Skip to content

Choosing a strategy

MAFFT's strategy choice trades speed against accuracy. The right pick depends on (i) how many sequences you have, (ii) how diverged they are, and (iii) whether you suspect large insertions or alignable regions embedded in unrelated flanks.

Quick decision table

Sequences Divergence Insertions / unalignable flanks Strategy CLI flag
any any no FFT-NS-2 (default) (none — default)
many, fast iter any no FFT-NS-i --maxiterate 1000
< 200 any no L-INS-i --localpair --maxiterate 1000
< 200 similar lengths no G-INS-i --globalpair --maxiterate 1000
< 200 mixed yes E-INS-i --genafpair --maxiterate 1000
10 000+ any no PartTree --parttree

Cheat sheet

use mafft::AlignmentMode;

// Default
AlignmentMode::FftNs2

// Iterative refinement
AlignmentMode::FftNsi { iterations: 1000 }

// Most accurate (< 200 seqs)
AlignmentMode::LInsi  { iterations: 1000 }   // local pair
AlignmentMode::GInsi  { iterations: 1000 }   // global pair
AlignmentMode::EInsi  { iterations: 1000 }   // generalized affine

pymafft.align(seqs)                              # FFT-NS-2
pymafft.align(seqs, strategy="fftnsi", maxiterate=1000)
pymafft.align(seqs, strategy="linsi", maxiterate=1000)
pymafft.align(seqs, strategy="ginsi", maxiterate=1000)
pymafft.align(seqs, strategy="einsi", maxiterate=1000)

mafft-rs in.fa                                      # FFT-NS-2
mafft-rs --maxiterate 1000 in.fa                    # FFT-NS-i
mafft-rs --localpair  --maxiterate 1000 in.fa       # L-INS-i
mafft-rs --globalpair --maxiterate 1000 in.fa       # G-INS-i
mafft-rs --genafpair  --maxiterate 1000 in.fa       # E-INS-i

Pragmatic guidance

  • You don't know which to pick → start with FFT-NS-2. It's the default for a reason; on most inputs it's within 1–2 SP-score points of the iterative methods.
  • You have time and < 200 sequences → run L-INS-i. Highest average accuracy on BAliBASE.
  • Your sequences clearly differ in length but the alignable region is obvious → L-INS-i (local pairwise reduces noise from flanks).
  • You have 10 000+ sequences → PartTree is the only practical option; O(N log N) distance estimation instead of O(N²).

The MAFFT official docs go into more depth on the algorithmic differences. rust-MAFFT is byte-identical to the C reference for every strategy listed above, so guidance written for upstream MAFFT applies verbatim here.