choose the results using different parameters for CAFE3.0 (gene family expansion analysis)

Question

[qwzhang0601]

Hello:

I am using the CAFE3 to identify gene family expansion/contraction in a rodent species that I am studying. I used "caferror.py" to search the error model and get the following score. Score with global errormodel: 216055.690765

As you may know, when we use "caferror.py" to estimate the error model, we can only set a uniform lamada. If we know different branches of the phylogenetic tree have different evolution rat we must run CAFE again by setting different lamada for different branches. So using this optimal error model from caferror.py I ran CAFE again and set two different lambda among two branches. Below are the results from 5 runs.

Run1 Lambda : 0.00176567769707,0.00214836810164 & Score: 216604.415668 Run2 Lambda : 0.00178123755558,0.00220871098435 & Score: 216268.293872 Run3 Lambda : 0.00176567624910,0.00214836293012 & Score: 216604.411813 Run4 Lambda : 0.00176572814301,0.00214852212537 & Score: 216604.409239 Run5 Lambda : 0.00176567624910,0.00214836293012 & Score: 216604.411813

Here are my questions. (1) The score a with single Lambda is lower than that with two Lambda. So the results with single Lambda are more reliable? (2) Compare among the results with two Lambda, four runs got similar parameters and scores, but the left one (i.e. Run2) has lower score. If I want to choose one result from the five runs, should I select the one with the lowest score (i.e., Run2) or the one from the four more stable results (e.g., Run3)?

Thanks

[Ben Fulton]

How did you decide which branches to change the lambdas on? Do these all represent searches for two lambdas based on different lambda trees?

[qwzhang0601]

In my phylogenetic tree, there are rodent species and primates. So I just simply consider the rodent species have the same lambda, and the primates species have another lambda. When I were setting the two lambda, I am considering different branches (primate and rodent) in the same tree.

thanks

Answer 1

[Ben Fulton]

The score reported is the calculated log of the likelihood of the tree being correct, so lower scores should represent more accurate values. However, your run 2 is surprisingly different from the other runs and I would be wary of trusting it without more information on what exactly the differences are from the other runs..