Acc. Chem. Res.

2002, 35, 455-464

electronic degrees of freedom are integrated out, ab initio

The Role and Perspective of Ab MD treats the electronic states explicitly and follows their
Initio Molecular Dynamics in the evolution as the nuclei move.
This finer level of description demands a much larger
Study of Biological Systems computational cost, which severely restricts the length and
time scale of the systems that can be described with the
PAOLO CARLONI,*,†
ab initio approach. This goes against the well-justified
URSULA ROTHLISBERGER,‡ AND trend of biomolecular simulations to study larger systems
MICHELE PARRINELLO‡,§ for longer times. Nonetheless, the experience accumulated
International School for Advanced Studies (SISSA/ISAS) and
Istituto Nazionale per la Fisica della Materia (INFM), via
in recent years has shown that ab initio MD can play a
Beirut 4, I-34014, Trieste, Italy, Department of Chemistry, significant and unique role in biomolecular simulations.
ETH-Hönggerberg HCI, CH-8093 Zurich, Switzerland, and This is due to the fact that there is a large class of
CSCS-Centro Svizzero di Calcolo Scientifico, Via Cantonale, phenomena that depend on the electronic states in such
CH-6928 Manno, Switzerland an intricate way that they cannot be modeled via effective
Received October 10, 2001 potentials.
Despite the usefulness of effective potentials, they are
ABSTRACT sometimes based on rather drastic approximations, such
Ab initio molecular dynamics (MD) allows realistic simulations to as the average way in which polarization effects are
be performed without adjustable parameters. In recent years, the treated. This leads to difficulties in the presence of strongly
technique has been used on an increasing number of applications
to biochemical systems. Here we describe the principles on which polarizing fields, such as those encountered in the sugar-
ab initio MD is based. We focus on the most popular implementa- phosphate backbone of nucleic acids or in the presence
tion, based on density functional theory and plane wave basis set. of highly charged counterions. In empirical models, the
By a survey of recent applications, we show that despite the current intramolecular conformation is determined by potentials
limitations of size and time scale, ab initio MD (and hybrid ab initio
MD/MM approaches) can play an important role for the modeling
that are a combination of pairwise angular and dihedral
of biological systems. Finally, we provide a perspective for the forces. Such parametrization does not lead to stable
advancement of methodological approaches which may further aromatic ring structures. These have to be enforced with
expand the scope of ab initio MD in biomolecular modeling. the use of, for example, improper potentials that are
meant to mimic the stabilization effect of aromatic π
systems. This type of constraint might at times introduce
1. Introduction excessive rigidity. This, together with the problem of
Molecular dynamics (MD) simulations based on the correctly capturing polarization effects, might play a role
description of interatomic interactions via empirical force for the well-known difficulty of effective potentials in
fields have a long and successful record in the study of describing systems, such as RNA and DNA. Similar
biomolecular systems.1 Over the years, ever larger systems problems are expected in the simulation of ion channels
have been studied for simulation times that now reach where large electric fields are present. The continuous
the 100-ns range. In fact, the development of realistic force increase in computational power makes ab initio MD an
fields is considered as one of the major achievements of increasingly attractive approach to help the theoretical
present-day theoretical chemistry. investigations of these systems.
Still, despite the successes, there are many areas in Another area in which traditional modeling has dif-
which the use of effective potentials may be not appropri- ficulty is in the description of metallic centers, such as
ate, and more sophisticated and accurate approaches are those that are encountered in many enzymes and metal/
needed. An alternative is offered by ab initio molecular DNA adducts. Here subtle chemical phenomena come
dynamics,2 which was developed in the early eighties and into play, such as the ligand field, and the fact that the
has been intensively applied in materials science and in metal ion ligand bond has a partially covalent nature. In
chemical problems. At variance with the standard ap- a dynamical environment, the number of ligands and the
proach, in ab initio MD, the interatomic forces are not metal oxidation state can change. This complexity cannot
empirically derived, but are evaluated from electronic easily be modeled, and no general consensus has yet been
structure calculations as the simulation proceeds. Typi- reached on the functional form to be used in order to
cally, use is made of density functional theory,3 but other describe metal centers.4 In the case of most transition
electronic structure methods have also been suggested. metal ions, several spin states are possible, and the ion
In contrast to the standard approach, in which the size and its chemistry can correspondingly vary consider-
ably. Subtle chemical effects can lead to configurations
that have very similar structural properties but rather
Michele Parrinello is Professor of Computational Sciences at ETH, Zurich, and different energies. Since the interaction potentials are
Director of the Swiss Center for Scientific Computing in Manno, Switzerland.
Both Paolo Carloni and Ursula Roethlisberger are former postdoctoral associates. † International School for Advanced Studies and Istituto Nazionale per
Ursula Rothlisberger is now Assistant Professor of Inorganic Chemistry, also at la Fisica della Materia.
ETH. Paolo Carloni is now Associate Professor of Chemistry at the International ‡ ETH-Hönggerberg.

School for Advanced Studies in Trieste, Italy. § CSCS-Centro Svizzero di Calcolo Scientifico.

10.1021/ar010018u CCC: $22.00  2002 American Chemical Society VOL. 35, NO. 6, 2002 / ACCOUNTS OF CHEMICAL RESEARCH 455
Published on Web 05/15/2002
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

functions of the nuclear coordinates only, it would be overwhelming majority of ab initio MD are based on DFT.
extremely difficult to discriminate among these effects. In the Kohn and Sham (KS) formulation of DFT,3
These are certainly important motivations for using ab
initio MD, but an even more cogent reason for its use BI) ) minψiEKS[ψi;R
E0(R BI] (1)
comes from the need to describe properly chemical bond
breaking-bond forming processes. Examples include BI] is the KS energy density functional which
where EKS[ψi;R
enzymatic reactions, for which, indeed, several investiga- depends on both the electronic [ψi] and the nuclear [R BI]-
tions have already been reported in the literature,5 and degrees of freedom.
proton-transfer processes.6 We recall also that crucial to
modeling is the determination of the protonation state, 1
information that can only rarely be obtained from the BI ] ) -
EKS[ψi;R
2
∑i ∫ dbr ψ*i(b)∇ψ
r i(b)
r +
experiments. Ab initio MD has the accuracy needed to 1 1
discriminate among different protonation states and ∫ dbr VN(b)F(
r b) r + ∫ db
2
r db′
r F(b)
r
|b
r - b′|
r
F(b′)
r +
naturally allows it to vary.
An important bonus of ab initio MD is that it allows Exc[F(b)]
r (2)
close contact to be made with the experimental data, such
as IR, Raman,7 and NMR8 spectroscopy data. These In eq 2, EKS[ψi;RBI] is decomposed into a kinetic energy
quantities can be calculated from the available informa- term, a term that takes into account the interaction with
tion on the electronic structure without the need for the nuclei; the Hartree term; and Exc, the exchange
making additional assumptions. Furthermore, all nonhar- correlation energy functional, which contains all of the
monic effects can be taken into account. Recent develop- intricacies of the many-body problem.3
ments in the theory, such as the introduction of DFT- The electronic density F(r b) is written in terms of the
based methods for excited states such as ROKS9 or occupied KS orbitals ψi(r b) as
TDDFT,3 are beginning to permit the study of photo-
chemical processes. F(b)
r )2 ∑i ψ*i(b)ψ
r i(b)
r (3)
In view of these potential advantages, as the compu-
tational resources have increased, so have the number of The exact form of Exc[F(r b)] is not known, but excellent
applications of ab initio MD to biological problems. Since approximations are available. The minimum has to be
exhaustive reviews of applications of ab initio MD work sought subject to the orthogonality condition ∫ dr b
to biochemical systems have already been published,5 we ψ/i (r
b)ψj(r
b) ) δi,j.
will discuss only a few paradigmatic examples that il- To solve numerically the minimum problem of eq 1,
lustrate the usefulness of the ab initio MD approach in we need to introduce a basis set in which to expand the
biology. orbitals. Schemes for ab initio MD based on the traditional
quantum chemistry (QC) approach of using Gaussian
2. The Method basis sets have been implemented, but generally, the most
In the course of an MD run, the interatomic forces need popular choice until now has been to expand the ψi(r b) in
to be evaluated a rather large number of times. It is, plane waves (PW).10 Such a basis set is unusual in QC,
therefore, imperative in any type of MD to make this part but it is very popular in condensed matter theory. There
of the calculation as efficient as possible. This is even more are many advantages to using a PW basis set. PW are
important in ab initio MD, where the interatomic potential totally unbiased, they automatically enforce the periodic
is calculated as boundary conditions that are usually employed in ab initio
MD, they do not introduce basis set superposition errors,
BI) ) E0(R
Φ(R BI) + VNN(R
B I) and FFT routines allow very efficient numerical imple-
mentation. Their major drawback is the very large number
where VNN(R BI) is the direct internuclear interaction and of PWs that is needed in order to achieve a satisfactory
BI) is the electronic ground-state energy evaluated at
E0(R b). Very great help in reducing the
representation of the ψi(r
fixed nuclear positions R BI. Here, we are assuming the number of PW comes from the realization that there are
validity of the Born-Oppenheimer approximation. Clearly two length scales in the problem, a short one associated
computing E0(R BI) is the most demanding part of the with the core electrons, which vary very rapidly, and a
calculation. Since an exact evaluation of E0(R BI) is not larger one associated with the valence electrons. The latter
possible except for the smallest systems, it is necessary is the one determining the chemical behavior, while the
to resort to some form of approximation. If one wishes to core electrons remain essentially unchanged during chemi-
apply ab initio MD to large systems, one needs to strike a cal processes and their effect can be expressed in terms
balance between accuracy and computational efficiency, of relatively smooth pseudopotentials acting only on the
since high accuracy such as that provided by high-level valence electrons.10 Thus, in this approach, the core is
quantum chemical approaches, such as MP2 or coupled treated by these pseudopotentials, and only the valence
cluster theory, is too expensive for use in ab initio MD. orbitals are treated with plane waves. The number of the
One scheme that provides useful accuracy at affordable latter depends on the type of the pseudopotential and the
cost is density functional theory (DFT),3 and in fact, the size of the simulation box that is used.10
456 ACCOUNTS OF CHEMICAL RESEARCH / VOL. 35, NO. 6, 2002
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

During the simulation, the KS minimum has to be treated in an accurate and consistent way, in particular
determined several times. To this effect, efficient optimi- in combination with a plane-wave-based Car-Parrinello
zation methods have been put into place.10 Alternatively, scheme. Several mixed QM/MM Car-Parrinello methods
a very stable dynamics can be obtained by propagating have recently been implemented. In the fully Hamiltonian
ions and electrons simultaneously according to the coupling scheme of ref 13, bonds between the QM and
Lagrangian,3 MM parts of the system are treated with specifically
designed monovalent pseudopotentials, whereas the re-
1 1
L) µ
2
∑i ∫ dbr ψ̇*i(b) r + ∑ MIR
r ψ̇j(b)
2 I
B˙ 2I - EKS[ψI;R
BI] - maining bonding interactions of the interface region, that
is, the angle bending and dihedral distortions, are de-

∑ Λi,j(∫ db
scribed on the level of the classical force field. The same
r ψ*i(b)ψ
r j(b)
r - δi,j) (5)
i,j
holds for the van der Waals interactions between the QM
and MM parts of the system. The electrostatic effects of
where µ is a fictitious classical kinetic energy term, MI the the classical environment, on the other hand, are taken
nuclear masses, and the Lagrange parameters Λi,j enforce into account in the quantum mechanical description as
orthogonality. The ensuing second-order dynamics is very an additional contribution to the external field of the
stable, and provided that the frequency xEG/µ, where EG quantum system,
is the HOMO-LUMO gap, is much higher than the
nuclear frequency, the nuclear dynamics is very close to
the Born-Oppenheimer evolution. These conditions guar-
Hele
QM-MM ) ∑ qi ∫ dr F(r)vi(|r - ri|)
i∈ MM
(8)
antee an adiabatic decoupling of the nuclear and elec-
tronic degrees of freedom, and the electrons follow the where qi is the classical point charge located at ri and
nuclear motion while remaining all the time close to the vi(|r - ri|) is a Coulombic interaction potential modified
ground state. at short range in such a way as to avoid spill-out of the
The overall scaling of this algorithm is O(N3), where N electron density to nearby positively charged classical
is the number of electrons. This scaling is less than that point charges.13 In the context of a plane-wave-based
of more elaborate QC schemes, but it still presents an Car-Parrinello scheme, a direct evaluation of eq 10 is
enormous barrier toward application to systems as large prohibitive, because it involves of the order of NrNMM
as those of interest to biology. To circumvent the O(N3) operations, where Nr is the number of real space grid
barrier, a host of O(N) methods have been introduced that points (typically, ∼1003), and NMM is the number of
exploit the locality of the chemical bond.11 The usefulness classical atoms (usually of the order of 10 000 or more in
of such methods for large-scale accurate calculations, systems of biochemical relevance). Therefore, the inter-
however, remains to be demonstrated. action between the QM system and the more distant MM
An alternative approach that is increasingly being atoms is included with a Hamiltonian term explicitly
applied is that of partitioning the whole system into a coupling the multipole moments of the quantum charge
chemically active region and the rest, which is treated with distribution with the classical point charges. In this way,
empirical potentials.12 This is the so-called quantum efficient and consistent QM/MM Car-Parrinello simula-
mechanics/molecular mechanics (QM/MM) approach tions of complex extended systems can be performed in
based on a mixed Hamiltonian of the form which the steric and electrostatic effects of the surround-
ings are taken explicitly into account.
H ) HQM + HMM + HQM-MM (6)
HQM represents the ab initio Hamiltonian, and HMM is the 3. Applications
purely classical Hamiltonian, which is described by a
In the following Sections, we present a few selected
standard biomolecular force field,
examples of Car-Parrinello applications to biological
HMM ) HbMM + Hnb (7) systems of special current interest. These are meant to
MM
whet your appetite and illustrate some of the particular
HbMM takes into account intramolecular interactions advantages and possibilities of such an ab initio approach
(stretching, bending and torsional terms), whereas Hnb MM to biology. The first is an investigation of the binding of
includes the van der Waals and electrostatic interactions. copper ions to the murine prion protein, which provides
Such a hierarchical approach has the advantage that a convincing illustration of the intricate nature of the
the computational effort can be concentrated on the part transition metal-protein interactions that would be dif-
of the system where it is most needed, whereas the effects ficult to capture in a simplified empirical interaction
of the surroundings are taken into account with a more potential. The second illustrates the precious benefit due
expedient model. In this way, large biomolecular systems to the explicit treatment of the electronic degrees of
can be treated in a computationally efficient manner and freedom, which allows for a direct simulation of the bond-
for longer time scales. The pitfalls of QM/MM methods breaking and -forming processes that occur in chemical
lie in the challenge of finding an appropriate and rigorous reactions. This special feature can be used to study, for
treatment of the coupling between QM and MM regions, example, the occurrence of low-barrier hydrogen bond
as described by the interaction Hamiltonian HQM-MM. (LBHB) or enzymatic mechanisms of the anti-AIDS target
Special care has to be taken that the QM/MM interface is HIV-1 protease. The last two parts give a short glimpse
VOL. 35, NO. 6, 2002 / ACCOUNTS OF CHEMICAL RESEARCH 457
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

FIGURE 1. Sequence of snapshots from a mixed QM/MM Car-Parrinello simulation of a potential Cu2+ binding site in the mouse prion
protein.17 The QM part is shown in cylinders, and the classical environment is given in a stick representation.18 The copper ion is shown in
yellow. The total length of the simulation is ∼5 ps. Note the significant dynamical changes that occur in the copper coordination sphere.

into recent extensions of the Car-Parrinello approach for of the bonding properties of the metal necessitates the
the in situ calculation of NMR chemical shifts and the use of a quantum mechanical electronic structure ap-
investigation of excited states. proach. At the same time, the protein environment and
1. Copper-Binding Sites in Prions. Prions are infec- solvent have to be taken into account as well as dynamical
tious agents that play a central role for a group of finite temperature effects. A QM/MM approach appears,
invariably fatal, neurodegenerative diseases affecting ani- therefore, to be the method of choice. However, since it
mals such as sheep (scrapie), cattle (BSE), and humans is a priori not known which ligands are involved in
(e.g. variants of the Creutzfeldt-Jacob disease). It is now binding, the necessary partitioning of the system into QM
widely established that these diseases are caused by an and MM regions is far from trivial. To circumvent this
abnormal isoform PrPSc of the normal cellular prion problem, we first performed a statistical analysis of all
protein PrPC..14 Recent experiments, both in vivo and in available high-resolution (e2 Å) crystal structures of
vitro, indicate that prions are able to bind metal ions, in copper proteins to determine the relative propensity of
particular Cu2+.15 It has generally been assumed that Cu2+ different amino acids for copper ligation. Using this
ions bind to the histidine-rich unstructured part.15 How- probability map, we scanned the structure of the mouse
ever, recent pulse EPR and ENDOR experiments15 suggest prion protein for the most likely Cu2+ binding regions. The
that copper binds with higher affinity in the structured
resulting candidate structures were ranked by increasing
part (amino acid residues 125-228). Three pH-dependent
probability, and their stability was subsequently tested in
copper binding sites have been identified together with
QM/MM Car-Parrinello simulations of several picosec-
some information about the nature of the immediate
onds around room temperature (300 K).
copper coordination sphere. The exact position of the
metal binding sites within the protein, however, could not The starting point of the simulations was the experi-
be localized experimentally. mentally determined NMR structure, which was immersed
An identification of the copper binding sites in the in a box of water and equilibrated by classical MD
structured part of prion proteins and a detailed charac- simulations based on the GROMOS96 force field.1c Sub-
terization of the metal protein interactions are of great sequently, a Cu2+ ion was placed in the vicinity of selected
interest, because it has been suggested that metal ions residues of the QM region, and the exact location of the
could influence the structural stability and, potentially, Cu2+ ion and its coordination were determined in mixed
also the transition to the infectious scrapie form. More- QM/MM Car-Parrinello simulations. Characteristic snap-
over, the interplay with metal ions is a common denomi- shots of one of these dynamics runs are shown in Figure
nator for several other neurodegenerative disorders such 1. This possible binding site involves a Met, a His, and an
as Alzheimer’s, Parkinson’s, and amylotrophic sclerosis, Asp residue. As is apparent from Figure 1, the system
and there is increasing evidence that the interaction with undergoes a number of changes in the coordination
metal ions plays a crucial role in all of them.16 sphere and the assumed ligand geometry. In fact, in
Starting from the available NMR structure of the metal- several simulations, the QM region of the system had to
free mouse PrPC,17 we have performed mixed QM/MM be adapted dynamically in such a way as to allow for
Car-Parrinello simulations with the aim of identifying coordination with alternative ligands whose potential
likely locations for Cu2+ binding.18 participation emerged only during the course of the
Locating (transition) metal ions in a protein structure simulation. This illustrates the highly complex and
is a particularly challenging task, since the intricate nature dynamical nature of the metal binding properties, which
458 ACCOUNTS OF CHEMICAL RESEARCH / VOL. 35, NO. 6, 2002
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

FIGURE 2. (a) Three-dimensional structure of HIV-1 protease.19 This enzyme, which cleaves peptide segments at specific locations, is essential
for the virus’ metabolism.19 The enzyme is composed of two identical subunits. A C2 symmetry axis passes through the cleavage site, which
consists of the Asp catalytic dyad (in red). (b) LBHB between the Asp pair in the free enzyme, as obtained by ab initio MD simulations on
model systems.21
are not easily captured in a simplified empirical force field pears to be consistent with site-directed mutagenesis
description. experiments on this position.21
2. HIV-1 Protease Cleavage Site. The protease from From the ab initio MD simulations, we have an
the human immunodeficiency virus type 1 (HIV-1 PR) is indication that several ingredients, such as the polarization
one of the major targets against the AIDS epidemic.19 The forces, the treatment of bond-forming/breaking processes,
enzyme is a dimer composed of two identical subunits and temperature effects, play a crucial role in the HIV-1
(Figure 2a). The cleavage site consists of two aspartyl PR active site. On the basis of these findings, specific force
residues (Asp25 and Asp25′) located at the subunit/ fields could now be developed for this system, which in
subunit interface. The two groups are almost coplanar and turn might allow for a more accurate modeling of HIV-1
rather close to each other. PR-drug interactions with empirical force-fields.
Force-field-based MD methods have encountered un- 3. Reaction Mechanism of HIV-1 PR. The catalytic
expected difficulties in describing the conformational power of enzymes is believed to result from the presence
properties of the cleavage site,20 which are believed to be of a properly engineered electric field23 at the active site.
crucial for drug binding and enzymatic function. As a
In HIV-1 PR, this view appears to be challenged by the
result, ad-hoc models imposing geometric constraints
emergence of “compensatory” drug-resistant mutations,
have been introduced.20 These models succeed in repro-
which may be located either close to or far from the active
ducing the structural properties of the cleavage site,
site. These mutations, which are usually associated with
although they do not provide the physicochemical origin
mutations in the region of the active site, are believed to
of the stability in the active site.
enhance the enzymatic reaction.19 In many cases, these
Ab initio MD simulations on models of the active site
mutations involve chemically similar residues (such as
have been used to provide an understanding of the
M46I, L47V, L63P, A71T), which cannot significantly
molecular interactions between the Asp pair.21
change the electrostatic field.
The calculations suggest that the close proximity of the
cleavage site is achieved by the formation of a LBHB6 Attempts at explaining the catalytic activity of HIV-1
connecting the two Asp groups (Figure 2b). This also helps PR based on a static picture of the protein derived from
explain the chemical equivalence of the two Asp carboxyl the X-ray structures have been unable to account for the
carbons in the free enzyme, as observed in a recent 13C mutagenic behavior described above.
NMR study.22 Furthermore, it accounts for the inverse In a recent investigation,24 we departed from this
solvent isotope effect in the free enzyme, because LBHB conventional approach and investigated the role of large-
have a very low fractionation factor.6 Finally, it permits scale protein motions for enzymatic activity. To this aim,
answering many mechanistic questions about aspartic we have adopted a two-step strategy.
proteases.6 First, we have carried out classical MD calculations on
The calculations further suggest that the electrostatic a protein/substrate complex in aqueous solution. The
interaction between the aspartyl negative charge and the simulation reveals large-scale protein motions on the
rather rigid Thr 26(26′)-Gly 27(27′) dipoles provide a very nanosecond time-scale involving the flaps of the protein
important contribution to the stabilization of the coplanar (Figure 3). The substrate motion turns out to be coupled
orientation of the two carboxy groups. This finding ap- with the lowest frequency motion of the protein.
VOL. 35, NO. 6, 2002 / ACCOUNTS OF CHEMICAL RESEARCH 459
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

FIGURE 3. (a) Reaction mechanism and drug resistance in HIV-1 PR. The complex between the enzyme and a model substrate used in ab
initio and classical MD calculations is shown.24 Selected positions undergoing compensatory mutations are also shown.19 (b) Classical MD
simulation: (left) Schematic representation of the lowest-frequency motion of the protein, on the nanosecond time scale. The motion is a
rotation of the flaps (blue) and other regions of the protein (gray), coupled to a movement of the substrate (green) along the protein C2 axis.
View from the top and from a plane containing the C2 symmetry axis. (Right) Distance (measured on the four CR’s, see ref 24) between the
substrate and Asp dyad CR carbons plotted as a function of the simulated time. (c) Ab initio MD simulation: selected snapshots at the
shortest substrate/Asp dyad distance investigated (labeled as C in the plot of part B of this figure).

As the second step, we have investigated the chemical average value of the classical simulation, the activation
reactivity of the enzyme at different configurations of the free energy is by far larger (several tens of kcal/mol) than
protein. We have focused on the first step of the catalytic that obtained experimentally (16-18 kcal26). In contrast,
reaction (from ES to INT, Chart 1) and estimated the at very short substrate/Asp dyad distances (indicated with
activation free energy with ab initio MD.25 The quantum- C in Figure 5b), a concerted double proton transfer among
chemical calculations show that at conformations in the Asp dyad, the catalytic water and the substrate can
which the substrate/enzyme distance corresponds to the take place (Figure 3c). This stabilizes the highly nucleo-
460 ACCOUNTS OF CHEMICAL RESEARCH / VOL. 35, NO. 6, 2002
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

FIGURE 4. HIV-1 PR/pepstatin complex.28 The protomer A1, postulated on the basis of 13C NMR data,22 evolves to A2 during the ab initio MD.
The diprotonated form B1, proposed in ref 28, is stable during the dynamics and exhibits two clearly distinct chemical shifts.

Chart 1 Thus, the protein appears to act as an “elaborate

nutcracker” that brings the substrate into a favorable
geometrical position for the appropriate mechanical
fluctuations of the protein frame. This mechanism vali-
dates the proposal of Lumry for HIV-1 PR in particular
and mechanical catalysis in general.27 Furthermore, it
offers a clue to how the influence of drug-resistant
mutations far from the active site can emerge. Indeed,
these mutations are expected to enhance the catalytic rate
of HIV-1 PR mutants by affecting the flexibility of the
protein, which in turn may affect the enzymatic activity.
4. NMR Properties of HIV-1 PR Cleavage Site. In the
HIV-1PR/pepstatin adduct (Figure 4), a recent 13C NMR
experiment22 measured two separate resonance lines. On
the basis of these results and the X-ray structure of the
complex, the Asp dyad was suggested to be monoproto-
nated (pattern A1 in Figure 4). The assignment appeared
to be surprising, because the chemical shifts are reversed
with respect to the adduct with the chemically similar
inhibitor KNI-272.
Ab initio MD on models of the pepstatin/enzyme
adduct28 (A1, Figure 4) have provided evidence that the
proposed pattern is not stable. It evolves to protomer A2
(Figure 4), whose calculated 13C chemical shifts are almost
equivalent, in disagreement with the experimental results,
philic agent OH-, which readily reacts to form a gem-diol whereas the other possible patterns of the monoproto-
intermediate INT (Chart 1) with a calculated activation nated form exhibited NMR properties that disagree with
energy comparable to the experimental value. experiment or are highly unstable.
VOL. 35, NO. 6, 2002 / ACCOUNTS OF CHEMICAL RESEARCH 461
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

transformed into a neuronal signal that is detected in the

brain. In the first step of this biochemical reaction chain,
the transmembrane protein rhodopsin, which is located
in the retina, plays the predominant role.29 Rhodopsin
consists of seven transmembrane helices accommodating
in their middle the covalently bound chromophore, the
protonated Schiff base of 11-cis retinal. Upon exposure
to light, retinal isomerizes from 11-cis to all-trans in
an extremely fast and efficient process, which is com-
pleted within 200 fs with a quantum yield of 0.67. This
induces a structural change of the intracellular part of the
protein and triggers the signal transduction to another
receptor protein of the signal cascade. A mechanistic
understanding of this primary reaction is not only es-
sential with regard to the cure of some eye diseases but
also of great interest for the development of optical
switches.
The recent X-ray structure of rhodopsin30 has paved
the way to a molecular investigation of the activation
mechanism.
Detailed mechanistic understanding has remained an
open question. The role of the protein environment that
promotes the isomerization reaction is not well-under-
stood. In addition, the conformational changes that ac-
company the accommodation of all-trans retinal within
the tight binding pocket are unknown. From a theoretical
point of view, the investigation of these questions requires
a mixed quantum/classical approach, because the excited-
state dynamics of retinal (50 atoms) is a quantum chemi-
cal process, but the description of the whole protein
(about 5600 atoms) is only feasible within a classical
framework.
We have recently combined a mixed QM/MM Car-
Parrinello approach18 with the restricted open-shell Kohn-
Sham (ROKS)9 method for describing the dynamics in the
first excited-singlet state. This combination allows the
investigation of photochemical processes in complex
biological systems. As a first test case study, we are
currently investigating the cis-trans photoisomerization
in rhodopsin. Our model system (24 000 atoms, see Figure
5) contains the protein, a hydrophobic membrane-
mimetic environment consisting of octane and a saline
FIGURE 5. Simulation of rhodopsin: (above) photoinduced cis- solution. This setup is based on the most recent crystal
trans isomerization of the protontated Schiff base of 11-cis retinal; structure of rhodopsin and results in a stable classical MD
(below) simulation system for rhodopsin. The protein (blue) contains of several nanoseconds without imposing any type of
seven transmembrane helices that accommodate the chromophore structural constraint. The key configurations of these
(yellow) in the middle.30 The membrane is mimicked by an octane
solution (turquoise) surrounded by water (red). classical preequilibration runs are currently used for mixed
QM/MM simulations of the initial reactant state and the
We suggest, therefore, that the Asp dyad is doubly ROKS-QM/MM Car-Parrinello simulations of the isomer-
protonated. The diprotonated form B1 (Figure 4) turns out ization in the excited state.
to be stable in the time scale investigated (few picosec-
onds) and exhibits different 13C chemical shifts, in qualita-
tive agreement with experimental data22 (Figure 4). 4. Concluding Remarks
We expect that the powerful combination of ab initio In this Account, we have shown that despite the current
MD with calculations of NMR chemical shifts will be limitations of size and time scale, ab initio MD can play
widely used in the future to interpret spectral data in an important role in the modeling of biological systems.
biomolecular systems. Particularly useful is the QM/MM approach that allows
5. Cis/Trans Photoisomerization in Rhodopsin. Hu- local chemical accuracy to be combined with a proper
man vision is a highly efficient process in which light is description of the environmental effects. As is clearly
462 ACCOUNTS OF CHEMICAL RESEARCH / VOL. 35, NO. 6, 2002
 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

demonstrated by the examples discussed, it is now pos- their dynamics be treated quantum mechanically. This is
sible through the use of the ab initio approach to tackle a formidable problem which, despite the progress achieved,
problems of great current interest that cannot be solved has not been seen solved except for relatively simple
by a different approach. Nevertheless these successes models.35 Finally, we wish to emphasize that many of the
cannot hide the fact that much needs to be done in order applications so far presented have used the Born-
to further expand the scope of ab initio MD. Oppenheimer approximation. Going beyond this ap-
The accuracy of the potential energy surface in ab initio proximation requires a method not only for calculating
MD is determined by the exchange and correlation accurately and efficiently the potential energy surfaces of
functionals used. These are quite accurate but have known excited states, but also for evolving the dynamics of the
limitations, such as approximate transition state energies. coupled nonadiabatic electron ion motions.
Some of these limitations can be overcome by locally From this rather cursory analysis, it is clear that the
using higher-order quantum chemical methods. Less challenges before us are manifold and difficult. We
localized correlation effects such as van der Waals inter- nonetheless wish to close on a positive note by pointing
actions are more difficult to include in an ab initio way31 out that we have already reached the stage where ab initio
and have to be added empirically whenever they are MD has a sufficient level of maturity to make it the
expected to play a significant role. method of choice for a very large class of problems.
As discussed earlier, size is an issue. Whenever a small
region where quantum chemical effects are important can
be identified and this region is clearly separated from the References
(1) (a) Karplus, M. Molecular dynamics simulations in biology. Nature
rest, the QM/MM approach can be of great help. However, 1990, 18, 347, 631-639. (b) van Gunsteren, W. F.; Luque, F. J.;
a large class of problems requires the explicit quantum Timms, D.; Torda, A. E. Molecular mechanics in biology: from
chemical description of rather large systems. For instance, structure to function, taking account of solvation. Annu. Rev.
Biophys. Biomol. Struct. 1994, 23, 847-863.
electron transfer can occur over quite large distances in (2) Car, R.; Parrinello, M. Unified Approach for Density Functional
electron-transfer proteins32 and proton pumps induce Theory and Molecular Dynamics. Phys. Rev. Lett. 1985, 55, 2471-
2474.
long-range proton motions.33 Given the difficulty of (3) Hohenberg, P.; Kohn, W. Inhomogeneous electron gas. Phys. Rev.
modeling nucleic acids with classical potentials, we expect B 1964, 136, 864-871. Kohn, W.; Sham, L. J. Self-consistent
this to be an area in which ab initio MD will have a great equations including exchange and correlation effects. Phys. Rev.
A 1965, 140, 1133-1138. Parr, R. G.; Yang, W. Density Functional
impact. The sheer size of these systems has until now been Theory of Atoms and Molecules, Oxford University Press: New
an impediment to wider applications. Although techno- York: 1989.
(4) Banci, L., Comba, P., Eds.; Molecular modeling and dynamics of
logical progress can, indeed, push forward the limit of bioinorganic compounds; Kluwer Academic Publishers: Dordere-
what is presently possible, a decisive impulse in this cht, Boston, London 1997.
(5) See, for example, Carloni, P.; Rothlisberger, U. Simulations of
direction will come when O(N) methods reach maturity. enzymatic systems: perspectives from Car-Parrinello molecular
The limited time scale over which a system can be dynamics simulations. In Theoretical biochemistrysprocesses
and properties of biological systems. Eriksson L., Ed. Elsevier
simulated is a general problem of all MD simulations. This Science: New York, 2000.
is all the more cogent in ab initio MD, in which the (6) Northrop, D. Follow the protons: a low-barrier hydrogen bond
computational cost of a single time step is so much higher. unifies the mechanism of the aspartic proteases. Acc. Chem. Res.
2001, 34, 790-797.
QM/MM methods improve matters, because they reduce (7) Silvestrelli, P. L.; Parrinello, M. Water molecule dipole in the gas
the time step cost, but they are not a solution for modeling phase and in the liquid phase. Phys. Rev Lett. 1999, 82, 3308-
3311.
processes that take place over time scales of nanoseconds (8) Sebastiani, D.; Parrinello, M. A new ab initio approach for NMR
or more. Extending the time scale of MD simulations is chemical shifts in periodic systems. J. Phys. Chem. A 2001, 105,
an area of great current interest and important progress 1951-1958.
(9) Frank, I.; Hutter, J.; Marx, D.; Parrinello, M. Molecular dynamics
is being made wherever the relevant free energy surface in low-spin excited states. J. Chem. Phys. 1998, 108, 4060-4069.
is composed of local minima separated by large barriers. (10) Marx, D.; Hutter, J. Ab initio Molecular Dynamics: Theory and
Implementation. In Modern Methods and Algorithms of Quantum
In this case, the system spends most of its time in these Chemistry; Grotendorst, J., Ed.; John von Neumann Institute for
minima, but occasionally it jumps very quickly to another Computing: Julich, 2000; 301-409.
(11) Goedecker, S. Linear scaling electronic structure methods. Rev.
minimum. These jumps are infrequent but crucial, since Mod. Phys. 1999, 71, 1085-1123.
they lead to important changes in the system. For in- (12) Aqvist, J.; Warshel, A. Simulation of enzyme reactions using
stance, they can correspond to a chemical reaction or to valence bond force fields and other hybrid quantum/classical
approaches. Chem. Rev. 1993, 93, 2523-2544.
a change in conformation. The locality in time of these (13) Laio, A.; VandeVondele, J.; Rothlisberger, U. A Hamiltonian
rare but important events can be exploited, and a variety Electrostatic Coupling Scheme for Hybrid Car-Parrinello Molec-
ular Dynamics Simulations, J. Chem. Phys., 2002, 116, 6941-6947.
of powerful methods have been suggested.34 There are, (14) Horwich, A. L.; Weissmann, J. S. Deadly conformationssProtein
however, problems in which the different time scales misfolding in prion disease. Cell 1997, 89, 499-510.
cannot be separated, one example out of the many being (15) Van Doorslaer, S.; Cereghetti, G. M.; Glockshuber, R.; Schweiger,
A. Unraveling the Cu2+ binding sites in the C-terminal domain of
actin dynamics. For this class of difficult problems, novel the murine prion protein: A pulse EPR and ENDOR study. J. Phys.
solutions need to be found, since it is not to be expected Chem. 2001, 105, 1631-1639.
(16) Price, D. L.; Sisodia, S. S.; Borchelt, D. R. Genetic neurodegen-
that technology can deliver the necessary computational erative diseases: The human illness and transgenic models.
power in the foreseeable future. Science 1998, 282, 1079-1083.
(17) Riek, R.; Hornemann, S.; Wider, G.; Billeter, M.; Glockshuber, R.;
We briefly mention the fact that a proper description Wuthrich, K. NMR structure of the mouse prion protein domain
of the proton motion and other light particles require that PrP(121-231). Nature 1996, 382, 180-182.

VOL. 35, NO. 6, 2002 / ACCOUNTS OF CHEMICAL RESEARCH 463

 Ab Initio Molecular Dynamics of Biological Systems Carloni et al.

(18) Colombo, M. C.; Guidoni, L.; Laio, A.; Magistrato, A.; Maurer, P.; (27) Lumry, R. On the interpretation of data from isothermal processes.
Piana, S.; Roehrig, U.; Spiegel, K.; VandeVondele, J.; Zumstein, Methods Enzymol. 1995, 259, 628-720.
M.; Rothlisberger, U. Hybrid QM/MM Car-Parrinello Simulations (28) Piana, S.; Sebastiani, D.; Carloni, P.; Parrinello, M. An ab initio
of Catalytic and Enzymatic Reactions. Chimia 2002, 56, 11-17. molecular dynamics-based assignment of the protonation state
(19) Ogden, R. C., Flexner, C. W., Eds.; Protease Inhibitors in AIDS of pepstatin A/HIV-1 protease cleavage site. J. Am. Chem. Soc.
Therapy, Marcel Dekker: New York: 2001. 2001, 123, 8730-8737.
(20) Straatsma, T. P.; Zacharias, M.; McCammon, J. A. Free energy (29) For a review, see: Gether U. Uncovering molecular mechanisms
difference calculations in biomolecular systems. In van Gunsteren, involved in activation of G protein-coupled receptor. Endocr. Rev.
W. F., Weiner, P. K., Wilkinson, A. J., Eds. Computer simulations 2000, 21, 90-113.
of biomolecular systems. ESCOM.: Leiden, The Netherlands, (30) Palczewski, K.; Kumasaka, T.; Hori, T.; Behnke, C. A.; Motoshima,
1993. H.; Fox, B. A.; Le Trong, I.; Teller, D. C.; Okada, T.; Stenkamp, R.
(21) Piana, S.; Carloni, P. Conformational Flexibility of the Catalytic E.; Yamamoto, M.; Miyano, M. Crystal structure of rhodopsin: A
Asp Dyad in HIV-1 Protease: an Ab Initio Study on the Free G protein-coupled receptor. Science 2000, 289, 739-745.
Enzyme. Proteins: Struct. Funct. Genet. 2000, 39, 26-36. (31) Meijer, E. J.; Sprik, M. A density-functional study of the inter-
(22) Smith, R.; Brereton, I. M.; Chai, R. Y.; Kent, S. B. H. Ionization molecular interactions of benzene. J. Chem. Phys. 1996, 105,
states of the catalytic residues in HIV-1 protease. Nature Struct. 8684-8689.
Biol. 1996, 3, 946-950. (32) Warshel, A. Protein Control of Redox Potentials of Iron-Sulfur
(23) Warshel, A. Electrostatic origin of the catalytic power of enzymes Proteins, Chem. Rev. 1996, 96 (7), 2491-2514.
and the role of preorganized active sites. J. Biol. Chem. 1998, 273, (33) Hofacker, I.; Schulten, K. Oxygen and proton pathways in cyto-
27035-27038. chrome c oxidase. Proteins: Struct. Funct. Genet. 1998, 30, 100-
(24) Piana, S.; Carloni, P.; Parrinello, M. The role of conformational 107.
fluctuations in the enzymatic mechanism of HIV-1 protease, J. (34) VandeVondele, J.; Rothlisberger, U. Efficient Multidimensional
Mol. Biol., in press. Free Energy Calculations for ab initio Molecular Dynamics using
(25) The free energy is calculated using constraint dynamics (Carter, Classical Bias Potentials, J. Chem. Phys. 2000, 113, 4863-4868.
E. A.; Ciccotti, G.; Hynes, J. T.; Kapral, R. Constrained reaction Passerone, A.; Parrinello, M. Action-derived molecular dynamics
coordinate dynamics for the simulation of rare events. Chem. in the study of rare events. Phys. Rev. Lett. 2001, 8710, 8302.
Phys. Lett. 1989, 156, 472-477). The reaction constraint is the (35) Webb, S. P.; Agarwal, P. K.; Hammes-Schiffer, S. Combining
distance between the catalytic water and the substrate carbonyl electronic structure methods with the calculation of hydrogen
carbon. vibrational wave functions: application to hydride transfer in liver
(26) Polgár, L.; Szeltner, Z.; Boros, I. Substrate-dependent mechanism alcohol dehydrogenase. J. Phys. Chem. 2000, 104, 8884-8894.
in the catalysis of human immunodeficiency virus protease.
Biochemistry 1994, 33, 9351-9357. AR010018U

464 ACCOUNTS OF CHEMICAL RESEARCH / VOL. 35, NO. 6, 2002