Article

pubs.acs.org/jcim

Generalized Potential Energy Finite Elements for Modeling Molecular

Nanostructures
Stavros Chatzieleftheriou,† Matthew R. Adendorff,‡ and Nikos D. Lagaros*,†
†
Institute of Structural Analysis & Antiseismic Research, Department of Structural Engineering, School of Civil Engineering, National
Technical University of Athens, 9 Heroon Polytechniou Street, Zografou Campus, GR-15780 Athens, Greece
‡
Laboratory of Computational Biology & Biophysics, Department of Biological Engineering, Massachusetts Institute of Technology,
77 Massachusetts Avenue, Cambridge, Massachusetts 02139-4307, United States
*
S Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: The potential energy of molecules and

nanostructures is commonly calculated in the molecular
Downloaded via UNIV DA CORUNA on February 23, 2022 at 16:08:11 (UTC).

mechanics formalism by superimposing bonded and non-

bonded atomic energy terms, i.e. bonds between two atoms,
bond angles involving three atoms, dihedral angles involving
four atoms, nonbonded terms expressing the Coulomb and
Lennard-Jones interactions, etc. In this work a new,
generalized numerical simulation is presented for studying
the mechanical behavior of three-dimensional nanostructures
at the atomic scale. The energy gradient and Hessian matrix of
such assemblies are usually computed numerically; a potential energy finite element model is proposed herein where these two
components are expressed analytically. In particular, generalized finite elements are developed that express the interactions
among atoms in a manner equivalent to that invoked in simulations performed based on the molecular dynamics method. Thus,
the global tangent stiffness matrix for any nanostructure is formed as an assembly of the generalized finite elements and is directly
equivalent to the Hessian matrix of the potential energy. The advantages of the proposed model are identified in terms of both
accuracy and computational efficiency. In the case of popular force fields (e.g., CHARMM), the computation of the Hessian
matrix by implementing the proposed method is of the same order as that of the gradient. This analysis can be used to minimize
the potential energy of molecular systems under nodal loads in order to derive constitutive laws for molecular systems where the
entropy and solvent effects are neglected and can be approximated as solids, such as double stranded DNA nanostructures. In this
context, the sequence dependent stretch modulus for some typical base pairs step is calculated.

1. INTRODUCTION was the discovery of CNT (carbon nanotubes), a famous

During the last 20 years, significant progress in the field of
nanotechnology has been established. Despite the fact that
nanotchnology covers a very broad range of scientific topics
and applications, the principal characteristic of extremely small
dimension(s) (1−100 nm) unifies this diverse field. The
applications are numerous, steaming from various fields, such as
surface science, organic chemistry, molecular biology, semi-
conductor physics, microfabrication, and many others. This
inherent diversity and the goal to manipulate objects on the
nanoscale often demands a multidisciplinary approach dissolv-
ing the boundaries between the various fields of traditional
science.
In the field of structural engineering, as nanostructures find
numerous applications in forming new categories of composites
(i.e., nanocomposites), methods developed by chemists,
chemical engineers, and biophysicists to study chemical and
biological systems (i.e., molecular mechanics and dynamics)
have been incorporated in the traditional engineering analyses
leading to new hybrid methods, such as the multiscale FEM
(finite element method), MSM (molecular structural mechan-
ics), and others. One important milestone for this development
carbon allotrope cylindrical nanostructure with extraordinary
properties which finds applications in nanotechnology,
electronics, optics, materials science, and others. Efforts to
simulate the material properties at the nanoscale stem from
various fields,1−11 such as molecular dynamics, molecular
mechanics, molecular structural mechanics equivalent contin-
uum formulations, and others.
In biological space, nucleic acid nanotechnology has
developed into a mature synthetic technique for the design of
hierarchical molecular assemblies,12−18 and these constructs
have found application to a number of contemporary research
endeavors, including, but not limited to, biomimicry,19
macromolecular sensing,20 energy transfer/photonics,21,22
drug delivery,23,24 and actuated nanoreactor design.25 Con-
tinuum model representations of the nucleic acid building
blocks that comprise these assemblies, based on Kirchoff’s
theory of elastic rods, have been successfully applied to
studying DNA molecules for many years.26−34 For large DNA
Received: June 16, 2016
Published: September 21, 2016

© 2016 American Chemical Society 1963 DOI: 10.1021/acs.jcim.6b00356

J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling
nanostructures, the finite element-based35 computational
modeling framework CanDo has been the most successful.36−38
In its original implementation36 DNA origami nanostructures
were modeled as bundles of isotropic elastic rods that were
rigidly constrained to their nearest neighbors at specific
crossover positions. Later modifications included the introduc-
tion of nicked-duplex and single-stranded DNA modeling,38
and extension to lattice-free designs via parametrization of
junction alignment elements.37 The introduction of alignment
elements required parametrization of the mechanical junction
models by all-atom simulations, an example of the possible
synergy between the finite element and all-atom molecular
dynamics methods. Very recently, a Brownian Dynamics (BD)
model has been added to the CanDo framework to simulate the
equilibrium dynamics of large nanostructures in solution,39
providing a powerful predictive tool for understanding the
dynamic behavior of these large assemblies. The success of the
aforementioned continuum models is strongly dependent on
the assigned material properties of the coarse-grained elements
used to represent a particular nanostructure, and efficient fine-
grained models are needed that can accurately parametrize
them.
In light of these new developments, nanostructure analysis
and design can benefit from an intedisciplinary approach
incorporating traditional structural and mechanical engineering
principles for buildnig structures and machines in the
nanoscale.
Energy minimization is a very important procedure in the
analysis of molecular systems. In order to perform molecular
dynamics simulations, the system has to be at a point not far
from a local minimum of the potential energy,40 and for specific
applications, e.g. normal-mode analysis,41 the system has to be
at a minimum. From a structural mechanics point of view, the
minimization of energy suggests equilibrium between nodal
loads and internal forces of the structure. A fundamental
principle of structural mechanics and elasticity theory is
subdomain equilibrium under general equilibrium.42 Consider-
ing equilibrium of a structure suggests equilibrium for every
subpart of that structure, and even if the structure is not in
equilibrium under zero loads, fictitious nodal forces can be
considered, that put the structure in equilibrium. This property
allows partitioning of the structure into an appropriate number
of elements (one element for each energy term). The analysis
of the internal forces is equivalent to the computation of the
energy gradient, and the global tangent stiffness matrix
assembly is equivalent to the Hessian matrix calculation.
In order to verify that argument, butane and DNA molecules
were analyzed, and the numerically computed Hessian was
compared to the structure global stiffness matrix by the
introduction of appropriate metrics and found to be equivalent.
Furthermore, as was indicated from the numerical examples
examined herein but also from a combinatorial analysis, the
analytical protocol proposed here was several orders of
magnitude faster than the conventional numerical computation.
For molecular systems that can be approximated as solids
revolving in a potential energy minimum, where the entropic
effects can be neglected, this analysis can incorporate nodal
loads and, via minimization, derive constitutive laws. In this
context, the sequence dependent stretch modulus for indicating
two base pair double stranded DNA molecules was derived.
1964
Article
2. METHODOLOGY
2.1. Generalized Formulation of Potential Energy
Finite Elements. 2.1.1. Force Fields as a Reference. In the
context of molecular modeling, a force field refers to the
functional form and parameter sets used to calculate the
potential energy of a system of atoms or coarse-grained
particles in molecular mechanics and molecular dynamics
simulations. Potential energy force fields provide parameters for
every type of atom in a system, including hydrogen. The basic
functional form of potential energy in molecular mechanics
includes bonded terms for interactions of atoms that are linked
by covalent bonds (intramolecular bonds, i.e. internal terms)
and nonbonded (intermolecular interactions, i.e. external terms
also called “noncovalent”) terms. The intramolecular portion of
the potential energy function includes terms for the bonds,
valence (in plane) angles, dihedral (torsion) angles, improper
dihedral (out of plane bending) angles, and the Urey−Bradley
1,3-term, while the intermolecular terms include the long-range
electrostatic and van der Waals forces. The general form for the
total potential energy can be written as the sum of bonded and
nonbonded energy terms, where the components of the
covalent energy term are the bond, angle, dihedral, and
Urey−Bradley contributions while those of the noncovalent
energy term are the electrostatic and van der Waals ones:
V = Vbonded + Vnonbonded
= Vbonds + Vangles + Vdihedrals + Vimpropers + Vnonbonded (1)
The bond and angle terms are usually modeled by quadratic
energy functions that do not allow bond breaking, while the
functional form for dihedral energy is highly variable.
Additional, improper torsional terms can also be added to
enforce planarity and “cross-terms” (Urey−Bradley 1,3-term)
that describe coupling of different internal variables. Some force
fields also include explicit terms for hydrogen bonds.
Nonbonded energy terms are most computationally intensive,
and a popular choice is to limit interactions to pairwise
energies. The van der Waals (vdW) term is usually computed
with a Lennard-Jones (LJ) potential and the electrostatic term
with the Coulombic (CL) potential law. CHARMM (Chem-
istry at HARvard Macromolecular Mechanics)43,44 is a widely
used set of potential energy force fields for molecular dynamics
and is the one that was adopted herein in order to implement
the proposed formulation of the generalized potential energy
finite element.
2.1.2. Generalized Internal and Nodal Forces for Energy
Terms Assuming Equilibrium. Based on the proposed
generalized formulation, for each component of the energy
terms (both bonded and nonbonded ones), a special finite
element35 is formulated composed of nodes equal to the
number of atoms (a node is located at each atom) expressing
the contribution of the specific energy term. Every energy term
implies a generalized internal force. At an arbitrary config-
uration, fictitious nodal forces can be considered that put the
structure in equilibrium.42 These nodal forces satisfy equili-
brium with the generalized internal force. Through the
equilibrium equation implemented for the ith generalized finite
element (corresponding to any energy term), the reaction
forces f i are computed according to the following expression:
fi = ni ·LTi ·cTi (2)
DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

where f i is the vector of element forces for element i, ni is the 2 the interaction between three atoms (represented by nodes I,
internal generalized force (force Ni or moment Mi) for this J, and K) is described. While the expressions corresponding to
element, LTi is the matrix of transformation to the global
coordinate system for element i, and cTi is the equilibrium
vector. The proposed formulation is general, since it is
applicable to any force field.
Bond Stretches and Nonbonded Interaction. The first term
in the energy function (eq 1) accounts for the bond stretches,
while the last term represents nonbonded interactions between
pairs of atoms (i, j). For the case of the bonded and nonbonded Figure 2. In plane angle bond interaction.
components of the energy terms, as shown in Figure 1 the

the energy function along with the internal force (Min) and
stiffness (k̅θ) are given below, together with the transformation
Figure 1. Bond stretches and nonbonded interaction.
matrix L and the corresponding equilibrium vector c:
1
Vangle(θ ) = kθ ·(θ − θο)2 ⎡ λT 0 01 × 3 ⎤
interaction between two atoms (represented by nodes I and J) 2 ⎢ 1 1×3 ⎥
is described. The expressions corresponding to the energy dVangle ⎢ 01 × 3 λ 2T 01 × 3 ⎥
function along with the internal force (Nin) and tangent Min = = kθ ·(θ − θο), L = ⎢ ⎥,
dθ ⎢0 λ1T ⎥⎥
stiffness (k̅b) of the interaction are given below, together with ⎢ 1 × 3 01 × 3
the transformation matrix L and the corresponding equilibrium 2
∂ Vangle ⎢ ⎥
vector c: kθ̅ = ⎣ 01 × 3 01 × 3 λ 2T ⎦
2
∂θ ⎡ 1/la ⎤
⎧1 2 ⎢ ⎥
⎪ kb·(b‐bο) ⎢−1/lβ ⎥
⎪ 2
cT = ⎢ ⎥
⎪ q ·q ⎡⎛ R ⎛ R min, ij ⎞6 ⎤ ⎢ −1/la ⎥
min, ij ⎞
12
⎪ i j
⎢
+ εij · ⎜ ⎟ − 2⎜ ⎟⎥ ⎢ ⎥
V (b) = ⎨ 
4·π ·D·b ⎢⎣⎝ b ⎠ ⎝ b ⎠ ⎥⎦ ⎢⎣ 1/lβ ⎥⎦
⎪ 
⎪Coulombic (4)
⎪ Lennard ‐Jones
where kθ is the angle force constant and while θ0 and θ are the
⎪
⎪ D ·(1 − e−a(b − bο))2
⎩ e reference and current in-plane angles, respectively. λ1 and λ2
define directional vectors perpendicular to the vectors along
(3)
axes defined between nodes I-J and J-K (unit vectors, see Figure
∂Vb ∂ 2Vb 2), respectively; and la, lβ are the distances between nodes J-K
Nin = , kb̅ = and I-J, respectively.
∂b ∂b2
Dihedral Angle Bond Interaction (Proper and Improper).
⎡ λT 0 ⎤ The third term is used for the dihedrals (i.e., torsion angles),
⎡−1⎤
L=⎢
1 1×3⎥
, cT = ⎢ ⎥ while the next one accounts for the improper dihedrals that
⎢0 T ⎥ ⎣1⎦
⎣ 1 × 3 λ1 ⎦ represent out of plane bending. For the case of the dihedral
angle bond interaction component of the energy terms, as
where kb is the bond force constant, De is the well depth, a shown in Figure 3, the interaction between four atoms
controls the width of the potential, while b0 and b are the (represented by nodes I, J, K, and L) is described. While the
reference and current bond lengths, respectively. In expression expressions corresponding to the energy function along with
for the energy function, the top term is for the bonded term, the internal force (Min) and stiffness (k̅φ) are given below:
the middle term is for the nonbonded term, and the bottom
term is the Morse potential for hydrogen bonds. By definition, ⎧ kφ·[1 + cos(n·φ − δ)] for the proper term
⎪
the nonbonded forces are only applied to atom pairs separated Vdihedral(φ) = ⎨
⎪ kφ·(φ − φ )2 for the improper term
by at least three bonds. The van der Waals energy is calculated ⎩ 0
with a standard 12−6 Lennard-Jones potential (VLJ) and the
electrostatic energy with a Coulombic potential (VCL). In the ∂Vdihedral ∂ 2Vdihedral
nonbonded term of the energy function, qi and qj are the partial Min = , kφ̅ =
∂φ ∂φ 2
atomic charges of atoms i and j, respectively, εij is the well
(5a)
depth, Rmin,ij is the radius in the Lennard-Jones 6−12 term used
to treat the van der Waals interactions, and b (also denoted where kφ is the dihedral force constant for the proper term
with rij in the literature) is the distance between atoms i and j. while for the case of the improper one it represents the
λ1 is the unit vector of the axis defined between nodes I and J improper force constant. In the proper term, n and δ are
(see Figure 1). constants (n is the multiplicity of the energy function and δ is
In Plane Angle Bond Interaction. The second term in eq 1 the phase shift) while φ is the dihedral angle. In the improper
accounts for the bond angles. For the case of the bonded angle term, φ0 and φ are the reference and current out of plane
interaction component of the energy terms, as shown in Figure angles, respectively; and
1965 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

Figure 3. Dihedral angle bond interaction (proper and improper).

⎡ λT 01 × 3 01 × 3 01 × 3 ⎤ order to restrain the motions of the bonds involved in the

⎢ 1 ⎥ angle. This potential function is often referred to as the Urey−
⎢0 λ1T 01 × 3 01 × 3 ⎥ Bradley potential. The Urey−Bradley component (cross-term
⎢ 1×3 ⎥ accounting for angle bending using I−K nonbonded
⎢0 01 × 3 λ1T 01 × 3 ⎥⎥ interactions) comprises the fifth term in the energy function
L=⎢
1×3
,
⎢ ⎥ (eq 1). The Urey−Bradley energy term describes the
⎢ 01 × 3 01 × 3 01 × 3 λ4T ⎥ interaction between two atoms (represented by nodes I and
⎢ ⎥ K). The expressions corresponding to the energy function
⎢ 01 × 3 λ4T 01 × 3 01 × 3 ⎥
along with the internal force (Nin) and stiffness (k̅UB) of the
⎢ ⎥ interaction are given below, together with the transformation
⎣ 01 × 3 01 × 3 λ4T 01 × 3 ⎦
matrix L and the corresponding equilibrium vector c:
⎡ 1 ⎤
⎢ ⎥ 1
⎢ la·sin θa ⎥ VUB(b) = kUB· (b − bο)2 ⎡ λT 0 ⎤
2 ⎡− 1⎤
L=⎢
1×3⎥
⎢ 1 ⎥
1
1 , cT = ⎢ ⎥
⎢− − ⎥ ∂Vb 2
∂ VUB ⎢0 T ⎥ ⎣1⎦
Nin = , kUB
̅ = ⎣ 1 × 3 λ1 ⎦
⎢ la·sin θα lb·tan θa ⎥ ∂b ∂b2
⎢ 1 ⎥ (6)
⎢ ⎥
T
⎢ lb·tan θa ⎥ where kUB is the respective force constant, while b0 and b are
c =⎢ ⎥ the reference and current bond lengths between atoms I and K
⎢ 1 ⎥ in the harmonic potential (of the in plane angle bond
⎢ lc·sin θc ⎥
⎢ ⎥ interaction configuration of the three atoms I, J, and K). λ1 is
⎢ 1 ⎥ the unit vector of the axis defined between nodes I and K.
⎢ lb·tan θc ⎥ Superposition of Potential Terms through Forces. With
⎢ ⎥ appropriate summation of the nodal forces at the correspond-
⎢ 1 1 ⎥ ing degrees of freedom for each element, the energy gradient
⎢− − ⎥ can be obtained:
⎣ lc·sin θc lb tan θc ⎦ (5b)
nele
are the transformation matrix L and the corresponding ∂V
Fglob = ∑ fi =
equilibrium vector c, where λ1 and λ4 represent directional i=1
∂x (7)
vectors perpendicular to planes defined by nodes J-K-L and I-J-
K, respectively (unit vectors, see Figure 3), and la, lβ, lc, θa, θβ, where V is the total potential energy of the system and nele is the
and θc are the distances between the four nodes (I, J, K, and L) number of all the potential energy finite elements.
and the angles shaped by the axes defined between the four 2.2. Tangent Stiffness Matrix of the Generalized Finite
atoms as depicted in Figure 3. Elements. If vector X represents the initial configuration and
Urey−Bradley Bond Interaction. In the CHARMM force vector x denotes the final configuration, then the displacement
field an additional potential function is added to the angle, in field can be considered as
1966 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

U=x−X projections of the infinitesimal displacements at the global

dU = d x (8) coordinate system of nodes 1 and 2 (i.e., J and I) along the axis
45
that is defined with the unit vector λ1, du1 and du2 are the
In order to derive the tangent stiffness matrix for every infinitesimal displacement vectors at the global coordinate
energy term (i.e., for every potential energy finite element system of nodes 1 and 2 (i.e., J and I), and dU is the
described in the previous section), the force vector fi is infinitesimal displacement vector at the global coordinate
differentiated as follows: system. For the second part of eq 9, it can be written that
r m
d fi ∂f ∂n ∂fi ∂ck ∂fi ∂λj r
Ki = = i i + ∑ + ∑ ∂fi ∂ck
dU ∂ni ∂U ∂ck ∂U ∂λj ∂U (9)
∑ =0
k=1 j=1
k=1
∂ck ∂U (11)
th
where Ki is the global tangent stiffness matrix for the i
element, ni is the generalized internal force (moment Mi or For the third part of eq 9, two infinitesimal rotations are
axial force Ni) for the ith element, ck is the corresponding term considered. As it is well established, the unit vector derivative
of the r × 1 vector c, and r is the number of DOFs of the expresses change in orientation and is perpendicular to the unit
element’s local system. Depending on the potential energy vector. An orthonormal basis matrix [λ1 λ2 λ3] is constructed
finite element (two atom, angle, dihedral, etc.) the number of and the projection of the global displacements at the
DOFs varies (r = 2 for the case of two atom interaction appropriate axes λ3 and λ2 is considered to cause rotations
element, r = 4 for the in plane angle element, and r = 6 for the dθ2 and dθ3, respectively. Therefore, it can be written that
dihedral angle element); these terms express the equilibrium
dλ1 = (λ 2 × λ1) ·dθ2 + (λ3 × λ1) ·dθ3
between internal and nodal forces for each element. λj is the
corresponding element axis vector, m is the number of element ⎛1 T T
= ⎜ (λ × λ ) · [(λ × λ ) − (λ × λ ) ]
axis vectors (m = 1 for the inline element, m = 2 for the in- ⎝b 2 1 1 2 1 2
plane angle element, and m = 2 for the dihedral angle element), 1 ⎞
and U is the element’s displacement vector at the global system. + (λ3 × λ1) ·[(λ1 × λ3)T − (λ1 × λ3)T ]⎟ ·dU
b ⎠
The tangent stiffness matrix for the ith element is comprised of
three parts; the first one is equivalent to the structural = B ·dU
mechanics stiffness matrix when considering small displace- ⎡−I3 × 3 ⎤ ∂λj
ments. Here the projection of the global displacements to the
m
∂fi ∂λj ⎡− B ⎤
∑ = Ni·⎢ ⎥· = Ni·⎢ ⎥
conjugate forces comes into play. The other two terms come ∂λj ∂U ⎣ I3 × 3 ⎦ ∂U ⎣B⎦
j=1 (12)
into play when there are significant internal forces present
(which is the case for molecular assemblies). The second one where ( × ) denotes the vector cross product; dλ1 is the
expresses changes in equilibrium due to variation of the infinitesimal change of the λ1 axis, considered to be a
displacement field, and the third one changes to geometry. superposition of the two infinitesimal rotations dθ2 and dθ3
2.2.1. Bond, Nonbonded, and Urey−Bradley Potential around axes λ2 and λ3, respectively; and B is the 3 × 6 matrix
Energy Finite Elements. For the case of interaction between expressing the element axis derivative (λ1) with respect to the
two atoms (bond, nonbonded and Urey−Bradley terms) the global displacements of nodes I and J. Therefore, based on the
differentiation of the corresponding internal force is expressed three parts of eq 9 provided by eqs 10b, 11, and 12, the global
as follows: tangent stiffness matrix for the two-atom interaction (bonded
dNi = kb̅ ·(dub2 − dub1) or nonbonded) is given by the following expression:
3 3 ⎡− B ⎤
= kb̅ ·(∑ λ1i ·du 2i − ∑ λ1i ·du1i) K i = LT ·(kb̅ ·c T ·c) ·L + Ni·⎢ ⎥
⎣B⎦ (13)
i=1 i=1 (10a)
while dNi can also be written as 2.2.2. In-Plane Angle Potential Energy Finite Element. For
the angle bond interaction (three atoms) there are two element
⎡ du11 ⎤ axis, namely λ1 and λ2. For the case of interaction between
⎢ ⎥ three atoms (in plane angle term), the differentiation of the
⎢ du12 ⎥
⎢ ⎥ corresponding internal force is expressed as follows:
⎡ λ11 λ12 λ13 0 0 0 ⎤ ⎢ du13 ⎥
dNi = kb̅ ·[− 1 1]·⎢ ⎥·⎢ ⎥ dMi = kθ̅ ·dθ = kθ̅ ·c ·L ·d U
⎢⎣ 0 0 0 λ11 λ12 λ13 ⎥⎦ ⎢ du 21 ⎥
⎢ du ⎥ dMi
= kθ̅ ·c ·L
⎢ 22 ⎥ dU (14)
⎢⎣ du 23 ⎥⎦
where dU is the infinitesimal displacement vector at the global
= kb̅ ·c ·L ·d U coordinate system. The first part of the stiffness matrix is given
dNi ∂f ∂Ni dN by
= kb̅ ·c ·L and i = LΤ·c Τ · i
dU ∂Ni ∂U dU ∂fi ∂ni dM
Τ Τ Τ
= LΤ·c Τ · = LΤ·(kθ̅ ·c Τ ·c) ·L = LΤ·KLoc·L
= L ·(kb̅ ·c ·c)·L = L ·KLoc·L (10b) ∂ni ∂U dU
(15)
The first part of eq 9 is derived in eqs 10, where dNi is the
infinitesimal change of the internal force, dub1 and dub2 are the The second term can be derived by differentiating c:
1967 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

Figure 4. Part of the stiffness matrix for the in-plane angle element due to rotations of the plane.

∂Fi ⎛ ∂ck ∂lα ∂c ∂lβ ⎞ ∂Fi ⎛ ∂λj ∂φ ⎞

r r m m 2 ∂λ
∂Fi ∂ck ∂Fi ∂λj ∂θ ∂λ ∂ω
∑ = ∑ ⎜⎜ + k ⎟⎟ ∑ = ∑ ⎜⎜∑ j l + j + ⎟
k=1
∂ck ∂U k=1
∂ck ⎝ ∂lα ∂U ∂lβ ∂U ⎠ j=1
∂λj ∂U j=1
∂λj ⎝ l = 1 ∂θl ∂U ∂ω ∂U ∂φ ∂U ⎟⎠
∂Fi ∂l ⎡ c1·I3 × 3 ⎤ ⎡ 03 × 3 ⎤
= M ·LT , a = [ λαT 01 × 3 −λαT ], ⎢ ⎥ ∂Fi ⎢ ⎥
∂c ∂U ∂Fi
∂lβ = ⎢ 03 × 3 ⎥ , = ⎢ c 2·I3 × 3 ⎥ ,
= [01 × 3 λβΤ −λβΤ ] ∂λ1 ⎢ ⎥ ∂λ 2 ⎢ ⎥
∂U ⎣ c3·I3 × 3 ⎦ ⎣ c4·I3 × 3 ⎦
∂c ∂c ∂λ1 ∂λ ∂λ ∂λ
= [−1/lα2 0 1/lα2 0 ], = [0 1/lβ2 0 −1/lβ2 ] = (R1 × λ1), 1 = 0, 2 = (R1 × λ 2), 2 = 0
∂la ∂lβ ∂θ1 ∂θ2 ∂θ2 ∂θ1
(16a) (18a)
and in a more compact form: and
r
∂Fi ∂ck ⎡ λΤ 01 × 3 01 × 3 ⎤
∑ = M ·LT ·KELOC ·L 2 ∂θ1
= [−1/la 1/la ]·⎢ ⎥
1
∂ck ∂U
k=1
∂U ⎢0 Τ ⎥
⎣ 1×3 01 × 3 λ1 ⎦
⎡0 Τ
01 × 3 ⎤
∂θ2 ⎢ 1×3 λ2 ⎥
= [−1/lβ 1/lβ ]·
KELOC = ∂U ⎢0 Τ ⎥
⎣ 1 × 3 01 × 3 λ 2 ⎦
⎡ −1/l 2 0 1/lα2 0 ⎤
⎢ α ⎥ ∂λ1 ∂θ1 ∂λ ∂θ
⎢ 0 2 ⎥ B1 = , B2 = 2 2
1/lβ2 0 −1/lβ ∂θ1 ∂U ∂θ2 ∂U
⎢ ⎥,
⎢ 2 ⎥ ⎡ ⎤
⎢ 1/lα 0 −1/lα2 0 ⎥ c1·B1
∂Fi ⎛ ⎞ ⎢ ⎥
m 2 ∂λ
⎢ ⎥ j ∂θl
−1/lβ2 1/lβ2 ⎦ ∑ ⎜⎜∑ ⎟⎟ = M ·⎢ c 2· B2 ⎥
⎣ 0 0
∂λj ⎝ l = 1 ∂θl ∂U ⎠ ⎢ ⎥
⎡ λT 0 ⎤ j=1
⎣ c3·B1 + c4·B2 ⎦
⎢ α 1 × 3 01 × 3 ⎥ (18b)
⎢0 λ T 01 × 3 ⎥ For the λ1 and λ2 axis derivatives the definition of
⎢ 1×3 β ⎥ infinitesimal rotations can be considered; the infinitesimal
L2 = ⎢
T ⎥
⎢ 01 × 3 01 × 3 λα ⎥ rotations dθ1 and dθ2 are caused by the global displacement
⎢ T ⎥
projections on the λ1 and λ2 axes at the start and end nodes of
⎣ 01 × 3 01 × 3 λβ ⎦ (16b) vectors JI and JK, respectively (see Figure 2). For the second
and third parts, the rotation of axes λ1 and λ2 due to rotations of
where lα and lβ are the length of vectors λα and λβ, respectively. the plane is considered; thus, the displacements perpendicular
For the third term, the change of the orientation of axis λ1 and to the plane uKproj, uIproj, and uJproj are of importance. These
λ2 is of importance. As this is an in-plane angle element, at an displacements can be decomposed into three parts, one
imposed displacement field, the plane will have a new position. translation and two rotations along two perpendicular axes,
One way to express the changes of the element axis is relative i.e. dφ and dω along R2 (unit vector KI; see Figures 2 and 4)
to the plane. The change is considered in two terms; the first and R3, respectively, where
term is due to in plane displacements. These displacement
cause λ1 and λ2 to rotate in plane along the plane vector, R1 × R 2
R3 =
|R1 × R 2| (19)
λβ × λα
R1 =
|λβ × λα| (17) which are considered to have as origin the center C0 of the
parallelogram of the triangle’s ABC coordinates (see Figure 4).
by dθ1 and dθ2, respectively. The second term is due to The second and third parts are calculated by the following
rotations ω and φ of the plane. relations:
1968 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

∂λ1 ∂λ ∂Fi ∂C ⎡ 1 1 ⎤T
= (R 2 × λ1), 1 = (R3 × λ1), = M · LT , = ⎢− 2 2
0 0 0 0⎥
∂φ ∂ω ∂C ∂la ⎣ la sin θa la sin θa ⎦
∂λ 2 ∂λ
= (R 2 × λ 2), 2 = (R3 × λ 2) ∂C ⎡ 1 1 ⎤
T
∂φ ∂ω = ⎢0 0 0 − 2 0 2 ⎥
∂lc ⎣ lc sin θc lc sin θc ⎦
∂λ1 ∂φ ∂λ ∂ω ∂λ ∂φ ∂λ ∂ω
+ 1 = Bl1 and 2 + 2 = Bl 2 ⎡ ⎤T
∂φ ∂U ∂ω ∂U ∂φ ∂U ∂ω ∂U ∂C 1 1 1 1
= ⎢0 2 − 2 0 − 2 ⎥
(20a) ∂lβ ⎣ lb tan θa lb tan θa lb tan θc lb 2 tan θc ⎦

thus ∂la
= [ λαΤ − λαΤ 01 × 3 01 × 3 ],
∂U
⎡ c1·Bl1 ⎤ ∂lc
m
∂Fi ⎛ ∂λj ∂ω ∂λj ∂φ ⎞ ⎢ ⎥ = [ 01 × 3 01 × 3 − λcΤ λcΤ ]
∑ ⎜ + ⎟ = M·⎢ c 2·Bl 2 ⎥ ∂U
j=1
∂λj ⎝ ∂ω ∂U ∂φ ∂U ⎠ ⎢ ⎥ ∂lβ
⎣ c3·Bl1 + c4·Bl 2 ⎦ = [01 × 3 λβΤ − λβΤ 01 × 3 ]
∂U
(20b) (23b)
For an analytical calculation of matrices Bl1 and Bl2, see the
Supporting Information. The global tangent stiffness matrix for
the angle bond interaction is given by the following expression:

K i = LΤ·(K theta·CΤ·C) ·L + M ·LT ·KELOC ·L 2 and

⎡ c1·(B1 + Bl1) ⎤
⎢ ⎥
+ M·⎢ c 2·(B2 + Bl 2) ⎥
⎢ ⎥ ∂C
⎢⎣ c3·(B1 + Bl1) + c4·(B2 + Bl 2)⎥⎦ (21) ∂θa
2.2.3. Dihedral Angle Potential Energy Finite Element. For ⎡ cos θa cos θa 1 1 ⎤T
= ⎢− 2 2
+ 2
− 2
0 0 0⎥
the case of interaction between four atoms (proper and ⎢⎣ la sin θa la sin θa lβ sin θa lβ sin θa ⎥⎦
improper dihedral terms), the differentiation of the correspond-
ing internal force is expressed as follows: There are two
element axes normal to each plane, namely λ1 and λ2. The
interatomic potential, the corresponding internal force ∂C
(torque), and the first term of the stiffness matrix are ∂θc
⎡ cos θc 1 cos θc 1 ⎤T
∂ 2Vdihedral = ⎢ 0 0 0 − − + ⎥
dM = dchi = Kchi·C·L ·dU ⎢⎣ lc sin 2 θc lβ sin 2 θc lc sin 2 θc lβ sin 2 θc ⎥⎦
∂chi 2
∂M
= Kchi·C·L
∂U ∂θa Τ Τ
= [(1/la)λΝΑ1 − (1/ la)λΝΑ1 01 × 3 01 × 3 ]
∂Fi ∂ni ∂UKL
= LΤ·(Kchi·CΤ·C) ·L = LΤ·KLoc ·L
∂ni ∂U (22) ∂θc Τ Τ
= [ 01 × 3 01 × 3 − (1/lc)λΝΑ4 (1/ lc)λΝΑ4 ]
∂UIJ
where dU is the element infinitesimal global displacement. The
(23c)
second term can be derived by differentiating C with respect to
la, lβ, lc, θa, and θc or by realizing that, for displacements along
each plane (see Figure 3), the internal torque should not
change and should appropriately define the changes dla, dlβ, dlc,
dθa, and dθc. The second term is divided into two terms: one
for changes of la, lβ, and lc and one for changes of θa and θc (see and in a more compact form:
Figure 3). Due to the fact that the change of the level arm of
the forces at nodes L and I (see Figure 3) is what is of interest
herein, the angle derivatives should be expressed regarding the
r
displacements of the appropriate arm of each angle. ∂Fi ∂ck
∑ = M ·LT ·Θ
∂ck ∂U
∂Fi ⎛ ∂ck ∂la
r r
∂Fi ∂ck ∂c ∂lβ ∂c ∂l k=1
∑ = ∑ ⎜⎜ + k + k c
∂ck ∂U ∂ck ⎝ ∂la ∂U ∂lβ ∂U ∂lc ∂U Θ = KELOC1·LA + KELOC 2·LB (23d)
k=1 k=1

∂ck ∂θa ∂c ∂θβ ⎞

+ + k ⎟
∂θa ∂UKL ∂θβ ∂UIJ ⎟⎠ (23a)

where where
1969 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

⎡−KE1 KE1 0 0 0 0 ⎤ For the third term the unit vectors define the two planes of
⎢ ⎥ the dihedral angle. The unit vector derivatives are derived from
⎢ KE1 −KE1 KE2 −KE2 0 0 ⎥ the dihedral unit vector (element’s axis) definition. The unit
⎢ ⎥ vector derivative expresses orientation change, and the
⎢ 0 0 −KE2 KE2 0 0 ⎥
KELOC1 = ⎢ , rotations of each of the two planes are of importance; thus,
0 0 0 0 KE3 −KE3 ⎥
⎢ ⎥ displacements perpendicular to each plane are considered.
⎢ 0 0 −KE4 KE4 0 0 ⎥ ⎡ C1·I3 × 3 ⎤ ⎡ 03 × 3 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 KE4 −KE4 −KE3 KE3 ⎥⎦
∂Fi ⎢C2·I3 × 3 ⎥ ∂Fi ⎢C5·I3 × 3 ⎥
⎡ λT = M·⎢ ⎥, = M⎢ ⎥
01 × 3 01 × 3 01 × 3 ⎤ ∂λ1 ⎢ C3·I3 × 3 ⎥ ∂λ4 ⎢C6·I3 × 3 ⎥
⎢ a ⎥
⎢0 ⎢ 0 ⎥ ⎢C ·I ⎥
λa 01 × 3 01 × 3 ⎥
T
⎣ 3×3 ⎦ ⎣ 4 3×3 ⎦ (25a)
⎢ 1×3 ⎥
⎢0 λb 01 × 3 01 × 3 ⎥⎥
T also
LA = ⎢
1×3
⎢ ⎥ 1 1
⎢ 01 × 3 01 × 3 λbT 01 × 3 ⎥ dλ1 = − b1′ × b2′ + b1 × b2
′ ′
|b1 × b2| |b1 × b2|
⎢ ⎥
⎢ 01 × 3 01 × 3 λcT 01 × 3 ⎥ 1
⎢ ⎥ =− ((b1 + u11 − u 21) × (b2 + u31 − u 21))
⎣ 01 × 3 01 × 3 01 × 3 λcT ⎦ (23e) |b1′ × b2′|
1
+ b1 × b2
and |b1 × b2|
1 1 1 1
KE1 = , KE2 = 2 , KE3 = 2 , ≅− [b1 × (u31 − u 21) + (u11 − u 21) × b2]
la2 sin(θα) lb tan(θα) lc sin(θc) |b1 × b2|
1 1
KE4 = 2 = (M1·λ1·( −[ 01 × 6 λ1T 01 × 3 ]·dU
lb tan(θc) |b1 × b2|
+ [ 01 × 3 λ1T 01 × 6 ]·dU ) −M 2 ·λ1·
KELOC2 = ( −[ λ1T 01 × 9 ]·dU − [ 01 × 3 λ1T 01 × 6 ]·dU ))
⎡ − KEB1 KEB1 0 0 ⎤
⎢ ⎥ = Bl1·dU
⎢(KEB1+KEB2 ) − (KEB1+KEB2 ) 0 0 ⎥
⎢ ⎥ (25b)
⎢ − KEB2 KEB2 0 0 ⎥
⎢ ⎥ and
⎢ 0 0 KEB3 − KEB3 ⎥
⎢ 0 0 KEB4 − KEB4 ⎥ 1
⎢ ⎥ dλ4 ≅ [b2 × (u44 − u34) + (u34 − u 24) × b3]
⎢⎣ |b2 × b3|
0 0 − (KEB3+KEB4 ) (KEB3+KEB4 )⎥⎦
1
(23f) = (M 2·λ4 ·([01 × 9 λ4T ]·dU − [01 × 6 λ4T 01 × 3 ]·dU )
|b2 × b3|
cos(θα) 1 − M3·λ4 ·([01 × 6 λ4T 01 × 3 ]·dU − [01 × 3 λ4T 01 × 6 ]·dU ))
KEB1 = , KEB2 = ,
la2 sin 2(θα) lalb sin 2(θα) = Bl4 ·dU
(25c)
cos(θc) 1
KEB3 = , KE4 = thus
lc2 sin 2(θc) lclb sin 2(θc)
⎡ c1·Bl1 ⎤
⎡ λT ⎢ ⎥
01 × 3 01 × 3 01 × 3 ⎤ m
∂Fi ∂λj
T
⎢ c 2·Bl1 + c5·Bl4 ⎥
⎢ NA1 ⎥ ∑ = M·⎢ ⎥
⎢0 λNA1 01 × 3 01 × 3 ⎥
T ∂λj ∂U ⎢ c3·Bl1 + c6·Bl4 ⎥
LB = ⎢ ⎥
1×3 j=1
⎢0 ⎢ ⎥
⎢ 1×3 01 × 3 λNA 4 01 × 3 ⎥⎥
T ⎣ c4·Bl4 ⎦ (25d)
⎢ T ⎥ where b1, b2, and b3 are the vectors defined by the 4 atoms (b1 =
⎣ 01 × 3 01 × 3 01 × 3 λNA 4⎦
−λα, b2 = −λβ, and b3 = λc) in the initial configuration and M1,
where λα is the vector KL, λβ is the vector JK, and λc is the M2, and M3 are the matrices expressing the cross product of
vector JI (see Figure 3): each of these vectors (skew-symmetric matrices), b1′ , b2′ , and b3′
are the vectors defined by the 4 atoms in the final configuration,
λ1 × λα and uXY are the projections of the global displacements of node
λNA1 = X (1 for node L, 2 for K, 3 for J, and 4 for I) at vector Y (1 for
|λ1 × λα|
λ1, 4 for λ4), while the matrices Bl1 and Bl4 express the unit
and vector derivatives λ1 and λ4, respectively. Finally, as dU is small,
λ4 × λc it is considered that |b′1 × b′2| ≅ |b1 × b2| and |b′2 × b′3| ≅ |b2 ×
λNA 4 = b3|. The global tangent stiffness matrix for the dihedral bond
|λ4 × λc| (24) interaction is given by
1970 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

Figure 5. Schematic representation for the element assembly of the global stiffness matrix for the nonbonded and Urey−Bradley interactions.

⎡ c1·Bl1 ⎤ 2.3. On the Calculation of the Hessian Matrix as a

⎢ ⎥
Τ Τ T
⎢ c 2·Bl1 + c5·Bl4 ⎥
K i = L ·(Κchi·C ·C) ·L + M ·L ·Θ + M ·⎢ ⎥
⎢ c3·Bl1 + c6·Bl4 ⎥
⎢ ⎥
⎣ c4·Bl4 ⎦ of atoms (a node is located at the position of each atom)
(26)
With appropriate summation at the corresponding degrees of
freedom for each element,35 the energy Hessian matrix is
obtained:
Finite Element Assembly. Based on the proposed
generalized formulation, for each component of the energy
terms (both bonded and nonbonded ones), a specialized finite
element is formulated composed by nodes equal to the number
expressing the specific component of the energy terms. The
total potential energy can be expressed as a summation of the
energy terms:
nele

nele
V= ∑ Vi
∂ 2V i=1
K glob = ∑ Ki =
i=1
∂x 2 (27) Vi = Vi (zi) = Vi (Xi)
nele
The element assembly for the formation of the global ∂V ∂Vi
stiffness matrix is not restricted only to the bonded atoms. The
= ∑
∂X i=1
∂X
same summation procedure is followed for the nonbonded
terms as well as for the Urey−Bradley terms involved in the ⎧ ∂Vi ∂zi
∂Vi ⎪ , for x = xi
angle bond interaction among three atoms. A schematic = ⎨ ∂zi ∂Xi
representation of the global element assembly for such cases ∂X ⎪
is presented in Figure 5. ⎩ 0, otherwise (28)

1971 DOI: 10.1021/acs.jcim.6b00356

J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

where nele is the number of energy terms (or energy elements) burden of a naive numerical computation, the connectivity of
and zi is the corresponding internal degree of freedom for the each degree of freedom was taken into account when
ith energy term, while Xi are the corresponding Cartesian computing energy pertrubations. In order to validate the
coordinates for the ith energy term (zi is a function of Xi). As proposed modeling approach and asses its efficiency in the case
can be recognized for the expressions of eq 28, the gradient can of energy minimization, the following molecular systems were
be obtained as the proper summation for all energy elements. studied:
Likewise, for the Hessian matrix it can be written that 2.4.1. Single Butane Molecule. A single butane molecule
nele (Figure 6) is used as a proof-of-concept system, where the
∂ 2V ∂ 2Vi
= ∑
∂X2 i=1
∂X2
⎧ ∂Vi ∂zi
∂ Vi2 ⎪ ∂
⎪ ∂zi ∂Xi ( )
= ⎨ , for x = xi
∂X2 ⎪ ∂X
⎪
⎩ 0, otherwise

∂ ( ∂Vi ∂zi
∂zi ∂X i ) = ⎛⎜ ∂ V ∂z ⎞⎟
i i
T 2
∂zi
+ i
∂V ( )
∂
∂zi
∂X i
2
∂X ⎝ i
∂z ∂X ⎠ ∂X i ∂zi ∂X
2 ⎛ T ⎞ 2
∂ Vi ∂zi ∂zi ∂V ∂ zi
= 2
⎜ ⎟+ i
∂zi ⎝ ∂Xi ∂Xi ⎠ ∂zi ∂Xi2
thus
⎧ ∂ 2V ⎛ ∂z T ∂z ⎞ ∂V ∂ 2z Figure 6. Butane molecule.
⎪ i⎜ i i
⎟+ i i
, for
⎪ ∂ z i
2
⎝ ∂X i ∂X i ⎠ ∂ z i ∂X i
2
number of nodes totalled 14 (4 carbon atoms and 10 hydrogen
2
∂ Vi ⎪
= ⎨ ⎛ ∂ Vi ⎞
2 atoms) and the corresponding degrees of freedom of the
∂X 2
⎪ ⎜ 2
⎟ ∈ Xi ⊗ Xi proposed FE model are 42. The number of bonds are 13, those
⎪ ⎝ ∂X ⎠mn of angle and Urey−Bradley bonds are 24, those of dihedral
⎪ angle bonds 27, and those of nonbonded terms 54. The single
⎩ 0, otherwise (29) butane molecule case was considered in order to validate the
A combinatorial comparison to a conventional numerical analytical expressions presented herein in order to calculate the
comutation for full nonbonded interaction cases is presented energy gradient and Hessian matrix. Atomic parameters from
below. The main argument is to demonstrate that, in this case, the CHARMM2743,44 field were used.
the computation cost is of the same order of that of the energy 2.4.2. Double Stranded B-DNA. The second system
gradient. In the full connectivity case (upper limit), each atom investigated was a canonical B-form DNA duplex. The number
is connected to at least (n − 3)/3 other atoms via nonbonded of nodes (or atoms) for AT (adenine-thymine) and TA bases
interactions, where n is the total number of configurational are 64, and those for CG (cytosine-guanine) and GC bases are
degrees of freedom. In this case, the number of energy terms 63 (Figure 7 denotes graphically the nucleobases of DNA).
(i.e., number of elements) is nele ∼ n2 (where ∼ denotes that nele However, in the model used here, we used 41 for AT and CG
is proportional to n2). Let O(NumHess) denote the number of as the hydrogen atoms where excluded and the hydrogen bonds
operations for the conventional numerical computation of the among the bases described explicitly using a Morse potential.
Hessian Matrix. If k operations are performed for each The number of bonds are 61 for CG and 62 for AT, the
component of the Hessian matrix, hence the total number of number of angle bonds are 58 for CG and AT, and the number
operations: O(NumHess) ∼ kn2. The number of operations k of dihedral angle bonds are 59 for CG and 56 for AT; finally,
(energy perturbations) are not fixed when considering the full the number of improper dihedral angle bonds are 4 for CG and
nonbonded interaction. The operations for the calculation of AT. Regarding the nonbonded terms that express Coulombic
the energy difference of the connected atoms is proportional to and Lenard-Jones interactions, a cut off radius of 8 Å was used.
n; thus, O(NumHess) ∼ n3. If O(PEEM) are the number of DNA cases of 2, 10, and 30 base pairs were used to compare
operations for the proposed method where the Hessian matrix the performance of the method when incorporated in energy
is computed as an element assembly of nele ∼ n2 energy terms, it minimization. The solution of equilibrium equations was used
is proved that according to the proposed modeling approach, in order to derive the sequence dependent stretch modulus for
Hessian matrix computing requirements O(PEEM) ∼ n2 are of indicative base pair steps. Atomic parameters from the
the same order to that of gradient and energy computations for CHARMM27 field43,44 were used. In the case of DNA, one
a full nonbonded connections case. In this context the use of an of the authors has seen successful use of the CHARMM field to
analytically computed Hessian matrix should be reconsidered study DNA nanostructure conformations.37,39,46
when dealing with energy minimization problems, since fast
and efficient algorithms can be developed. 3. RESULTS AND DISCUSSION
2.4. Computational Details and Test Systems. All the 3.1. Validation. For the butane molecule case, the energy
computations were performed in Matlab version 15b.45 The gradient and Hessian matrix were computed both numerically
Hessian numerical computation was also done in Matlab by and analytically, implementing the proposed modeling
implementing central differences, and in order to reduce the approach. For the numerical differentiation, the centered
1972 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

Figure 7. DNA bases (a) adenine, (b) thymine, (c) cytosine and (d)
guanine.

differencing technique was used, aiming to achieve increased

accuracy. The difference between the gradient calculated
numerically and that achieved by means of the assembly of
internal force vector is practically equal to zero (the larger
difference is equal to 5 × 10−07%), proving the equivalence of
the internal force analysis with the energy differentiation. The
energy Hessian matrix computed numerically and the tangent
global stiffness matrix are identical (the differences are of the
order of E−06% while the large elements of the matrices are of
the order of E+03; see Figures 8a and 8b for the graphical
presentation of the form of the Hessian matrix and that of the
differences). In order to verify the equivalence of the two
matrices (calculated numerically and analytically), the following
energy comparison was performed. The energy is approximated
locally as a quadratic function of displacement. Thus, in an
arbitrary direction, the energy can be computed by Figure 8. Butane molecule-performance of the proposed modeling
approach: (a) graphical representation of the Hessian Matrix and (b)
1 element by element difference of Hessian matrices (analytical
V1 = V + g T ·U + UT ·H ·U calculation obtained based on the new modeling approach and
2 (30)
numerically computed).
where V and V1 are the potential energy at the initial and final
configurations, respectively, U is an arbitrary displacement
vector while g and H are the energy gradient vector and
Hessian matrix, respectively. The measure for comparing the effort required for computing the Hessian matrix. The CPU
Hessian matrices considered herein is based on the maximum time required for computing the Hessian matrix based on the
difference (ρ) calculated along every possible direction that is proposed analytical calculation and that of the numerical
defined according to the following expression: computation for the case of double stranded DNA molecules
(varying on the number of base pairs considered) is depicted in
UT ·H ·U − UT ·K·U Figure 9. As was mentioned previously for the numerical
ρ = max ,∀U computation, the central differencing technique for each DOF
UT ·H ·U (31)
was applied while the energy differences were computed by
where K is the tangent global stiffness matrix. The value of ρ for involving only the relevant energy terms. As is already
the butane molecule case is equal to 1.26 × 10−08%, proving the mentioned in section 4 and depicted in Figure 9, the proposed
equivalence of the proposed method. As already mentioned, the method, as well as any method that assembles the Hessian
proposed modeling approach resulting in analytical calculation matrix as a summation from the matrices of second derivatives
of the energy gradient vector and Hessian matrix was for every energy term, is significantly faster than the numerical
significantly faster than the case of the numerical computation. computation. The method was integrated in the energy
3.2. Performance in Terms of Speed and Energy minimization algorithm in order to assess its performance
Minimization. Initially, the performance of the proposed when compared with the BFGS minimization scheme, which is
modeling approach was assessed in terms of computational one of the standards for this kind of system. The algorithm
1973 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

BFGS minimization does not use the entire Hessian matrix

but rather approximates it by the results obtained through the
successive iterations. In the BFGS scheme used for this
comparative study, the analytical gradients (residual forces)
were provided, so the method could be comparable in terms of
speed (numerical gradients demand an appropriate energy
computation in each DOF involving only the relevant energy
terms; otherwise, they are much slower). The results of energy
minimization implementing both schemes in three double
stranded DNA molecules (composed by 2, 10, and 30 base
pairs) are presented in Figure 10. Both methods were
approximately equivalent in terms of the speed of each
iteration. However, the BFGS minimization scheme did not
manage to reach a satisfactory minimization level (the
minimized configuration contains large remaining forces). For
both minimization schemes, the maximum number of iterations
allowed was 1000. In all cases, BFGS terminated when the
number of maximum iterations was reached. The trust-region
Figure 9. DNA molecules-performance of the proposed modeling algorithm, integrating the Hessian matrix and gradient vectors
approach (i.e., analytical calculation based on the proposed FE type calculated in accordance with the proposed method, performed
approach) vs numerical calculation of the Hessian matrix in terms of significantly better in terms of minimization level (the
computation effort. remaining forces were almost equal to zero in all three cases)
as well as in the rate of convergence (number of iterations
integrating the proposed method is of the trust-region family.47 needed to reach the minimum).
In general, the problem is stated as In all three cases examined here, the Hessian matrix was
storable; however, in cases where such an option is not feasible,
min { 12 s Hs + s g , such that Ds ≤ Δ}
T T
(32)
a sparse version of the Hessian matrix was also examined. For
the three DNA cases, a banded version of the tangent stiffness
where Δ is the trust-region dimension. In the particular matrix (Hessian) was used. The 11 main diagonals were stored
variant48 used, the trust-region subproblem is restricted into a and used in the minimization process. This amount of memory
two-dimensional subspace S {s1, s2} where s1 is the direction of is, for the 2 base pair case, 4.8% of the matrix, in the 10 base
the gradient g and s2 is defined by pair case, 0.97% of the matrix, and in the 30 base pair case,
0.32% of the matrix. In all the cases, the scheme implemented
Hs2 = −g using the banded form of the Hessian matrix performed better
or in terms of residual forces than the one using the BFGS
s2T Hs2 < 0 method, as can be seen in Tables 1−3.
(33)
3.3. Equivalent Finite Element Simulation and
and obtained by solving iteratively via the preconditioned Modulus of Elasticity of Double Stranded DNA. The
conjugate gradient algorithm. proposed analysis is in essence a structural molecular model.

Figure 10. Thirty base pair double stranded DNA: (a) minimized; (b) minimized under load on the top in y−y′ direction and fixed at the base.

1974 DOI: 10.1021/acs.jcim.6b00356

J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

Table 1. Minimization Results50 for Double Stranded DNA (gradient equals 0). Then external forces (F) are applied at the
of DNA Molecule of 2 Base Pairs (CT) top two edges of the DNA molecule backbone (see Figure 11
and Table 4) while the DOFs of the bottom base pair are fixed
Trust-Region
Trust-Region (Banded along the direction of F along with all the DOFs (x,y,z) of the
BFGS (Full Hessian) Hessian) connecting to the backbone atoms of the bottom base pair. In
No. of Iterations 1001 237 1001 order to calculate the mean distance (l ̅) between the base pairs
Potential energy value 2.86 × 1004 2.85 × 1004 2.86 × 1004 and the mean displacement (u)̅ of the top base pair, two
Max. Residual Force 13.425 9.21 × 10−06 6.689 schemes are implemented. In the first one, l ̅ is the difference of
(First order the mean values of the z-coordinate of the base pairs while u̅ is
optimality) the mean value of the z-direction displacement of the top base
pair (with the bottom base pair being fixed on this direction).
Table 2. Minimization Results50 for Double Stranded DNA In the second, l ̅ is the mean value of the z-coordinate of the
of DNA Molecule of 10 Base Pairs (AGTCTAGTCT) distance between the atoms connecting the backbone for top
Trust-Region and bottom base pairs; for example, the atoms which are fixed
Trust-Region (Banded and the atoms where the external forces are applied and as
BFGS (Full Hessian) Hessian) displacement the mean value of the z-direction displacement of
No. of Iterations 1001 285 1001 the two atoms where the external forces are applied. As the
Potential energy value 1.466 × 1005 1.462 × 1005 1.466 × 1005 stretch modulus is of interest here, there is no need for
Max. Residual Force 68.71 7.43 × 10−06 16.10 nonlinear minimization; thus, solving the linear system in eq 36
(First order
optimality) is straightforward.
K mU = F (36)
Table 3. Minimization Results50 for Double Stranded DNA
of DNA Molecule of 30 Base Pairs where Km is the reordered stiffness matrix. The stretch modulus
(AAGGGCCTACGTACGATTCGAGTCTAGTCT) EA is computed by the following expression:

Trust-Region Fl ̅
EA =
Trust-Region (Banded u̅ (37)
BFGS (Full Hessian) Hessian)
No. of Iterations 1001 491 1001 The values for the stretch modulus for four different base
Potential energy value 4.518 × 1005 4.500 × 1005 4.518 × 1005 pair steps are depicted in Table 2. The exact stretch modulus is
Max. Residual Force 162.01 2.53 × 10−05 35.10 not independent of the profile of the external forces; for the
(First order four cases examined herein, the external forces were applied at
optimality) the edges of the backbone of the top base pair.

The force and stiffness matrix analysis was used in the double
stranded DNA model using the CHARMM force field. The
deformed configuration for a double stranded DNA chain
under external loads can be derived (neglecting entropic
effects) by implementing the following analysis: As the tangent
stiffness matrix is a structural model, arbitrary DOFs can be
“fixed” by eliminating the corresponding lines and columns and
reordering. Nodal loads are considered, and the potential
energy for minimization is
E = V − FT ·U
where E is the potential energy, V is the internal potential
energy, F is the nodal loads vector, and U is the displacement
vector. At equilibrium, as the potential energy exists as a
function of displacement and as Castigliano’s first theorem
implies:
∇E = ∇V − F = 0
The double stranded DNA sequence was minimized at
extensive loads of increasing magnitude. The minimization
procedure is nonlinear and equivalent to energy minimization
usually performed before molecular dynamics analysis, though
much more rigorous, as the energy must be at an absolute
minimum (see Figures 9a−9b).
3.4. Sequence Dependent DNA Stretch Modulus. For
the derivation of the modulus of elasticity in vacuum and low
temperature (neglecting entropic effects), a set of 4 base pair
steps is examined. Initially, energy minimization is performed in
order to render the global stiffness matrix positive-definite
(eigenvalues greater or equal to 0) and vanish external forces
4. CONCLUSIONS
In this work a new modeling approach for simulating the
potential energy interactions of molecular nanostructures is
proposed. Finite element method based approaches are usually
employed in solving differential equations in continuum media.
The proposed numerical modeling approach devises the finite
element method in order to develop a discrete model that
accurately captures the interactions at the atomic level. The
new modeling approach relies on the development of
generalized potential energy finite elements and allows taking
(34) into account all types of force fields and parameter sets used to
calculate the potential energy of a system of atoms or coarse-
grained particles. The analytical expression of both gradient
vector and Hessian matrix is the main advantage of the
proposed modeling approach, leading to drastic reduction of
the computational effort required. Two molecular systems are
considered in order to assess the performance of the proposed
(35)
modeling approach: a single butane molecule and double
stranded B-DNA sequences (composed of 2, 10, and 30 base-
pairs).
The fast and efficient performance of the method can find
applications in directly improving the energy minimization
algorithms, especially in cases of rigorous minimization.
Furthermore, the sequence dependent stretching modulus for
four DNA base pair steps was calculated. Finally, one other
application of the proposed formulation when coupled with an
appropriate coarse graining orientation scheme (orientation
and displacement of groups of atoms) is the development of a
systematic definition of coarse grained models. This could be
possibly achieved by coupling already developed methodologies
1975 DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

Figure 11. DNA base pair steps along with deformed configurations (a) AA, (b) AT, (c) GC, and (d) GG.

Table 4. Stretch Modulus for Four Different DNA Base Pair

Steps along with Deformed Configurations51
■ AUTHOR INFORMATION
Corresponding Author
DNA Base Pair Stretch modulus (1st Stretch modulus (2nd *E-mail: nlagaros@central.ntua.gr.
Step method) (pN) method) (pN) Notes
AA 1182 533 The authors declare no competing financial interest.
AT
GC
GG
2785
2605
3525
1108
1235
429
■ ACKNOWLEDGMENTS
The first and third authors would like to thank the people of
the Laboratory for Computational Biology and Biophysics of
of static reduction in the field of structural mechanics with the Department of Biological Engineering at MIT for sharing
existing methods of absolute orientation49 that can help important aspects of the fascinating field of DNA nano-
determine the “stiff” parts of the nanostructure to be technology and for their help in the development of the
presented work.

■
condensed.

■
*
ASSOCIATED CONTENT
S Supporting Information
REFERENCES
(1) Agrawal, P. M.; Sudalayandi, B. S.; Raff, L. M.; Komanduri, R.
The Supporting Information is available free of charge on the Molecular Dynamics (MD) Simulations of the Dependence of C-C
ACS Publications website at DOI: 10.1021/acs.jcim.6b00356. Bond Lengths and Bond Angles on the Tensile Strain in Single-Wall
Carbon Nanotubes (SWCNT). Comput. Mater. Sci. 2008, 41, 450−
Analytical calculation of matrices Bl1 and Bl2 used for 456.
deriving the in-plane ange potential energy finite element (2) Chen, W. H.; Cheng, H. C.; Hsu, Y. C. Mechanical Properties of
(PDF) Carbon Nanotubes Using Molecular Dynamics Simulations with the

1976 DOI: 10.1021/acs.jcim.6b00356

J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling
Inlayer Van Der Waals Interactions. Comput. Model. Eng. Sci. 2007, 20,
123−145.
(3) Cheng, H.-C.; Liu, Y.-L.; Hsu, Y.-C.; Chen, W.-H. Atomistic-
Continuum Modeling for Mechanical Properties of Single-Walled
Carbon Nanotubes. Int. J. Solids Struct. 2009, 46, 1695−1704.
(4) Gao, G.; Ç aǧin, T.; Goddard, W. A., III Energetics, Structure,
Mechanical and Vibrational Properties of Single-Walled Carbon
Nanotubes. Nanotechnology 1998, 9, 184−191.
(5) Hu, N.; Nunoya, K.; Pan, D.; Okabe, T.; Fukunaga, H. Prediction
of Buckling Characteristics of Carbon Nanotubes. Int. J. Solids Struct.
2007, 44, 6535−6550.
(6) Krasheninnikov, A. V.; Banhart, F.; Li, J. X.; Foster, A. S.;
Nieminen, R. M. Stability of Carbon Nanotubes Under Electron
Irradiation: Role of Tube Diameter and Chirality. Phys. Rev. B:
Condens. Matter Mater. Phys. 2005, 72, 125428.
(7) Nasdala, L.; Ernst, G. Development of a 4-Node Finite Element
for the Computation of Nano-Structured Materials. Comput. Mater. Sci.
2005, 33, 443−458.
(8) Odegard, G. M.; Gates, T. S.; Nicholson, L. M.; Wise, K. E.
Equivalent-Continuum Modeling of Nano-Structured Materials.
Compos. Sci. Technol. 2002, 62, 1869−1880.
(9) Wang, Y.; Sun, C.; Sun, X.; Hinkley, J.; Odegard, G. M.; Gates, T.
S. Nano-Scale Finite Element Analysis of Polymer Networks. In 43rd
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and
Materials Conference, Denver, Colorado, USA, April 22−25, 2002.
(10) Yakobson, B. I.; Brabec, C. J.; Bernholc, J. Nanomechanics of
Carbon Tubes: Instabilities Beyond Linear Response. Phys. Rev. Lett.
1996, 76, 2511−2514.
(11) Merli, R.; Lázaro, C.; Monleón, S.; Domingo, A. A Molecular
Structural Mechanics Model Applied to the Static Behavior of Single-
Walled Carbon Nanotubes: New General Formulation. Comput. Struct.
2013, 127, 68−87.
(12) Seeman, N. C.; Kallenbach, N. R. Design of Immobile Nucleic
Acid Junctions. Biophys. J. 1983, 44, 201−209.
(13) Rothemund, P. W. K. Folding DNA to Create Nanoscale Shapes
and Patterns. Nature 2006, 440, 297−302.
(14) Douglas, S. M.; Dietz, H.; Liedl, T.; Hoegberg, B.; Graf, F.; Shih,
W. M. Self-Assembly of DNA into Nanoscale Three-Dimensional
Shapes. Nature 2009, 459, 414−418.
(15) Dietz, H.; Douglas, S. M.; Shih, W. M. Folding DNA into
Twisted and Curved Nanoscale Shapes. Science 2009, 325, 725−730.
(16) Han, D.; Pal, S.; Nangreave, J.; Deng, Z.; Liu, Y.; Yan, H. DNA
Origami with Complex Curvatures in Three-Dimensional Space.
Science 2011, 332, 342−346.
(17) Zhang, F.; Jiang, S.; Wu, S.; Li, Y.; Mao, C.; Liu, Y.; Yan, H.
Complex Wireframe DNA Origami Nanostructures with Multi-Arm
Junction Vertices. Nat. Nanotechnol. 2015, 10, 779−784.
(18) Geary, C.; Rothemund, P. W. K.; Andersen, E. S. A Single-
Stranded Architecture for Cotranscriptional Folding of RNA
Nanostructures. Science 2014, 345, 799−804.
(19) Delebecque, C. J.; Lindner, A. B.; Silver, P. A.; Aldaye, F. A.
Organization of Intracellular Reactions with Rationally Designed RNA
Assemblies. Science 2011, 333, 470−474.
(20) Wei, R.; Martin, T. G.; Rant, U.; Dietz, H. DNA Origami
Gatekeepers for Solid-State Nanopores. Angew. Chem., Int. Ed. 2012,
51, 4864−4867.
(21) Dutta, P. K.; Varghese, R.; Nangreave, J.; Lin, S.; Yan, H.; Liu, Y.
DNA-Directed Artificial Light-Harvesting Antenna. J. Am. Chem. Soc.
2011, 133, 11985−11993.
(22) Pan, K.; Boulais, E.; Yang, L.; Bathe, M. Structure-Based Model
for Light-Harvesting Properties of Nucleic Acid Nanostructures.
Nucleic Acids Res. 2014, 42, 2159−2170.
(23) Mohri, K.; Nishikawa, M.; Takahashi, N.; Shiomi, T.; Matsuoka,
N.; Ogawa, K.; Endo, M.; Hidaka, K.; Sugiyama, H.; Takahashi, Y.
Design and Development of Nanosized DNA Assemblies in Polypod-
Like Structures as Efficient Vehicles for Immunostimulatory CpG
Motifs to Immune Cells. ACS Nano 2012, 6, 5931−5940.
1977
Article
(24) Davis, M. E.; Chen, Z.; Shin, D. M. Nanoparticle Therapeutics:
An Emerging Treatment Modality for Cancer. Nat. Rev. Drug Discovery
2008, 7, 771−782.
(25) Liu, M.; Fu, J.; Hejesen, C.; Yang, Y.; Woodbury, N. W.;
Gothelf, K.; Liu, Y.; Yan, H. A DNA Tweezer-Actuated Enzyme
Nanoreactor. Nat. Commun. 2013, 4, 2127.
(26) Wadati, M.; Tsuru, H. Elastic Model of Looped DNA. Phys. D
1986, 21, 213−226.
(27) Westcott, T. P.; Tobias, I.; Olson, W. K. Elasticity Theory and
Numerical Analysis of DNA Supercoiling: An Application to DNA
Looping. J. Phys. Chem. 1995, 99, 17926−17935.
(28) Purohit, P. K.; Nelson, P. C. Effect of Supercoiling on
Formation of Protein-Mediated DNA Loops. Phys. Rev. E 2006, 74,
061907.
(29) Balaeff, A.; Mahadevan, L.; Schulten, K. Elastic Rod Model of a
DNA Loop in the Lac Operon. Phys. Rev. Lett. 1999, 83, 4900−4903.
(30) Balaeff, A.; Koudella, C. R.; Mahadevan, L.; Schulten, K.
Modelling DNA Loops Using Continuum and Statistical Mechanics.
Philos. Trans. R. Soc., A 2004, 362, 1355−1371.
(31) Balaeff, A.; Mahadevan, L.; Schulten, K. Modelling DNA Loops
Using the Theory of Elasticity. Phys. Rev. E 2006, 73, 031919.
(32) Villa, E.; Balaeff, A.; Mahadevan, L.; Schulten, K. Multiscale
Method for Simulating Protein-DNA Complexes. Multiscale Model.
Simul. 2004, 2, 527−553.
(33) Starostin, E. L. Symmetric Equilibria of a Thin Elastic Rod with
Self-Contacts. Philos. Trans. R. Soc., A 2004, 362, 1317−1334.
(34) Coleman, B. D.; Swigon, D. Theory of Self-Contact in Kirchhoff
Rods with Applications to Supercoiling of Knotted and Unknotted
DNA Plasmids. Philos. Trans. R. Soc., A 2004, 362, 1281−1299.
(35) Finite Element Procedures; Bathe, K.-J., Ed.; Prentice Hall Inc.:
Upper Saddle River, New Jersey, NJ, 1996.
(36) Castro, C. E.; Kilchherr, F.; Kim, D.-N.; Shiao, E. L.; Wauer, T.;
Wortmann, P.; Bathe, M.; Dietz, H. A Primer to Scaffolded DNA
Origami. Nat. Methods 2011, 8, 221−229.
(37) Pan, K.; Kim, D.-N.; Zhang, F.; Adendorff, M. R.; Yan, H.;
Bathe, M. Lattice-Free Prediction of Three-Dimensional Structure of
Programmed DNA Assemblies. Nat. Commun. 2014, 5, 5578.
(38) Kim, D.-N.; Kilchherr, F.; Dietz, H.; Bathe, M. Quantitative
Prediction of 3D Solution Shape and Flexibility of Nucleic Acid
Nanostructures. Nucleic Acids Res. 2012, 40, 2862−2868.
(39) Sedeh, R. S.; Pan, K.; Adendorff, M. R.; Hallatschek, O.; Bathe,
K.-J.; Bathe, M. Computing Nonequilibrium Conformational Dynam-
ics of Structured Nucleic Acid Assemblies. J. Chem. Theory Comput.
2016, 12, 261−273.
(40) Understanding Molecular Simulation: From Algorithms to
Applications; Frenkel, D., Smit, B., Eds.; Academic Press: San Diego,
CA, 1996.
(41) Hayward, S.; de Groot, B. L. Normal Modes and Essential
Dynamics. Methods Mol. Biol. 2008, 443, 89−106.
(42) Elasticity, Theory, Applications and Numerics; Sadd, M. H., Ed.;
Elsevier: Amsterdam, 2009.
(43) Foloppe, N.; MacKerell, A. D. All-Atom Empirical Force Field
for Nucleic Acids: I. Parameter Optimization Based on Small Molecule
and Condensed Phase Macromolecular Target Data. J. Comput. Chem.
2000, 21, 86−104.
(44) MacKerell, A. D.; Banavali, N. K. All-Atom Empirical Force
Field for Nucleic Acids: II. Application to Molecular Dynamics
Simulations of DNA and RNA in Solution. J. Comput. Chem. 2000, 21,
105−120.
(45) MATLAB Primer, R2015b; The Math Works Inc.: 3 Apple Hill
Drive Natick, MA 01760-2098, 2015.
(46) Dhakal, S.; Adendorff, M. R.; Liu, M.; Yan, H.; Bathe, M.;
Walter, N. G. Rational Design of DNA-actuated Enzyme Nanoreactors
Guided by Single Molecule Analysis. Nanoscale 2016, 8, 3125−3137.
(47) Moré, J. J.; Sorensen, D. C. Computing a Trust Region Step.
SIAM J. Sci. Stat. Comp. 1983, 4 (3), 553−572.
(48) Byrd, R. H.; Schnabel, R. B.; Shultz, G. A. Approximate Solution
of the Trust Region Problem by Minimization Over Two-Dimensional
Subspaces. Math. Program. 1988, 40, 247−263.
DOI: 10.1021/acs.jcim.6b00356
J. Chem. Inf. Model. 2016, 56, 1963−1978
Journal of Chemical Information and Modeling Article

(49) Horn, B. K. P. Closed-Form Solution of Absolute Orientation

Using Unit Quaternions. J. Opt. Soc. Am. A 1987, 4 (4), 629−642.
(50) The results are obtained implementing the BFGS method, the
trust region method where the Hessian matrix is computed using the
proposed finite element formulation, and the trust region method
where the Hessian matrix stored is a banded form (11 main diagonals)
of that computed using the proposed finite element formulation (the
potential energy values are in kcal/mol, while the forces are in kcal/(A
mol) = 4.1868E25 pN/mol).
(51) Stretch moduli are derived from static equilibrium by assuming
external forces at the DNA backbone of the top base pair and fixing
the z-direction DOFs of all the bottom base pair atoms. The x,y-
directions for the atoms connected to the backbone of the bottom
base pairs were also fixed. The first number of the stretch modulus is
derived by assuming as initial length the difference of the mean values
of the z-coordinate of the base pairs and as displacement the mean
value of the z-direction displacement of the top base pair (with the
bottom base pair being fixed in this direction). The second number in
parentheses is the stretch modulus derived assuming as initial length
the mean value of the z-coordinate of the distance between the atoms
connecting the backbone for top and bottom base pairs, e.g. the atoms
which are fixed and the atoms where the external forces are applied
and as displacement the mean value of the z-direction displacement of
the two atoms where the external forces are applied.

1978 DOI: 10.1021/acs.jcim.6b00356

J. Chem. Inf. Model. 2016, 56, 1963−1978