Hi Viktor,
Sorry I can't seem to be of much help, either--the dihedrals are a Fourier
series but I've never heard of fitting them by doing a Fourier transform.
Simple enough to do the least squares optimization of each coefficient. If
you want to try out mdgx, it should be able to work with your structures
and input energies, then produce parameters. I did a lot of work to try
and make it user-friendly, let you explore what new parameters you want to
add, but it's still not a walk in the park. I helped another group awhile
back and while I appreciate their tenacity and interest in the project, it
reinforced my belief that force field development takes as long as you are
willing to give it.
I will continue to try and improve mdgx and the Amber modeling process, but
there are so many different opinions on how to proceed, each of which is
deep-seated and not necessarily compatible with the others. The modeling
programs tend to reflect the opinions of the people who wrote them. I will
suggest that a broader goal of force field development should be higher
fidelity to the quantum benchmarks so that better quantum benchmarks can
shine through the noise that our crude approximations inject. Without a
guiding light and without better molecular mechanics approximations, fights
over our personal opinions are all that we're going to get over the
meanings of those shadows on the wall of the cave.
Dave
On Thu, Nov 16, 2017 at 7:26 PM, Robin Betz <robin.robinbetz.com> wrote:
> Hi Viktor,
>
> Paramfit's energies are all calculated using the full AMBER equation, and
> do not represent the pure dihedral energy profile.
> If you want to fit to this sort of profile only, your fitting procedure
> generates a frcmod suitable for further use. Like you mention though,
> quantum calculations do incorporate nonbonded and other bonded
> contributions to the total energy.
>
> What do you mean by physically nonsense values when conducting a fit with
> Paramfit? If you're seeing poor quality results, try fitting fewer
> variables (for example, force constants only), adding more representative
> or diverse input structures, or adding more dihedral terms. I usually fit
> only force constants and set all phases to 0 or 180 to be consistent with
> the rest of the forcefield.
>
> Sorry the code can be kind of difficult to use. I wrote it many years ago
> and haven't had the time to make it more user-friendly. The usage tutorial
> (
> http://ambermd.org/tutorials/advanced/tutorial23/index.html) is probably
> the best available reference for what the code can do.
>
> Hope this helps,
> Robin
>
> On Wed, Nov 15, 2017 at 7:19 AM, viktor drobot <linux776.gmail.com> wrote:
>
> > I try to understand what energy paramfit returns. If I zeroed Kp for my
> > term in frcmod, performed generation of rotational trajectory with this
> > frcmod and then calculated energies with paramfit - will these energies
> > contain all terms (including vdw and electrostatic) with the only
> > exception? In other words, I want to obtain pure dihedral energy profile
> > for my rotating group and apply FFT to it to extract significant dihedral
> > terms and use them as initial assumptions in future paramfit work.
> >
> > 2017-11-15 15:14 GMT+03:00 viktor drobot <linux776.gmail.com>:
> >
> > > No, I just want to make my fit more physically determined. I expect
> that
> > > parameters what paramfit returns (Kp and phases) should be in agreement
> > > with those obtained via direct FFT decomposition (because of the nature
> > of
> > > dihedral terms - they are the Fourier series, aren't they?). But params
> > > returned by paramfit shows some strange and ugly results.
> > >
> > > 2017-11-15 11:08 GMT+03:00 David Cerutti <dscerutti.gmail.com>:
> > >
> > >> If I understand correctly, you are trying to circumvent what paramfit
> > was
> > >> telling you. Paramfit stipulates that the parameters it fits must
> > bridge
> > >> the gap between the molecular mechanics energies it knows and the
> > quantum
> > >> energy profile of the system. I don't fully understand your
> alternative
> > >> procedure, but Paramfit did take into account the 1:4 interactions and
> > any
> > >> other terms as it fit the dihedral force constants. So, take the
> > results
> > >> from paramfit--that frcmod is producing a reliable model. You can
> also
> > >> use
> > >> the mdgx program to accomplish this--it will produce tighter fits,
> > ceteris
> > >> paribus, and a rich description of how the parameters it optimized
> > operate
> > >> in the training set--but the mechanism of each code is essentially the
> > >> same.
> > >>
> > >> Dave
> > >>
> > >>
> > >> On Tue, Nov 14, 2017 at 4:32 PM, viktor drobot <linux776.gmail.com>
> > >> wrote:
> > >>
> > >> > Hello community!
> > >> > We are working on parameterizing of some molecule
> (6-aminopenicillanic
> > >> > acid) and for these moment we are trying to find right dihedral
> terms
> > >> for
> > >> > hydrogens attached to nitrogen (we're rotating NH2 group). Because
> of
> > >> > tricky nature of the 6-APA we're decided to 'decompose' molecule to
> > set
> > >> of
> > >> > smaller fragments. One of these molecules:
> > C1(C(=O)N(C1=O)[H])N([H])[H]
> > >> > (SMILES)
> > >> > After geomeptry optimization in Gaussian and RESP charge fitting we
> > >> rotated
> > >> > the NH2 group around N-C bond with stepping of 5.0 degrees (full
> > circle)
> > >> > and conducted single-point energies run at Gaussian (according to
> the
> > >> > "paramfit" manual at ambermd.org). So we have potential energy
> > profile
> > >> for
> > >> > NH2 rotating.
> > >> >
> > >> > As we understand it these energy profile contains not only dihedral
> > >> > interactions of hydrogens on amino group but also van der Waals and
> > >> > electrostatics with other atoms in molecule. We assigned basic
> > >> parameters
> > >> > from GAFF field with parmchk2 and decided to fit energy term for
> > H-N-C-C
> > >> > dihedral but not with plain paramfit.
> > >> >
> > >> > As soon as we know the Amber energy is: E = bonds + angles +
> > dihedrals +
> > >> > vdw + el
> > >> > So for the first step we set Kp value for H-N-C-C dihedral term to 0
> > (in
> > >> > hope that in result we obtain all energy terms with these one
> > excepted:
> > >> E =
> > >> > bonds + angles + dihedrals* + vdw + el). For the sake of simplicity
> > >> we're
> > >> > conducted dummy fit of the K value with paramfit and get MM energies
> > >> from
> > >> > the output file (WRITE_ENERGY is set in job control file). After
> that
> > we
> > >> > subtracted these energies (blue line) from quantum data (orange
> line)
> > in
> > >> > hope that now we will have pure dihedral energy dependency on the
> > angle
> > >> of
> > >> > NH2 rotation (yellow line). All these dependencies are on attached
> > plot
> > >> > [image: Встроенное изображение 2]
> > >> >
> > >> > As soon as we obtained pure dihedral energy profile we conducted the
> > FFT
> > >> > transform on these data and obtained Fourier coefficients and phase
> > >> shifts.
> > >> > We are decided to get the first 4 harmonics of Fourier series
> (because
> > >> of
> > >> > high amplitude density) and discarded null coefficient because it's
> > only
> > >> > matter of fitting the K value soon:
> > >> > Np Kp Phase
> > >> > 1 1.06514231 -111.231865
> > >> > 2 0.32663696 137.269480
> > >> > 3 0.13145744 7.266503
> > >> > 4 0.08701518 -77.208482
> > >> > Then we constructed energy profile back with this formula:
> > >> > E(phi) = Kp_1 * (1 + cos(phi + phase_1)) + Kp_2 * (1 + cos(2 * phi +
> > >> > phase_2)) + Kp_3 * (1 + cos(3 * phi + phase_3) + Kp_4 * (1 + cos(4 *
> > >> phi +
> > >> > phase_4))
> > >> > The result is shown on the image above (green line; compare with the
> > >> yellow
> > >> > line - only vertical shift is needed...). Seems like the results
> > >> obtained
> > >> > are in good agreement with our assumptions (brown line; compare with
> > the
> > >> > orange one). For the frcmod file we're inverting signs of phases (as
> > >> soon
> > >> > as Amber requires E(phi) = Vn * (1 + cos(n * phi - gamma_n)))
> > >> >
> > >> > After that we edited frcmod again and now replaced old term for
> > H-N-C-C
> > >> > dihedral with four new with parameters from FFT analysis. But next
> > >> things
> > >> > went bad. After fitting the K value we can't get right energy
> profile
> > as
> > >> > brown line on above plot! Things are messed up[image: Встроенное
> > >> > изображение 3]
> > >> > We are supposing that paramfit needs some transformed angles on its
> > >> input
> > >> > or we are need to apply shift to our phases (already tried pi, -pi
> and
> > >> sign
> > >> > inversion). Other thoughts that some mess are on the stage of
> getting
> > >> pure
> > >> > dihedral energies...
> > >> >
> > >> > Can you help us? We're completely stuck upon it. Blind fit of phases
> > >> and/or
> > >> > Kp's in paramfit gives us physically non-sense values so we want to
> do
> > >> that
> > >> > on concrete basis. Thank you!
> > >> >
> > >> >
> > >> > --
> > >> > С уважением,
> > >> > Дробот Виктор
> > >> >
> > >> > _______________________________________________
> > >> > AMBER mailing list
> > >> > AMBER.ambermd.org
> > >> > http://lists.ambermd.org/mailman/listinfo/amber
> > >> >
> > >> >
> > >> _______________________________________________
> > >> AMBER mailing list
> > >> AMBER.ambermd.org
> > >> http://lists.ambermd.org/mailman/listinfo/amber
> > >>
> > >
> > >
> > >
> > > --
> > > С уважением,
> > > Дробот Виктор
> > >
> >
> >
> >
> > --
> > С уважением,
> > Дробот Виктор
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber
> >
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Nov 16 2017 - 17:00:03 PST
