BLOSUM matrix

The BLOSUM matrix (BLOcks SUbstitution Matrix) is a family of scoring matrices used in protein sequence alignment, derived from empirical analysis of conserved protein regions rather than extrapolated evolutionary models. Developed by Steven Henikoff and Jorja Henikoff in 1992, BLOSUM matrices addressed a critical limitation in the PAM matrices: PAM matrices extrapolate from closely observed mutations to distant evolutionary relationships, accumulating error with each multiplication step. BLOSUM matrices, by contrast, derive substitution frequencies directly from aligned blocks of distantly related proteins.

The numbering convention inverts intuition: BLOSUM62, the most commonly used matrix, is derived from blocks of sequences with at most 62% identity. BLOSUM80 uses closely related sequences; BLOSUM45 uses more distantly related ones. This direct derivation from observed data made BLOSUM matrices more accurate for detecting distant evolutionary relationships, and they have largely replaced PAM matrices in database search tools like BLAST.

The shift from PAM to BLOSUM reflects a broader methodological transition in bioinformatics from model-based to data-driven inference. Where Dayhoff's PAM matrices embodied a theoretical evolutionary model, the Henikoffs' BLOSUM matrices embodied empirical observation. Both approaches remain valid, but the field's preference for BLOSUM reveals a disciplinary bet on data over theory that has only intensified with the rise of machine learning in protein structure prediction.