November 27, 2018

What happens if the Scoring Matrix changes?

In general, when the Scoring matrix changes, then the optimal alignment also changes

For example, if mismatches are expensive then the method may use more gaps

But if we multiply every score by a positive constant, then the alignment does not change

What happens if the Scoring Matrix changes?

For example, if we align DNA with

  • +10 for match and
  • -20 for mismatch,

every decision will be the same as when the score is

  • +1 for match and
  • -2 for mismatch

The alignment is invariant to scale changes

The alignment does not change if the scoring matrix \(S\) is replaced by another matrix \(\lambda S\)

(this is valid for any positive number \(\lambda\))

What is the best \(\lambda\)?

To answer that question we have to understand how the Scoring matrices are built

Reference alignments

As we have seen, scoring matrices are built based on a previous alignment

Margaret Dayhoff took 71 families of closely related proteins

Inside each family all sequences had at least 85% of identical residues

She built the alignments manually, and calculated the frequency of mutation for each pair of residues

Reference alignments

The same procedure was done by Henikoff and Henikoff, but using only the conserved domains of proteins

In both cases, they observed each residue

  • \(a\) on the first sequence, and
  • \(b\) on the second sequence

and they calculated the proportion \(r_{ab}\) of residues \(a\) aligned to \(b\)

We can “easily” do the same with DNA

Assuming that this is a good alignment

TAATACCGCATAATAATCAGTTTTCACATGGAGACTGATTTAAAGGAGTAATCCGCTT
TAATACCGCATAACATTAGTATTTCACATGAAGTACTAATTAAAGGAGTAATCCGCTT
   b
a       A     C     G     T
  A 29.31  1.72  1.72  3.45
  C  1.72 13.79  0.00  1.72
  G  1.72  0.00 12.07  1.72
  T  3.45  1.72  1.72 24.14

Probabilities

These observed proportions can be seen as an estimation of the probability of aligning two residues when they have a common ancestor

We want to compare this against the probability of aligning two unrelated residues. We will compare two hypothesis

\(H_0\): there is no common ancestor

\(H_1\): there is a common ancestor

Probabilities for one pair

As we said, when there is a common ancestor, we have \[\Pr(a,b|H_1) = r_{ab}\] In the other case, if there is no common ancestor, then \(a\) and \(b\) are independent, thus \[\Pr(a, b|H_0) = \Pr(a|H_0)\Pr(b|H_0)=p_{a}p_{b}\] where \(p_a\) is simply the proportion of residue \(a\) in all the sequences

(independent probabilities means that we can multiply them)

Odds Ratio

To compare these two probabilities we can take their ratio \[\frac{\Pr(a,b|H_1)}{\Pr(a, b|H_0)}=\frac{r_{ab}}{p_a p_b}\] If this value is greater than 1, then \(H_1\) is more plausible than \(H_0\)

It can be proved that this value is proportional to \[\Pr(H_1| a\text{ aligns with }b)\] which is really what we are looking for

Log-Ratio

In practice it is much easier to work with the logarithm of it \[\log\frac{\Pr(a,b|H_1)}{\Pr(a, b|H_0)}=\log \frac{r_{ab}}{p_a p_b}\] This number is positive when \(H_1\) is more plausible than \(H_0\), and negative otherwise

Extending

We want to extend this to the complete sequences \(q\) and \(s\)

Assuming that each residue is independent, we have \[\Pr(q,s|H_1)=\Pr(q_1,s_1|H_1)\cdot\Pr(q_2,s_2|H_1)\cdots\Pr(q_n,s_n|H_1)\] The same formula works for \(H_0\), so \[\frac{\Pr(q,s|H_1)}{\Pr(q,s|H_0)}=\frac{\Pr(q_1,s_1|H_1)}{\Pr(q_1,s_1|H_0)}\cdot\frac{\Pr(q_2,s_2|H_1)}{\Pr(q_2,s_2|H_0)}\cdots\frac{\Pr(q_n,s_n|H_1)}{\Pr(q_n,s_n|H_0)}\]

(here we assume that \(q\) and \(s\) may contain gaps so they have equal length \(n\))

Now with logarithm

\[\log\frac{\Pr(q,s|H_1)}{\Pr(q,s|H_0)}=\log\frac{\Pr(q_1,s_1|H_1)}{\Pr(q_1,s_1|H_0)}+\log\frac{\Pr(q_2,s_2|H_1)}{\Pr(q_2,s_2|H_0)}+\cdots+\log\frac{\Pr(q_n,s_n|H_1)}{\Pr(q_n,s_n|H_0)}\] That is \[\log\frac{\Pr(q,s|H_1)}{\Pr(q,s|H_0)}=\sum_{i=1}^n\log\frac{\Pr(q_i, s_i|H_1)}{\Pr(q_i, s_i|H_0)} = \sum_{i=1}^n \log \frac{r_{q_i s_i}}{p_{q_i}p_{s_i}}\] This sum is a score of the alignment!

Score of alignment

Let’s assume that we have an alignment of \(q\) and \(s\), that is, we have gaps inside \(q\) and \(s\) so they have the same length \(n\)

If we have a scoring matrix \(S\) We can find its score as the sum \[\sum_{i=1}^n S[q_i, s_i]\] Comparing this with the previous slide we find that \[\log \frac{r_{q_i s_i}}{p_{q_i}p_{s_i}}\] can be used as a score.

General formula

Remember that the score can be multiplied by \(\lambda\) without any effect on the alignment.

Simplifying, we can say \[\lambda S[a,b]=\log \frac{r_{ab}}{p_{a}p_{b}}\]

This is how PAM and BLOSUM matrices are built

According to WikiPedia, they used the formula \[\lambda S[a,b]=\text{round}\left(2\cdot\log_2 \frac{r_{ab}}{p_{a}p_{b}}\right)\] that is, rounded to the nearest integer

Computers work faster with integer numbers, so PAM and BLOSUM drop the decimals

The factor 2 is used to get more information before rounding

Going backwards

We have seen the process from probabilities to score matrix

In the other way,

  • if we know the matrix \(S\)
    • (such as PAM30 or BLOSUM62),
  • then we can find the probabilities

Example with DNA

Let’s say that the score for match is 1, and we will try different values for mismatch, like -1, -2,…, -5

In this case we have 4 letters, that we assume have the same probability. Thus \[p_a = \frac{1}{4}\quad\text{ for all }a\in \{A,C,G,T\}\]

What about \(r_{ab}\)?

To find \(r_{ab}\) we need to find \(\lambda\), because \[\lambda S[a,b]=\log \frac{r_{ab}}{p_{a}p_{b}}\] and therefore \[r_{ab}= p_a p_b \exp(\lambda S[a,b])\]

Notice that so far we did not say what was the logarithm’s base

Here we chose natural logarithms, but in theory it can be any base

\(r_{ab}\) is a probability

In this case we also know that \[\sum_{a,b} r_{ab}=1\] therefore \[\sum_{ab} p_a p_b \exp(\lambda S[a,b]) = 1\] This condition allows us to find \(\lambda\)

Looking for \(\lambda\)

Finding \(\lambda\)

Let’s consider the function \[f(x)=\sum_{a,b}p_a\, p_b\, \exp(x\, S_{a,b})\] so \(f(\lambda)=1\). When \(x=0\) we have \[f(0)=\sum_{a,b}p_a\, p_b\, \exp(0) = \sum_{a}p_a\, \sum_{b}p_b = 1\] because \(p_a\) and \(p_b\) are probabilities. But we want \(\lambda>0\)

Let’s plot \(f(x)\)

It can be proved that always \(f(x)\) looks like this:

Each matrix has its own \(\lambda\)

It is easy to find \(\lambda\) with a computer

Example for match = 1 and mismatch = -2

For match = 1 and mismatch = -2 we have \(\lambda=1.333\)

The \(r_{ab}\) probability matrix is

       A      C      T      G
A 23.691  0.435  0.435  0.435
C  0.435 23.691  0.435  0.435
T  0.435  0.435 23.691  0.435
G  0.435  0.435  0.435 23.691

Example for match = 1 and mismatch = -5

For match = 1 and mismatch = -5 we have \(\lambda=1.385\)

The \(r_{ab}\) probability matrix is

         A        C        T        G
A 24.97400  0.00614  0.00614  0.00614
C  0.00614 24.97400  0.00614  0.00614
T  0.00614  0.00614 24.97400  0.00614
G  0.00614  0.00614  0.00614 24.97400

Average identity

Once we know \(\lambda\) we can determine the probabilities of mutation \[q_{a,b}= p_a\, p_b\, \exp(\lambda S_{a,b})\] Now we can, for example, determine the average identity between two aligned sequences \[\sum_a p_a\, q_{a,a}\] that is, the average probability that a residue is conserved

Some examples of average identity

match mismatch lambda Avg.ident
1 -1 1.098 75
1 -2 1.333 94.8
1 -3 1.375 98.8
1 -4 1.383 99.7
1 -5 1.385 99.9
2 -3 0.6333 47.1
4 -5 0.3013 33.8