VMD-L Mailing List
From: Pawel Kedzierski (pawel.kedzierski_at_pwr.edu.pl)
Date: Thu Nov 19 2020 - 03:25:43 CST
- Next message: Raman Preet Singh: "Trajectory analysis without loading all frames"
- Previous message: Daniel Fellner: "Re: charge opt not converged"
- In reply to: Daniel Fellner: "Re: charge opt not converged"
- Next in thread: Lokendra Poudel: "Re: charge opt not converged"
- Reply: Lokendra Poudel: "Re: charge opt not converged"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] [ attachment ]
Dear Lokendra,
As you wrote that your ligand is hydrophobic, the divergence of
distances you observe is rather expected. I guess that most of the atoms
of your ligand which are exposed to direct interactions with water
molecules are aliphatic hydrogens. By the CGenFF parametrization
procedure, these hydrogens should have standard CHARMM atom type HA,
HA1, HA2 or HA3 and a constant charge of +0.09e. So, non-bonding
parameters on these atoms are not optimized at all and they will never
"improve". Instead, you only optimize charges of the buried carbons
which are poorly determined. The numerical effect is that small
improvement in the optimized error function requires relatively big
changes in these charges, confusing the optimizer. Therefore every
optimization may finish in a different local minimum - fortunately, they
should not differ too much in the energy and dipole moment errors. It
may be actually useful for such cases to lower the weights of the
distances in the objective function (in advanced settings).
BTW, if someone select the aliphatic hydrogen type different than the
generic HA, FFTK do not recognize them and these hydrogens appear in the
table of atomic charges to be optimized. They have to be deleted from
the table manually. I consider this an overlooked bug in FFTK which may
confuse inexperienced users.
The apparent problem with hydrophobic interactions is that the
Hartree-Fock method used in calculation of the reference QM interactions
with water molecules by definition do not describe *at all* dispersion
interactions (i.e. "hydrophobic" or "van der Waals type"). Also the
polar interactions are too strong in H-F approximation. The choice of
H-F method and a limited basis set was done on purpose back in the
nineties of the previous century, because it was cheap enough, size
extensive, and the overestimation of interaction energies actually
cancelled well another shortcoming of the calculations in gas phase,
namely the lack of bulk phase polarization. But while the approximation
is reasonable for polar interactions, it is not for hydrophobic ones and
hence it is required to use constant predefined non-bonding parameters
for aliphatic hydrogens.
Quite often, there is no energy minimum to be located on Hartree-Fock
level for the interaction between aliphatic hydrogen and water. In such
cases, the water molecule drifts away much further than the range of
distance deviations considered by FFTK. In Gaussian log files, such
cases could be usually recognized by "Optimization completed on the
basis of negligible forces" convergence message, but a more certain way
is to check the final atom-water distance in VMD after loading the
Gaussian log file. If its larger than about 2.5A, such reference is not
very useful for charge optimization (too large distance, too weak
interaction). If you have enough other reference log files (at least 1
for an atom with optimized charge), I would recommend against using such
results for charge optimization.
If you have several results of charge optimization I would use the one
with lowest energy and dipole moment errors, ignoring the divergent
distances for aliphatic hydrogens. If your distances diverge for polar
atoms, then you may start with different Lennard-Jones parameters for
these atoms.
With greetings,
Pawel
W dniu 18.11.2020 o 23:34, Daniel Fellner pisze:
> Please CC the VMD mailing list when replying.
>
> The distances will vary as much as you let them (with the water shift
> settings) to achieve the best fit. The energies are more important.
>
> *Daniel Fellner BSc(Hons)*
> PhD Candidate
> School of Chemical Sciences
> University of Auckland
> Ph +64211605326
>
>
> On Thu, Nov 19, 2020 at 11:16 AM Lokendra Poudel <poudel_2039_at_tamu.edu
> <mailto:poudel_2039_at_tamu.edu>> wrote:
>
> Hi Daniel,
>
> I checked the energy and distance of individual atoms. The
> energies look ok but distances have highly fluctuated. Do have any
> thought about this?
> Moreover, the total energy and dipole seem fine but not distant.
>
> Thanks
> Lokendra Poudel, PhD
>
> On Wed, Nov 18, 2020 at 3:45 PM Daniel Fellner
> <dfel694_at_aucklanduni.ac.nz <mailto:dfel694_at_aucklanduni.ac.nz>> wrote:
>
> You'll probably rarely get convergence as good as the tutorials.
>
> A few things:
>
> 1. You probably don't need to optimise charges with low
> penalty scores (<10 ~ 20). If you're still having issues with
> validation and you've tried everything else, that's when I'd
> go back and reoptimise these low penalty charges
> 2. Check the output after your runs and you can find out which
> interactions are causing the problem (large QME-MME deviation)
> and weight them lower accordingly. This will improve your fit.
> Also, ensure there are at least as many interactions loaded as
> there are atoms you're trying to optimise.
> 3. You're probably closer to a good solution at the higher
> water shift settings. I'm not sure why this is, but it should
> be reflected in the objective function.
>
>
> *Daniel Fellner BSc(Hons)*
> PhD Candidate
> School of Chemical Sciences
> University of Auckland
> Ph +64211605326
>
>
> On Thu, Nov 19, 2020 at 9:37 AM Gumbart, JC
> <gumbart_at_physics.gatech.edu
> <mailto:gumbart_at_physics.gatech.edu>> wrote:
>
> Hi Lokendra,
>
> I think we usually run SA first and then downhill. But
> it’s quite possible there is no unique solution. Instead,
> you likely have a very rugged energy landscape sprinkled
> with local minima that you are hopping between. This is
> going to be compounded if you have too many charges to
> optimize at once. So you’ll just have to decide what
> interactions are most important to get right (use COLP to
> help).
>
> Best,
> JC
>
>> On Nov 18, 2020, at 11:12 AM, Lokendra Poudel
>> <poudel_2039_at_tamu.edu <mailto:poudel_2039_at_tamu.edu>> wrote:
>>
>> Hi James,
>>
>> I am trying to develop parameter and topology files for a
>> hydrophobic ligand PDBu (76 atoms) using Force Field
>> Toolkit (FFTK). I am in charge optimization step in fftk.
>> The quantum mechanical calculations for water interaction
>> sites run fine. After this, I tried to optimize
>> the charge but the optimized charges are not converged
>> .i.e charges are changing each iteration during optimization.
>> During optimization, I did the following steps
>> 1/ in charge constraints: removed all non-polar Hydrogen
>> and those atoms which have zero penalty score (CGENFF
>> Program)
>> 2/ in QM Target data: excluded those QM water
>> interactions which clashing with other atoms and fly away
>> from an atom.
>> 3/ in advance setting: used simulated annealing with a
>> different combinations such as
>> water shift setting +/-0.4, +/- 1.0, +/- 1.5, +/-2.0 and
>> delta 0.05 or 0.1
>> optimize setting tolerance 0.005 0.001 0.01 distance
>> weight 0.5/1.0 etc
>>
>> After several attempts, my optimized charges are not
>> converged. I spent almost more than two weeks on this but
>> couldn't fix the problem. I would appreciate it if you
>> could help me with it.
>> /
>> /
>> Thanks
>> Lokendra Poudel, PhD
>
- text/x-vcard attachment: pawel_kedzierski.vcf
- Next message: Raman Preet Singh: "Trajectory analysis without loading all frames"
- Previous message: Daniel Fellner: "Re: charge opt not converged"
- In reply to: Daniel Fellner: "Re: charge opt not converged"
- Next in thread: Lokendra Poudel: "Re: charge opt not converged"
- Reply: Lokendra Poudel: "Re: charge opt not converged"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] [ attachment ]