Re: [AMBER] Questions about TI

From: <>
Date: Wed, 18 Nov 2015 09:35:58 +0000

I think a few things here should be clarified.

The first question to answer is whether the transformation is an "absolute" one or a relative one. Those two should be clearly distinguished in this discussion.

For absolute free energy a "one-step" protocol seems to be working fine. By "one-step" I mean the simultaneous transformations of all force field terms at the same time with the same lambda value (in Gromacs-speak they all follow the same lambda path). I am not sure what a "three-step" protocol in this context would mean. What is done in practice though is to separate the electrostatic transformation from the vdW transformation which would make it a "two-step" protocol in the definition above. Setup for absolute transformation is trivial with AMBER. For the two-step protocol the user would necessarily need to set crgmask to the TI mask.

When I say that the one-step protocol fails I mean this in the context of _relative_ transformations. Personally, I have found this to produce wildly wrong results in case of rather small organic molecules. A two-step protocol, i.e. transform charges from V0 to V1 first and next transformation of vdW+bonded(*) terms (or vice versa depending on which state has the vanishing atoms), seems to be particularly useful. This does not, however, work when both states have vanishing/appearing atoms. In this case a three-step protocol can come in use. The practical question, however, is how to deal with the charge transformation. I am not sure why setting the charge of the whole molecule would mean that convergence is slow. After all it is a simple linear transformation and the gradients tend to be rather smooth. The problem with only setting parts of the charges to zero is that this means that you will most likely create a fractional total charge on the molecule. While the PME code principally neutralises the simulation system, it still feels unnatural.

(*) One could also combine the bonded terms with the electrostatics but this is considerably more difficult to set up with leap and would involve a post-processing step of the parmtop (parmed is your friend here).

The number of windows in a TI calculation will necessarily depend on the shape of the resulting gradient. If the gradient varies too much in a region you will probably want to add additional lambda points (and that's the advantage of TI). For convergence you will need to look at the actual data. Prescribing a certain amount of simulation time is not too useful as this may be very much system-dependent and also lambda-dependent. The crucial point here is to look at the correlation which can be very much lambda-dependent too.

Thanks for pointing out these shortcomings from the tutorial. I will see to get them fixed.

From: []
Sent: 18 November 2015 08:17
To: AMBER Mailing List
Subject: Re: [AMBER] Questions about TI

Hi Pengfei,

On Thu, Nov 12, 2015 at 8:45 AM, lipengfei_mail <>

> Dear all,
> When I do TI calculations, I encountered some questions as below:
> 1.In the "Final Remarks"part of AMBER TI Tutorial A9, there is a word
> that"The author of this tutorial has found that the one¡Vstep protocol may
> fail in some circumstances".
> What does it mean? Could you give me some examples in which one-step
> protocol may fail?

   Are there some rules about how to choose one-step or three-step protocol?

Here are some one-step protocol papers:

J. Chem. Phys. 127, 214108 (2007); doi: 10.1063/1.2799191


JCC. DOI: 10.1002/jcc.21909

We just have another accepted article discussed about this topic. I am
more than happy to keep you posted when it is available online (hopefully
in 2-3 weeks).

In brief, if computational resources are not a limitation, three-step TI is
recommended since it should converge better. Vice versa. If you run a
three-step/one-step TI, you should make a plot of final dG along MD time
frame. This simple method could make sure the dG results are converged.
This is particularly important for one-step TI since the one-step TI might
not converged as well as three-step TI.

> 2.After I accomplished this tutorial, I got much the same results as the
> tutorial gave.
> But I think that for ligands, £GGligands (solution)= -38.99 kcal/mol,
> this seems unreasonable.
> As a rough estimation, the solvation free energy for benzene and phenol
> are 0.59 and -4.68 kcal/mol,respectively.
> The energy variation from benzene to phenol in gas phase is about
> -23.22 kcal/mol.
> Then the total energy variation from benzene to phenol in the solvation
> can be estamitated as dG=-0.59-4.68 -23.22=-28.49 kcal/mol,
> which has large difference with the result given out by the A9
> tutorial (-38.99 kcal/mol ).

How many windows and how long the simulation did you run? To get converged
and accurate results, longer TI simulation (usually > 1-3 ns) and and more
windows (>6-11 windows) might be required. For such a 'light' system, I
would probably run 21 windows or more.

> 3.In the course of decharge, if only decharge the O1 and H6 of benzene
> where crgmask = ':1.O1,H6',
> the rest atoms of benzene for MD will be charged which is obviously
> unreasonable.
> One way is to set the crgmask=¡¥¡G1' . But if 1 refers to a large ligand,
> then there is less overlap between the initial and final state.

crgmask = ':1.O1,H6' is correct. If we choose crgmask=¡¥¡G1', the TI
transformation region is too large and can hardly converge.

   Could any advice be given out about the two choices as above or some
> better ways instead?
> Best,
> Thank you very much !
> --
> -------------------------------------------------------------------------
> Pengfei Li & Fengjiao Liu
> _______________________________________________
> AMBER mailing list


Pin-Chih Su (Henry Su)


Center for Pharmaceutical Biotechnology (MC 870)

College of Pharmacy, University of Illinois at Chicago

900 South Ashland Avenue, Room 1052

Chicago, IL 60607-7173

office 312-996-5388

fax 312-413-9303
AMBER mailing list
AMBER mailing list
Received on Wed Nov 18 2015 - 02:00:07 PST
Custom Search