Hi Ross,
Thanks so much for the suggestions. They are very helpful. ... in addition,
I would like to ask another question about if pmemd.cuda and/or
pmemd.cuda.MPI support TI computations?
It looks to me ... yes they should according to your webpage info: poster
2013" Ross C. Walker - "Fast and Reliable Ligand Binding Free Energies -
Thermodynamic Integration on GPUs", Gordon Research Conference on Computer
Aided Drug Design, Mount Snow Resort, Vermont, USA, Jul 2013."
However, on amber pages or amber14 manual, the information is not clear to
me. ... the information are mostly about regular MD instead of free energy
calculation.
Sorry that, I could have been compiling myself and tested but the software
installation support here is not so handy, and we are in a process
evaluating buying (and use up budget in time) GPU machine(s) now. So I ask
for help here.
Thanks very much again.
Ying-chieh
-----Original Message-----
From: Ross Walker [mailto:ross.rosswalker.co.uk]
Sent: Wednesday, May 14, 2014 11:06 PM
To: AMBER Mailing List
Subject: Re: [AMBER] thermodynamic integration (TI) simulation: pmemd of
amber14 vs sander of amber11- vdw transformation
Hi Ying-Chieh,
To add some more to what Jason said - accepting the fact that your
simulation parameters appear to be very different (are you definitely using
softcore in amber11 for example) this also looks to me like a case of the
results being hopelessly under converged. 600,000 steps is only 1.2ns which
likely means your error bars on the sampling are huge. So a good chunk of
that difference could be coming from there. Try doing 5 or so repeats of
your AMBER 11 calculations with different random seeds / initial structures
and see what the variation is in the results you get.
As with any paper I receive where people do free energy calculations I
always ask them to do multiple repeats or ideally plot the free energy
obtained as a function of cumulative sampling length - the last one people
rarely do - I think because when they try it it is a real eye opener into
how unconverged the calculations typically are. You try this yourself to
see.
Take you 600K steps from the AMBER 11 simulation. Throw away all but the
first 10K and calculate the delta delta G and then plot that on a graph
where the X axis is sample points and the Y axis the DDG obtained. Then
repeat throwing away all but the first 20K, add that to the plot. Then 30K,
then 40K etc all the way up to the full 600K - it's probably best to script
this. See what the plot looks like and it should give you a nice visual idea
of how the run is converging.
All the best
Ross
On 5/14/14, 4:37 AM, "Jason Swails" <jason.swails.gmail.com> wrote:
>On Wed, 2014-05-14 at 18:21 +0800, Ying-Chieh Sun wrote:
>> Hi,
>>
>>
>>
>> We are learning TI computation for protein-ligand binding using pmemd
>>of amber14. We used 3-step protocol ((1)switch charge off, (2) vdw
>>transformation, (3)switch charge back on) to compute Delta Delta G of
>>two ligand binding with the same protein.
>>
>>
>>
>> The results of the vdw transformation are summarized below:
>>
>>
>>
>> Delta G (Complex) Delta G(Solution)
>>
>> -1.6746+-4.7307 -0.4455+-4.0797 (pmemd of amber14: ligand charge:
>>BCC
>> charge!)
>>
>> -10.8895+-6.0608 -13.1352+-5.4417 (sander of amber11: ligand charge:
>>RESP
>> charge!)
>>
>>
>>
>> Because we used BCC charge this time for pmemd, we have not used RESP
>>charge yet. But, in the vdw transformation, the charges are off.
>>Therefore, I think the Delta G (not Delta Delta G) should be similar.
>>But they differ by
>>~10
>> kcal/mol in the complex and solution computations, respectively. The
>>Delta Delta G¹s are indeed in the similar range.
>
>The TI code in pmemd is quite new and using it is radically different
>than using the TI code in sander. Therefore, I would suggest changing
>as _few_ things as possible when carrying out your comparisons. (i.e.,
>make sure you are using exactly the same parameters in both cases to
>carry out your simulations for comparison). Any attempt to explain the
>discrepancy will most likely be unsubstantiated speculation.
>
>>
>>
>>
>> Any advice/suggestions are appreciated.
>>
>>
>>
>> We also found a typo, I think, in the output file. As shown below, it
>>should
>> be ³A V E R A G E S O V E R 600000 S T E P S² instead of 3000 S T
>>E P
>> S.
>
>No, this is correct. pmemd does not compute the energy every step (it
>only computes the energy on the steps that it needs to print the
>energy). This is done because computing the energy introduces an
>unnecessary computational cost (and if you don't need the energy, why
>compute it?). So my guess is that you printed every 200 steps for
>600,000 total steps...
>
>HTH,
>Jason
>
>--
>Jason M. Swails
>BioMaPS,
>Rutgers University
>Postdoctoral Researcher
>
>
>_______________________________________________
>AMBER mailing list
>AMBER.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed May 14 2014 - 21:00:02 PDT
