I can't address your own results, but I can tell you my experience:

To find a native like structure I created 100 models for 5 different template combinations with and without loop modelling and with and without MD refinement, and I found that no loop modelling and no MD refinement gave me the best models.

I took my 5-10 best models as ranked by dope and the objective function and then submitted them to PROsa.  Then I further analyzed the 5 or so best via procheck.  IMO brute force and sheer numbers are the best way to determine emperically what combinations of settings will work best for you.

I would definitely try no loop model and no MD and always do 100 models minimum.

Cheers.

--- On Tue, 6/15/10, modeller_usage-request@salilab.org modeller_usage-request@salilab.org wrote:

From: modeller_usage-request@salilab.org modeller_usage-request@salilab.org Subject: modeller_usage Digest, Vol 9, Issue 88 To: modeller_usage@salilab.org Date: Tuesday, June 15, 2010, 8:25 AM

Send modeller_usage mailing list submissions to     modeller_usage@salilab.org

To subscribe or unsubscribe via the World Wide Web, visit     https://salilab.org/mailman/listinfo/modeller_usage or, via email, send a message with subject or body 'help' to     modeller_usage-request@salilab.org

You can reach the person managing the list at     modeller_usage-owner@salilab.org

When replying, please edit your Subject line so it is more specific than "Re: Contents of modeller_usage digest..."

Today's Topics:

1. structure refinement and loop optimization protocol       (Thomas Evangelidis)    2. Using my own initial structure to model a loop       (Thomas Evangelidis)

----------------------------------------------------------------------

Message: 1 Date: Tue, 15 Jun 2010 02:29:37 +0100 From: Thomas Evangelidis tevang3@gmail.com Subject: [modeller_usage] structure refinement and loop optimization     protocol To: modeller_usage@salilab.org Message-ID:     AANLkTil1lD1Yf9cWtV5O1ID9kxv7QZkzaDcz3b4n0iBo@mail.gmail.com Content-Type: text/plain; charset="iso-8859-1"

Dear Modellers,

I've read previous posts on the same topic and concluded that it is better to generate multiple models with moderate refinement and loop optimization level, rather that a few with very thorough parameterization. I've also noticed myself that with the thorough parameterization parts of the secondary structure are distorted.

I have concluded about the optimum alignment after a lot of experimentation and would like to set up a very effective optimization process. However I'm not sure about the output files. My code looks like this:

a = MyLoopModel(env, alnfile=alignment, >                       knowns=known_templates, > assess_methods=(assess.DOPEHR,assess.normalized_dope), >                       sequence='target') >             a.starting_model = 1 >             a.ending_model = 2 >             # Normal VTFM model optimization: >             a.library_schedule = autosched.normal >             a.max_var_iterations = 200 ## 200 by default >             # Very thorough MD model optimization: >             a.md_level = refine.slow >             a.repeat_optimization = 1 > >             a.loop.starting_model = 1           # First loop model >             a.loop.ending_model   = 5          # Last loop model >             a.loop.md_level       = refine.slow # Loop model refinement > level >

Which generates the following pdb files:

target.B99990001.pdb  target.B99990002.pdb  target.BL00040002.pdb > target.IL00000001.pdb  target.IL00000002.pdb >

I thought the above should perform model refinement twice and write 5 different conformations (loop optimization) for each. So my questions are the following:

1) Can you explain what's happening with the .pdb files?

2) I 'd like to ask your opinion about the most effective way to find a near-native protein conformation in low sequence identity levels. How should the parameters shown above be set? I don't care if it's running a day or so as long as I get good results.

3) I also attempted to cluster the models with a.cluster(cluster_cut=1.5), which generated a representative structure with the parts of the protein that remained similar in most of the models but without the variable parts (files cluster.ini and cluster.opt). Does it make sense to select the model that is closer to that consensus structure? If yes is there a way to do it with Modeller? I know it can been found with Maxcluster program. Or alternatively, do you reckon it is better to select the based model based on the normalized DOPE z-score?

Hope to get some answers on these question cause I've been strangling to find the best refinement/optimization protocol for several weeks.

thanks, Thomas