Hi Sergei,
I am running the YangNielsenBranSite2005 analysis and I am continually getting worse scores for my Alternate model than for I am for Null for Test 2.  I have tried increasing the optimisation precision and the number of iterations per variable to

OPTIMIZATION_PRECISION = 0.00001;
MAXIMUM_ITERATIONS_PER_VARIABLE=1000;

but I am now not getting convergence for my alternate hypothesis.

I have run the PAML branch site test on the same data and it is reporting higher likelihood scores than HYPHY for both the Alternate hypothesis and the Null for Test 2. For other tests I usually find the programs agree to within a few decimal places as you have stated in my previous post so I am very interested in figuring out why the results are off this time.

I have two questions.

1) If I am getting a worse likelihood score for the Alternate hypothesis should I take that to mean I shouldn't reject the Null hypothesis. Or should I take that to mean the function is not converging and I need to increase the number of iterations and the precision further?

2) What precision level should I be using to get the same results as PAML and how many iterations should be necessary to get convergence for the likelihood function? Also is there anyway to speed up the convergence?

As always thank you for the help.

Sincerely,
Andrew

Dear Andrew,

Hard to say what the difference is due to without looking at the dataset in question. Quite likely it is a convergence problem (if you read the latest branch-site paper by Anisimova and Yang, they also caution that PAML may have convergence issues for some datasets, and suggest repeated runs with different starting conditions). HyPhy and PAML use different optimization techniques, so it is logical that PAML will have better convergence properties in some cases.

If you send me your data file and what branch you were testing for (spond at ucsd dot edu) I can see what may be causing the issue.

Cheers,
Sergei