From: Jerome Henin (jhenin_at_cmm.chem.upenn.edu)
Date: Fri May 01 2009 - 14:28:24 CDT
On Fri, May 1, 2009 at 1:05 PM, Christopher Hartshorn
> Chris and Jerome,
> Thank you again for the journal links, etc. So, I reran the tutorial with
> the new parameters in 2.7 and the pllot "looked" better, but nothing like
> the 2.6 type from tutorial. I then reran using NAMD2.6 and got exactly the
> same graph as in the tutorial (nice smooth increase up to 20kcal/mol then
> sharrp/smooth drop near the end. I guess this was to be expected because of
> the lack of the soft-core approach in 2.6. So, I have three questions this
> 1) Besides the fact that the softcore approach is tunable, what are the
> other advantages over the old? Is it computational cost or more realistic,
Much better convergence at the end-points, so you don't need to add
quite as many windows anymore. Actually the "traditional" scaling for
Lennard-Jones potentials has a singularity at the end-points, so
sometimes convergence can be next to impossible to achieve - you
essentially have to get lucky.
> 2) Since the latest 2.7 is so much faster (150%) on all of my systems I
> would, of course, prefer to keep using it. What config file lines would I
> put to start an FEP calculation in 2.7 that ran exactly the same as 2.6 to
> test that extreme (e.g. softcore off) versus others with softcore on? Would
> this just be "decouple off?"
You can go back to the old behavior using the following settings:
a) decouple off (to go back to "vanishing mode" instead of decoupling)
b) fepVdwShiftCoeff 0 (to turn off the soft-core potential)
c) fepElecLambdaStart 0
d) fepVdwLambdaEnd 1
> 3) Since these runs can take up to 25 days each, which parameters would be
> best to change first in order to minimize the amount of runs I would need to
> find the ideal set of softcore parameters? In other words, what if anything
> would make me want to change, hypothetically, fepElecLambdaStart 0.5
> to fepElecLambdaStart 0.3 or fepVdwLambdaEnd 1.0 to fepVdwLambdaEnd 0.8 etc.
> etc. after looking at a plot of dG vs. lambda step?
Interesting question. I am not aware of a systematic study documenting
the effects of these parameters on convergence. I think most people
set them based on intuition and, to some degree, personal taste.
Playing with small, inexpensive systems helps building that intuition.
I would not try any systematic tuning based on full-scale simulations.
> Thank you very much. I know I have a lot of questions. So, hopefully they
> will be a benefit to others as well.
The code has been changing a lot, and it is difficult to make the
documentation clear and complete, so yes, I think this discussion is
useful for everyone.
> On May 1, 2009, at 6:35 AM, Chris Harrison wrote:
> I just wanted to quickly let you know I got your email and will send a more
> complete answer later today. However, I did want to mention that the plot
> you got looked not dissimilar to plots we got during early stages of
> implementing softcore, so it may not be as bad as you feared and only
> require the proper softcore parameters. The tutorial is still based on FEP
> in NAMD2.6; we are currently working on an updated tutorial that will
> released shortly after the official NAMD2.7 release. Please let me know how
> the run using the parameters I sent turns out, as this will be a strong
> indicator to me if there is a bug. I'll send a more complete answer to your
> question(s) later today.
> Chris Harrison, Ph.D.
> Theoretical and Computational Biophysics Group
> NIH Resource for Macromolecular Modeling and Bioinformatics
> Beckman Institute for Advanced Science and Technology
> University of Illinois, 405 N. Mathews Ave., Urbana, IL 61801
> char_at_ks.uiuc.edu Voice: 217-244-1733
> http://www.ks.uiuc.edu/~char Fax: 217-244-6078
> On Thu, Apr 30, 2009 at 12:57 PM, Christopher Hartshorn
> <cmhartshorn_at_wsu.edu> wrote:
>> Thank you for the help. Much appreciated. As far as the plot I have
>> attached the graph to this email. I can not send it to the server.
>> As far as what I am plotting on this graph, I am plotting up the end line
>> of every Lambda step (which is at every 10000th line in my .fepout file
>> because I am doing 100Ksteps which are output every 10) which includes the
>> overall dG (which I am assuming is the sum of all dG's for each Lambda step
>> up to that point) up to that point. Not the n-iterations for each. I do
>> separately plot each set of n-iterations for each lambda step to verify if
>> it has converged (e.g. become relatively constant for a long period of time
>> or n-steps). I guess my question is does the plot seem normal and if it is
>> why would it be so far off of all the others I see out including the one in
>> the tutorial for the exact same simulation/calculation?
>> In this case where I ran the tutorial and got the attached plot of dG vs.
>> lambda (which looks nothing like the plot in the tutorial), I did exactly as
>> is told for mutation in a water box with about 20 lambda steps each of which
>> go for 100000 steps (or 100ps). I have verified that it having "converged"
>> for each step. I am running the exact same tutorial right now with the
>> three extra lines you mention at the beginning of your email to see if
>> anything changes. Ill let you know when done if it works for me.
>> Finally, is "doubly-decoupled" the same or something different then a
>> On Apr 30, 2009, at 8:23 AM, Chris Harrison wrote:
>>> fepElecLambdaStart 0.5
>>> fepVdwLambdaEnd 1.0
>>> decouple on
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:52:43 CST