I have posted this query in a Gromacs forum, but got no reply. Can anybody give an advice on the following question?
I am trying to apply the REMD to study the difference between the folding free energy of a 24 residue peptide with two different conformations. As I am short of required number of processors for REMD in explicit water, I am trying my luck with implicit solvent. I have come across many papers that have used implicit solvent in their simulation studies. So, after various trails of temp. distributions, I selected some 15 replicas with the following temp. range:
Repl 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Repl T 250.0 264.0 279.0 295.0 312.0 329.0 348.0 367.0 388.0 410.0 433.0 458.0 484.0 511.0 540.0
Repl
Repl exchange interval: 1000
Here's the result for a trial run of 2ns per replica:
Replica exchange statistics
Repl 999 attempts, 500 odd, 499 even
Repl average probabilities:
Repl 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Repl .34 .34 .21 .27 .39 .41 .40 .37 .41 .42 .41 .43 .46 .44
Repl number of exchanges:
Repl 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Repl 166 155 116 130 184 201 193 185 212 216 200 208 223 225
Repl average number of exchanges:
Repl 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Repl .33 .31 .23 .26 .37 .40 .39 .37 .42 .43 .40 .42 .45 .45
The average exchange ratio for each replica lies between 0.3-0.4. Is it okay ? . As the value between 0.2-0.3 is considered to be good.
Here's the graph for replica_index , replica_temp and PE of replicas.
https://docs.google.com/file/d/0B6ehLXK0eP7sQWRNb010MXNLM28/edit
I want to know whether, I can move forward with the large production runs or not ??
Hello Bharat,
I can't answer your question directly, but give you my two-cents worth. Just eye-balling your temperature, they don't appear to follow an exponential scale, which is what I think you want for REMD (a.k.a. parallel tempering). I have not followed the REMD literature, but I know quite a bit of research/application is done within the Amber community. Below are a few references that might help you answer your question - use them as a start for further literature searches (e.g. who do they site, and who has sited these papers).
Personally, with regard to simply sampling conformations I am not sure what benefits you gain for including temperatures below 310 K (i.e. body temperature) - perhaps it is necessary for your desire of obtaining a free energy landscape.
1. Patriksson, A. & van der Spoel, D. A temperature predictor for parallel tempering simulations. Phys Chem Chem Phys, 2008, 10, 2073-2077
2. Sindhikara, D.; Meng, Y. & Roitberg, A. E. Exchange frequency in replica exchange molecular dynamics. J. Chem. Phys., 128, 024103.
3. Garcia, A. E. Molecular dynamics simulations of protein folding. Methods Mol. Biol., Department of Physics, Appled Physics and Astronomy, 2008, 413, 315-330
Best regards,
Karl
Hello Bharat,
I can't answer your question directly, but give you my two-cents worth. Just eye-balling your temperature, they don't appear to follow an exponential scale, which is what I think you want for REMD (a.k.a. parallel tempering). I have not followed the REMD literature, but I know quite a bit of research/application is done within the Amber community. Below are a few references that might help you answer your question - use them as a start for further literature searches (e.g. who do they site, and who has sited these papers).
Personally, with regard to simply sampling conformations I am not sure what benefits you gain for including temperatures below 310 K (i.e. body temperature) - perhaps it is necessary for your desire of obtaining a free energy landscape.
1. Patriksson, A. & van der Spoel, D. A temperature predictor for parallel tempering simulations. Phys Chem Chem Phys, 2008, 10, 2073-2077
2. Sindhikara, D.; Meng, Y. & Roitberg, A. E. Exchange frequency in replica exchange molecular dynamics. J. Chem. Phys., 128, 024103.
3. Garcia, A. E. Molecular dynamics simulations of protein folding. Methods Mol. Biol., Department of Physics, Appled Physics and Astronomy, 2008, 413, 315-330
Best regards,
Karl
Exchange ratio for each replica between 0.3-0.4 is very good.
It seems to me that the exchange between replicas more often the better, so I do it every 16 step.
When using implicit water, integer charges may give too high energy. Perhaps they should be scaled. For example, in this simulation http://www.biomolecular-modeling.com/Abalone/Protein-folding-Trp-Cage.html they were halved.
One thing to consider is the fact that the outcomes of implicit solvent will be different from explicit solvent. In explicit solvent (which is more common), the potential energy of a system is largely dominated by water, which then dominates the calculation of exchange rates, whereas that is not an issue with your approach. Hence, you're getting a larger exchange acceptance rate than you might in explicit solvent. The 0.2 - 0.3 acceptance rate is sort of an empirical "golden rule," but I don't necessarily think that (1) you should inherently think that such an outcome is "correct" and (2) you should expect exchange rates in explicit solvent (which is where most of the 0.2 - 0.3 rates come from) to be the same in implicit solvent.
Somewhere, buried in the gmx-users archive, is a post from Mark Abraham that discusses REMD exchange acceptance rates and how one might interpret them. I am not an REMD expert, so I won't attempt any hand-waving explanation. But with a bit of Googling, I'm sure you can unearth Mark's nice explanation.
70 jobs on 56 processors will be incredibly inefficient. Not only are you running each replica in serial, but the jobs will fight each other for resources. REMD is very computationally expensive. It won't work well if you don't have the necessary resources.
In that case .. is there any way to scale down the energies , if I use implicit solvent ??
What I mean is that i case of explicit solvent, water can neutralize the charge of the charged residues, which is not possible in case of implicit solvent. That's why the energies are higher in implicit solvent if there are large number of charged residues.
Dear Bharat,
There is a very old practice of reducing charges on the amino acid residues. But I have not heard that it was used with implicit water. However, according to my tests it works (see above link).
Regarding the frequency of exchange. I have no clear evidence, however it can be shown that frequent exchange works at least not worse. 2 ps certainly enough for a single pair. However, we have a lot of pairs and to pass the entire range requires a lot of exchanges. So I try to exchange as often as possible to do it without degrading the overall efficiency. Overall efficiency depends on the implementation in a specific program. It seems to me that the more the rate of exchange diffusion, the better.
Regarding "scaled energies," that's really a force field issue. Some older models used neutralized groups for everything, but I think this is somewhat unphysical. There's a lot of literature out there on the topic, but I'm not well-versed in it.
Regarding exchanges, it is possible to exchange too frequently. Attempting an exchange every 1000 or so steps is fairly standard. Attempting exchanges too frequently will actually inhibit sampling because you're not allowing sufficient time for a given replica to equilibrate within its new surroundings and sample the new ensemble.
Hi,
After a large number of trails I was able to get a good PE overlap with 0.2-0.3 acceptance ratio for all replicas. I tried REMD for 10 replicas with the following temp. -
280.0 300.3 322.0 345.4 370.5 397.3 426.3 457.0 490.1 525.7. I ran each replica for 100ns i.e. 10 microseconds in total. Now, comes the difficult part, where I am having problem. Since I am studying the folding of a peptide (24 residue peptide), I would like to count to number of folding event and free of folding. I used demux.pl file to generate the trajectory of replica. Now, I need to know which replica I should analyze for studying such folding events. I read many papers, but none of them mentions about how they carried out the analysis. I am following this paper for the analysis:-
http://scitation.aip.org/content/aip/journal/jcp/126/1/10.1063/1.2404954
Here's the plot for PE overlap after the final production run...
https://drive.google.com/file/d/0B6ehLXK0eP7sOXJHMXBsd1RkNTg/edit?usp=sharing
Actually, my aim is to study the folding difference between the same peptide with two different conformations. I simulated the first peptide, now for the second peptide should I use the same temp. or should try again for correct set of temp. Since, I have to compare the difference, is it necessary that both the peptide should be simulated with same REMD temperatures. Can anybody cite some other good references that I can follow ....
Looking forward for your responses .....
Since this topic is still of importance, even if the question is old, here's just my cents on exchange probability:
It is not "the more, the better"
What is often unmentioned in my opinion is that a too high exchange rate causes the exchanges to be driven entirely by randomness, instead of e.g. solute conformation. Therefore the sampling will not converge into the temperature specific boltzmann weighted ensemble of conformations, but only to a random collection of structures with uninterpretable probabilities. To ever overcome this randomness, the sampling would have to be extensively elongated so that by many exchanges this noise would finally average out. But it would be still in question, if a larger protein could ever have enough time to fold, before being heated up too much again. Thats's why the optimal exchange rate is usually mentioned as about 20% - Random enough for convergence (fast temperature traverse to temperature of interest), specific enough to produce meaningful results (correct weights between states -> free energy difference).
- Badges
- Science topic