5.1 The claim a column makes #
5.1.1 Not “all the pairs” #
A multiple alignment places three or more sequences so that one column proposes positional homology across all of them. Pairwise alignments need not agree: A with B, B with C, and A with C can demand incompatible gaps. The multiple alignment has to choose. It is not the average of the pairs.
Exact dynamic programming scales with the product of the lengths and is impossible for twenty proteins, so practical methods are heuristics. A sum-of-pairs score adds every pair in a column. Progressive methods do not even optimise that sum globally. They follow an order, and the order is part of the result.
5.1.2 What a bad alignment does downstream #
A tree reads a column as substitutions, a logo as conservation, a homology model as one spatial position, and a codon model as homologous codons. A gap inside a codon makes the last of those nonsense. No later statistic repairs residues that do not belong together, and a high bootstrap on a broken alignment supports the artefact.
Repetitive sequence aligns in several registers, and one outlier shifts every column. The control is a residue already known to be conserved, such as a catalytic side chain. If it is split, the alignment is not finished.
5.2 Building one #
5.2.1 Progressive alignment and the guide tree #
The progressive method, which Clustal made standard, aligns the closest sequences first and then aligns alignments, following a guide tree. That tree is usually a fast distance clustering: a construction order, not a phylogeny. Publishing it as the family tree is circular, because the alignment used to test it was built by obeying it.
Orthomam V11A 2022 162Species Circular Tree
Classically, a gap once opened stays: “once a gap, always a gap.” If A1 and A2 share an insertion that B lacks, the right order records one gap in B. The wrong order gaps A1 against B and then forces A2 into that error. The column looks decisive and is an artefact of the order. A fragment looks artificially distant, and the alignment follows it.
5.2.2 Iterative and consistency-based methods #
Later programs repair that failure. They are named here as ideas, not as versions to review. Iterative refinement, as in MUSCLE and also in MAFFT, removes a sequence or a subtree, realigns it, and keeps the change if the score rises. The guide tree can be rebuilt, so an early gap is no longer fate. A higher score under the wrong objective is still a polished error.
Consistency-based alignment, as in T-Coffee, prefers a column that a library of pairwise alignments agrees on, so support can pass from A-with-B and B-with-C to A-with-C. MAFFT also finds candidate segments quickly, including by a transform of composition; that transform is an implementation detail. No method rescues sequences that should not have been aligned together. What to align is decided before the program runs.
5.3 Motifs and profiles #
5.3.1 Regular expressions #
A regular expression writes a short motif as required symbols and alternatives. N-linked glycosylation is often written as Asn, then not Pro, then Ser or Thr, then not Pro. The pattern matches or it does not. A site that fails one position and is perfect elsewhere is rejected, and a match in a meaningless context is accepted.
Expressions suit a short motif you will look at by eye. They do not suit a long domain of graded conservation. A regex hit is not an E-value: there is no near miss, and no account of how surprising the match is in a large database.
5.3.2 Position-specific scoring matrices #
A position-specific scoring matrix gives every letter its own log-odds in every column, counted from a multiple alignment. A near miss at a weak site can still rank high if the rest is exact. A regular expression cannot say that.
Pseudocounts stop an unseen letter scoring as impossible. Twenty leucines and no isoleucine do not prove isoleucine cannot occur. Too large a prior erases real conservation; too small a prior freezes a sampling accident. The matrix PSI-BLAST builds is only as good as the sequences allowed into it.
5.3.3 Profile hidden Markov models #
A profile HMM gives gaps their own states. Each consensus column has a match state that emits from that column’s distribution, an insert state that emits extra residues and can loop, and a silent delete state that skips the column. Match to match is “no indel”. Entering insert or delete is an opening. A self-loop is an extension. That is the affine gap of Chapter 3, with probabilities learned from the alignment rather than fixed at 11 and 1.
The drawing of an alignment is one best path through those states. These notes do not derive that search, or BaumโWelch re-estimation from unaligned sequences. Widely used implementations score significance by summing paths against a null model, while the picture is still one path. A family that can be aligned in many similar ways is more evidence than the picture alone.
Pfam stores one such model per family, each with its own threshold, so raw scores are not comparable across families. Several hits on one protein are its domain architecture: a kinase domain alone, and the same domain beside a lipid-binding domain, are different hypotheses. A profile hit is stronger family evidence than one pairwise alignment when the family is divergent. It is still not an experiment, and a function copied onto the family is the Chapter 2 problem.
5.3.4 Logos #
A sequence logo stacks letters in a column. The stack height is the information content, and each letter’s height is its share of that information. For DNA, H = โฮฃ p logโ p and the information is 2 โ H. A column that is always thymine carries 2 bits. Half A and half G gives H = 1, so 1 bit, split between A and G. Equal use of four bases carries 0. For protein the maximum is logโ 20, about 4.32 bits.
Chromopsin Consensus Sequence Logos
These bits describe a column. They are not BLAST bit scores. With few sequences an uncorrected logo overstates conservation, because a short column can be constant by chance. A bad alignment is inherited: the logo will draw a conserved letter at a position the alignment invented.
5.4 Key terms, summary, and self-check #
Key terms. Multiple alignment; column; progressive alignment; guide tree; once a gap, always a gap; iterative refinement; consistency; regular expression; position-specific scoring matrix; pseudocount; profile HMM; match; insert; delete; Pfam; sequence logo; information content.
Summary. A column claims positional homology across every sequence, which the pairwise alignments may contradict. Progressive alignment follows a guide tree and classically never reopens an early gap, so a bad order becomes a bad tree, logo, or model. Iterative methods realign; consistency methods prefer columns the pairs support. Expressions are binary. Position-specific matrices score every letter. Profile HMMs add match, insert, and delete states, and Pfam stores them with per-family thresholds. A logo’s height is column information, inflated when few sequences were used.
Self-check.
1. Why can three pairwise alignments fit no single multiple alignment?
2. What is a guide tree for, and why is publishing it as the phylogeny circular?
3. A known catalytic residue is split across three columns, and the bootstrap is high. What does that bootstrap support?
4. How can a position-specific matrix rank a site that a regular expression rejects?
5. A DNA column is half A and half G. What is its information content, and why is that not a bit score?
