How to Read a Multiple Sequence Alignment
6 min read · Updated October 30, 2026
An alignment is three outputs stacked on top of each other: the aligned rows, a consensus row under them, and a conservation value per column. Each answers a different question, and the most common mistake is to read the consensus as though it were a real sequence.
Here is a five-sequence, 400-column alignment of closely related sequences, and what each part of the output says about it.
The consensus is not one of your sequences
The consensus takes the commonest residue in each column and writes them in a row. Nothing guarantees that the result exists in nature, and in a set of real sequences it usually does not: it is a position-by-position majority, assembled without regard for whether those majorities ever occur together in one molecule.
That makes it a summary and a bad template. Use it to see at a glance where the set agrees, and never as the sequence you order, clone or submit — for that, pick a real member of the set and say which one.
Conservation is per column, and the low columns are the information
Conservation is the fraction of rows carrying the commonest residue in that column. In the worked example, 350 of 400 columns are identical in all five sequences, the mean conservation is 96.3%, and the lowest column scores 0.6 — three sequences agreeing against two.
Reading the mean alone tells you almost nothing here: 96.3% describes five sequences that are nearly the same, which you knew. The 50 columns that vary are where every downstream question lives — which group does a column separate, does it fall inside a coding region, does it change a residue.
A column at 0.6 in a five-sequence alignment is a 3-against-2 split. In a forty-sequence alignment the same 0.6 is a 24-against-16 split, which is a much stronger statement. Conservation values are only comparable within one alignment of one depth.
The identity matrix, and the two percentages in it
Pairwise identity over this alignment runs 97.5% between X1 and X2 (390 of 400 columns), 97.0% between X1 and X3, and 91.25% to 92.0% across the deeper split in the set. Those three numbers are the whole structure of the alignment in one table, and they are what a distance tree is built from.
Two percentages matter per pair and they are not the same: identity over the columns where both rows have a residue, and identity over all columns. For sequences of equal length with no gaps the two coincide. For a set where one member is a fragment they diverge wildly — two sequences overlapping in 50 residues out of 2,000 columns can read 100% on the first measure and 2.5% on the second.
That is how an alignment of unrelated fragments gets described as identical, so always check how many columns a pairwise identity was computed over before quoting it.
A half-point of identity can become a clade
In this set X1 and X2 are 97.5% identical and X1 and X3 are 97.0%. Half a percentage point, which is two columns out of 400. Build a tree from the same alignment and that half-point becomes a grouping of X1 with X2 carrying 94% bootstrap support.
Nothing is wrong with either number. The alignment is reporting a small difference and the tree is reporting that the small difference is consistent. But it means a tree drawn from a very similar set can look much more decisive than the underlying identities do, which is worth remembering before describing two sequences as belonging to different groups.
Gaps are a claim the aligner made
Every gap in an alignment was inserted by the aligner to make the columns line up; none of them was measured. A region where several sequences carry gaps in different places is a region the aligner was guessing in, and the columns there should not be read as homologous positions.
If the sequences are protein-coding, aligning the protein and projecting the alignment back onto the codons avoids most of this, because an aligner working on nucleotides will happily open a one- or two-base gap that no reading frame can accommodate.
And an alignment with no gap characters anywhere, on rows that are all the same length, may never have been aligned at all — that is what an unaligned set of equal-length sequences looks like.
Frequently asked questions
What does the consensus sequence in an alignment mean?
It is the commonest residue in each column, written as a row. It summarises where the set agrees and is not necessarily a sequence that exists: each position is a separate majority, so the combination can be one no member carries. It is a reading aid, not a template to order from.
What is a good percent identity between sequences?
There is no universal threshold, because identity depends on what is being compared and over how much. What matters more is the denominator: identity measured over shared columns and identity measured over the whole alignment can differ by an order of magnitude when one sequence is a fragment. Quote both, or quote the number of columns compared.
Should I align DNA or protein?
Protein, for coding sequences that have diverged, then project the alignment back onto the codons. A nucleotide aligner can open a gap of one or two bases, which is biologically impossible inside a coding sequence and silently corrupts every column after it. For closely related sequences — clones of one gene, for instance — nucleotide alignment is fine and keeps synonymous differences visible.
Why do my sequences look identical in the alignment but split in the tree?
Because the tree is built from the differences, however few. In the worked example, 350 of 400 columns are identical, the pairwise identities differ by half a percentage point, and the tree reports a 94%-supported grouping off two columns. The alignment and the tree are both right; they are reporting different things.
Related references
Related tools
Align several DNA or protein sequences and view a colored alignment with a consensus.
Build a neighbour-joining tree with bootstrap support, or open a Newick file somebody sent you.
Align two sequences with Needleman-Wunsch or Smith-Waterman.
Convert between FASTA, GenBank and tab-separated formats, and extract CDS or protein sequences.
Align a query to a reference and list substitutions, insertions and deletions with effects.