4.5 Sequence Similarity

Sequence similarity is a simple extension of amino acid or nucleotide similarity. To determine it, sum up the individual pair-wise scores in an alignment. For example, the raw score of the following BLAST alignment under the BLOSUM62 matrix is 72. Converting 72 to a normalized score is as simple as multiplying by lambda. (Note that for BLAST statistical calculations, the normalized score is lS - lnk.)

Query:   885 QCPVCHKKYSNALVLQQHIRLHTGE 909
             +C VC K ++    L++H RLHTGE
Sbjct:   267 ECDVCSKSFTTKYFLKKHKRLHTGE 291

Recall from Chapter 3 that the score of each pair of letters is considered independently from the rest of the alignment. This is the same idea. There is a convenient synergy between alignment algorithms and alignment scores. However, when treating the letters independently of one another, you lose contextual information. Can you assume that the probability of A followed by G is the same as the probability of G followed by A? In a natural language such as English, you know that this doesn't make sense. In English, Q is always followed by U. If you treat these letters independently, you lose this restriction. The context rules for biological sequences aren't as strict as for English, but there are tendencies. For example, low entropy sequences appear by chance much more frequently than expected. To avoid becoming sidetracked by the details, accept that you're using an approximation, and note that in practice, it works well.