I have simulated a protein and a protein ligand simulation for 100 ns each. Now I want to know if the simulation is converged. For that I have learnt that block averaging is a good method. Can anyone please tell me how should I proceed with it. The salt bridge distance and secondary structure elements are conclusive parameters of my study. Please help.
Grossfield and Zuckermann (2009) Annu Rep Comput Chem5: 23–48.
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2865156/
---
But I think that molecular dynamics simulation is a controlled explosion. Can a controlled explosion converge? I sense a contradiction in my question.
If you find out that the convergence has not been achieved in your MD simulation, what will you do? Will you extend the existing MD simulation or run another one from different starting coordinates?
If you find out that the convergence has been achieved, does it mean that the system finds itself in a native conformation? And if yes, what if there are other native conformations?
I understand your point :) It is really difficult to represent it and most importantly convergence of a simulation in which we are studying an IDP. However, Convergence indicates that sampling is almost complete or left with few sub spaces; so we should not expect that system is 100% converged. Anyhow, I have studied a lot of reports in which people say that their system is "converged" where, most probably, they mean least variance in a conclusive parameter.
In particuar, there is nothing wrong in limiting the time scale to a certain time point, as you can find many of the MD simulation paper to have conformational sampling within range of 100-1000 ns. The extent of simulation time mainly depends on your point of interest and availability of computational resource in your lab.
The point of finding convergence is generally used by researcher to ascertain that the time scale adopted for MD simulation is sufficient enough to answer the "problem statement". Some of the common method used for assessing the convergence of your MD simulation include:
(i) RMSD convergence - Check if the final 50ns of your MD simulation is having a RMSD deviation ~2Å.
(ii) PCA analysis - Identify the correlated data points for first 2/3 eigen vectors. If the points are not scattered over a larger area, they denote convergence of your MD simulation.
(iii) Statistical inefficiency is another method, in which you can use avergage blocking method. Here you need to refer to the variance of series/blocks (of defined length) in average method. Again you can use the distance, RMSD, variation in secondary structure percentage as variable points.
(i) Sample-averaged measurement (SAM) and (ii) cumulative average measurement (CAM) are two effective methods that are useful in validating the macroscopic values from MD simulation (due to chaotic nature of molecular collisions and the random path of trajectory). Often these macroscopic values deviates from the average values, which in turn gives statistical errors.
Thus, SAM and CAM are useful methods in finding the convergence in liquid simulation.
@Rajiv Kumar Kar
Thank you very much . This was indeed very helpful. I am thinking to do PCA first and check the results. Then definityl I will study Block average, SAM and CAM. Can you provide a link for paper or any capter/book etc. describing SAM and CAM.
Thank you
@Martin Please elaborate. You mean, for example, cumulative average of RMSD and then?
@Manmeet: . . . and then I would plot the cumulative average vs MD snapshots; once the curve flattens and stays flat for the remainder of the MD simulation, one can assume that the convergence has been reached. Although, the past is not the predictor of the future [which is why the technical analysis of stock prices does not work (on its own)].
If your MD simulations "converge" in this way very fast I would consider running multiple short MD simulations (e.g., 10 x 10 ns; 5 x 20 ns) starting from different initial coordinates rather than running one long MD simulation.
@Martin Yes the concept of multiple short MD is quite good but as I mentioned I have already completed one 200 ns MD. So for the time being I will put effort into that. Thanx a lot for clear explanation
@Rajiv Kumar Kar
Sir I have performed PCA and obtained following graph for autocorrelation between first 2 eigen vectors. Can we comment if it indicates convergence or enough sampling? well for instance, I have just collected eigen vectors for 25 ns.
- Badges
- Science topic