wehs7661 / lambda_metad_questions Goto Github PK

This issue is for the discussion of the second problem, which is mainly about the strategies for deciding simulation parameters for 2D alchemical metadynamics for the CB8-G3 system.

Briefly, the questions described in the corresponding jupyter notebook (any commit after 545046e) can be summarized as follows:

• Are the strategies I used for deciding metadynamics parameters appropriate?
• Can the bias be regarded as quasi-stationary if the Gaussian height in the unphysical region is still very high?
• What are the general suggestions you would give about deciding the parameters for a wall potential? How large should the energy penalty typically be?
• What should I modify in my protocol for free energy calculations to consider the reweighting of the wall potential?
• Is there a way to prohibit the exploration of regions such as $\lambda=40$ and $n=0$?

@wehs7661 @mrshirts as usual, questions via GH issues :-)

I am focusing on Problem 1 now. Didn't go through the full analysis yet, but I did some quick check on the results and I agree, they are different and this is unexpected. I mean:

1. Different from a reference result might imply some theoretical error, unexpected but still possible.
2. Different between each other is a non-sense since (I checked) both are extensively sampling all lambdas and all angles).

We are in case 2 now.

Little doubt: I am not very familiar with the WL implemented in GROMACS, so I cannot decipher this in the log file:

   nstexpanded                    = 10
   lmc-stats                      = wang-landau
   lmc-move                       = metropolized-gibbs
   lmc-weights-equil              = wl-delta
   weight-equil-wl-delta          = 0.001
   lmc-seed                       = 1000
   mc-temperature                 = 298
   lmc-repeats                    = 1
   lmc-gibbsdelta                 = -1
   lmc-forced-nstart              = 0
   symmetrized-transition-matrix  = true
   nst-transition-matrix          = 100000
   mininum-var-min                = 100
   weight-c-range                 = 0
   wl-scale                       = 0.8
   wl-ratio                       = 0.8
   init-wl-delta                  = 0.5
   wl-oneovert                    = false

Does it imply you are using WL as well? I though you were expecting METAD to help sampling lambda, and thus disabled WL (maybe this is what you did... can you confirm?). If you are using WL + METAD, we should take that into account when you do the analysis (it's possible). A quick check in the gromacs manual tells me that lmc-weights-equil = wl-delta implies that you are changing the weights until some point. If the weights at that point (and then used in the following part of the simulation) are different in the two simulations, this would explain your weird behavior. By knowing these weights (if they are frozen for a significant part of the simulation) we could easily correct the result.

If instead the input implies that you are not using WL, forgive my ignorance on GROMACS input, I will continue to hunt for the reason of the discrepancy somewhere else