAST | MUSCLE's Super5 mode — faster, for large datasets. |
| PAGAN | Phylogeny-Aware graph alignments. Not thoroughly tested. |
Evaluation methodology¶
Five curated LANL reference alignments of HIV-1 env (subtype M, no recombinants) were split by subtype and by batch (200 sequences max per file) into 134 reference sub-alignments, gaps stripped, and realigned using 22 aligner/parameter combinations across the 9 tools above.
Two kinds of comparison were made:
- Against the reference —
qscorewas used to compute fD, fM, CSS, and TC scores between each inferred alignment and its curated reference. - Between aligners — the
d_posdistance (Blackburne & Whelan 2012) was computed pairwise between alignments inferred from the same input sequences (31,767 pairwise comparisons), and the result visualized with a Principal Coordinate Analysis (PCoA) to produce the clustering described above.
Caveats¶
- VIRULIGN and MACSE alter their input during alignment, removing or translating
sequence content beyond the addition of gaps; VIRULIGN can discard entire sequences it
judges to be frameshifted. This makes reference-free (
d_pos) comparison unreliable for both tools, and VIRULIGN's benchmark numbers rest on a smaller, less variable dataset than the other tools. Their scores should therefore be treated as indicative rather than definitive. - The benchmark was conducted using HIV-1 env sequences only, and results may not generalize to other organisms or gene regions.