7
$\begingroup$

Below is the free energy profile obtained after a histogram analysis of an umbrella sampling simulation. Generally, one is interested in the free energy difference (highest energy minus lowest energy). Let's assume that we have a good assumption where the lowest and highest are on the reaction coordinate.

  • Should one run an umbrella simulation on the entire reaction coordinate?
  • Isn't sampling the lowest and highest energy region enough to derive the free energy difference?

enter image description here

Image source: http://www.mdtutorials.com/gmx/umbrella/07_analysis.html

Improve this question
$\endgroup$
0

1 Answer 1

Reset to default
7
$\begingroup$

The short answer is: yes you need to sample the whole reaction coordinate if you want accurate free energy values.

If you are simply interested in qualitative behaviour, then you might (key word being might) get away with fewer windows (but they still need to be distributed fairly uniformly), but I would only recommend this for a dry/test run and never for a production run. Otherwise, you need to be absolutely sure you get good phase space overlap between all distributions, or you will either get huge deviations from the correct free energy values or your analysis will crash. The reason for this is quite simple: the method that umbrella sampling uses for analysis, the weighted histogram analysis method (WHAM), relies on data from all windows in order to calculate relative populations. If you have gaps in your phase space, this in turn creates gaps in the PMF (potential of mean force) and WHAM cannot calculate the relative heights, meaning that it will either converge to a strange result or not converge at all.

As a final note, I would say that one of the biggest disadvantages of umbrella sampling (apart from the cost) is the fact that the efficiency is directly dependent on the shape of the PMF, bearing in mind that the WHAM estimator is exponentially sensitive to the phase space overlap. For example, if you have a very high kinetic barrier, you will need to concentrate more windows around this point in order to get some overlap. This means that while you don't "need to know the answer to get the answer", you do "need to know the answer to be certain in your convergence". In other words, you might have to add extra windows after your analysis to make sure your distributions overlap and that the extra windows don't change your results, because they shouldn't.

Edit: I just noticed that the artifact at around 0.8 nm can be readily explained by the lack of phase space overlap, so this gives you quite a good idea of what to expect when you have insufficient data. The authors of the tutorial also seem to acknowledge that and they also state that you need to make sure you always have adequate overlap.

Improve this answer
$\endgroup$
5
  • $\begingroup$ I've just made the place where WHAM is defined more easy to find, and defined PMF. $\endgroup$
    – Nike Dattani
    Jul 25, 2020 at 0:59
  • $\begingroup$ Thank you for this very helpful answer. Three more ones. 1) Can you please elaborate about the concept of convergence? 2) Will oversampling (if that exist) change result. For example, in the profile above, we have three almost overlapping windows at the beginning, does that have any effect? 3) The last one is about the little drop in energy in the beginning. I have noticed it on many profiles. What is the physical interpretation? $\endgroup$
    – BND
    Jul 25, 2020 at 11:51
  • $\begingroup$ What exactly would you like to know about convergence, so that I know how to modify my answer? $\endgroup$
    – Godzilla
    Jul 25, 2020 at 12:08
  • $\begingroup$ @Godzilla123 I had to do a bit of reading. Correct me if wrong, the converge is about the profile of the PMF. It is fully converged when there is no gap in the profile (like the one at 0.8 nm). $\endgroup$
    – BND
    Jul 27, 2020 at 9:58
  • $\begingroup$ A gap in the profile certainly indicates that something might be wrong but it's always difficult to say for sure. To my knowledge, the best metric is looking at the overlap between all distributions you get - all neighbouring distributions should have some degree of overlap. And you can see that both the profile and the overlap have a discontinuity at 0.8 nm. Also, you can never "oversample" - if your answer changes with additional windows, this means that it was never converged in the first place. $\endgroup$
    – Godzilla
    Jul 27, 2020 at 11:41

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .