[gmx-users] Umbrella Sampling

chris.neale at utoronto.ca chris.neale at utoronto.ca
Fri Apr 1 17:44:42 CEST 2011

1. how do you know that data is "not good"? It is not good science to  
remove data just because it doesn't match with your expected results.  
For now, there is no compelling reason to remove that data.

2. My first step in checking the convergence is to block average the  
time. If you have 10 ns total per umbrella, then extract the first 0-1  
ns  and make a PMF based on that. Then extract the 1-2 ns data and  
make a PMF based on that. Continue until you make a PMF based on the  
9-10 ns data. Now check and see if this PMF is drifting over time. You  
can do this by looking at the PMFs or by integrating the bound state  
and plotting the binding free energy as a function of equilibration  
time. If the binding free energy is still drifiting as a function of  
increasing equilibration time then you are certainly unconverged. Note  
that there will always be noise in this measure, but what you are  
looking for is unidirectional drift in the value. -- One could write a  
textbook on this topic. i suggest that you seek out alternative  
resources to understand how to check convergence properly.


Cheers Chris

Is the best way to check for convergence; to keep adding in more
histograms until the curves converge. Also your comment  'don't remove
any data', do you mean to keep histograms that are not so good.

chris.neale at utoronto.ca wrote:
> your comment:
> which should be centred around 0.80nm
> is flawed, as i mentioned earlier. also, it is not g_wham that is
> sensitive but the convergence and sampling of phase space that is
> sensitive. don`t remove any data. do evaluate your convergence.
> without convergence measures, a pmf is worse than useless.
> chris.
> -- original message --
> Cheers Chris
> If I remove the red histogram (the first of the wider distributions),
> which should be centred around 0.80nm but is actually centred around
> 0.78 nm; and add in some more histograms with higher force constants the
> profile changes slightly. It seems that  g_wham is very sensitive to
> these subtleties. How do I know which curve is correct? I have about six
> such curves that differ slightly in this manner.
> Gavin
> chris.neale at utoronto.ca wrote:
> [Hide Quoted Text]
> looks fine to me, no need to do that extra sampling that I suggested
> since it appears that you already did this -- benefits of seeing real
> data ;). If you want to understand why your histograms are not always
> centered at r0 (note that this is just fine) then you should read more
> about US, WHAM, and how to bias/debias the data for US (I am sure that
> there are textbooks around that explain this). The only case in which
> all of your histograms will be centered at their respective r0 is when
> the underlying PMF is exactly flat.
> Chris.
> -- original message --
> Hi Chris many thanks again for the advise. I have, or at least I thought
> have sampled my barrier region to death, but as I say some histograms
> may not be centred around r0. I will proceed with what you suggest.
> Please find attached a picture of the histograms, the corresponding
> profile, and a sample mdp file that I use.

More information about the gromacs.org_gmx-users mailing list