Content uploaded by Monique Tirion
Author content
All content in this area was uploaded by Monique Tirion on Nov 19, 2014
Content may be subject to copyright.
Content uploaded by Monique Tirion
Author content
All content in this area was uploaded by Monique Tirion on Oct 08, 2014
Content may be subject to copyright.
VOLUME 77, NUMBER 9 PHYSICAL REVIEW LETTERS 26AUGUST 1996
Large Amplitude Elastic Motions in Proteins from a Single-Parameter, Atomic Analysis
Monique M. Tirion*
Department of Membrane Research and Biophysics, Weizmann Institute of Science, Rehovot 76100, Israel
(Received 22 April 1996)
Normal mode analysis (NMA) is a leading method for studying long-time dynamics and elasticity
of biomolecules. The method proceeds from complex semiempirical potentials characterizing the
covalent and noncovalent interactions between atoms. It is widely accepted that such detailed potentials
are essential to the success of NMA’s. We show that a single-parameter potential is sufficient to
reproduce the slow dynamics in good detail. Costly and inaccurate energy minimizations are eliminated,
permitting direct analysis of crystal coordinates. The technique can be used for new applications, such
as mapping of one crystal form to another by means of slow modes, and studying anomalous dynamics
of large proteins and complexes. [S0031-9007(96)01063-0]
PACS numbers: 87.15.By, 87.15.He
Thermal equilibrium fluctuations of the x-ray crystal
coordinates of proteins provide a basis for understanding
the complex dynamics and elasticity of biological macro-
molecules [1]. Analysis of the normal modes of globu-
lar proteins shows an interesting anomaly. The density
of the slow vibrational modes is proportional to their fre-
quency, gsvd,v, rather than gsvd,v
2
as predicted
by Debye’s theory [2]. Yet, the atoms in globular pro-
teins are packed as tightly as in solids. We show that
a single-parameter potential reproduces the slow elastic
modes of proteins obtained with vastly more complex
empirical potentials. The simplicity of the potential per-
mits greater insight and understanding of the mechanisms
that underlie the slow, anomalous motions in biological
macromolecules such as proteins.
To date, normal modes of globular proteins have been
used to reproduce crystallographic temperature factors [3]
and diffuse scatter [4]. Normal mode analyses (NMA’s)
shed light on shear and hinge motions necessary for cata-
lytic reactions, and have been used with some success to
map one crystal form of a protein into another [5]. Finally,
NMA’s yield macroscopic elastic moduli of large protein
assemblies, based on their microscopic structure [6].
NMA studies of macromolecules are handicapped, how-
ever, by the complex phenomenological potentials used to
model the covalent and nonbonded interactions between
atom pairs. The necessary initial energy minimization re-
quires a great deal of computer time and memory, and
is virtually impossible for even moderately large proteins
(with typically thousands of degrees of freedom) with a
reasonable degree of accuracy. This inevitably leads to
unstable modes which must be eliminated through elabo-
rate methods, and which cast doubts on the validity of the
analysis. Moreover, partly because the minimization is
carried out in vacuo, the final configuration disagrees with
the known crystallographic structure, complicating the in-
terpretation of the results of NMA.
A typical example of a semiempirical potential used in
molecular dynamics studies and NMA’s has the form [7]
E
p
1
2
X
bonds
K
b
sb 2 b
0
d
2
1
1
2
X
angles
K
u
su2u
0
d
2
1
1
2
X
dihedrals
K
f
f1 1 cossnf2ddg
1
X
nonbonded pairs
∑
A
r
12
2
B
r
6
1
q
1
q
2
Dr
∏
. (1)
The first three terms describe the energy cost in the
distortion of bond lengths, bond angles, and dihedral
angles, and the last term represents steric repulsions,
van der Waals attractions, and electrostatic interactions
between nonbonded atoms. The various bonded con-
stants, K
b
, b
0
, K
u
, etc., are specific for each type of co-
valent interaction, and the nonbonded constants, A and
B, are specific for every type of interacting atom pairs.
These constants are carefully determined from extensive
theoretical and experimental studies (see Ref. [7]).
In this work, I show that in some cases the usual sophis-
ticated potentials may be replaced by a far simpler pair-
wise Hookean potential, controlled by a single parameter.
Such a formulation is sufficient to fully describe the
anomalous low-frequency motion of large globular pro-
teins, including time scales and eigenfrequencies, as well
as displacements of atoms as predicted by eigenvectors.
The simplified potential provides a very attractive alter-
native for the NMA of macromolecular assemblies. The
derivation of the eigenvalue equation is simple, rapid, and
accurate. Time-consuming and structure-distorting en-
ergy minimization are circumvented, preventing unphysi-
cal instabilities (negative eigenvalues), and the reduction
in the number of fitting parameters yields to the new for-
mulation a stronger predictive value.
The size of the computation depends on the number
of internal coordinates considered. A molecule consist-
ing of N atoms possesses 3N degrees of freedom, defined
by bond lengths, bond angles, and rotations about bonds
(or, alternatively, by the Cartesian coordinates of each of
the constituent atoms). For long-chain molecules such as
0031-9007y96y77(9)y1905(4)$10.00 © 1996 The American Physical Society 1905
VOLUME 77, NUMBER 9 PHYSICAL REVIEW LETTERS 26AUGUST 1996
proteins, bond lengths and angles are constrained within
very narrow limits by the chemical bonding, while ro-
tations about (single) bonds are much less restricted.
Hence, for the analysis of slow modes one typically con-
siders the bond lengths and angles as fixed, permitting
only rotations about bonds. The latter are known as dihe-
dral angles and serve as a convenient set of generalized
degrees of freedom. The number of the dihedral angle co-
ordinates is about Ny2 and is determined by the specific
makeup of the protein in question.
Given a potential energy function E
p
and a set of
n generalized coordinates, q
0
, one first must find a
stable local minimum E
p,min
sq
0
q
0
0
d. The potential
energy is then approximated by the quadratic form: E
p
s1y2d
P
q
i
F
ij
q
j
s1y2dq
y
F
q
sq ; q
0
2 q
0
0
d, where F is
the generalized force matrix
F
i,j
≠
2
E
p
≠q
i
≠q
j
Ç
q0
.
Similarly, the kinetic energy is expressed as E
k
s1y2d 3
Ù
q
y
H
Ù
q, where the elements of the “mass” matrix H are
H
i,j
N
X
l1
m
l
≠r
l
≠q
i
?
≠r
l
≠q
j
.
r
l
are the Cartesian coordinates of atom l, and the sum-
mation runs over all the atoms of the molecule. The
h≠ry≠qj are moving derivatives which eliminate transla-
tional and rotational motion of the molecule as a whole.
Equations of motion are derived from Lagrange’s
equation, with the Lagrangian L E
k
2 E
p
. Writing
q
j
P
N
k
A
jk
a
k
cossv
k
t 1d
k
d, one obtains the eigen-
value problem
FA LHA , (2)
subject to the normalization condition A
y
HA I. (This
ensures that the eigenmodes diagonalize the system’s
Hamiltonian, H E
k
1 E
p
.) The eigenfrequencies v
i
are given by the elements of the diagonal matrix L, v
2
i
L
ii
, the eigenvectors are the columns of the matrix A, and
the amplitudes and phases, a
k
and d
k
, are determined by
initial conditions.
I replace the habitual detailed potentials, such as
the one in Eq. (1), by the Hookean pairwise potential
(between atoms a and b):
Esr
a
, r
b
d
C
2
sjr
a,b
j 2 jr
0
a,b
jd
2
. (3)
Here r
a,b
; r
a
2 r
b
denotes the vector connecting
atoms a and b, and the zero superscript indicates the
given initial configuration. Thus, the usual minimization
of the potential energy is eliminated.
Expanding to second order about r
0
a,b
yields
Esr
a
, r
b
d
C
2
µ
r
0
a,b
?Dr
a,b
jr
0
a,b
j
∂
2
, (4)
where Dr ; r 2 r
0
. The strength of the potential C is
a phenomenological constant, assumed to be the same for
all interacting pairs.
The potential energy within a molecule is then given by
E
p
X
sa,bd
Esr
a
, r
b
d . (5)
The sum is restricted to atom pairs separated by less
than R
vdW
sad 1 R
vdW
sbd 1 R
c
, where R
vdW
refers to
the van der Waals radii, and R
c
is an arbitrary cutoff
parameter which models the decay of interactions with
distance. R
c
determines the total number of interacting
atom pairs contributing to the potential energy of the
system, and is inversely related to the “bond strength” C.
We shall argue below that best results are obtained with
small cutoff distances. Fortunately, this tends to reduce
the size of the computation.
The proposed potential seems too simplistic. For ex-
ample, it is unclear how interactions between nonbonded
atoms could model the 3-state “knob” potential for ro-
tations about the backbone. The answer lies in the fact
that slow vibrational modes involve coherent motion of
large groups of atoms. The effective force opposing large
scale oscillations stems from the combined effect of nu-
merous interacting atom pairs. The sum of these interac-
tions approaches a universal form, governed by the central
limit theorem, regardless of the details of individual pair-
potentials. Hence, for slow vibrations these details could
be neglected.
To test this hypothesis, I compare the eigenfrequency
and eigenvector data obtained using the potential of
Eqs. (4) and (5) and the detailed potential of Levitt, L79
[7]. In all cases, I refer only to the slowest frequency
modes. The higher frequency modes, pertaining to rapid
oscillations of sidechains and small groups of atoms,
require an accurate analysis at the microscopic level and
could not be modeled by simplified potentials.
I performed extensive tests on the muscle protein,
G-actin. This system has a molecular weight of 44 kD
and contains 375 residues (3500 atoms) in a single
polypeptide chain [8]. The polypeptide chain is folded
so as to form two large domains joined by a narrow neck
region. These two domains are partly held together by
salt bridges and hydrogen bonds provided by a nucleotide
(ADP) and a cation sCa
11
d bound in the cleft between the
two domains. Complete eigenfrequency and eigenvector
data exist for this system using the L79 potential [9].
Focusing first on the density of modes, let us examine
the cumulative density of modes up to frequency v,
Gsvd n
21
R
v
0
dv
0
gsv
0
d. In globular proteins Gsvd
falls under a universal curve [2]. For small frequencies,
Gsvd,v
2
. The difference from regular crystals, where
Gsvd,v
3
, reflects the anomalous dynamics of slow
vibrations in proteins.
Figure 1 shows Gsvd against v for G-actin:ADP:Ca
11
for the slowest 10% of the modes (138 modes). The
dashed curve refers to data obtained using the standard
L79 potential. Superposed are curves obtained with
R
c
1.1, 1.5, 2.0, and 2.5 Å, resulting in 19248, 27310,
1906
VOLUME 77, NUMBER 9 PHYSICAL REVIEW LETTERS 26AUGUST 1996
FIG. 1. The fraction of the total number of modes up
to frequency v scm
21
d for the slowest 150 modes of the
G-actin:ADP:Ca
11
system. The dashed line pertains to data
obtained using the L79 potential, while the four solid curves
are obtained using R
C
values of 1.1, 1.5, 2.0, and 2.5 Å.
The R
c
1.1 Å curve is nearest the dashed line at higher
frequencies, with the fit progressively worsening for the higher
cutoff values.
38654, and 51020 nonbonded interactions, respectively.
To obtain optimal fits to the standard (dashed) curve,
the values of C need to be adjusted to 2.49, 1.29,
0.73, and 0.47 kJyÅ
2
mol, respectively. Curiously, C
seems to scale as 1yR
2
c
rather than 1yR
3
c
. This may
reflect the unusual spectral dimension, i.e., the effective
dimensionality of interatomic interactions, of globular
proteins, d
s
2 [2]. The product CR
2
c
results in a
universal “bond-strength” constant of about 3.0 kJymol.
Figure 1 also shows that larger cutoff values increase
the curvature of the curve, and hence smaller R
c
val-
ues better model the previous eigenvalue data. This em-
phasizes that nonbonded interactions are dominated by
nearest-neighbor interactions, presumably due to screen-
ing effects.
In addition to the frequency of the modes, one can
compute the root mean square (rms) displacements of
atoms from equilibrium at room temperature from the
eigenvector data. Figure 2 shows the rms fluctuations of
all the C
a
atoms as a function of the slowest 30 modes.
The dashed curve shows the data obtained using the
standard, L79 potential. The rms deviations decrease
rapidly with mode number, indicating that the correlation
length of the motion decreases as well. Superposed on
the dashed curve are the four curves obtained with the
same R
c
and C values used in Fig. 1. Values of C
predicted from the eigenfrequency data, Fig. 1, fit the
eigenvector data equally well. These data also indicate
no clear advantage to using larger values of the cutoff
parameter, R
c
. For the slowest modes, therefore, there is
overall consistency and a very good match of the current
FIG. 2. The rms deviation of all mainchain C
a
atoms per
mode, for the slowest 30 modes. The dashed line refers to data
obtained using the L79 potential, and the four solid curves are
obtained with the same cutoff values, R
c
, as in Fig. 1. The rms
fluctuations due to the first four modes contribute over 50% to
the total rms deviations of all C
a
due to all modes [8]. There
is no obvious improvement in choosing one particular cutoff
parameter according to these data.
results with those obtained previously, in terms of both
dispersion spectra as well as rms deviations.
Theoretical temperature factors, B, used to model
x-ray crystallographic temperature factors, are obtained
by computing the rms fluctuations at room temperature
of each C
a
due to a superposition of modes. Figure 3
shows the theoretical temperature factors for each C
a
of
FIG. 3. Comparison of theoretical temperature factors, B,
obtained with the L79 potential (dashed curve) and the
potential of Eq. (1), for the G-actin:ADP:Ca
11
system. The
contributions of the 30 slowest modes are included. The inset
shows the scatter plot of the two data sets: the standard potential
along the ordinate, and the current simplified potential along the
abscissa.
1907
VOLUME 77, NUMBER 9 PHYSICAL REVIEW LETTERS 26AUGUST 1996
TABLE I. CPU time requirements to compute generalized eigenvalue equations.
pdb # of # of # of nonbonded CPU time
Protein identifier residues coordinates interactions (min)
Crambin 1crn 46 139 4817 0.12
Trypsin inhibitor 5pti 58 208 6529 0.45
Ribonuclease A 5rsa 124 455 14946 2.50
Lysozyme 6lyz 129 471 15834 3.10
G-actin 1atn 372 1382 37951 66.0
Myosin (HC) 1mys 780 3010 88653 390
G-actin:ADP:Ca
11
, including the contribution of the 30
slowest modes. The data obtained with the L79 potential
are shown by the dashed curve. The solid line is
obtained using the current potential, with R
c
2.2 Å.
The very good fit argues against the need for additional
parametrization in Eqs. (3)–(5).
Finally, I have tested the efficiency of NMA with the
simple potential. Table I shows the central processing
unit (CPU) times required for reading in coordinates and
chemical formulas, indexing degrees of freedom and non-
bonded interactions, and setting up the generalized eigen-
value Eq. (2). The entries show the CPU requirements
on a Convex 220 for the following four proteins (the
Brookhaven protein database label is given in parenthe-
ses): bovine pancreatic trypsin inhibitor (5pti), ribonucle-
ase A (5rsa), G-actin bound with ADP and Ca
11
(1atn),
and myosin subfragment 1 bound with ADP (lmys, with
the sidechains kindly modeled by Michael Lorenz). These
proteins range in size from 58 residues (trypsin inhibitor)
to 780 residues (myosin S1). For this table, I used a cut-
off distance R
c
2.0 Å. These CPU times represent im-
provements of 2 to 3 orders of magnitudes over earlier
NMA. The main effect is due to the absence of mini-
mizations and the faster computation of the force matrix F
with the simple potential. It should be stressed that NMA
of systems as large and poorly resolved as myosin S1
could not be undertaken with “standard” potentials, due to
the accumulation of roundoff errors and distortions during
minimizations.
This work demonstrates the surprising result that a
single-parameter model can reproduce complex vibra-
tional properties of macromolecular systems. The simple
form of the potential dispenses with the need to perform
initial energy minimizations, which are especially detri-
mental for NMA’s due to the absence of solvent. Since
the analysis proceeds directly from the crystal coordinates,
it is now possible to quantitatively test whether two crystal
forms of a protein, as in an “open” and “closed” configu-
ration, are interconvertible using the slow modes as coor-
dinates. Tests performed on a periplasmic maltodextrin
binding protein indicate that the slowest modes do indeed
closely map the open form into the closed form [Tirion, in
preparation].
I thank Daniel ben-Avraham for useful discussions and
ideas, Ken Holmes for use of computer facilities at the
Max-Planck Institute for Medical Research in Heidelberg,
and Michael Lorenz for the full coordinates of myosin S1.
This material is based upon work supported by the
National Science Foundation under Grant No. MCB-
9316109.
*Electronic address: mmt@craft.camp.clarkson.edu
[1] M. Levitt, P. Stern, and C. Sander, J. Mol. Biol. 181,
423 (1985); A. Kidera and N. G
¯
o, J. Mol. Biol. 225, 457
(1992); B. Brooks and M. Karplus, Proc. Natl. Acad. Sci.
U.S.A. 80, 6571 (1983).
[2] D. ben-Avraham, Phys. Rev. B 47, 14559 (1993).
[3] R. Diamond, Acta Crystallogr. A46, 425 (1990);
A. Kidera and N. G¯o, J. Mol. Biol. 225, 457 (1992).
[4] P. Faure et al., Nature Struct. Biol. 1, 124 (1994).
[5] O. Marques and Y.-H. Sanejouand, Proteins 33, 557
(1995); M.M. Tirion, D. ben-Avraham, M. Lorenz, and
K.C. Holmes, Biophys. J. 68, 5 (1995).
[6] D. ben-Avraham and M. M. Tirion, Biophys. J. 68, 1231
(1995).
[7] S. J. Weiner et al., J. Am. Chem. Soc. 106, 765 (1984);
B.R. Brooks et al., J. Comp. Chem. 4, 187 (1983);
M. Levitt, J. Mol. Biol. 168, 595 (1983).
[8] W. Kabsch et al., Nature (London) 347, 37 (1990).
[9] M. M. Tirion and D. ben-Avraham, J. Mol. Biol. 230, 186
(1993).
1908
