HyPhy message board
Theoretical questions >> Sequence Analysis >> Branch-site REL and mean omega values

Message started by jstrohmg on Jun 17th, 2014 at 1:57pm

Title: Branch-site REL and mean omega values
Post by jstrohmg on Jun 17th, 2014 at 1:57pm
Hi there,

I'm trying to compute dN/dS for 48 species of fish. I'm computing these ratios for each of the 13 OXPHOS genes from the mitochondrial genome individually.

For some genes/ species I get a mean omega value of ten when the average from other species is ~0.05.

Also the means from omega 1,2,3 don't average to 10.

Ex: species/gene1 omega 1= 0 2= 0 3= 0 mean =10

species2 1= 1 2= 1 3= 12.69 mean =10

These values of 10 seem arbitrary as are never followed by fractions/decimals.

Is it safe to keep them in my analyses or should they be removed as outliers?

It seems to be more common in shorter genes.
Many thanks!


Title: Re: Branch-site REL and mean omega values
Post by Sergei on Jun 18th, 2014 at 10:56am
Hi Jeff,

In Branch-site REL, the mean omega is estimated using a simple evolutionary model (one omega per branch), and artificially capped at 10 (as you noticed). It is meant to serve as a starting value for the optimization for BS REL, and should not be used for inference.

The average of omega estimates from the multiple rate model is NOT going to equal the mean in general. You should compute the weighed average, if you want to get a more comparable value (omega1 * prob (omega = omega1) + omega2 * prob (omega = omega2) ...), but even then don't expect equality. This is because averaging point estimates does not take variance into account. If you have omega = 10000 inferred from 1% of sites, the confidence bounds on that 10000 are going to be very wide.

Generally, you should NOT use point estimates, but rather the results of hypothesis testing (for some discussion on the matter see  Multimedia File Viewing and Clickable Links are available for Registered Members only!!  You need to Login Login). If you wish to estimate the mean omega at a branch, it is better to do this directly (e.g. with a simple MG94 local model), and not by transforming point estimates of individual rate classes under a more complex model.


HyPhy message board » Powered by YaBB 2.5.2!
YaBB Forum Software © 2000-2023. All Rights Reserved.