1
$\begingroup$

I have % identity and % similarity scores for ~50K protein alignments, that I fetched from Ensembl Compara database. The issue is that I wanted to have divergence scores instead. So in order to calculate divergence scores, I first looked for the conventional method to calculate divergence score from protein alignments. But turns out, most of the methods/tools calculate % identity and % similarity scores instead of divergence scores.

(1) So I wonder what is the conventional method to calculate divergence scores from protein alignments.

(2) Could I simply calculate % sequence divergence as 100 - % identity or 100- % similarity score? If that's ok, should I prefer 100 - % identity or 100- % similarity?

Any suggestions from experts are welcome.

Update: As suggested in a comment below, I should mention that the average % identity score is more than 70. I hope this information would be helpful.

$\endgroup$
4
  • $\begingroup$ It depends on how similar your proteins are, the quality of the alignements. Measuring distance between distant sequences is harder than for very similar sequences. $\endgroup$ – reuns Feb 27 at 13:50
  • 1
    $\begingroup$ The sequences of the proteins very similar. Average % identity is >70. Thanks for suggesting, in the updated question, I mention this. Also, I started the bounty, in case you may know the answer! $\endgroup$ – Ramirez Feb 27 at 22:58
  • $\begingroup$ Since you don't want to use similarity or identity calculating divergence as 1-(similarity | identity) won't do what you want. What exactly you want to find? Why you can't use similarity or identity $\endgroup$ – Hachiloni Feb 28 at 6:49
  • $\begingroup$ @Hachiloni, yes, indeed the answer of @Mike Serfas below, suggests that the 100 - (% identity) may not be the right way to calculate the % divergence. However, similarity score could be used, if some condition is met. Well, that's what I was asking i.e. if I could calculate divergence score from similarity|identity scores. $\endgroup$ – Ramirez Feb 28 at 17:55
1
+50
$\begingroup$

This is a very tentative answer as I haven't done anything like this, but it's a learning experience for me and I hope it could be helpful to get a more knowledgeable response started.
I don't think you can do this. Per Wang, 2009 and Yona, 2002, divergence scores could be Kullback-Leibler divergence or Jensen-Shannon divergence, but both papers opt for the latter. Divergence scores are based on "empirical probability distributions between the 20 amino acids". BLOSUM or PAM matrices may be used. Simply counting the identical or similar residues wouldn't contain as much information. Wang, 2004 compared JensenShannon to numerous other scoring methods in the context of PSI-BLAST.

The similarity score is related to the divergence score: Score = 0.5(1-D)(1+S) where D is the divergence score and S is the significance score. It is equation #15 in Wang, 2009 paper. If you could safely assume the significance score for some alignments was near 1 (no chance similarities) then maybe you could say D = 1 - similarity score. Unfortunately, judging by the glossary because I haven't downloaded any of the datasets, for the Ensembl data similarity is merely "How well one sequence matches another determined by calculation by an alignment program of identical and conserved residues/nucleotides." So from that I take it they may just be counting up the "conserved" amino acids in the old fashioned way, but don't know that.

$\endgroup$
5
  • $\begingroup$ thanks for the answer, it indeed clears things out. Especially the equation i.e. Score = 0.5(1-D)(1+S) seems to be very relevant to what I wanted to do. I wonder if the condition i.e. significance score for some alignments was near 1 (no chance similarities) is satisfied in the case of the alignments I have. Could you please point to a reference where I can get more details regarding this equation? $\endgroup$ – Ramirez Feb 28 at 18:08
  • 1
    $\begingroup$ It's equation 15 from Wang, 2009 above, which goes through the derivation. But I don't know if/when you could actually assume significance score near 1, though small deviations from that wouldn't affect the value of D very much. $\endgroup$ – Mike Serfas Feb 28 at 22:48
  • $\begingroup$ Thanks! I just added source of the equation in your answer. I am indeed not sure if/when I could assume significance score near 1. I wonder what makes you believe that small deviations in that won't change D much. $\endgroup$ – Ramirez Mar 1 at 16:31
  • $\begingroup$ I don't see your edit - I'd thought my hyperlink to Wang 2009 would be enough, given the context. My comment about the significance score was trivial - since it's added to 1, a variation in that score only affects D half as much. $\endgroup$ – Mike Serfas Mar 1 at 17:53
  • $\begingroup$ I think my edit which may not be that helpful indeed, will be checked by moderators I think before it shows up publicly. Ok, indeed variation in D may not have much effect. Thanks for the clarification. $\endgroup$ – Ramirez Mar 1 at 23:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.