Question 7 of 20 Fundamentals Concept
What does a substitution matrix such as BLOSUM62 represent, and when would you change it?
Short answer
Each score is a log-odds ratio: how much more often two amino acids are aligned in real homologs than by chance. BLOSUM62 is a deep matrix tuned for distant relatives in full-length proteins. Switch to a shallower matrix for short queries or domains, or when looking only for close relatives.
Why
Aligning protein sequences needs a score for every pair of amino acids. Altschul showed that any matrix suitable for local alignment can be read as log-odds scores, the log of the frequency a pair is seen in homologs divided by the frequency expected by chance. A positive score means the pair is more common in related proteins, such as isoleucine with valine.
The Henikoffs built the BLOSUM matrices by counting replacements inside conserved, ungapped blocks of distantly related protein families, using a percent identity threshold to exclude close relatives. Pearson's guide to choosing matrices explains the practical consequence. Deep matrices such as BLOSUM62, the BLASTP default, target alignments of about 20 to 30 percent identity.
They are very sensitive but need long alignments and can extend an alignment past the true homologous region. Shallow matrices target alignments of about 50 to 90 percent identity and suit short domains, short queries, or searches meant to find recent orthologs. NCBI's BLAST page switches to PAM30 for short queries for the same reason.
Gap penalties belong to each matrix, so changing one usually means changing both.
A strong answer covers
- Scores are log-odds of pairing in homologs versus chance
- BLOSUM was built from conserved blocks of distant families
- Deep matrices find distant homologs but need long alignments
- Shallow matrices suit short sequences and close relatives
- Gap penalties are matched to the matrix
Common mistakes
- Thinking a higher BLOSUM number means a more sensitive search for distant homologs
- Using BLOSUM62 for a 20 residue peptide query and trusting the result
- Changing the matrix without adjusting the gap penalties
The interviewer might ask next
Why can a deep scoring matrix extend an alignment into a neighboring unrelated domain?