I am looking for ways to estimate whether my simulation has sampled for long enough and I came across the subspace overlap calculation (Hess, 2002) which compares the covariance matrices of two trajectories. If the two matrices are identical the overlap is 1 and if they are orthogonal the overlap value is 0. I've seen some publications that split up their simulation into two or three parts and compute the overlap. They also say that if they obtain a high overlap it means that it is unlikely that there are unexplored regions. Others say that a high overlap means you are sampling from the same energy minima and thus need to run longer simulations. So I am not very clear on how to interpret this.
In case this is useful, I am simulating a wild type and a point mutant of a 35 residue intrinsically disordered protein for 2 microseconds in GROMACS. Each simulation was run twice with different initial velocities. The simulations are in explicit water.
Thanks!
The subspace overlap can be explored by computing the Principal Component and find how the eigenvectors contribute to the overall variance of your trajectories.
The validity of the subspace overlap can be identified using Z-score analysis, in which individual pairwise comparison of models is needed. As for your system, you need to perform this to both the wild-type and mutant trajectories. Note that overlap means that simulation shows conserved dynamic behaviour, which may/may not be the characterisitcs of mentioned two system. However, you will find that replicate system will not show idenitical dot product overlap, which is becuase you have used different initial velocity.
For simplifying your data, you can divide the simulation, say 0ns-10ns; 10ns-50ns; and 50ns-100ns (or according to your resource availability). This will help in discussing how the dynamic variation is observed for wild-type and mutant during the course of simulation. Again, the dot products of inidividual system (scattering) must be conserved to a relatively "smaller area" in the eigenvector correlation plot (towards the final portion of your trajectory divide-up); which will indicate the convergence of your individual trajectories.
Thank you very much for your answer Rajiv that is very helpful.
I am just still not very clear on the second part of your answer with the dot products in the eigenvector correlation plot being conserved. Could you point me to an example in the literature that explains it or describes this technique?
Thanks again!
You can check one of my contribution in Plos one - Use of a Small Peptide Fragment as an Inhibitor of Insulin Fibrillation Process: A Study by High and Low Resolution Spectroscopy
- Badges
- Science topic