How would you test whether a weak pairwise protein alignment reflects real homology?

Why

Percent identity is a poor test of homology. Pearson notes that the common rule of 30 percent identity misses many easily detected homologs, and that short alignments reach much higher identity by chance. The test is excess similarity, measured by an E-value or bit score from an accurate statistical model.

For proteins, an E-value below 0.001 can reliably support homology, and about 50 bits is almost always significant for average-length proteins. Statistics can still go wrong for unusual sequences, so two checks help. The first is to align the query with proteins known to be unrelated, for example ones with different domains.

Accurate statistics should give them E-values near 1. If they score below 0.001 to 0.01, the new relationship is suspect. The second is to shuffle the target, keeping length and composition, and see how often the shuffled sequence scores as well as the real one.

Pearson recommends shuffles that preserve local composition, such as reversing the sequence or shuffling within windows, because a global shuffle can be too easy to beat for sequences with biased composition.

How to work it through

1. Run a search with an accurate statistical method and note the E-value and bit score.
2. Align the query with several proteins known to be unrelated and check that their E-values are near 1.
3. Generate a set of shuffled targets that preserve local composition, such as windowed shuffles or the reversed sequence.
4. Score the query against each shuffled target and see where the real score falls.
5. Check that the alignment is not confined to a low-complexity or repetitive segment.
6. Look for supporting evidence, such as conserved key residues or a profile search.

Where the answer stops holding

Shuffled sequences are an imperfect stand-in for real proteins, especially those with repeats. A non-significant score does not prove two proteins are unrelated, because very distant homologs can fall below detection. Structural similarity can sometimes reveal relationships that sequence tests cannot.