From: Pawel Kedzierski (pawel.kedzierski_at_pwr.edu.pl)
Date: Thu Apr 22 2021 - 09:13:08 CDT

W dniu 21.04.2021 o 21:33, Dr. Eddie pisze:
> Hello,
> This reply was very helpful! I was hoping I could ask for a few more
> details.
> 1) Is there a place that explains how to read the results of the COLP?
> I'm not exactly sure what it is showing me or what I should be looking
> for. What does a "good fit" look like vs what does a "bad fit" look
> like and what should you do?
COLP shows you how the errors change during optimization. Ideally they
should converge to 0 at the right side of the plot.
> 2) If there are atom(s) (Carbon) which you cannot find a HF min
> (despite trying dozens of different parameters) with gaussian, do you
> leave the file(s) out or set the weight(s) to 0?

Zero weight is like discarding the file. You may get rid of some such
results if the interactions are too weak to be meaningful anyway.
However, the number of Gaussian files representing interactions should
be at least equal the number of charges you optimize. Otherwise, the
number of fitted parameters (charges) is bigger than the number of data
points (gaussian interactions) and your fit will be underdetermined,
resulting in divergence.

A possible reason why there is no minimum could be that the atom was
placed on the donors list when it really should be an acceptor or vice
versa. Check the atom Mulliken charges at the end of the Gaussian log
file produced at geometry optimization step. If the partial charge is
negative, the atom should be an acceptor, if positive, it will work as a
donor.

> 3) How do you know your fit is "good"? I've run fits that end up
> giving very different charges than my original (from charmm-gui) and
> it seems each subsequent attempt gives every different results with
> some atoms near +/-1 and some near 0 which I would assume is incorrect.

I have never used charmm-gui charges as starting ones and I think they
may be result of some heuristic guessing. I would advice to start the
optimization from the atomic Mulliken charges. Please also check the
upper and lower boundaries, they should allow variation both on + and on
- from the starting charge. Also, the divergent results may be due to
underdetermined fit.

What is a "good fit" is relative. I am usually satisfied if I get the
final energy errors within +/- 1kcal/mole range in the COLP graphs.

With regards,

Paweł Kędzierski

>
> Any help would be greatly appreciated!
> Eddie
>
> On Thu, Nov 19, 2020 at 4:47 AM Pawel Kedzierski
> <pawel.kedzierski_at_pwr.edu.pl <mailto:pawel.kedzierski_at_pwr.edu.pl>> wrote:
>
> 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
>>
>
>
>
> --
> Eddie