Re: [External] Boost value in aMD simulation

From: Thomas Evangelidis (
Date: Wed Jul 23 2014 - 16:37:47 CDT

On 23 July 2014 23:32, James Starlight <> wrote:

> Hi Thomas,
> this is some receptor from the GPCR family but with no known experimental
> structure available at this moment. The reason of the refirement of the
> extracellular loops is that I need very good model for further md
> simulation to study initial ligand binding (which happens initially in
> these extracellular loops) with the receptor (like in this paper
> ). In case of this
> receptor modeller never predicted some secondary structure elements in the
> region of second extracellular loop (which always looks like very long coil
> (~30 aa) althought 2 disulphide bridges are located here as the possible
> constraints) although we changes templates (in other GPCRs for instance in
> Beta-adrenergic receptor) there are short helix in this region for instance
> which is the vestibule to the ligand-binding pocket.
In the paper you cited the authors ran multiple simulations, each one
lasting several microsecond, to see the binding events. So unless you have
access to Anton, the chances to see a binding event and transport are very
low with unbiased MD on conventional computer clusters. Even with
accelerated MD it is still difficult and you also have the complication
that you cannot reweight such big systems to get accurate unbiased
observables. I am working on a similar case of ligand binding and transport
across the membrane. I have run more that 1 microsecond aMD with distance
restraints between the ligands (5 copies of the same molecule) and the
protein and still haven't seen binding. Your best bet is to use a more
sophisticated method, like funnel metadynamics, which can give you all
binding sites (including the allosteric) on the extracellular surface.
However, this method is much more convoluted than aMD and hence I wouldn't
recommend pursuing it without the guidance of an expert.

As for the initial loop conformation, from what you say I wouldn't worry
that much and rather let it assume the hypothetical helical conformation
during the production run. However, freezing the transmembrane part of the
protein while doing loop prediction with MD is wrong. Transmembrane
proteins are crystallized in detergent, not in native conditions. There are
also crystal contacts effects that need to be corrected. Consequently, you
have to let them relax first before freezing them, especially in your case
that you start from a homology model.

> On other hand we never do *MANY* models. From your suggestion it seems
> that I need to generate many models (e.g 1000 pdbs) which will be with the
> same bundle but differ only in one extracellular loopin which I most
> interested. Than I need to cluster my 1000 pdbs according to the
> conformation of extracellulalar loop and chose some shared conformation
> which will be in most populated cluster. Could you provide me with the
> ideas of some soft which will be good for this which will have i)python
> interface ii) tutorial :-)?
No I mean, at least 10,000 loop models for the big one. Modeller have a
python API and can run in parallel, so than won't take much to finish. I
lately do these kind of tasks with Rosetta which provides more options
specifically for transmembrane proteins and does also run in parallel. Both
software have plenty of tutorials on their websites or source bundles.
Rosetta tends to give more loop models with helical structures (due to
fragment-based assembly) as in your case, yet their validity is
questionable. In fact, I am not aware of any loop modeling method that can
give accurate predictions for loops longer than 12 aa.

> Many thanks for suggestions,
> James
> 2014-07-24 0:02 GMT+04:00 Thomas Evangelidis <>:
>> On 23 July 2014 22:24, James Starlight <> wrote:
>>> Dear NAMD users,
>>> I'm very appologise, this question was adressed to both forums :-)
>>> My question is refirement of loops predicted by modeller
>>> the goal is
>>> 1- just refine loops but not refine rest of the protein (freezing it)
>>> 2- avoid to use the membrane, because I'd like to refine water expoised
>>> regions of the membrane protein only
>>> 3- enhansing sampling engine to refine it quickly
>>> 4- do several refirement simulation, use PCA to test coverage (found
>>> shared regions in PC projections of the loops conformations predicted by
>>> such refirement from seveal simulations)
>> Do you realize how much time this protocol needs to reach convergence on
>> the eigenvectors of the accumulated trajectory (if that ever happens at
>> all)? Why not just create tens of thousands of loop models, cluster them
>> and start a simulation from the representative conformation of the
>> predominant cluster?
>> On another note, are these extracellular and intracellular loops so
>> important to invest that amount of labour on them? What kind of protein is
>> this? Can you upload some pictures on Dropbox to show us the protein with
>> the loops coloured differently?
>>> 2014-07-23 23:14 GMT+04:00 Pino, James Christopher <
>>>> Your greeting implies this went to the wrong forum.
>>>> However, I am using aMD through amber also.
>>>> I am curious what your goal is? Loop refinement of de novo models?
>>>> James
>>>> Vanderbilt University
>>>> ________________________________________
>>>> From: [] on behalf of
>>>> James Starlight []
>>>> Sent: Wednesday, July 23, 2014 6:30 AM
>>>> To: Namd Mailing List
>>>> Subject: [External] namd-l: Boost value in aMD simulation
>>>> Dear Amber users!
>>>> In this topic I would like to talk about information obtained from the
>>>> amd.log concerning the reasonability of the values of boost potentials
>>>> applied to my system. For my case I'm simulating protein in explicit water
>>>> with the task to refine its loops appling position restraints on part of
>>>> the protein (which I'd like to keep unchanged). Firstly I've performed cMD
>>>> with no restraints to obtain all values needed to compute boost and alpha
>>>> according to the impirical formuli presented in manual. Then I run 2 boost
>>>> aMD with applied of the position restraints on the bigger part of the
>>>> protein and see amd.log. According to themd.log the value of dihedral
>>>> boost added to my system per step during amd simulation has been ~ 10
>>>> Kcal/mol on each step. I wounder if the dihe boost of this value have been
>>>> applied to the whole protein (including its restrained parts) or only to
>>>> its unrestrained (in my case loops) parts? What the reasonable dUdihe
>>>> should be expected in principle for the simulation of protein consisted of
>>>> ~ 300 amino acids? I guess that this value should be nuch biger than
>>>> several Kcal/ mol to obtain better sampling.
>>>> Thanks for suggestions,
>>>> James
>> --
>> ======================================================================
>> Thomas Evangelidis
>> PhD student
>> University of Athens
>> Faculty of Pharmacy
>> Department of Pharmaceutical Chemistry
>> Panepistimioupoli-Zografou
>> 157 71 Athens
>> email:
>> website:

Thomas Evangelidis
PhD student
University of Athens
Faculty of Pharmacy
Department of Pharmaceutical Chemistry
157 71 Athens


This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:22:39 CST