measure of residue dissimilarity in aligment (Phylogenetic discriminant)

Question

[michau]

Hi, Is there any probabilistic measure of position dissimilarity from rest of alignment? ie. method to assess which residues are responsible for the phylogenetic and functional differences.

I have alignment of ATP synthases, and I noticed that Mycobacterium is highly divergent from other Bacteria or even other Actinobacteria. I need some measure to statistically discriminate which residues are responsible for this divergence.

Any thoughts?

Thanks in advance

[Joe]

I'm not sure if I exactly understand what you're after, but here goes:

You might be interested in calculating the Shannon Entropy per column of your Sequence alignment. High entropy positions will be your more divergent ones see for example: https://gist.github.com/jrjhealey/130d4efc6260dd76821edc8a41d45b6a.

You may need to take this further and do a dN/dS analysis or similar, since it probably won't be enough to just determine sites that are variable. You will need to demonstrate that they are causing meaningful selection (i.e. non synonymous).

[michau]

Firstly: Thank you for answer

I was thinking to use dn/ds as a next step, but as far as I know it allows only for pairwise comparisons. I was looking for something more like measuring inside clade vs. outside clade (site specific) variation. Column-wide, like:

for each column:
    MANOVA where 
        independent var= inside / outside 
        dependent var  = freqs of aminoacids in a column.

Or am I thinking bullshit? (I started my bioinformatical adventure recently → I'm still green as a lime and trying to learn) Nevertheless I will go with dn/ds as it will answer my question.

[Joe]

I don’t know enough about the stats to speak to whether a manova approach would work.

The only limitation that strikes me with that approach, however, is that objectively clustering ‘clades’ is a very difficult problem. It’s often much more obvious to a person than to a computer.

I would think there are approaches which will work in a non pairwise fashion. A quick bit of googling bought this up, which sounds like it might fit the bill?

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2887424/