# MAFFT vs MUSCLE vs Clustal Omega: which aligner to use

> MAFFT, MUSCLE and Clustal Omega all align protein sequences, but "MAFFT" covers methods that differ more from each other than from its rivals. We ran all three on the 218 BAliBASE 3 reference families, measured speed, and tested how well MUSCLE 5 ensembles flag the columns it gets wrong.

## Which is better, MAFFT, MUSCLE or Clustal Omega?

[MAFFT](/app/mafft) in its L-INS-i mode, [MUSCLE 5](/app/muscle5) and [Clustal Omega](/app/clustal-omega) finish within 0.03 of each other in mean column score on BAliBASE 3, a standard protein benchmark, and none of them wins everywhere. MAFFT's fast default is the outlier, about 0.1 behind the other three. The mode you run matters more than the program you name.

A single family shows why no ranking settles it. Thirty-seven helix-turn-helix DNA-binding domains, from phage repressors to NarL-type regulators, went through all three programs. MAFFT and MUSCLE got none of the 17 reference core columns right, while Clustal Omega got 15. In another family of four sequences the order flips: MAFFT's L-INS-i and MUSCLE scored 0.94 and 0.96, and Clustal Omega scored 0.00.

Across all 218 BAliBASE 3 families, MUSCLE 5.3 had the strictly highest column score in 76, MAFFT L-INS-i in 52 and Clustal Omega in 36. Which one leads on your data depends on the kind of divergence, the number of sequences and the compute you can spend. None of them says which of its own columns are wrong unless you ask MUSCLE 5 for an ensemble.

## How the three aligners build an alignment

All three are progressive aligners at heart. They estimate how similar the sequences are, build a guide tree from those distances and merge sequences and groups in tree order, closest first. A gap placed wrongly in an early merge is carried into every later one, so the programs differ mostly in how they cut that risk and how they keep the work affordable.

MAFFT turns each residue into two numbers, volume and polarity, so a sequence becomes a signal, and uses a fast Fourier transform to find stretches that match between two sequences without filling a full dynamic programming matrix. What you get then depends on the strategy, and the fast default, FFT-NS-2, is purely progressive.

L-INS-i adds local pairwise alignments, consistency scores and iterative refinement, in which the alignment is split in two and the halves realigned until the score stops improving. The `--auto` option chooses for you. In the 7.526 script we read, fewer than 100 sequences whose longest member is under 3,000 residues get L-INS-i with up to 1,000 iterations, and from 500 sequences up it stops iterating.

MUSCLE 5 uses a different algorithm from the 2004 program. Its default alignment, called PPP, applies two rounds of consistency scoring and 100 rounds of refinement. Edgar built it so that perturbing the HMM probabilities and permuting the guide tree yields an ensemble of alignments, and a separate Super5 mode splits large sets into pieces so that tens of thousands of sequences stay tractable.

Clustal Omega spends its effort on scale. Its mBed step represents each sequence by its distances to a few reference sequences, which cuts the distance step from quadratic to O(N log N) time, groups those vectors with k-means and builds the guide tree from the groups. Merging then converts each group to a profile hidden Markov model and aligns pairs of them with HHalign.

![Three columns of three small drawings each. MAFFT: two residue signals with a matching stretch highlighted, a guide tree, and an alignment split in two groups with a refinement arrow. MUSCLE 5: two sequences linked by shared k-mers, two groups of sequences merged into one, and a stack of four alternative alignments. Clustal Omega: points with three reference points, three dashed clusters of points, and two profile HMMs merged into one.](/images/guides/mafft-vs-muscle-vs-clustal-omega/aligner-workflows.webp "**Figure 1. How each aligner builds an alignment.** Schematic of the main steps described by the authors, not drawn from real sequences, distances or trees. Sources: Katoh and Standley (2013), Edgar (2022), Sievers et al. (2011).")

MAFFT and MUSCLE spend their compute on consistency and refinement, Clustal Omega on cutting the cost of the distance step. What each choice costs in accuracy only shows on a benchmark with a known answer.

## How they score on BAliBASE 3

BAliBASE 3 holds 218 reference alignments whose scored regions, called core blocks, were fixed by superposing three-dimensional structures. Its six reference sets pose different problems: 38 families of sequences under 20% identity, 44 at 20 to 40%, 41 where a family comes with a few orphan sequences, 30 of divergent subfamilies, 49 with long N-terminal or C-terminal extensions and 16 with internal insertions.

Two scores summarize an alignment. The sum-of-pairs score (SP) is the share of aligned residue pairs in the core blocks that the program reproduced, and the column score (CS) is the share of core columns reproduced exactly. CS is the harsher: one misplaced residue in a column of 60 sequences loses the whole column.

The published rankings disagree, because they compared different versions. Sievers and colleagues measured a total column score of 0.554 for Clustal Omega, 0.475 for MUSCLE 3 and 0.458 for MAFFT's default on these 218 families, with MAFFT's Auto mode at 0.588.

Edgar's 2022 paper reports that MUSCLE 5 aligned 59% of columns correctly on a benchmark of 10,000 sequences per set, against 52% for Clustal Omega and 47% for MAFFT. The first comparison predates MUSCLE 5. The second uses a different benchmark, was run by MUSCLE's author, and its main text does not name the MAFFT mode.

So we ran them ourselves. We aligned each of the 218 families with MAFFT 7.526 in its fast default (FFT-NS-2) and with L-INS-i, with MUSCLE 5.3 in its default PPP mode and with Clustal Omega 1.2.4 at default settings, one thread each, using the full-length sequences from the BAliBASE download page.

We scored every alignment against the reference with our own SP and CS code. It gives 1.0 for each reference alignment, counts a core column as correct when all its reference residues share one column and reproduces the published figures: 0.559 for Clustal Omega against 0.554, 0.480 for MAFFT's default against 0.458 and 0.590 for L-INS-i against 0.588 for Auto.

MAFFT's Auto mode picks the L-INS-i settings for any set of fewer than 100 sequences, which covers 214 of the 218 families, and on the four larger ones we ran Auto itself; its column scores were 0.00 to 0.06 below L-INS-i. Table 1 and Figure 2 give the results.

  <caption><strong>Table 1. Four configurations on the 218 BAliBASE 3 families.</strong> Mean sum-of-pairs and column scores over all families, the number of families where each had the strictly highest column score, and runtime on one thread. Our run; times came from a shared 16-core machine, so read them as ratios.</caption>
  <thead>
    <tr>
      <th scope="col">Configuration</th>
      <th scope="col">Mean SP</th>
      <th scope="col">Mean CS</th>
      <th scope="col">Families won</th>
      <th scope="col">Median time</th>
      <th scope="col">All 218 families</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>MAFFT 7.526 default (FFT-NS-2)</td>
      <td>0.819</td>
      <td>0.480</td>
      <td>4</td>
      <td>0.9 s</td>
      <td>5 min</td>
    </tr>
    <tr>
      <td>MAFFT 7.526 L-INS-i</td>
      <td>0.868</td>
      <td>0.590</td>
      <td>52</td>
      <td>3.2 s</td>
      <td>63 min</td>
    </tr>
    <tr>
      <td>MUSCLE 5.3 PPP</td>
      <td>0.870</td>
      <td>0.581</td>
      <td>76</td>
      <td>8.0 s</td>
      <td>284 min</td>
    </tr>
    <tr>
      <td>Clustal Omega 1.2.4</td>
      <td>0.840</td>
      <td>0.559</td>
      <td>36</td>
      <td>0.7 s</td>
      <td>9 min</td>
    </tr>
  </tbody>

![Grouped bars of mean BAliBASE column score for four aligner configurations across six reference sets and all 218 families. MAFFT's default is lowest in every group. MUSCLE is highest in the two low-identity groups and the insertions group, and L-INS-i is highest in the orphan, subfamily and extension groups and overall.](/images/charts/mafft-muscle-clustal-balibase-column-score.webp "**Figure 2. Mean column score by BAliBASE reference set.** Column score of MAFFT 7.526 (default and L-INS-i), MUSCLE 5.3 and Clustal Omega 1.2.4 on the 218 BAliBASE 3 families, with the number of families in brackets. Our run, default settings, one thread.")

The headline gap is inside MAFFT. On the 38 families below 20% identity, the default scores 0.293 and L-INS-i 0.414, a difference of 0.121 that is six times the 0.020 between L-INS-i and MUSCLE (0.434). A methods section that says only "MAFFT" does not say which of those numbers it means.

The three accurate configurations trade places by family type. MUSCLE leads the two equidistant sets (0.434 and 0.862) and the set with internal insertions (0.628), the three sets without orphans, subfamilies or long flanks.

L-INS-i leads on orphans and long terminal extensions (0.461 and 0.606 against MUSCLE's 0.424 and 0.558), as a local alignment that can leave a flank unaligned should. Clustal Omega sits within 0.01 of L-INS-i on the orphan and subfamily sets but trails MUSCLE by 0.07 on the least similar sequences.

Size changes the order as well. Among the 107 families of fewer than 20 sequences MUSCLE averages 0.665, L-INS-i 0.650 and Clustal Omega 0.607. From 20 sequences up, L-INS-i is ahead of MUSCLE in every size band, by 0.039 at 20 to 49 and 0.023 at 50 to 99, and Clustal Omega is 0.030 ahead of MUSCLE at 20 to 49 and 0.015 behind it above that.

Two catches limit all of this. BAliBASE is small, with at most 142 sequences per family, and Sievers and colleagues note that benchmark sets like these are "heavily biased toward single domain globular proteins". It also scores only the core blocks, so it says nothing about the regions where real data are hardest to align.

## How speed scales with sequence count

A spread of 0.03 in accuracy sits on top of a spread of about 30 times in cost. Across the 218 families, Clustal Omega finished in 9 minutes in total, L-INS-i in 63 and MUSCLE 5.3 in 284, so MUSCLE's 0.022 lead over Clustal Omega in column score cost roughly 30 times the compute.

The gap widens with the number of sequences, as Table 2 shows. On the 142-sequence family BB30003, MAFFT's default took 1.6 seconds, Clustal Omega 8, L-INS-i 134 and MUSCLE's PPP mode 51 minutes. All of these ran on one thread; MUSCLE and MAFFT can use more.

  <caption><strong>Table 2. Median time per family by number of sequences.</strong> Seconds on one thread for the 218 BAliBASE 3 families, binned by sequence count, with the number of families in brackets. Our run on a shared machine, so the ratios matter more than the values.</caption>
  <thead>
    <tr>
      <th scope="col">Sequences per family</th>
      <th scope="col">MAFFT default</th>
      <th scope="col">MAFFT L-INS-i</th>
      <th scope="col">MUSCLE 5.3 PPP</th>
      <th scope="col">Clustal Omega</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Under 20 (107)</td>
      <td>0.7</td>
      <td>1.0</td>
      <td>1.1</td>
      <td>0.2</td>
    </tr>
    <tr>
      <td>20 to 49 (69)</td>
      <td>1.0</td>
      <td>7.3</td>
      <td>26.4</td>
      <td>2.0</td>
    </tr>
    <tr>
      <td>50 to 99 (38)</td>
      <td>1.3</td>
      <td>21.0</td>
      <td>171.9</td>
      <td>3.8</td>
    </tr>
    <tr>
      <td>100 to 142 (4)</td>
      <td>1.6</td>
      <td>70.0</td>
      <td>925.6</td>
      <td>7.9</td>
    </tr>
  </tbody>

Beyond a few hundred sequences the benchmark stops applying. Sievers and colleagues aligned more than 190,000 sequences with Clustal Omega on one processor in a few hours. On 18 HomFam families of 10,099 to 50,157 sequences, it reached a column score of 0.464 in 27,329 seconds in total, against 0.253 for MAFFT's PartTree in 6,119 seconds and 0.216 for MUSCLE 3, which needed 110,292.

MAFFT has aligned 13,822 existing sequences together with 138,210 added fragments in 5.9 hours on 10 cores. Both results date from 2011 and 2013, before MUSCLE 5 existed.

MUSCLE 5 added a Super5 mode for tens of thousands of sequences, and Edgar's 2022 comparison at 10,000 sequences per set puts MUSCLE 5 first and Clustal Omega second. Treat any large-set ranking as dated, and time your own data on a subsample before committing hours to a full run.

On ProteinIQ a MAFFT run is cut off after nine minutes, Clustal Omega after 25 and MUSCLE5 after about two hours. If the subsample says MUSCLE 5 PPP would run past that, Super5 or Clustal Omega is the realistic route.

## A family where the three disagree

Reference set 3 holds families of divergent subfamilies, and BB30016 is a clear case. It has 37 helix-turn-helix DNA-binding domains of about 290 residues, with two core blocks of 8 and 9 columns, 17 core columns in all. The sequences fall into three groups: 24 phage repressors, Cro proteins and POU-domain transcription factors, 9 NarL and LuxR type regulators and 4 Ner type proteins.

MAFFT's default, L-INS-i and MUSCLE aligned each group sensibly on its own, then put the groups side by side instead of on top of each other. The first core residue of the 9-sequence group landed 137, 143 and 141 columns away from the 24-sequence group, respectively. Their column scores were 0.00 and their pair scores 0.52, 0.46 and 0.45. Clustal Omega put all 37 first residues in one column, scoring 0.94 for pairs and 0.88 for columns.

![Four rows of circles on a column axis. For MAFFT default, MAFFT L-INS-i and MUSCLE 5.3, the 24-sequence group sits at column 0, the 4-sequence group within 22 columns of it and the 9-sequence group about 140 columns to the right, with column scores of 0.00. For Clustal Omega all three groups sit at column 0 with a column score of 0.88.](/images/guides/mafft-vs-muscle-vs-clustal-omega/subfamily-offsets.webp "**Figure 3. Where each aligner put the first core residue of three subfamilies in BAliBASE family BB30016.** Each circle is a group of sequences whose first core residue fell in one alignment column, plotted by that column relative to the 24-sequence group. Offsets are to scale; circle sizes grow with group size. SP is the sum-of-pairs score and CS the column score. Our run.")

This is one family. Over the 30 families of this set the three accurate configurations average the same, 0.586 for L-INS-i, 0.580 for MUSCLE and 0.579 for Clustal Omega, so BB30016 shows variance, not a rule that Clustal Omega aligns subfamilies better. The reverse happens too. In BB12020, four sequences of 126 residues, L-INS-i and MUSCLE scored 0.94 and 0.96, MAFFT's default 0.90 and Clustal Omega 0.00.

Agreement between aligners is not evidence. Three configurations from two programs agreed on the wrong answer in BB30016, so a match between two aligners shows shared difficulty with distant subfamilies, not correctness. The check has to come from outside the sequences: known catalytic residues landing in one column, or a structural alignment of two members with [US-align](/app/usalign), which would have put the two helices of these domains together.

## Which columns to trust

Every alignment in Table 1 is a single answer. Nothing in the file says which of its columns the program was sure about, and the three programs differ in whether they can tell you. Wong and colleagues showed why it matters: on genomic data from seven yeast species, different alignment methods led to different conclusions.

Markova-Raina and Petrov compared six aligners, MUSCLE and ClustalW among them, on about 6,690 genes from 12 Drosophila genomes. The aligner strongly changed which sites looked positively selected. By visual inspection most flagged sites were misaligned, for false positive rates of 48 to 82%.

MUSCLE 5 is the only one of the three built to answer the question. It can return an ensemble of alignments made with perturbed HMM parameters and permuted guide trees, and the fraction of members that reproduce a column is that column's confidence. Edgar reports that an alignment's mean confidence correlates with its accuracy on protein benchmarks (Pearson r = 0.80).

We tested the column-level version of that claim. For each of the 82 families in reference sets 1 and 2 we built a 16-member stratified ensemble and asked how often a column of the default alignment was right, given the share of members that reproduced it.

![Bars showing the share of BAliBASE core columns that MUSCLE 5.3 placed correctly in three confidence bands. Columns reproduced by fewer than half of the 16 ensemble members are correct 8.5 percent of the time, columns reproduced by half to under 90 percent are correct 27.2 percent, and columns reproduced by 90 percent or more are correct 91.5 percent.](/images/charts/mafft-muscle-clustal-ensemble-confidence.webp "**Figure 4. Ensemble confidence predicts which columns are right.** Share of BAliBASE 3 core columns (12,652 columns in the 82 families of reference sets 1 and 2) that MUSCLE 5.3 placed correctly, split by the fraction of a 16-member stratified ensemble that reproduced the column. The bands hold 1,237, 1,527 and 9,888 columns. Our calculation; a column is correct if all its reference residues share one column of the default alignment.")

The split is sharp. Of the 9,888 columns that at least 90% of members reproduced, 91.5% were correct. Of the 1,237 columns that fewer than half reproduced, 8.5% were.

The 90% line keeps 9,048 of the 9,569 correct columns and drops 2,243 of the 3,083 wrong ones. Masking the low-confidence columns before a [maximum likelihood tree](/guides/iqtree-vs-raxml-vs-fasttree) or a conservation analysis therefore removes most of the errors at the price of about 5% of the right columns. Across families, the share of columns at or above 90% confidence tracked the family's column score closely (Pearson r = 0.95).

Two limits apply. The test used reference sets 1 and 2, small families of 4 to 34 equidistant sequences, so it says nothing about large or multidomain sets. And confidence measures how stable a column is under the program's own perturbations, so a bias that every member shares goes unseen.

MUSCLE5 on ProteinIQ offers both ensemble modes and returns the confidence table, the highest-confidence member and a letter-confidence alignment, so the masking test above is one setting away. Neither Clustal Omega nor MAFFT writes a per-column confidence in the files the hosted tools return; for those, run a second aligner and distrust the regions where the two differ.

## Which aligner should you use?

Table 3 turns the results into starting points. In every row, state the exact program, version and mode in your methods, such as "MAFFT 7.526 L-INS-i" and not "MAFFT", because on this benchmark the mode moved the score more than the program.

  <caption><strong>Table 3. Which aligner to start with.</strong> Starting points drawn from the benchmark above; none replaces a check against structure or known residues.</caption>
  <thead>
    <tr>
      <th scope="col">Your situation</th>
      <th scope="col">Start with</th>
      <th scope="col">Add</th>
      <th scope="col">Check</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Under about 100 sequences of one domain family</td>
      <td>[MAFFT](/app/mafft) L-INS-i or [MUSCLE5](/app/muscle5) PPP</td>
      <td>The other one</td>
      <td>The columns where the two differ</td>
    </tr>
    <tr>
      <td>Hundreds to tens of thousands of sequences</td>
      <td>[Clustal Omega](/app/clustal-omega), or MUSCLE5 Super5 or MAFFT with its large-set strategies</td>
      <td>A subsample aligned with L-INS-i</td>
      <td>Whether the subsample's columns survive in the full alignment</td>
    </tr>
    <tr>
      <td>Very distant sequences or far-apart subfamilies</td>
      <td>MUSCLE5 PPP or MAFFT L-INS-i</td>
      <td>A structure-based alignment with [US-align](/app/usalign)</td>
      <td>Whether subfamilies sit side by side in distant columns</td>
    </tr>
    <tr>
      <td>Orphan sequences or long terminal extensions</td>
      <td>MAFFT L-INS-i</td>
      <td>Trimming to the shared domain</td>
      <td>Long gap blocks in the flanking regions</td>
    </tr>
    <tr>
      <td>An alignment that feeds a tree or a selection test</td>
      <td>MUSCLE5 ensemble</td>
      <td>A second aligner, then [IQ-TREE](/app/iq-tree) on both</td>
      <td>Whether the conclusion survives masking columns below 0.9 confidence</td>
    </tr>
  </tbody>

Whichever row fits, check that the sequences are homologous first, because every one of these programs aligns unrelated sequences without complaint. A search with [HMMER](/app/hmmer) or [BLAST](/app/blast-search) takes a minute and removes the worst failure.

## Alignment tools on ProteinIQ

A [multiple sequence alignment](/use-cases/sequence-alignment) project runs in this order: check that the sequences are homologous, align them with one or two of the programs, compare the results in a viewer, then build a tree or a conservation analysis from the columns you trust.

- [HMMER](/app/hmmer) and [BLAST](/app/blast-search): confirm that the sequences are homologs before aligning
- [MAFFT](/app/mafft): L-INS-i, Auto, FFT-NS-2 and PartTree strategies, and adding sequences to an existing alignment
- [MUSCLE5](/app/muscle5): PPP and Super5 alignment, with ensembles and column confidence
- [Clustal Omega](/app/clustal-omega): large sets, profile alignment, guide tree and distance matrix output
- [MSA viewer](/app/msa-viewer): inspect and compare alignments in the browser
- [US-align](/app/usalign): a structure-based reference when sequences are very distant
- [IQ-TREE](/app/iq-tree) and [FastTree](/app/fasttree): build a [phylogenetic tree](/guides/phylogenetic-tree-software) from the alignment

## Frequently asked questions

### Is MAFFT or MUSCLE more accurate?

It depends on the mode and the family. On the 218 BAliBASE 3 families, MAFFT's L-INS-i averaged a column score of 0.590 and MUSCLE 5.3 0.581, while MAFFT's fast default averaged 0.480. MUSCLE was ahead on the low-identity and insertion sets, L-INS-i on sets with orphans and terminal extensions.

### Is Clustal Omega still worth using?

Yes, for speed and for large sets. It averaged 0.559 on BAliBASE 3, 0.022 below MUSCLE 5.3, and finished all 218 families in about 9 minutes against 284 for MUSCLE. The Clustal Omega authors aligned more than 190,000 sequences on one processor in a few hours.

### Which aligner should I use for a phylogenetic tree?

Use one that gives you a way to check the alignment. MUSCLE 5 ensembles flag columns that its own perturbations do not reproduce, and in our test columns reproduced by at least 90% of 16 ensemble members were right 91.5% of the time. Different aligners can lead to different evolutionary conclusions from the same genes, so repeat the analysis with a second one.

### What does MAFFT Auto do?

Auto chooses a strategy from the size of the input. In MAFFT 7.526, fewer than 100 sequences with the longest under 3,000 residues gets L-INS-i with up to 1,000 iterations, and from 500 sequences up it stops iterating. That is why Auto and L-INS-i would give the same alignment on 214 of our 218 families.

### How many sequences can each aligner handle?

Clustal Omega has aligned more than 190,000 sequences, MAFFT 13,822 existing sequences plus 138,210 added fragments, and MUSCLE 5's Super5 mode is built for tens of thousands. Our BAliBASE test stops at 142 sequences per family, so it says nothing about those scales.
