Deriving bonded parameters with mdgx: Introduction
Parameter development need not be a black box. The key is to have tools
that speak in terms of coordinates and energies, recognize that the molecular
mechanics energy is merely a relative quantity, treat the physical phases of
matter correctly, and understand that the Newtonian parameters are prone to
over-fitting. There's some finesse to the process, but machines can take away
a huge amount of the tedium and handle much bigger problems (in sheer terms of
the number of parameters and size of the data sets) than a human being fitting
one parameter at a time. This is the right way to approach the problem:
parameters do not exist in isolation‐not at the level of approximation we
are taking wth molecular mechanics. Moreover, a sophisticated apparatus for
collecting and interpreting the data can be enhanced by a rich set of
post-processing tools to interpret the resulting parameters, place them in the
context of the data, probe their vulnerabilities to prevent over-fitting, and
ensure that the models will function in the intended conditions. This is what
mdgx aims to offer.
Stage 1: The molecule in many, many guises
Stage 2: Fitting parameters
Stage 3: Iterate
Stage 4: Use the new parameters
Proper credit should be given to ParamFit and its creators
Robin Betz and Ross Walker. A number of the features of the mdgx
force field tools are inspired by that project, particularly the methods for
generating conformations of the molecules of interest and dealing with data in
batches. In building the mdgx force field tools, I have tried to
extend and refine a lot of the process in
the ParamFit
tutorial. I hope that the following tutorial will convince you that
mdgx sets a high standard in this field, and that the wisdom of
numerous experts has been distilled into the tools.
This bonded parameter development is ideally used after fitting charges with
the IPolQ
scheme, but if you are happy with the charges you've got it can also be
taken standalone. Like the IPolQ that preceded it, this is an iterative
process, but in this case the iterations are meant to relieve human operators
from the burden of having to intercede to eliminate spurious parameters. We
feel that the automated approach is in fact better for this purpose in the way
that it fixes what can be widespread and subtle errors: humans cannot easily
identify these errors, much less correct them, but new parameter sets are rife
with them. The automated tuning is also just as effective at taking care of
the obvious catastrophic problems that human intuition and meticulous care have
solved in the past.
The parameter development scheme is detailed in these papers:
- D.S. Cerutti, J.E. Rice, W.C. Swope, and D.A. Case. (2013) "Derivation of
Fixed Partial Charges for Amino Acids Accommodating a Specific Water Model and
Implicit Polarization." J. Phys. Chem. B 117: 2328-2338.
link
- D.S. Cerutti, W.C. Swope, J.E. Rice, and D.A. Case. (2014) "ff14ipq: A
Self-Consistent Force Field for Condensed-Phase Simulations of Proteins."
J. Chem. Theory Comput. 10: 4515-4534.
link
- K.T. Debiec, D.S. Cerutti, L.R. Baker, A.M. Gronenborn, D.A. Case, and L.T.
Chong. (2016) "Further along the Road Less Traveled: AMBER ff15ipq, an
Original Protein Force Field Built on a Self-Consistent Physical Model."
J. Chem. Theory Comput. 12: 3926-3947.
link
|
