07. BLOSUM - BLOcks SUbstitution Matrix

An alternative to the PAM matrix is BLOSUM (BLocks SUbstitution Matrix), which was derived by Henikoff and Henikoff in 1992. NCBI uses BLOSUM62 as its the default matrix for protein BLAST.

Derivation of BLOSUM matrices

BLOSUM matrices are derived from comparisons of blocks of sequences from the Blocks database.

What are blocks and what is the blocks database?

A block is an ungapped multiple alignments of highly conserved, short regions. Here is what a sample block looks like:

Conserved blocks in alignment
Conserved blocks in alignment

The blocks database contains multiple alignments of conserved regions in protein families.

Derivation a BLOSUM matrix

The Henikoffs developed a database of "blocks" based on sequences with shared motifs. More than 2,000 blocks of aligned sequence segments were analyzed from more than 500 groups of related proteins. Within each block, they counted the relative frequencies of amino acids and their substitution probabilities

Why did they use Blocks?

The Henikoffs used blocks due to several reasons:

  • Need to have multiple alignments and it's easier to align with similar sequences.
  • They didn't want to complicate calculations with insertions/deletions.
  • Wanted to focus on conserved regions for computing the scoring matrix.

What does a BLOSUM matrix tell us?

A BLOSUM tells us the likelihood of occurrence of each pairwise substitution, and we can use these values to score a pairwise comparison.

Each scoring matrix is constructed based on how identical the ungapped multiple sequence alignments are. For example, BLOSUM62 is derived from blocks containing at most 62% identity in the ungapped sequence aligments.

Calculating a BLOSUM

Here we'll show you how to calculate a BLOSUM.

Removing redundancies

Before we start constructing a matrix BLOSUM r, we have to eliminate the sequences that are more than r% identical. This solves us from the bias we get from databases over-representing certain classes of proteins. To do this, we have two options:

  1. Remove sequences from the block.
  2. Replace the similar sequences with a new sequence that represents the cluster.

Since most databases today have an over-representation of proteins, the extraneous DCE sequences should be eliminated in order to make our database more representative.

Calculating possible pairwise combinations

Thus, after elminating redundancies, we look at the first vertical column in our block:


Let's find out how many possible pairwise combinations we can see for each possible pair.

For the AA pair, we have 2 possible combinations, for AB or BA we have 4. For AD we have 2. We continue these calculations until the occurrence of all possible pairs are found.

PairColumn 1 scoreColumn 2 scoreColumn 3 scoreTotal
AB or BA4004
AD or DA2002
BD or DB2024
DE or ED0044

Note that the total sum is 26, which we can use to normalize our matrix.


Calculate scores

To obtain integer values for our scoring matrix, we need to find the score per cell. We can do this with the following equation:

sij = log2(qij/eij)

Where qij is observed frequency and eij is the expected frequency.

qij = cij / T

cij is the cell value as calculated above.

To calculate the Total T:

T = w * n(n-1) / 2

Where w is the number of columns and n is the number of sequences. With T, we can calculate qij, which is the rate of change of residue i to residue j.

In our case, T = 30, so let's divide all our cells by 30.


Now pi can be found with the following equation:

pi = qii + ∑(qij/2)
pA = ( 1 + 6/2 ) / 30 = 0.133
pB = ( 1 + 8/2 ) / 30 = 0.167
pC = 10 / 30
pD = ( 1 + 10/2 ) /30 = 0.200
pE = ( 1 + 4/2 ) / 30 = 0.0133

The expected frequencies:

eii = pi2
eij = 2pipj (i ≠ j)

Notice how I didn't calculate cell values that had a value of 0 - you'll see that we don't need these values in the actual scoring matrix.

Now we have all we need! Just plug in values from the two matrices above into the equation below to obtain our scoring matrix.

sij = log2(qij/eij)

To obtain scores, we multiple sij by two and round.

sij = round (2 * log2(qij/eij))


What is a blocks database?
BLOSUM Matrices Lecture. IA State.
Columbia CS Department.

Take your Linux skills to the next level!

Command Line Kung Fu

Take your Linux skills to the next level! Try Linux & UNIX

Command Line Kung Fu is packed with dozens of tips and practical real-world examples. You won't find theoretical examples in this book. The examples demonstrate how to solve actual problems. The tactics are easy to find, too. Each chapter covers a specific topic and groups related tips and examples together.

$ Check price
14.9914.99Amazon 4.5 logo(27+ reviews)

More Linux & UNIX resources

Become a Bioinformatics Whiz!

Introduction to Bioinformatics Vol. 1

Become a Bioinformatics Whiz! Try Bioinformatics

If you're looking for a fun and easy entry point into bioinformatics algorithms, this book it just for you! Filled with graphics, and written in a light-hearted and humorous story-telling persona, Bioinformatics Algorithms guides you through the intricacies of the problems faced in biology, and the clever solutions used to solve them.

$ Check price
49.9949.99Amazon 4.5 logo(4+ reviews)

More Bioinformatics resources