Re: [AMBER] When I use fitted parameters as initial parameters in paramfit, I am getting different results?

From: David Cerutti <dscerutti.gmail.com>
Date: Fri, 6 Mar 2020 14:10:49 -0500

I can't advise on what exactly ParamFit is doing, but your hunch about the
"K" parameter that raises and lowers the waterline for all energies sounds
right. Let me take this opportunity, if you are just getting started with
your project in ParamFit, to point you to mdgx. There is another tutorial
on developing parameters with mdgx here:

http://ambermd.org/tutorials/advanced/tutorial32/index.php

This is the program that we will support going forward, and which will be
receiving updates in Amber20 and beyond. You don't need to do the IPolQ
charges to use the bonded parameter fitting module--whatever you think the
charges should be, it'll take that electrostatic surface as part of its MM
energy estimates.

The mdgx development cycle MIGHT also give you different answers if you
start from different parameter settings. This will be the case if you are
using certain types of regularization (the fitted parameters are tethered
towards an absolute value like zero, or towards the values in the initial
guess). The mdgx cycle is designed to have parameters evolve by adding
more and more fitting data to the training set, a learning process whereby
we apply each successive model to manipulate molecules via energy
minimization and then test whether the true QM energy surface is reflected
in the results.

The first advantage you'll find with mdgx is that you don't have to
hand-prune each of your structures. It creates things that are by
definition energy minima on the MM surface of the initial guess model
(i.e. an antechamber-interpolated GAFF residue prepi file), and
automatically prunes things that fail to meet certain sanity checks. The
tutorial above uses the &configs module to do this, but a more detailed
tutorial on that one module is here:

http://ambermd.org/tutorials/basic/tutorial6/index.php

The second clear advantage is that mdgx doesn't bother you with fitting the
K parameter--it has the same kind of math, but it's all folded into the
least squares problem so you automatically get the best possible K for
every molecule in your training set. It tells you what those were in the
output, if you're interested, but there's not much to read in the tea
leaves. I've agonized over the math and I'm certain that it is the correct
way to go about this. Paramfit has the right idea--other modeling
workflows miss it entirely--but with mdgx you have peace of mind that there
is no better way to make the relative MM energies match the QM surface.

As always, parameter fitting will take as long as you give it. mdgx is
written to smooth out a lot of the process, and there's more I can do, but
this is all geared towards making it easier to manipulate a large amount of
data and synthesize it into a coherent MM model. You will have to decide
at what point you are satisfied with the convergence. In my own work, I'm
starting to ask fundamental questions like whether it is possible to get
better MM models with customized partial charges on one chemical group in
different contexts (e.g. the natural amino acids' backbone), or whether
such improvements are just a mirage. With 64 or more conformations of each
amino acid in the training data and proper care to avoid undesirable
features like hydrogen bonding between side chain and backbone atoms in the
structures, it's becoming apparent that unique sets of charges for each
amino acid are a negligible benefit relative to a single, consensus charge
set. Plus, the consensus charge set makes it easier to get torsion
parameters (a la ParamFit, mdgx &param module) to put the keystone into the
overall energy surface.

Dave



On Fri, Mar 6, 2020 at 7:12 AM Erdem Yeler <erdemyeler.gmail.com> wrote:

> Hello amberers,
> Let's say I am optimizing X dihedral by using paramfit. After running
> paramfit I obtained fitted.frcmod. And fit_output_energy.out file shows
> that: molecular mechanic energies obtained by using "fitted.frcmod" file
> are, very close to quantum mechanic calculations. No problem so far,
> everything happened as we expected.
> But, when I use fitted.frcmod file as initial parameters and running
> paramfit with these parameters (quantum_A.dat, prmtop and mdcrd files exc..
> are the same, only inital frcmod file was changed), this time
> fit(2)_output_energy.out file says that : my "initial parameters" are very
> far away from quantum mechanic calculations (even if I referenced "relative
> energies" (relative energy on the chart = energy-minimum energy, so there
> is no need to be use Ka)). But how can this happen? Paramfit found some
> parameters and it says these parameters are good but when I re-use them as
> initial, this time how can exactly the same parameters could be "bad
> parameters" as initial? I am really confused. Am I doing something wrong?
> or this result is normal? (I tested many many times).
> Thank you so much...
> _______________________________________________
> 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 Fri Mar 06 2020 - 11:30:02 PST
Custom Search