[Zurück]

F. Buchner, C. Jochum, G. Kahl, A. Singraber, C. Dellago:
"Free energy landscapes for dendrimer-like DNAs via neural networks";
Vortrag: CECAM and IUPAP workshop on High density DNA arrays: models, theories and multiscale simulations, Ljubljana (eingeladen); 24.07.2019 - 26.07.2019; in: "CECAM and IUPAP workshop on High density DNA arrays: models, theories and multiscale simulations", CIP - Katalozni zapis o publikaciji Narodna in univerzitetna knjiznica, Ljubljana, (2019), ISBN: 978-961-6104-45-6; S. 45.

Free energy landscapes for dendrimer-like DNAs
via neural networks
Florian Buchner 1 , Clemens Jochum 1 , Gerhard Kahl 2 , Andreas
Singraber 1,3 , Christoph Dellago 3
1 Institute
for Theoretical Physics, TU Wien, Wiedner Hauptstraße 8-10, A-1040 Vienna, Austria
for Theoretical Physics and Center for Computational Material Science (CMS), TU
Wien, Wiedner Hauptstraße 8-10, A-1040 Vienna, Austria
3 Faculty of Physics, University of Vienna, Boltzmanngasse 5, A-1090 Vienna, Austria
e-mail: clemens.jochum@tuwien.ac.at
2 Institute
We consider dendrimer-like DNAs (DL-DNAs, as thoroughly discussed
in the contribution by Natasa Adzić) [1] which we model within a monomeric
bead-spring model: the DNA base-pairs are modeled by charged monomers
and their interactions are chosen to mimic the equilibrium properties of
DNA correctly [2].
The huge number of internal degrees of freedom of this model drastically
limits its use in simulations of concentrated solutions: even with the use of
optimized code packages (such as GROMACS, ESPResSo MD, LAMMPS or
oxDNA) simulations of ensembles of typically 50 (or at most 100 DL-DNAs)
are within reach.
This limitation calls for suitable strategies to cope with this problem:
one standard route is the use of effective potentials obtained by suitably
integrating over the degrees of freedom of the monomeric entities. Here we
represent such effective potentials with artificial neural networks,relying on
their capacity to accurately reproduce complex high-dimensional functions.
Once properly trained with reference data calculated via the monomeric
model, neural networks provide access to effective potential energies and
analytically derived forces, which can then be used in molecular dynamics
simulations. The latter ones can be performed for our system at a fraction
of computational costs as compared to the simulations of the monomeric
model. In this contribution we report about first results of these investiga-
tions. Recently, Luo and his co-workers at Cornell University synthesized
dendrimer-like DNA (DL-DNA) via enzymatic ligation of Y-shaped DNA
building blocks.
[1] C. Jochum, N. Adzić, E. Stiakakis, T. L. Derrien, D. Luo, G. Kahl, and C. N. Likos, Nanoscale,
11, 1604-1617, (2019)
[2] A. Wynveen and C. N. Likos, Soft Matter, 6, 163-171, (2009)

https://publik.tuwien.ac.at/files/publik_280771.pdf

Projektleitung Gerhard Kahl:
DFS