I'm new to pairwise and multiple sequence alignment in general, but I thought I understood how clustal works -- k-tuple distance is a cheap metric that's 'probably approximately' monotonic in the real global pairwise-alignment score, so we can just calculate it between all C(n, 2) sequences and do a simple UPGMA. Clustalw algorithm description here
But I'm observing odd results. Suppose I have a file with 60 sequences and run clustal on it, and observe that seqA and seqB are paired terminal child nodes. If I then delete a handful (say 10) of the sequences, but keep seqA and seqB, then in the new clustal alignment, seqA and seqB may not remain paired.
This is strange to me, because if Clustal was actually doing the k-tuple distance + UPGMA thing I thought it was doing, then all terminal pairings should always be stable under deletions of other sequences.
So... what am I missing?