Re: ABF

From: Giacomo Fiorin (giacomo.fiorin_at_gmail.com)
Date: Thu Apr 04 2013 - 16:09:15 CDT

The details of that picture may be different from your case, but I'll
assume everything to be the same for the sake of simplicity. (If I write
something wrong, you're responsible for it because you took a figure out of
context without citing its source).

In the figure, the total number of force measurements changes between the
three sets. Which means that the total simulation time was different.
 That is not a rigorous comparison between the three. A good comparison
would be to run all three simulations for the same number of steps. (In
your previous email, that was 100 million, right?).

If the three simulations (two in your case) run for the same amount of time
and converge to the same PMF, then everything is good. To estimate the
minimum amount of time required, look at which simulation converges faster.

If the three simulations run for the same amount of time, and do not
converge or converge to different PMFs, then the simulation time is too
short, no matter what fullSamples is.

If fullSamples is large enough, the PMF should be reproducible and
independent from fullSamples itself.

In other words, the #1 most important parameter is the simulation length.
 fullSamples is much less important.

You have tested three different values for fullSamples: that's enough. Try
with different simulation lengths, now.

Giacomo

On Thu, Apr 4, 2013 at 1:47 PM, karthik kumar <karthik3327_at_gmail.com> wrote:

> Thanks Giacomo.
>
> while going through literature.. I got one image attached.
>
> fullsamples were 5000, 20000, 80000. FE profiles are quite different with
> changes about 10 kcal/mol..
>
> This can kind of behavior even I do experienced in my simulations.
>
> How do I decide for good value of fullsamples.
>
> Thanks
>
> Karteek Kumar
>
>
> On Thu, Apr 4, 2013 at 9:54 PM, Giacomo Fiorin <giacomo.fiorin_at_gmail.com>wrote:
>
>> Those are two extremes: like in many situations, the best choice lies in
>> the middle.
>>
>> fullSamples is not the number of steps after which convergence is
>> guaranteed on one bin of the PMF (and it shouldn't be).
>>
>> In case I, you clearly don't have a very accurate estimate of the biasing
>> force after fullSamples steps. But if you run for longer, your statistics
>> improves, and eventually you converge.
>>
>> In case II, your estimate of the biasing force is very accurate, but you
>> barely cover the entire range of the reaction coordinate once. Ideally,
>> you want to sweep this range multiple times.
>>
>> If you're concerned that case I is too inaccurate and case II is too
>> slow, how about choosing a fullSamples that is the geometric average of the
>> two cases (i.e. fullSamples = 10000)?
>>
>> Giacomo
>>
>>
>> On Thu, Apr 4, 2013 at 11:53 AM, karthik kumar <karthik3327_at_gmail.com>wrote:
>>
>>> Hi NAMD users,
>>>
>>> I'm very much new to ABF method. I would like to know few things.
>>>
>>>
>>> *case I* : fullsamples 100
>>> width 0.1 A
>>>
>>> reaction coordinate length 10 A
>>>
>>>
>>> so there will be 100 bins approx and it will take roughly
>>> 100*100 timesteps to scan the full reaction coordinate..
>>>
>>> At this point , PMF will no be converged. For the convergence of PMF we
>>> will be extending the trajectory
>>>
>>>
>>> After some timesteps , say 100 million time steps PMF has converged
>>>
>>> *case II* : fullsamples is 1 million timesteps
>>>
>>> width 0.1 A
>>>
>>> Approximately after 100 miliion timesteps , full reaction coordinate
>>> is reached.
>>>
>>> Can I expect to get converged PMF similar to PMF obtained from case I
>>>
>>>
>>>
>>> comparing case I and case II.. which has advantage and why??
>>>
>>>
>>> Thanks,
>>>
>>> Karteek Kumar
>>>
>>
>>
>

This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:21:06 CST