[gmx-users] some questions about umbrella sampling

[chris.neale at utoronto.ca]

> I am doing umbrella sampling to calculate energy profile of an drug  
> passing through a ion channel by GROMACS3.3.

why not gmx 4?

> In order to get right result, I need to choose a reasonable force  
> constant in my simulation, right? I know that, a reasonable force  
> constant should produce enough "overlap" to reduce WHAM errors, But  
> I'm wondering "how much overlap" are reasonable overlap?Does there  
> any rules I can use to judge whether my "overlap" is enough?

I aim for 20%, but that's because I also conduct REMD-like exchanges.  
To be sure, you can do twice as many (or otherwise increase the  
overlap) and if the profile doesn't change then you had enough to  
begin with. I realize that this isn't a fantastically efficient  
solution, but I don't know of a better one.

> My second question is that, how to choose "number of bins", when I  
> using g_wham command? Does it decided by the numbers of windows I  
> used? I found that, different numbers will produce slightly  
> different results (a difference of 0.5kcal/mol in my condition).  
> What does it mean? Does it mean that, my simulation is not enough?

It could mean a lot of things. Based on what you mention below, I  
suspect that you are using too bins that are too small when the energy  
changes. WHAM works with probability histograms, and so you need  
enough counts per bin that if you were to run the entire simulation  
again the difference would be a small fraction of the average.

> the third question: I found someone say that, to get results in a  
> good quality, one must make sure that, "individual histograms should  
> look smooth". What does it mean? Does it mean that,the distribution  
> of the positions of the drug in one window should be "smooth"

Yes
> or should opproximately be an Gauss distribution

No. They could easily be bimodal.

> (I found that, whether it is smooth is decided by the number of  
> bins, if a very large number of bins is used, it is not smooth)?

See my comment above regarding probability histograms

> The last question: if my simulation produced enough "overlap", can I  
> believe that, my simulation is "right" theoretically?

No. Your sampling within one or many or all umbrellas could still be  
trapped -- hence the reason that I do equilibrium exchange. Consider  
this:

    |------ Sidechain A
    |
---|       <drug in umbrella N>

versus:

---|       <drug in umbrella N>
    |
    |------ Sidechain A

Regular US does not offer an efficient mechanism for the sidechain A  
to switch sides of the drug while the drug is restrained to a given  
position when the drug is blocking the rotamerization of that  
sidechain. Allowing the constituent ensembles to interconvert using  
e.g. REMD alleviated this sampling deficiency.

Chris.